Loading...
1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/sizes.h>
4#include <linux/list_sort.h>
5#include "misc.h"
6#include "ctree.h"
7#include "block-group.h"
8#include "space-info.h"
9#include "disk-io.h"
10#include "free-space-cache.h"
11#include "free-space-tree.h"
12#include "volumes.h"
13#include "transaction.h"
14#include "ref-verify.h"
15#include "sysfs.h"
16#include "tree-log.h"
17#include "delalloc-space.h"
18#include "discard.h"
19#include "raid56.h"
20#include "zoned.h"
21#include "fs.h"
22#include "accessors.h"
23#include "extent-tree.h"
24
25#ifdef CONFIG_BTRFS_DEBUG
26int btrfs_should_fragment_free_space(struct btrfs_block_group *block_group)
27{
28 struct btrfs_fs_info *fs_info = block_group->fs_info;
29
30 return (btrfs_test_opt(fs_info, FRAGMENT_METADATA) &&
31 block_group->flags & BTRFS_BLOCK_GROUP_METADATA) ||
32 (btrfs_test_opt(fs_info, FRAGMENT_DATA) &&
33 block_group->flags & BTRFS_BLOCK_GROUP_DATA);
34}
35#endif
36
37/*
38 * Return target flags in extended format or 0 if restripe for this chunk_type
39 * is not in progress
40 *
41 * Should be called with balance_lock held
42 */
43static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
44{
45 struct btrfs_balance_control *bctl = fs_info->balance_ctl;
46 u64 target = 0;
47
48 if (!bctl)
49 return 0;
50
51 if (flags & BTRFS_BLOCK_GROUP_DATA &&
52 bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
53 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
54 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
55 bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
56 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
57 } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
58 bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
59 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
60 }
61
62 return target;
63}
64
65/*
66 * @flags: available profiles in extended format (see ctree.h)
67 *
68 * Return reduced profile in chunk format. If profile changing is in progress
69 * (either running or paused) picks the target profile (if it's already
70 * available), otherwise falls back to plain reducing.
71 */
72static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
73{
74 u64 num_devices = fs_info->fs_devices->rw_devices;
75 u64 target;
76 u64 raid_type;
77 u64 allowed = 0;
78
79 /*
80 * See if restripe for this chunk_type is in progress, if so try to
81 * reduce to the target profile
82 */
83 spin_lock(&fs_info->balance_lock);
84 target = get_restripe_target(fs_info, flags);
85 if (target) {
86 spin_unlock(&fs_info->balance_lock);
87 return extended_to_chunk(target);
88 }
89 spin_unlock(&fs_info->balance_lock);
90
91 /* First, mask out the RAID levels which aren't possible */
92 for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
93 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
94 allowed |= btrfs_raid_array[raid_type].bg_flag;
95 }
96 allowed &= flags;
97
98 /* Select the highest-redundancy RAID level. */
99 if (allowed & BTRFS_BLOCK_GROUP_RAID1C4)
100 allowed = BTRFS_BLOCK_GROUP_RAID1C4;
101 else if (allowed & BTRFS_BLOCK_GROUP_RAID6)
102 allowed = BTRFS_BLOCK_GROUP_RAID6;
103 else if (allowed & BTRFS_BLOCK_GROUP_RAID1C3)
104 allowed = BTRFS_BLOCK_GROUP_RAID1C3;
105 else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
106 allowed = BTRFS_BLOCK_GROUP_RAID5;
107 else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
108 allowed = BTRFS_BLOCK_GROUP_RAID10;
109 else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
110 allowed = BTRFS_BLOCK_GROUP_RAID1;
111 else if (allowed & BTRFS_BLOCK_GROUP_DUP)
112 allowed = BTRFS_BLOCK_GROUP_DUP;
113 else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
114 allowed = BTRFS_BLOCK_GROUP_RAID0;
115
116 flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
117
118 return extended_to_chunk(flags | allowed);
119}
120
121u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
122{
123 unsigned seq;
124 u64 flags;
125
126 do {
127 flags = orig_flags;
128 seq = read_seqbegin(&fs_info->profiles_lock);
129
130 if (flags & BTRFS_BLOCK_GROUP_DATA)
131 flags |= fs_info->avail_data_alloc_bits;
132 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
133 flags |= fs_info->avail_system_alloc_bits;
134 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
135 flags |= fs_info->avail_metadata_alloc_bits;
136 } while (read_seqretry(&fs_info->profiles_lock, seq));
137
138 return btrfs_reduce_alloc_profile(fs_info, flags);
139}
140
141void btrfs_get_block_group(struct btrfs_block_group *cache)
142{
143 refcount_inc(&cache->refs);
144}
145
146void btrfs_put_block_group(struct btrfs_block_group *cache)
147{
148 if (refcount_dec_and_test(&cache->refs)) {
149 WARN_ON(cache->pinned > 0);
150 /*
151 * If there was a failure to cleanup a log tree, very likely due
152 * to an IO failure on a writeback attempt of one or more of its
153 * extent buffers, we could not do proper (and cheap) unaccounting
154 * of their reserved space, so don't warn on reserved > 0 in that
155 * case.
156 */
157 if (!(cache->flags & BTRFS_BLOCK_GROUP_METADATA) ||
158 !BTRFS_FS_LOG_CLEANUP_ERROR(cache->fs_info))
159 WARN_ON(cache->reserved > 0);
160
161 /*
162 * A block_group shouldn't be on the discard_list anymore.
163 * Remove the block_group from the discard_list to prevent us
164 * from causing a panic due to NULL pointer dereference.
165 */
166 if (WARN_ON(!list_empty(&cache->discard_list)))
167 btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
168 cache);
169
170 kfree(cache->free_space_ctl);
171 btrfs_free_chunk_map(cache->physical_map);
172 kfree(cache);
173 }
174}
175
176/*
177 * This adds the block group to the fs_info rb tree for the block group cache
178 */
179static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
180 struct btrfs_block_group *block_group)
181{
182 struct rb_node **p;
183 struct rb_node *parent = NULL;
184 struct btrfs_block_group *cache;
185 bool leftmost = true;
186
187 ASSERT(block_group->length != 0);
188
189 write_lock(&info->block_group_cache_lock);
190 p = &info->block_group_cache_tree.rb_root.rb_node;
191
192 while (*p) {
193 parent = *p;
194 cache = rb_entry(parent, struct btrfs_block_group, cache_node);
195 if (block_group->start < cache->start) {
196 p = &(*p)->rb_left;
197 } else if (block_group->start > cache->start) {
198 p = &(*p)->rb_right;
199 leftmost = false;
200 } else {
201 write_unlock(&info->block_group_cache_lock);
202 return -EEXIST;
203 }
204 }
205
206 rb_link_node(&block_group->cache_node, parent, p);
207 rb_insert_color_cached(&block_group->cache_node,
208 &info->block_group_cache_tree, leftmost);
209
210 write_unlock(&info->block_group_cache_lock);
211
212 return 0;
213}
214
215/*
216 * This will return the block group at or after bytenr if contains is 0, else
217 * it will return the block group that contains the bytenr
218 */
219static struct btrfs_block_group *block_group_cache_tree_search(
220 struct btrfs_fs_info *info, u64 bytenr, int contains)
221{
222 struct btrfs_block_group *cache, *ret = NULL;
223 struct rb_node *n;
224 u64 end, start;
225
226 read_lock(&info->block_group_cache_lock);
227 n = info->block_group_cache_tree.rb_root.rb_node;
228
229 while (n) {
230 cache = rb_entry(n, struct btrfs_block_group, cache_node);
231 end = cache->start + cache->length - 1;
232 start = cache->start;
233
234 if (bytenr < start) {
235 if (!contains && (!ret || start < ret->start))
236 ret = cache;
237 n = n->rb_left;
238 } else if (bytenr > start) {
239 if (contains && bytenr <= end) {
240 ret = cache;
241 break;
242 }
243 n = n->rb_right;
244 } else {
245 ret = cache;
246 break;
247 }
248 }
249 if (ret)
250 btrfs_get_block_group(ret);
251 read_unlock(&info->block_group_cache_lock);
252
253 return ret;
254}
255
256/*
257 * Return the block group that starts at or after bytenr
258 */
259struct btrfs_block_group *btrfs_lookup_first_block_group(
260 struct btrfs_fs_info *info, u64 bytenr)
261{
262 return block_group_cache_tree_search(info, bytenr, 0);
263}
264
265/*
266 * Return the block group that contains the given bytenr
267 */
268struct btrfs_block_group *btrfs_lookup_block_group(
269 struct btrfs_fs_info *info, u64 bytenr)
270{
271 return block_group_cache_tree_search(info, bytenr, 1);
272}
273
274struct btrfs_block_group *btrfs_next_block_group(
275 struct btrfs_block_group *cache)
276{
277 struct btrfs_fs_info *fs_info = cache->fs_info;
278 struct rb_node *node;
279
280 read_lock(&fs_info->block_group_cache_lock);
281
282 /* If our block group was removed, we need a full search. */
283 if (RB_EMPTY_NODE(&cache->cache_node)) {
284 const u64 next_bytenr = cache->start + cache->length;
285
286 read_unlock(&fs_info->block_group_cache_lock);
287 btrfs_put_block_group(cache);
288 return btrfs_lookup_first_block_group(fs_info, next_bytenr);
289 }
290 node = rb_next(&cache->cache_node);
291 btrfs_put_block_group(cache);
292 if (node) {
293 cache = rb_entry(node, struct btrfs_block_group, cache_node);
294 btrfs_get_block_group(cache);
295 } else
296 cache = NULL;
297 read_unlock(&fs_info->block_group_cache_lock);
298 return cache;
299}
300
301/*
302 * Check if we can do a NOCOW write for a given extent.
303 *
304 * @fs_info: The filesystem information object.
305 * @bytenr: Logical start address of the extent.
306 *
307 * Check if we can do a NOCOW write for the given extent, and increments the
308 * number of NOCOW writers in the block group that contains the extent, as long
309 * as the block group exists and it's currently not in read-only mode.
310 *
311 * Returns: A non-NULL block group pointer if we can do a NOCOW write, the caller
312 * is responsible for calling btrfs_dec_nocow_writers() later.
313 *
314 * Or NULL if we can not do a NOCOW write
315 */
316struct btrfs_block_group *btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info,
317 u64 bytenr)
318{
319 struct btrfs_block_group *bg;
320 bool can_nocow = true;
321
322 bg = btrfs_lookup_block_group(fs_info, bytenr);
323 if (!bg)
324 return NULL;
325
326 spin_lock(&bg->lock);
327 if (bg->ro)
328 can_nocow = false;
329 else
330 atomic_inc(&bg->nocow_writers);
331 spin_unlock(&bg->lock);
332
333 if (!can_nocow) {
334 btrfs_put_block_group(bg);
335 return NULL;
336 }
337
338 /* No put on block group, done by btrfs_dec_nocow_writers(). */
339 return bg;
340}
341
342/*
343 * Decrement the number of NOCOW writers in a block group.
344 *
345 * This is meant to be called after a previous call to btrfs_inc_nocow_writers(),
346 * and on the block group returned by that call. Typically this is called after
347 * creating an ordered extent for a NOCOW write, to prevent races with scrub and
348 * relocation.
349 *
350 * After this call, the caller should not use the block group anymore. It it wants
351 * to use it, then it should get a reference on it before calling this function.
352 */
353void btrfs_dec_nocow_writers(struct btrfs_block_group *bg)
354{
355 if (atomic_dec_and_test(&bg->nocow_writers))
356 wake_up_var(&bg->nocow_writers);
357
358 /* For the lookup done by a previous call to btrfs_inc_nocow_writers(). */
359 btrfs_put_block_group(bg);
360}
361
362void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
363{
364 wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
365}
366
367void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
368 const u64 start)
369{
370 struct btrfs_block_group *bg;
371
372 bg = btrfs_lookup_block_group(fs_info, start);
373 ASSERT(bg);
374 if (atomic_dec_and_test(&bg->reservations))
375 wake_up_var(&bg->reservations);
376 btrfs_put_block_group(bg);
377}
378
379void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg)
380{
381 struct btrfs_space_info *space_info = bg->space_info;
382
383 ASSERT(bg->ro);
384
385 if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
386 return;
387
388 /*
389 * Our block group is read only but before we set it to read only,
390 * some task might have had allocated an extent from it already, but it
391 * has not yet created a respective ordered extent (and added it to a
392 * root's list of ordered extents).
393 * Therefore wait for any task currently allocating extents, since the
394 * block group's reservations counter is incremented while a read lock
395 * on the groups' semaphore is held and decremented after releasing
396 * the read access on that semaphore and creating the ordered extent.
397 */
398 down_write(&space_info->groups_sem);
399 up_write(&space_info->groups_sem);
400
401 wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
402}
403
404struct btrfs_caching_control *btrfs_get_caching_control(
405 struct btrfs_block_group *cache)
406{
407 struct btrfs_caching_control *ctl;
408
409 spin_lock(&cache->lock);
410 if (!cache->caching_ctl) {
411 spin_unlock(&cache->lock);
412 return NULL;
413 }
414
415 ctl = cache->caching_ctl;
416 refcount_inc(&ctl->count);
417 spin_unlock(&cache->lock);
418 return ctl;
419}
420
421void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
422{
423 if (refcount_dec_and_test(&ctl->count))
424 kfree(ctl);
425}
426
427/*
428 * When we wait for progress in the block group caching, its because our
429 * allocation attempt failed at least once. So, we must sleep and let some
430 * progress happen before we try again.
431 *
432 * This function will sleep at least once waiting for new free space to show
433 * up, and then it will check the block group free space numbers for our min
434 * num_bytes. Another option is to have it go ahead and look in the rbtree for
435 * a free extent of a given size, but this is a good start.
436 *
437 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
438 * any of the information in this block group.
439 */
440void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
441 u64 num_bytes)
442{
443 struct btrfs_caching_control *caching_ctl;
444 int progress;
445
446 caching_ctl = btrfs_get_caching_control(cache);
447 if (!caching_ctl)
448 return;
449
450 /*
451 * We've already failed to allocate from this block group, so even if
452 * there's enough space in the block group it isn't contiguous enough to
453 * allow for an allocation, so wait for at least the next wakeup tick,
454 * or for the thing to be done.
455 */
456 progress = atomic_read(&caching_ctl->progress);
457
458 wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
459 (progress != atomic_read(&caching_ctl->progress) &&
460 (cache->free_space_ctl->free_space >= num_bytes)));
461
462 btrfs_put_caching_control(caching_ctl);
463}
464
465static int btrfs_caching_ctl_wait_done(struct btrfs_block_group *cache,
466 struct btrfs_caching_control *caching_ctl)
467{
468 wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
469 return cache->cached == BTRFS_CACHE_ERROR ? -EIO : 0;
470}
471
472static int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
473{
474 struct btrfs_caching_control *caching_ctl;
475 int ret;
476
477 caching_ctl = btrfs_get_caching_control(cache);
478 if (!caching_ctl)
479 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
480 ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
481 btrfs_put_caching_control(caching_ctl);
482 return ret;
483}
484
485#ifdef CONFIG_BTRFS_DEBUG
486static void fragment_free_space(struct btrfs_block_group *block_group)
487{
488 struct btrfs_fs_info *fs_info = block_group->fs_info;
489 u64 start = block_group->start;
490 u64 len = block_group->length;
491 u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
492 fs_info->nodesize : fs_info->sectorsize;
493 u64 step = chunk << 1;
494
495 while (len > chunk) {
496 btrfs_remove_free_space(block_group, start, chunk);
497 start += step;
498 if (len < step)
499 len = 0;
500 else
501 len -= step;
502 }
503}
504#endif
505
506/*
507 * Add a free space range to the in memory free space cache of a block group.
508 * This checks if the range contains super block locations and any such
509 * locations are not added to the free space cache.
510 *
511 * @block_group: The target block group.
512 * @start: Start offset of the range.
513 * @end: End offset of the range (exclusive).
514 * @total_added_ret: Optional pointer to return the total amount of space
515 * added to the block group's free space cache.
516 *
517 * Returns 0 on success or < 0 on error.
518 */
519int btrfs_add_new_free_space(struct btrfs_block_group *block_group, u64 start,
520 u64 end, u64 *total_added_ret)
521{
522 struct btrfs_fs_info *info = block_group->fs_info;
523 u64 extent_start, extent_end, size;
524 int ret;
525
526 if (total_added_ret)
527 *total_added_ret = 0;
528
529 while (start < end) {
530 if (!find_first_extent_bit(&info->excluded_extents, start,
531 &extent_start, &extent_end,
532 EXTENT_DIRTY | EXTENT_UPTODATE,
533 NULL))
534 break;
535
536 if (extent_start <= start) {
537 start = extent_end + 1;
538 } else if (extent_start > start && extent_start < end) {
539 size = extent_start - start;
540 ret = btrfs_add_free_space_async_trimmed(block_group,
541 start, size);
542 if (ret)
543 return ret;
544 if (total_added_ret)
545 *total_added_ret += size;
546 start = extent_end + 1;
547 } else {
548 break;
549 }
550 }
551
552 if (start < end) {
553 size = end - start;
554 ret = btrfs_add_free_space_async_trimmed(block_group, start,
555 size);
556 if (ret)
557 return ret;
558 if (total_added_ret)
559 *total_added_ret += size;
560 }
561
562 return 0;
563}
564
565/*
566 * Get an arbitrary extent item index / max_index through the block group
567 *
568 * @block_group the block group to sample from
569 * @index: the integral step through the block group to grab from
570 * @max_index: the granularity of the sampling
571 * @key: return value parameter for the item we find
572 *
573 * Pre-conditions on indices:
574 * 0 <= index <= max_index
575 * 0 < max_index
576 *
577 * Returns: 0 on success, 1 if the search didn't yield a useful item, negative
578 * error code on error.
579 */
580static int sample_block_group_extent_item(struct btrfs_caching_control *caching_ctl,
581 struct btrfs_block_group *block_group,
582 int index, int max_index,
583 struct btrfs_key *found_key)
584{
585 struct btrfs_fs_info *fs_info = block_group->fs_info;
586 struct btrfs_root *extent_root;
587 u64 search_offset;
588 u64 search_end = block_group->start + block_group->length;
589 struct btrfs_path *path;
590 struct btrfs_key search_key;
591 int ret = 0;
592
593 ASSERT(index >= 0);
594 ASSERT(index <= max_index);
595 ASSERT(max_index > 0);
596 lockdep_assert_held(&caching_ctl->mutex);
597 lockdep_assert_held_read(&fs_info->commit_root_sem);
598
599 path = btrfs_alloc_path();
600 if (!path)
601 return -ENOMEM;
602
603 extent_root = btrfs_extent_root(fs_info, max_t(u64, block_group->start,
604 BTRFS_SUPER_INFO_OFFSET));
605
606 path->skip_locking = 1;
607 path->search_commit_root = 1;
608 path->reada = READA_FORWARD;
609
610 search_offset = index * div_u64(block_group->length, max_index);
611 search_key.objectid = block_group->start + search_offset;
612 search_key.type = BTRFS_EXTENT_ITEM_KEY;
613 search_key.offset = 0;
614
615 btrfs_for_each_slot(extent_root, &search_key, found_key, path, ret) {
616 /* Success; sampled an extent item in the block group */
617 if (found_key->type == BTRFS_EXTENT_ITEM_KEY &&
618 found_key->objectid >= block_group->start &&
619 found_key->objectid + found_key->offset <= search_end)
620 break;
621
622 /* We can't possibly find a valid extent item anymore */
623 if (found_key->objectid >= search_end) {
624 ret = 1;
625 break;
626 }
627 }
628
629 lockdep_assert_held(&caching_ctl->mutex);
630 lockdep_assert_held_read(&fs_info->commit_root_sem);
631 btrfs_free_path(path);
632 return ret;
633}
634
635/*
636 * Best effort attempt to compute a block group's size class while caching it.
637 *
638 * @block_group: the block group we are caching
639 *
640 * We cannot infer the size class while adding free space extents, because that
641 * logic doesn't care about contiguous file extents (it doesn't differentiate
642 * between a 100M extent and 100 contiguous 1M extents). So we need to read the
643 * file extent items. Reading all of them is quite wasteful, because usually
644 * only a handful are enough to give a good answer. Therefore, we just grab 5 of
645 * them at even steps through the block group and pick the smallest size class
646 * we see. Since size class is best effort, and not guaranteed in general,
647 * inaccuracy is acceptable.
648 *
649 * To be more explicit about why this algorithm makes sense:
650 *
651 * If we are caching in a block group from disk, then there are three major cases
652 * to consider:
653 * 1. the block group is well behaved and all extents in it are the same size
654 * class.
655 * 2. the block group is mostly one size class with rare exceptions for last
656 * ditch allocations
657 * 3. the block group was populated before size classes and can have a totally
658 * arbitrary mix of size classes.
659 *
660 * In case 1, looking at any extent in the block group will yield the correct
661 * result. For the mixed cases, taking the minimum size class seems like a good
662 * approximation, since gaps from frees will be usable to the size class. For
663 * 2., a small handful of file extents is likely to yield the right answer. For
664 * 3, we can either read every file extent, or admit that this is best effort
665 * anyway and try to stay fast.
666 *
667 * Returns: 0 on success, negative error code on error.
668 */
669static int load_block_group_size_class(struct btrfs_caching_control *caching_ctl,
670 struct btrfs_block_group *block_group)
671{
672 struct btrfs_fs_info *fs_info = block_group->fs_info;
673 struct btrfs_key key;
674 int i;
675 u64 min_size = block_group->length;
676 enum btrfs_block_group_size_class size_class = BTRFS_BG_SZ_NONE;
677 int ret;
678
679 if (!btrfs_block_group_should_use_size_class(block_group))
680 return 0;
681
682 lockdep_assert_held(&caching_ctl->mutex);
683 lockdep_assert_held_read(&fs_info->commit_root_sem);
684 for (i = 0; i < 5; ++i) {
685 ret = sample_block_group_extent_item(caching_ctl, block_group, i, 5, &key);
686 if (ret < 0)
687 goto out;
688 if (ret > 0)
689 continue;
690 min_size = min_t(u64, min_size, key.offset);
691 size_class = btrfs_calc_block_group_size_class(min_size);
692 }
693 if (size_class != BTRFS_BG_SZ_NONE) {
694 spin_lock(&block_group->lock);
695 block_group->size_class = size_class;
696 spin_unlock(&block_group->lock);
697 }
698out:
699 return ret;
700}
701
702static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
703{
704 struct btrfs_block_group *block_group = caching_ctl->block_group;
705 struct btrfs_fs_info *fs_info = block_group->fs_info;
706 struct btrfs_root *extent_root;
707 struct btrfs_path *path;
708 struct extent_buffer *leaf;
709 struct btrfs_key key;
710 u64 total_found = 0;
711 u64 last = 0;
712 u32 nritems;
713 int ret;
714 bool wakeup = true;
715
716 path = btrfs_alloc_path();
717 if (!path)
718 return -ENOMEM;
719
720 last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
721 extent_root = btrfs_extent_root(fs_info, last);
722
723#ifdef CONFIG_BTRFS_DEBUG
724 /*
725 * If we're fragmenting we don't want to make anybody think we can
726 * allocate from this block group until we've had a chance to fragment
727 * the free space.
728 */
729 if (btrfs_should_fragment_free_space(block_group))
730 wakeup = false;
731#endif
732 /*
733 * We don't want to deadlock with somebody trying to allocate a new
734 * extent for the extent root while also trying to search the extent
735 * root to add free space. So we skip locking and search the commit
736 * root, since its read-only
737 */
738 path->skip_locking = 1;
739 path->search_commit_root = 1;
740 path->reada = READA_FORWARD;
741
742 key.objectid = last;
743 key.offset = 0;
744 key.type = BTRFS_EXTENT_ITEM_KEY;
745
746next:
747 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
748 if (ret < 0)
749 goto out;
750
751 leaf = path->nodes[0];
752 nritems = btrfs_header_nritems(leaf);
753
754 while (1) {
755 if (btrfs_fs_closing(fs_info) > 1) {
756 last = (u64)-1;
757 break;
758 }
759
760 if (path->slots[0] < nritems) {
761 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
762 } else {
763 ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
764 if (ret)
765 break;
766
767 if (need_resched() ||
768 rwsem_is_contended(&fs_info->commit_root_sem)) {
769 btrfs_release_path(path);
770 up_read(&fs_info->commit_root_sem);
771 mutex_unlock(&caching_ctl->mutex);
772 cond_resched();
773 mutex_lock(&caching_ctl->mutex);
774 down_read(&fs_info->commit_root_sem);
775 goto next;
776 }
777
778 ret = btrfs_next_leaf(extent_root, path);
779 if (ret < 0)
780 goto out;
781 if (ret)
782 break;
783 leaf = path->nodes[0];
784 nritems = btrfs_header_nritems(leaf);
785 continue;
786 }
787
788 if (key.objectid < last) {
789 key.objectid = last;
790 key.offset = 0;
791 key.type = BTRFS_EXTENT_ITEM_KEY;
792 btrfs_release_path(path);
793 goto next;
794 }
795
796 if (key.objectid < block_group->start) {
797 path->slots[0]++;
798 continue;
799 }
800
801 if (key.objectid >= block_group->start + block_group->length)
802 break;
803
804 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
805 key.type == BTRFS_METADATA_ITEM_KEY) {
806 u64 space_added;
807
808 ret = btrfs_add_new_free_space(block_group, last,
809 key.objectid, &space_added);
810 if (ret)
811 goto out;
812 total_found += space_added;
813 if (key.type == BTRFS_METADATA_ITEM_KEY)
814 last = key.objectid +
815 fs_info->nodesize;
816 else
817 last = key.objectid + key.offset;
818
819 if (total_found > CACHING_CTL_WAKE_UP) {
820 total_found = 0;
821 if (wakeup) {
822 atomic_inc(&caching_ctl->progress);
823 wake_up(&caching_ctl->wait);
824 }
825 }
826 }
827 path->slots[0]++;
828 }
829
830 ret = btrfs_add_new_free_space(block_group, last,
831 block_group->start + block_group->length,
832 NULL);
833out:
834 btrfs_free_path(path);
835 return ret;
836}
837
838static inline void btrfs_free_excluded_extents(const struct btrfs_block_group *bg)
839{
840 clear_extent_bits(&bg->fs_info->excluded_extents, bg->start,
841 bg->start + bg->length - 1, EXTENT_UPTODATE);
842}
843
844static noinline void caching_thread(struct btrfs_work *work)
845{
846 struct btrfs_block_group *block_group;
847 struct btrfs_fs_info *fs_info;
848 struct btrfs_caching_control *caching_ctl;
849 int ret;
850
851 caching_ctl = container_of(work, struct btrfs_caching_control, work);
852 block_group = caching_ctl->block_group;
853 fs_info = block_group->fs_info;
854
855 mutex_lock(&caching_ctl->mutex);
856 down_read(&fs_info->commit_root_sem);
857
858 load_block_group_size_class(caching_ctl, block_group);
859 if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
860 ret = load_free_space_cache(block_group);
861 if (ret == 1) {
862 ret = 0;
863 goto done;
864 }
865
866 /*
867 * We failed to load the space cache, set ourselves to
868 * CACHE_STARTED and carry on.
869 */
870 spin_lock(&block_group->lock);
871 block_group->cached = BTRFS_CACHE_STARTED;
872 spin_unlock(&block_group->lock);
873 wake_up(&caching_ctl->wait);
874 }
875
876 /*
877 * If we are in the transaction that populated the free space tree we
878 * can't actually cache from the free space tree as our commit root and
879 * real root are the same, so we could change the contents of the blocks
880 * while caching. Instead do the slow caching in this case, and after
881 * the transaction has committed we will be safe.
882 */
883 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE) &&
884 !(test_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags)))
885 ret = load_free_space_tree(caching_ctl);
886 else
887 ret = load_extent_tree_free(caching_ctl);
888done:
889 spin_lock(&block_group->lock);
890 block_group->caching_ctl = NULL;
891 block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
892 spin_unlock(&block_group->lock);
893
894#ifdef CONFIG_BTRFS_DEBUG
895 if (btrfs_should_fragment_free_space(block_group)) {
896 u64 bytes_used;
897
898 spin_lock(&block_group->space_info->lock);
899 spin_lock(&block_group->lock);
900 bytes_used = block_group->length - block_group->used;
901 block_group->space_info->bytes_used += bytes_used >> 1;
902 spin_unlock(&block_group->lock);
903 spin_unlock(&block_group->space_info->lock);
904 fragment_free_space(block_group);
905 }
906#endif
907
908 up_read(&fs_info->commit_root_sem);
909 btrfs_free_excluded_extents(block_group);
910 mutex_unlock(&caching_ctl->mutex);
911
912 wake_up(&caching_ctl->wait);
913
914 btrfs_put_caching_control(caching_ctl);
915 btrfs_put_block_group(block_group);
916}
917
918int btrfs_cache_block_group(struct btrfs_block_group *cache, bool wait)
919{
920 struct btrfs_fs_info *fs_info = cache->fs_info;
921 struct btrfs_caching_control *caching_ctl = NULL;
922 int ret = 0;
923
924 /* Allocator for zoned filesystems does not use the cache at all */
925 if (btrfs_is_zoned(fs_info))
926 return 0;
927
928 caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
929 if (!caching_ctl)
930 return -ENOMEM;
931
932 INIT_LIST_HEAD(&caching_ctl->list);
933 mutex_init(&caching_ctl->mutex);
934 init_waitqueue_head(&caching_ctl->wait);
935 caching_ctl->block_group = cache;
936 refcount_set(&caching_ctl->count, 2);
937 atomic_set(&caching_ctl->progress, 0);
938 btrfs_init_work(&caching_ctl->work, caching_thread, NULL);
939
940 spin_lock(&cache->lock);
941 if (cache->cached != BTRFS_CACHE_NO) {
942 kfree(caching_ctl);
943
944 caching_ctl = cache->caching_ctl;
945 if (caching_ctl)
946 refcount_inc(&caching_ctl->count);
947 spin_unlock(&cache->lock);
948 goto out;
949 }
950 WARN_ON(cache->caching_ctl);
951 cache->caching_ctl = caching_ctl;
952 cache->cached = BTRFS_CACHE_STARTED;
953 spin_unlock(&cache->lock);
954
955 write_lock(&fs_info->block_group_cache_lock);
956 refcount_inc(&caching_ctl->count);
957 list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
958 write_unlock(&fs_info->block_group_cache_lock);
959
960 btrfs_get_block_group(cache);
961
962 btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
963out:
964 if (wait && caching_ctl)
965 ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
966 if (caching_ctl)
967 btrfs_put_caching_control(caching_ctl);
968
969 return ret;
970}
971
972static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
973{
974 u64 extra_flags = chunk_to_extended(flags) &
975 BTRFS_EXTENDED_PROFILE_MASK;
976
977 write_seqlock(&fs_info->profiles_lock);
978 if (flags & BTRFS_BLOCK_GROUP_DATA)
979 fs_info->avail_data_alloc_bits &= ~extra_flags;
980 if (flags & BTRFS_BLOCK_GROUP_METADATA)
981 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
982 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
983 fs_info->avail_system_alloc_bits &= ~extra_flags;
984 write_sequnlock(&fs_info->profiles_lock);
985}
986
987/*
988 * Clear incompat bits for the following feature(s):
989 *
990 * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
991 * in the whole filesystem
992 *
993 * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
994 */
995static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
996{
997 bool found_raid56 = false;
998 bool found_raid1c34 = false;
999
1000 if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
1001 (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
1002 (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
1003 struct list_head *head = &fs_info->space_info;
1004 struct btrfs_space_info *sinfo;
1005
1006 list_for_each_entry_rcu(sinfo, head, list) {
1007 down_read(&sinfo->groups_sem);
1008 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
1009 found_raid56 = true;
1010 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
1011 found_raid56 = true;
1012 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
1013 found_raid1c34 = true;
1014 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
1015 found_raid1c34 = true;
1016 up_read(&sinfo->groups_sem);
1017 }
1018 if (!found_raid56)
1019 btrfs_clear_fs_incompat(fs_info, RAID56);
1020 if (!found_raid1c34)
1021 btrfs_clear_fs_incompat(fs_info, RAID1C34);
1022 }
1023}
1024
1025static int remove_block_group_item(struct btrfs_trans_handle *trans,
1026 struct btrfs_path *path,
1027 struct btrfs_block_group *block_group)
1028{
1029 struct btrfs_fs_info *fs_info = trans->fs_info;
1030 struct btrfs_root *root;
1031 struct btrfs_key key;
1032 int ret;
1033
1034 root = btrfs_block_group_root(fs_info);
1035 key.objectid = block_group->start;
1036 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
1037 key.offset = block_group->length;
1038
1039 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1040 if (ret > 0)
1041 ret = -ENOENT;
1042 if (ret < 0)
1043 return ret;
1044
1045 ret = btrfs_del_item(trans, root, path);
1046 return ret;
1047}
1048
1049int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
1050 struct btrfs_chunk_map *map)
1051{
1052 struct btrfs_fs_info *fs_info = trans->fs_info;
1053 struct btrfs_path *path;
1054 struct btrfs_block_group *block_group;
1055 struct btrfs_free_cluster *cluster;
1056 struct inode *inode;
1057 struct kobject *kobj = NULL;
1058 int ret;
1059 int index;
1060 int factor;
1061 struct btrfs_caching_control *caching_ctl = NULL;
1062 bool remove_map;
1063 bool remove_rsv = false;
1064
1065 block_group = btrfs_lookup_block_group(fs_info, map->start);
1066 BUG_ON(!block_group);
1067 BUG_ON(!block_group->ro);
1068
1069 trace_btrfs_remove_block_group(block_group);
1070 /*
1071 * Free the reserved super bytes from this block group before
1072 * remove it.
1073 */
1074 btrfs_free_excluded_extents(block_group);
1075 btrfs_free_ref_tree_range(fs_info, block_group->start,
1076 block_group->length);
1077
1078 index = btrfs_bg_flags_to_raid_index(block_group->flags);
1079 factor = btrfs_bg_type_to_factor(block_group->flags);
1080
1081 /* make sure this block group isn't part of an allocation cluster */
1082 cluster = &fs_info->data_alloc_cluster;
1083 spin_lock(&cluster->refill_lock);
1084 btrfs_return_cluster_to_free_space(block_group, cluster);
1085 spin_unlock(&cluster->refill_lock);
1086
1087 /*
1088 * make sure this block group isn't part of a metadata
1089 * allocation cluster
1090 */
1091 cluster = &fs_info->meta_alloc_cluster;
1092 spin_lock(&cluster->refill_lock);
1093 btrfs_return_cluster_to_free_space(block_group, cluster);
1094 spin_unlock(&cluster->refill_lock);
1095
1096 btrfs_clear_treelog_bg(block_group);
1097 btrfs_clear_data_reloc_bg(block_group);
1098
1099 path = btrfs_alloc_path();
1100 if (!path) {
1101 ret = -ENOMEM;
1102 goto out;
1103 }
1104
1105 /*
1106 * get the inode first so any iput calls done for the io_list
1107 * aren't the final iput (no unlinks allowed now)
1108 */
1109 inode = lookup_free_space_inode(block_group, path);
1110
1111 mutex_lock(&trans->transaction->cache_write_mutex);
1112 /*
1113 * Make sure our free space cache IO is done before removing the
1114 * free space inode
1115 */
1116 spin_lock(&trans->transaction->dirty_bgs_lock);
1117 if (!list_empty(&block_group->io_list)) {
1118 list_del_init(&block_group->io_list);
1119
1120 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
1121
1122 spin_unlock(&trans->transaction->dirty_bgs_lock);
1123 btrfs_wait_cache_io(trans, block_group, path);
1124 btrfs_put_block_group(block_group);
1125 spin_lock(&trans->transaction->dirty_bgs_lock);
1126 }
1127
1128 if (!list_empty(&block_group->dirty_list)) {
1129 list_del_init(&block_group->dirty_list);
1130 remove_rsv = true;
1131 btrfs_put_block_group(block_group);
1132 }
1133 spin_unlock(&trans->transaction->dirty_bgs_lock);
1134 mutex_unlock(&trans->transaction->cache_write_mutex);
1135
1136 ret = btrfs_remove_free_space_inode(trans, inode, block_group);
1137 if (ret)
1138 goto out;
1139
1140 write_lock(&fs_info->block_group_cache_lock);
1141 rb_erase_cached(&block_group->cache_node,
1142 &fs_info->block_group_cache_tree);
1143 RB_CLEAR_NODE(&block_group->cache_node);
1144
1145 /* Once for the block groups rbtree */
1146 btrfs_put_block_group(block_group);
1147
1148 write_unlock(&fs_info->block_group_cache_lock);
1149
1150 down_write(&block_group->space_info->groups_sem);
1151 /*
1152 * we must use list_del_init so people can check to see if they
1153 * are still on the list after taking the semaphore
1154 */
1155 list_del_init(&block_group->list);
1156 if (list_empty(&block_group->space_info->block_groups[index])) {
1157 kobj = block_group->space_info->block_group_kobjs[index];
1158 block_group->space_info->block_group_kobjs[index] = NULL;
1159 clear_avail_alloc_bits(fs_info, block_group->flags);
1160 }
1161 up_write(&block_group->space_info->groups_sem);
1162 clear_incompat_bg_bits(fs_info, block_group->flags);
1163 if (kobj) {
1164 kobject_del(kobj);
1165 kobject_put(kobj);
1166 }
1167
1168 if (block_group->cached == BTRFS_CACHE_STARTED)
1169 btrfs_wait_block_group_cache_done(block_group);
1170
1171 write_lock(&fs_info->block_group_cache_lock);
1172 caching_ctl = btrfs_get_caching_control(block_group);
1173 if (!caching_ctl) {
1174 struct btrfs_caching_control *ctl;
1175
1176 list_for_each_entry(ctl, &fs_info->caching_block_groups, list) {
1177 if (ctl->block_group == block_group) {
1178 caching_ctl = ctl;
1179 refcount_inc(&caching_ctl->count);
1180 break;
1181 }
1182 }
1183 }
1184 if (caching_ctl)
1185 list_del_init(&caching_ctl->list);
1186 write_unlock(&fs_info->block_group_cache_lock);
1187
1188 if (caching_ctl) {
1189 /* Once for the caching bgs list and once for us. */
1190 btrfs_put_caching_control(caching_ctl);
1191 btrfs_put_caching_control(caching_ctl);
1192 }
1193
1194 spin_lock(&trans->transaction->dirty_bgs_lock);
1195 WARN_ON(!list_empty(&block_group->dirty_list));
1196 WARN_ON(!list_empty(&block_group->io_list));
1197 spin_unlock(&trans->transaction->dirty_bgs_lock);
1198
1199 btrfs_remove_free_space_cache(block_group);
1200
1201 spin_lock(&block_group->space_info->lock);
1202 list_del_init(&block_group->ro_list);
1203
1204 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1205 WARN_ON(block_group->space_info->total_bytes
1206 < block_group->length);
1207 WARN_ON(block_group->space_info->bytes_readonly
1208 < block_group->length - block_group->zone_unusable);
1209 WARN_ON(block_group->space_info->bytes_zone_unusable
1210 < block_group->zone_unusable);
1211 WARN_ON(block_group->space_info->disk_total
1212 < block_group->length * factor);
1213 }
1214 block_group->space_info->total_bytes -= block_group->length;
1215 block_group->space_info->bytes_readonly -=
1216 (block_group->length - block_group->zone_unusable);
1217 block_group->space_info->bytes_zone_unusable -=
1218 block_group->zone_unusable;
1219 block_group->space_info->disk_total -= block_group->length * factor;
1220
1221 spin_unlock(&block_group->space_info->lock);
1222
1223 /*
1224 * Remove the free space for the block group from the free space tree
1225 * and the block group's item from the extent tree before marking the
1226 * block group as removed. This is to prevent races with tasks that
1227 * freeze and unfreeze a block group, this task and another task
1228 * allocating a new block group - the unfreeze task ends up removing
1229 * the block group's extent map before the task calling this function
1230 * deletes the block group item from the extent tree, allowing for
1231 * another task to attempt to create another block group with the same
1232 * item key (and failing with -EEXIST and a transaction abort).
1233 */
1234 ret = remove_block_group_free_space(trans, block_group);
1235 if (ret)
1236 goto out;
1237
1238 ret = remove_block_group_item(trans, path, block_group);
1239 if (ret < 0)
1240 goto out;
1241
1242 spin_lock(&block_group->lock);
1243 set_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags);
1244
1245 /*
1246 * At this point trimming or scrub can't start on this block group,
1247 * because we removed the block group from the rbtree
1248 * fs_info->block_group_cache_tree so no one can't find it anymore and
1249 * even if someone already got this block group before we removed it
1250 * from the rbtree, they have already incremented block_group->frozen -
1251 * if they didn't, for the trimming case they won't find any free space
1252 * entries because we already removed them all when we called
1253 * btrfs_remove_free_space_cache().
1254 *
1255 * And we must not remove the chunk map from the fs_info->mapping_tree
1256 * to prevent the same logical address range and physical device space
1257 * ranges from being reused for a new block group. This is needed to
1258 * avoid races with trimming and scrub.
1259 *
1260 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1261 * completely transactionless, so while it is trimming a range the
1262 * currently running transaction might finish and a new one start,
1263 * allowing for new block groups to be created that can reuse the same
1264 * physical device locations unless we take this special care.
1265 *
1266 * There may also be an implicit trim operation if the file system
1267 * is mounted with -odiscard. The same protections must remain
1268 * in place until the extents have been discarded completely when
1269 * the transaction commit has completed.
1270 */
1271 remove_map = (atomic_read(&block_group->frozen) == 0);
1272 spin_unlock(&block_group->lock);
1273
1274 if (remove_map)
1275 btrfs_remove_chunk_map(fs_info, map);
1276
1277out:
1278 /* Once for the lookup reference */
1279 btrfs_put_block_group(block_group);
1280 if (remove_rsv)
1281 btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
1282 btrfs_free_path(path);
1283 return ret;
1284}
1285
1286struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1287 struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1288{
1289 struct btrfs_root *root = btrfs_block_group_root(fs_info);
1290 struct btrfs_chunk_map *map;
1291 unsigned int num_items;
1292
1293 map = btrfs_find_chunk_map(fs_info, chunk_offset, 1);
1294 ASSERT(map != NULL);
1295 ASSERT(map->start == chunk_offset);
1296
1297 /*
1298 * We need to reserve 3 + N units from the metadata space info in order
1299 * to remove a block group (done at btrfs_remove_chunk() and at
1300 * btrfs_remove_block_group()), which are used for:
1301 *
1302 * 1 unit for adding the free space inode's orphan (located in the tree
1303 * of tree roots).
1304 * 1 unit for deleting the block group item (located in the extent
1305 * tree).
1306 * 1 unit for deleting the free space item (located in tree of tree
1307 * roots).
1308 * N units for deleting N device extent items corresponding to each
1309 * stripe (located in the device tree).
1310 *
1311 * In order to remove a block group we also need to reserve units in the
1312 * system space info in order to update the chunk tree (update one or
1313 * more device items and remove one chunk item), but this is done at
1314 * btrfs_remove_chunk() through a call to check_system_chunk().
1315 */
1316 num_items = 3 + map->num_stripes;
1317 btrfs_free_chunk_map(map);
1318
1319 return btrfs_start_transaction_fallback_global_rsv(root, num_items);
1320}
1321
1322/*
1323 * Mark block group @cache read-only, so later write won't happen to block
1324 * group @cache.
1325 *
1326 * If @force is not set, this function will only mark the block group readonly
1327 * if we have enough free space (1M) in other metadata/system block groups.
1328 * If @force is not set, this function will mark the block group readonly
1329 * without checking free space.
1330 *
1331 * NOTE: This function doesn't care if other block groups can contain all the
1332 * data in this block group. That check should be done by relocation routine,
1333 * not this function.
1334 */
1335static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1336{
1337 struct btrfs_space_info *sinfo = cache->space_info;
1338 u64 num_bytes;
1339 int ret = -ENOSPC;
1340
1341 spin_lock(&sinfo->lock);
1342 spin_lock(&cache->lock);
1343
1344 if (cache->swap_extents) {
1345 ret = -ETXTBSY;
1346 goto out;
1347 }
1348
1349 if (cache->ro) {
1350 cache->ro++;
1351 ret = 0;
1352 goto out;
1353 }
1354
1355 num_bytes = cache->length - cache->reserved - cache->pinned -
1356 cache->bytes_super - cache->zone_unusable - cache->used;
1357
1358 /*
1359 * Data never overcommits, even in mixed mode, so do just the straight
1360 * check of left over space in how much we have allocated.
1361 */
1362 if (force) {
1363 ret = 0;
1364 } else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1365 u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1366
1367 /*
1368 * Here we make sure if we mark this bg RO, we still have enough
1369 * free space as buffer.
1370 */
1371 if (sinfo_used + num_bytes <= sinfo->total_bytes)
1372 ret = 0;
1373 } else {
1374 /*
1375 * We overcommit metadata, so we need to do the
1376 * btrfs_can_overcommit check here, and we need to pass in
1377 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1378 * leeway to allow us to mark this block group as read only.
1379 */
1380 if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1381 BTRFS_RESERVE_NO_FLUSH))
1382 ret = 0;
1383 }
1384
1385 if (!ret) {
1386 sinfo->bytes_readonly += num_bytes;
1387 if (btrfs_is_zoned(cache->fs_info)) {
1388 /* Migrate zone_unusable bytes to readonly */
1389 sinfo->bytes_readonly += cache->zone_unusable;
1390 sinfo->bytes_zone_unusable -= cache->zone_unusable;
1391 cache->zone_unusable = 0;
1392 }
1393 cache->ro++;
1394 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1395 }
1396out:
1397 spin_unlock(&cache->lock);
1398 spin_unlock(&sinfo->lock);
1399 if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1400 btrfs_info(cache->fs_info,
1401 "unable to make block group %llu ro", cache->start);
1402 btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1403 }
1404 return ret;
1405}
1406
1407static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1408 struct btrfs_block_group *bg)
1409{
1410 struct btrfs_fs_info *fs_info = bg->fs_info;
1411 struct btrfs_transaction *prev_trans = NULL;
1412 const u64 start = bg->start;
1413 const u64 end = start + bg->length - 1;
1414 int ret;
1415
1416 spin_lock(&fs_info->trans_lock);
1417 if (trans->transaction->list.prev != &fs_info->trans_list) {
1418 prev_trans = list_last_entry(&trans->transaction->list,
1419 struct btrfs_transaction, list);
1420 refcount_inc(&prev_trans->use_count);
1421 }
1422 spin_unlock(&fs_info->trans_lock);
1423
1424 /*
1425 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1426 * btrfs_finish_extent_commit(). If we are at transaction N, another
1427 * task might be running finish_extent_commit() for the previous
1428 * transaction N - 1, and have seen a range belonging to the block
1429 * group in pinned_extents before we were able to clear the whole block
1430 * group range from pinned_extents. This means that task can lookup for
1431 * the block group after we unpinned it from pinned_extents and removed
1432 * it, leading to a BUG_ON() at unpin_extent_range().
1433 */
1434 mutex_lock(&fs_info->unused_bg_unpin_mutex);
1435 if (prev_trans) {
1436 ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1437 EXTENT_DIRTY);
1438 if (ret)
1439 goto out;
1440 }
1441
1442 ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1443 EXTENT_DIRTY);
1444out:
1445 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1446 if (prev_trans)
1447 btrfs_put_transaction(prev_trans);
1448
1449 return ret == 0;
1450}
1451
1452/*
1453 * Process the unused_bgs list and remove any that don't have any allocated
1454 * space inside of them.
1455 */
1456void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1457{
1458 LIST_HEAD(retry_list);
1459 struct btrfs_block_group *block_group;
1460 struct btrfs_space_info *space_info;
1461 struct btrfs_trans_handle *trans;
1462 const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1463 int ret = 0;
1464
1465 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1466 return;
1467
1468 if (btrfs_fs_closing(fs_info))
1469 return;
1470
1471 /*
1472 * Long running balances can keep us blocked here for eternity, so
1473 * simply skip deletion if we're unable to get the mutex.
1474 */
1475 if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1476 return;
1477
1478 spin_lock(&fs_info->unused_bgs_lock);
1479 while (!list_empty(&fs_info->unused_bgs)) {
1480 u64 used;
1481 int trimming;
1482
1483 block_group = list_first_entry(&fs_info->unused_bgs,
1484 struct btrfs_block_group,
1485 bg_list);
1486 list_del_init(&block_group->bg_list);
1487
1488 space_info = block_group->space_info;
1489
1490 if (ret || btrfs_mixed_space_info(space_info)) {
1491 btrfs_put_block_group(block_group);
1492 continue;
1493 }
1494 spin_unlock(&fs_info->unused_bgs_lock);
1495
1496 btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1497
1498 /* Don't want to race with allocators so take the groups_sem */
1499 down_write(&space_info->groups_sem);
1500
1501 /*
1502 * Async discard moves the final block group discard to be prior
1503 * to the unused_bgs code path. Therefore, if it's not fully
1504 * trimmed, punt it back to the async discard lists.
1505 */
1506 if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1507 !btrfs_is_free_space_trimmed(block_group)) {
1508 trace_btrfs_skip_unused_block_group(block_group);
1509 up_write(&space_info->groups_sem);
1510 /* Requeue if we failed because of async discard */
1511 btrfs_discard_queue_work(&fs_info->discard_ctl,
1512 block_group);
1513 goto next;
1514 }
1515
1516 spin_lock(&space_info->lock);
1517 spin_lock(&block_group->lock);
1518 if (btrfs_is_block_group_used(block_group) || block_group->ro ||
1519 list_is_singular(&block_group->list)) {
1520 /*
1521 * We want to bail if we made new allocations or have
1522 * outstanding allocations in this block group. We do
1523 * the ro check in case balance is currently acting on
1524 * this block group.
1525 */
1526 trace_btrfs_skip_unused_block_group(block_group);
1527 spin_unlock(&block_group->lock);
1528 spin_unlock(&space_info->lock);
1529 up_write(&space_info->groups_sem);
1530 goto next;
1531 }
1532
1533 /*
1534 * The block group may be unused but there may be space reserved
1535 * accounting with the existence of that block group, that is,
1536 * space_info->bytes_may_use was incremented by a task but no
1537 * space was yet allocated from the block group by the task.
1538 * That space may or may not be allocated, as we are generally
1539 * pessimistic about space reservation for metadata as well as
1540 * for data when using compression (as we reserve space based on
1541 * the worst case, when data can't be compressed, and before
1542 * actually attempting compression, before starting writeback).
1543 *
1544 * So check if the total space of the space_info minus the size
1545 * of this block group is less than the used space of the
1546 * space_info - if that's the case, then it means we have tasks
1547 * that might be relying on the block group in order to allocate
1548 * extents, and add back the block group to the unused list when
1549 * we finish, so that we retry later in case no tasks ended up
1550 * needing to allocate extents from the block group.
1551 */
1552 used = btrfs_space_info_used(space_info, true);
1553 if (space_info->total_bytes - block_group->length < used) {
1554 /*
1555 * Add a reference for the list, compensate for the ref
1556 * drop under the "next" label for the
1557 * fs_info->unused_bgs list.
1558 */
1559 btrfs_get_block_group(block_group);
1560 list_add_tail(&block_group->bg_list, &retry_list);
1561
1562 trace_btrfs_skip_unused_block_group(block_group);
1563 spin_unlock(&block_group->lock);
1564 spin_unlock(&space_info->lock);
1565 up_write(&space_info->groups_sem);
1566 goto next;
1567 }
1568
1569 spin_unlock(&block_group->lock);
1570 spin_unlock(&space_info->lock);
1571
1572 /* We don't want to force the issue, only flip if it's ok. */
1573 ret = inc_block_group_ro(block_group, 0);
1574 up_write(&space_info->groups_sem);
1575 if (ret < 0) {
1576 ret = 0;
1577 goto next;
1578 }
1579
1580 ret = btrfs_zone_finish(block_group);
1581 if (ret < 0) {
1582 btrfs_dec_block_group_ro(block_group);
1583 if (ret == -EAGAIN)
1584 ret = 0;
1585 goto next;
1586 }
1587
1588 /*
1589 * Want to do this before we do anything else so we can recover
1590 * properly if we fail to join the transaction.
1591 */
1592 trans = btrfs_start_trans_remove_block_group(fs_info,
1593 block_group->start);
1594 if (IS_ERR(trans)) {
1595 btrfs_dec_block_group_ro(block_group);
1596 ret = PTR_ERR(trans);
1597 goto next;
1598 }
1599
1600 /*
1601 * We could have pending pinned extents for this block group,
1602 * just delete them, we don't care about them anymore.
1603 */
1604 if (!clean_pinned_extents(trans, block_group)) {
1605 btrfs_dec_block_group_ro(block_group);
1606 goto end_trans;
1607 }
1608
1609 /*
1610 * At this point, the block_group is read only and should fail
1611 * new allocations. However, btrfs_finish_extent_commit() can
1612 * cause this block_group to be placed back on the discard
1613 * lists because now the block_group isn't fully discarded.
1614 * Bail here and try again later after discarding everything.
1615 */
1616 spin_lock(&fs_info->discard_ctl.lock);
1617 if (!list_empty(&block_group->discard_list)) {
1618 spin_unlock(&fs_info->discard_ctl.lock);
1619 btrfs_dec_block_group_ro(block_group);
1620 btrfs_discard_queue_work(&fs_info->discard_ctl,
1621 block_group);
1622 goto end_trans;
1623 }
1624 spin_unlock(&fs_info->discard_ctl.lock);
1625
1626 /* Reset pinned so btrfs_put_block_group doesn't complain */
1627 spin_lock(&space_info->lock);
1628 spin_lock(&block_group->lock);
1629
1630 btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1631 -block_group->pinned);
1632 space_info->bytes_readonly += block_group->pinned;
1633 block_group->pinned = 0;
1634
1635 spin_unlock(&block_group->lock);
1636 spin_unlock(&space_info->lock);
1637
1638 /*
1639 * The normal path here is an unused block group is passed here,
1640 * then trimming is handled in the transaction commit path.
1641 * Async discard interposes before this to do the trimming
1642 * before coming down the unused block group path as trimming
1643 * will no longer be done later in the transaction commit path.
1644 */
1645 if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1646 goto flip_async;
1647
1648 /*
1649 * DISCARD can flip during remount. On zoned filesystems, we
1650 * need to reset sequential-required zones.
1651 */
1652 trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1653 btrfs_is_zoned(fs_info);
1654
1655 /* Implicit trim during transaction commit. */
1656 if (trimming)
1657 btrfs_freeze_block_group(block_group);
1658
1659 /*
1660 * Btrfs_remove_chunk will abort the transaction if things go
1661 * horribly wrong.
1662 */
1663 ret = btrfs_remove_chunk(trans, block_group->start);
1664
1665 if (ret) {
1666 if (trimming)
1667 btrfs_unfreeze_block_group(block_group);
1668 goto end_trans;
1669 }
1670
1671 /*
1672 * If we're not mounted with -odiscard, we can just forget
1673 * about this block group. Otherwise we'll need to wait
1674 * until transaction commit to do the actual discard.
1675 */
1676 if (trimming) {
1677 spin_lock(&fs_info->unused_bgs_lock);
1678 /*
1679 * A concurrent scrub might have added us to the list
1680 * fs_info->unused_bgs, so use a list_move operation
1681 * to add the block group to the deleted_bgs list.
1682 */
1683 list_move(&block_group->bg_list,
1684 &trans->transaction->deleted_bgs);
1685 spin_unlock(&fs_info->unused_bgs_lock);
1686 btrfs_get_block_group(block_group);
1687 }
1688end_trans:
1689 btrfs_end_transaction(trans);
1690next:
1691 btrfs_put_block_group(block_group);
1692 spin_lock(&fs_info->unused_bgs_lock);
1693 }
1694 list_splice_tail(&retry_list, &fs_info->unused_bgs);
1695 spin_unlock(&fs_info->unused_bgs_lock);
1696 mutex_unlock(&fs_info->reclaim_bgs_lock);
1697 return;
1698
1699flip_async:
1700 btrfs_end_transaction(trans);
1701 spin_lock(&fs_info->unused_bgs_lock);
1702 list_splice_tail(&retry_list, &fs_info->unused_bgs);
1703 spin_unlock(&fs_info->unused_bgs_lock);
1704 mutex_unlock(&fs_info->reclaim_bgs_lock);
1705 btrfs_put_block_group(block_group);
1706 btrfs_discard_punt_unused_bgs_list(fs_info);
1707}
1708
1709void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1710{
1711 struct btrfs_fs_info *fs_info = bg->fs_info;
1712
1713 spin_lock(&fs_info->unused_bgs_lock);
1714 if (list_empty(&bg->bg_list)) {
1715 btrfs_get_block_group(bg);
1716 trace_btrfs_add_unused_block_group(bg);
1717 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1718 } else if (!test_bit(BLOCK_GROUP_FLAG_NEW, &bg->runtime_flags)) {
1719 /* Pull out the block group from the reclaim_bgs list. */
1720 trace_btrfs_add_unused_block_group(bg);
1721 list_move_tail(&bg->bg_list, &fs_info->unused_bgs);
1722 }
1723 spin_unlock(&fs_info->unused_bgs_lock);
1724}
1725
1726/*
1727 * We want block groups with a low number of used bytes to be in the beginning
1728 * of the list, so they will get reclaimed first.
1729 */
1730static int reclaim_bgs_cmp(void *unused, const struct list_head *a,
1731 const struct list_head *b)
1732{
1733 const struct btrfs_block_group *bg1, *bg2;
1734
1735 bg1 = list_entry(a, struct btrfs_block_group, bg_list);
1736 bg2 = list_entry(b, struct btrfs_block_group, bg_list);
1737
1738 return bg1->used > bg2->used;
1739}
1740
1741static inline bool btrfs_should_reclaim(struct btrfs_fs_info *fs_info)
1742{
1743 if (btrfs_is_zoned(fs_info))
1744 return btrfs_zoned_should_reclaim(fs_info);
1745 return true;
1746}
1747
1748static bool should_reclaim_block_group(struct btrfs_block_group *bg, u64 bytes_freed)
1749{
1750 const struct btrfs_space_info *space_info = bg->space_info;
1751 const int reclaim_thresh = READ_ONCE(space_info->bg_reclaim_threshold);
1752 const u64 new_val = bg->used;
1753 const u64 old_val = new_val + bytes_freed;
1754 u64 thresh;
1755
1756 if (reclaim_thresh == 0)
1757 return false;
1758
1759 thresh = mult_perc(bg->length, reclaim_thresh);
1760
1761 /*
1762 * If we were below the threshold before don't reclaim, we are likely a
1763 * brand new block group and we don't want to relocate new block groups.
1764 */
1765 if (old_val < thresh)
1766 return false;
1767 if (new_val >= thresh)
1768 return false;
1769 return true;
1770}
1771
1772void btrfs_reclaim_bgs_work(struct work_struct *work)
1773{
1774 struct btrfs_fs_info *fs_info =
1775 container_of(work, struct btrfs_fs_info, reclaim_bgs_work);
1776 struct btrfs_block_group *bg;
1777 struct btrfs_space_info *space_info;
1778
1779 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1780 return;
1781
1782 if (btrfs_fs_closing(fs_info))
1783 return;
1784
1785 if (!btrfs_should_reclaim(fs_info))
1786 return;
1787
1788 sb_start_write(fs_info->sb);
1789
1790 if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_BALANCE)) {
1791 sb_end_write(fs_info->sb);
1792 return;
1793 }
1794
1795 /*
1796 * Long running balances can keep us blocked here for eternity, so
1797 * simply skip reclaim if we're unable to get the mutex.
1798 */
1799 if (!mutex_trylock(&fs_info->reclaim_bgs_lock)) {
1800 btrfs_exclop_finish(fs_info);
1801 sb_end_write(fs_info->sb);
1802 return;
1803 }
1804
1805 spin_lock(&fs_info->unused_bgs_lock);
1806 /*
1807 * Sort happens under lock because we can't simply splice it and sort.
1808 * The block groups might still be in use and reachable via bg_list,
1809 * and their presence in the reclaim_bgs list must be preserved.
1810 */
1811 list_sort(NULL, &fs_info->reclaim_bgs, reclaim_bgs_cmp);
1812 while (!list_empty(&fs_info->reclaim_bgs)) {
1813 u64 zone_unusable;
1814 int ret = 0;
1815
1816 bg = list_first_entry(&fs_info->reclaim_bgs,
1817 struct btrfs_block_group,
1818 bg_list);
1819 list_del_init(&bg->bg_list);
1820
1821 space_info = bg->space_info;
1822 spin_unlock(&fs_info->unused_bgs_lock);
1823
1824 /* Don't race with allocators so take the groups_sem */
1825 down_write(&space_info->groups_sem);
1826
1827 spin_lock(&bg->lock);
1828 if (bg->reserved || bg->pinned || bg->ro) {
1829 /*
1830 * We want to bail if we made new allocations or have
1831 * outstanding allocations in this block group. We do
1832 * the ro check in case balance is currently acting on
1833 * this block group.
1834 */
1835 spin_unlock(&bg->lock);
1836 up_write(&space_info->groups_sem);
1837 goto next;
1838 }
1839 if (bg->used == 0) {
1840 /*
1841 * It is possible that we trigger relocation on a block
1842 * group as its extents are deleted and it first goes
1843 * below the threshold, then shortly after goes empty.
1844 *
1845 * In this case, relocating it does delete it, but has
1846 * some overhead in relocation specific metadata, looking
1847 * for the non-existent extents and running some extra
1848 * transactions, which we can avoid by using one of the
1849 * other mechanisms for dealing with empty block groups.
1850 */
1851 if (!btrfs_test_opt(fs_info, DISCARD_ASYNC))
1852 btrfs_mark_bg_unused(bg);
1853 spin_unlock(&bg->lock);
1854 up_write(&space_info->groups_sem);
1855 goto next;
1856
1857 }
1858 /*
1859 * The block group might no longer meet the reclaim condition by
1860 * the time we get around to reclaiming it, so to avoid
1861 * reclaiming overly full block_groups, skip reclaiming them.
1862 *
1863 * Since the decision making process also depends on the amount
1864 * being freed, pass in a fake giant value to skip that extra
1865 * check, which is more meaningful when adding to the list in
1866 * the first place.
1867 */
1868 if (!should_reclaim_block_group(bg, bg->length)) {
1869 spin_unlock(&bg->lock);
1870 up_write(&space_info->groups_sem);
1871 goto next;
1872 }
1873 spin_unlock(&bg->lock);
1874
1875 /*
1876 * Get out fast, in case we're read-only or unmounting the
1877 * filesystem. It is OK to drop block groups from the list even
1878 * for the read-only case. As we did sb_start_write(),
1879 * "mount -o remount,ro" won't happen and read-only filesystem
1880 * means it is forced read-only due to a fatal error. So, it
1881 * never gets back to read-write to let us reclaim again.
1882 */
1883 if (btrfs_need_cleaner_sleep(fs_info)) {
1884 up_write(&space_info->groups_sem);
1885 goto next;
1886 }
1887
1888 /*
1889 * Cache the zone_unusable value before turning the block group
1890 * to read only. As soon as the blog group is read only it's
1891 * zone_unusable value gets moved to the block group's read-only
1892 * bytes and isn't available for calculations anymore.
1893 */
1894 zone_unusable = bg->zone_unusable;
1895 ret = inc_block_group_ro(bg, 0);
1896 up_write(&space_info->groups_sem);
1897 if (ret < 0)
1898 goto next;
1899
1900 btrfs_info(fs_info,
1901 "reclaiming chunk %llu with %llu%% used %llu%% unusable",
1902 bg->start,
1903 div64_u64(bg->used * 100, bg->length),
1904 div64_u64(zone_unusable * 100, bg->length));
1905 trace_btrfs_reclaim_block_group(bg);
1906 ret = btrfs_relocate_chunk(fs_info, bg->start);
1907 if (ret) {
1908 btrfs_dec_block_group_ro(bg);
1909 btrfs_err(fs_info, "error relocating chunk %llu",
1910 bg->start);
1911 }
1912
1913next:
1914 if (ret)
1915 btrfs_mark_bg_to_reclaim(bg);
1916 btrfs_put_block_group(bg);
1917
1918 mutex_unlock(&fs_info->reclaim_bgs_lock);
1919 /*
1920 * Reclaiming all the block groups in the list can take really
1921 * long. Prioritize cleaning up unused block groups.
1922 */
1923 btrfs_delete_unused_bgs(fs_info);
1924 /*
1925 * If we are interrupted by a balance, we can just bail out. The
1926 * cleaner thread restart again if necessary.
1927 */
1928 if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1929 goto end;
1930 spin_lock(&fs_info->unused_bgs_lock);
1931 }
1932 spin_unlock(&fs_info->unused_bgs_lock);
1933 mutex_unlock(&fs_info->reclaim_bgs_lock);
1934end:
1935 btrfs_exclop_finish(fs_info);
1936 sb_end_write(fs_info->sb);
1937}
1938
1939void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info)
1940{
1941 spin_lock(&fs_info->unused_bgs_lock);
1942 if (!list_empty(&fs_info->reclaim_bgs))
1943 queue_work(system_unbound_wq, &fs_info->reclaim_bgs_work);
1944 spin_unlock(&fs_info->unused_bgs_lock);
1945}
1946
1947void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg)
1948{
1949 struct btrfs_fs_info *fs_info = bg->fs_info;
1950
1951 spin_lock(&fs_info->unused_bgs_lock);
1952 if (list_empty(&bg->bg_list)) {
1953 btrfs_get_block_group(bg);
1954 trace_btrfs_add_reclaim_block_group(bg);
1955 list_add_tail(&bg->bg_list, &fs_info->reclaim_bgs);
1956 }
1957 spin_unlock(&fs_info->unused_bgs_lock);
1958}
1959
1960static int read_bg_from_eb(struct btrfs_fs_info *fs_info, struct btrfs_key *key,
1961 struct btrfs_path *path)
1962{
1963 struct btrfs_chunk_map *map;
1964 struct btrfs_block_group_item bg;
1965 struct extent_buffer *leaf;
1966 int slot;
1967 u64 flags;
1968 int ret = 0;
1969
1970 slot = path->slots[0];
1971 leaf = path->nodes[0];
1972
1973 map = btrfs_find_chunk_map(fs_info, key->objectid, key->offset);
1974 if (!map) {
1975 btrfs_err(fs_info,
1976 "logical %llu len %llu found bg but no related chunk",
1977 key->objectid, key->offset);
1978 return -ENOENT;
1979 }
1980
1981 if (map->start != key->objectid || map->chunk_len != key->offset) {
1982 btrfs_err(fs_info,
1983 "block group %llu len %llu mismatch with chunk %llu len %llu",
1984 key->objectid, key->offset, map->start, map->chunk_len);
1985 ret = -EUCLEAN;
1986 goto out_free_map;
1987 }
1988
1989 read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
1990 sizeof(bg));
1991 flags = btrfs_stack_block_group_flags(&bg) &
1992 BTRFS_BLOCK_GROUP_TYPE_MASK;
1993
1994 if (flags != (map->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1995 btrfs_err(fs_info,
1996"block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
1997 key->objectid, key->offset, flags,
1998 (BTRFS_BLOCK_GROUP_TYPE_MASK & map->type));
1999 ret = -EUCLEAN;
2000 }
2001
2002out_free_map:
2003 btrfs_free_chunk_map(map);
2004 return ret;
2005}
2006
2007static int find_first_block_group(struct btrfs_fs_info *fs_info,
2008 struct btrfs_path *path,
2009 struct btrfs_key *key)
2010{
2011 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2012 int ret;
2013 struct btrfs_key found_key;
2014
2015 btrfs_for_each_slot(root, key, &found_key, path, ret) {
2016 if (found_key.objectid >= key->objectid &&
2017 found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
2018 return read_bg_from_eb(fs_info, &found_key, path);
2019 }
2020 }
2021 return ret;
2022}
2023
2024static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2025{
2026 u64 extra_flags = chunk_to_extended(flags) &
2027 BTRFS_EXTENDED_PROFILE_MASK;
2028
2029 write_seqlock(&fs_info->profiles_lock);
2030 if (flags & BTRFS_BLOCK_GROUP_DATA)
2031 fs_info->avail_data_alloc_bits |= extra_flags;
2032 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2033 fs_info->avail_metadata_alloc_bits |= extra_flags;
2034 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2035 fs_info->avail_system_alloc_bits |= extra_flags;
2036 write_sequnlock(&fs_info->profiles_lock);
2037}
2038
2039/*
2040 * Map a physical disk address to a list of logical addresses.
2041 *
2042 * @fs_info: the filesystem
2043 * @chunk_start: logical address of block group
2044 * @physical: physical address to map to logical addresses
2045 * @logical: return array of logical addresses which map to @physical
2046 * @naddrs: length of @logical
2047 * @stripe_len: size of IO stripe for the given block group
2048 *
2049 * Maps a particular @physical disk address to a list of @logical addresses.
2050 * Used primarily to exclude those portions of a block group that contain super
2051 * block copies.
2052 */
2053int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
2054 u64 physical, u64 **logical, int *naddrs, int *stripe_len)
2055{
2056 struct btrfs_chunk_map *map;
2057 u64 *buf;
2058 u64 bytenr;
2059 u64 data_stripe_length;
2060 u64 io_stripe_size;
2061 int i, nr = 0;
2062 int ret = 0;
2063
2064 map = btrfs_get_chunk_map(fs_info, chunk_start, 1);
2065 if (IS_ERR(map))
2066 return -EIO;
2067
2068 data_stripe_length = map->stripe_size;
2069 io_stripe_size = BTRFS_STRIPE_LEN;
2070 chunk_start = map->start;
2071
2072 /* For RAID5/6 adjust to a full IO stripe length */
2073 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2074 io_stripe_size = btrfs_stripe_nr_to_offset(nr_data_stripes(map));
2075
2076 buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
2077 if (!buf) {
2078 ret = -ENOMEM;
2079 goto out;
2080 }
2081
2082 for (i = 0; i < map->num_stripes; i++) {
2083 bool already_inserted = false;
2084 u32 stripe_nr;
2085 u32 offset;
2086 int j;
2087
2088 if (!in_range(physical, map->stripes[i].physical,
2089 data_stripe_length))
2090 continue;
2091
2092 stripe_nr = (physical - map->stripes[i].physical) >>
2093 BTRFS_STRIPE_LEN_SHIFT;
2094 offset = (physical - map->stripes[i].physical) &
2095 BTRFS_STRIPE_LEN_MASK;
2096
2097 if (map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2098 BTRFS_BLOCK_GROUP_RAID10))
2099 stripe_nr = div_u64(stripe_nr * map->num_stripes + i,
2100 map->sub_stripes);
2101 /*
2102 * The remaining case would be for RAID56, multiply by
2103 * nr_data_stripes(). Alternatively, just use rmap_len below
2104 * instead of map->stripe_len
2105 */
2106 bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
2107
2108 /* Ensure we don't add duplicate addresses */
2109 for (j = 0; j < nr; j++) {
2110 if (buf[j] == bytenr) {
2111 already_inserted = true;
2112 break;
2113 }
2114 }
2115
2116 if (!already_inserted)
2117 buf[nr++] = bytenr;
2118 }
2119
2120 *logical = buf;
2121 *naddrs = nr;
2122 *stripe_len = io_stripe_size;
2123out:
2124 btrfs_free_chunk_map(map);
2125 return ret;
2126}
2127
2128static int exclude_super_stripes(struct btrfs_block_group *cache)
2129{
2130 struct btrfs_fs_info *fs_info = cache->fs_info;
2131 const bool zoned = btrfs_is_zoned(fs_info);
2132 u64 bytenr;
2133 u64 *logical;
2134 int stripe_len;
2135 int i, nr, ret;
2136
2137 if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
2138 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
2139 cache->bytes_super += stripe_len;
2140 ret = set_extent_bit(&fs_info->excluded_extents, cache->start,
2141 cache->start + stripe_len - 1,
2142 EXTENT_UPTODATE, NULL);
2143 if (ret)
2144 return ret;
2145 }
2146
2147 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2148 bytenr = btrfs_sb_offset(i);
2149 ret = btrfs_rmap_block(fs_info, cache->start,
2150 bytenr, &logical, &nr, &stripe_len);
2151 if (ret)
2152 return ret;
2153
2154 /* Shouldn't have super stripes in sequential zones */
2155 if (zoned && nr) {
2156 kfree(logical);
2157 btrfs_err(fs_info,
2158 "zoned: block group %llu must not contain super block",
2159 cache->start);
2160 return -EUCLEAN;
2161 }
2162
2163 while (nr--) {
2164 u64 len = min_t(u64, stripe_len,
2165 cache->start + cache->length - logical[nr]);
2166
2167 cache->bytes_super += len;
2168 ret = set_extent_bit(&fs_info->excluded_extents, logical[nr],
2169 logical[nr] + len - 1,
2170 EXTENT_UPTODATE, NULL);
2171 if (ret) {
2172 kfree(logical);
2173 return ret;
2174 }
2175 }
2176
2177 kfree(logical);
2178 }
2179 return 0;
2180}
2181
2182static struct btrfs_block_group *btrfs_create_block_group_cache(
2183 struct btrfs_fs_info *fs_info, u64 start)
2184{
2185 struct btrfs_block_group *cache;
2186
2187 cache = kzalloc(sizeof(*cache), GFP_NOFS);
2188 if (!cache)
2189 return NULL;
2190
2191 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
2192 GFP_NOFS);
2193 if (!cache->free_space_ctl) {
2194 kfree(cache);
2195 return NULL;
2196 }
2197
2198 cache->start = start;
2199
2200 cache->fs_info = fs_info;
2201 cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
2202
2203 cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
2204
2205 refcount_set(&cache->refs, 1);
2206 spin_lock_init(&cache->lock);
2207 init_rwsem(&cache->data_rwsem);
2208 INIT_LIST_HEAD(&cache->list);
2209 INIT_LIST_HEAD(&cache->cluster_list);
2210 INIT_LIST_HEAD(&cache->bg_list);
2211 INIT_LIST_HEAD(&cache->ro_list);
2212 INIT_LIST_HEAD(&cache->discard_list);
2213 INIT_LIST_HEAD(&cache->dirty_list);
2214 INIT_LIST_HEAD(&cache->io_list);
2215 INIT_LIST_HEAD(&cache->active_bg_list);
2216 btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
2217 atomic_set(&cache->frozen, 0);
2218 mutex_init(&cache->free_space_lock);
2219
2220 return cache;
2221}
2222
2223/*
2224 * Iterate all chunks and verify that each of them has the corresponding block
2225 * group
2226 */
2227static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
2228{
2229 u64 start = 0;
2230 int ret = 0;
2231
2232 while (1) {
2233 struct btrfs_chunk_map *map;
2234 struct btrfs_block_group *bg;
2235
2236 /*
2237 * btrfs_find_chunk_map() will return the first chunk map
2238 * intersecting the range, so setting @length to 1 is enough to
2239 * get the first chunk.
2240 */
2241 map = btrfs_find_chunk_map(fs_info, start, 1);
2242 if (!map)
2243 break;
2244
2245 bg = btrfs_lookup_block_group(fs_info, map->start);
2246 if (!bg) {
2247 btrfs_err(fs_info,
2248 "chunk start=%llu len=%llu doesn't have corresponding block group",
2249 map->start, map->chunk_len);
2250 ret = -EUCLEAN;
2251 btrfs_free_chunk_map(map);
2252 break;
2253 }
2254 if (bg->start != map->start || bg->length != map->chunk_len ||
2255 (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
2256 (map->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2257 btrfs_err(fs_info,
2258"chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
2259 map->start, map->chunk_len,
2260 map->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
2261 bg->start, bg->length,
2262 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
2263 ret = -EUCLEAN;
2264 btrfs_free_chunk_map(map);
2265 btrfs_put_block_group(bg);
2266 break;
2267 }
2268 start = map->start + map->chunk_len;
2269 btrfs_free_chunk_map(map);
2270 btrfs_put_block_group(bg);
2271 }
2272 return ret;
2273}
2274
2275static int read_one_block_group(struct btrfs_fs_info *info,
2276 struct btrfs_block_group_item *bgi,
2277 const struct btrfs_key *key,
2278 int need_clear)
2279{
2280 struct btrfs_block_group *cache;
2281 const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
2282 int ret;
2283
2284 ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
2285
2286 cache = btrfs_create_block_group_cache(info, key->objectid);
2287 if (!cache)
2288 return -ENOMEM;
2289
2290 cache->length = key->offset;
2291 cache->used = btrfs_stack_block_group_used(bgi);
2292 cache->commit_used = cache->used;
2293 cache->flags = btrfs_stack_block_group_flags(bgi);
2294 cache->global_root_id = btrfs_stack_block_group_chunk_objectid(bgi);
2295
2296 set_free_space_tree_thresholds(cache);
2297
2298 if (need_clear) {
2299 /*
2300 * When we mount with old space cache, we need to
2301 * set BTRFS_DC_CLEAR and set dirty flag.
2302 *
2303 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
2304 * truncate the old free space cache inode and
2305 * setup a new one.
2306 * b) Setting 'dirty flag' makes sure that we flush
2307 * the new space cache info onto disk.
2308 */
2309 if (btrfs_test_opt(info, SPACE_CACHE))
2310 cache->disk_cache_state = BTRFS_DC_CLEAR;
2311 }
2312 if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
2313 (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
2314 btrfs_err(info,
2315"bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2316 cache->start);
2317 ret = -EINVAL;
2318 goto error;
2319 }
2320
2321 ret = btrfs_load_block_group_zone_info(cache, false);
2322 if (ret) {
2323 btrfs_err(info, "zoned: failed to load zone info of bg %llu",
2324 cache->start);
2325 goto error;
2326 }
2327
2328 /*
2329 * We need to exclude the super stripes now so that the space info has
2330 * super bytes accounted for, otherwise we'll think we have more space
2331 * than we actually do.
2332 */
2333 ret = exclude_super_stripes(cache);
2334 if (ret) {
2335 /* We may have excluded something, so call this just in case. */
2336 btrfs_free_excluded_extents(cache);
2337 goto error;
2338 }
2339
2340 /*
2341 * For zoned filesystem, space after the allocation offset is the only
2342 * free space for a block group. So, we don't need any caching work.
2343 * btrfs_calc_zone_unusable() will set the amount of free space and
2344 * zone_unusable space.
2345 *
2346 * For regular filesystem, check for two cases, either we are full, and
2347 * therefore don't need to bother with the caching work since we won't
2348 * find any space, or we are empty, and we can just add all the space
2349 * in and be done with it. This saves us _a_lot_ of time, particularly
2350 * in the full case.
2351 */
2352 if (btrfs_is_zoned(info)) {
2353 btrfs_calc_zone_unusable(cache);
2354 /* Should not have any excluded extents. Just in case, though. */
2355 btrfs_free_excluded_extents(cache);
2356 } else if (cache->length == cache->used) {
2357 cache->cached = BTRFS_CACHE_FINISHED;
2358 btrfs_free_excluded_extents(cache);
2359 } else if (cache->used == 0) {
2360 cache->cached = BTRFS_CACHE_FINISHED;
2361 ret = btrfs_add_new_free_space(cache, cache->start,
2362 cache->start + cache->length, NULL);
2363 btrfs_free_excluded_extents(cache);
2364 if (ret)
2365 goto error;
2366 }
2367
2368 ret = btrfs_add_block_group_cache(info, cache);
2369 if (ret) {
2370 btrfs_remove_free_space_cache(cache);
2371 goto error;
2372 }
2373 trace_btrfs_add_block_group(info, cache, 0);
2374 btrfs_add_bg_to_space_info(info, cache);
2375
2376 set_avail_alloc_bits(info, cache->flags);
2377 if (btrfs_chunk_writeable(info, cache->start)) {
2378 if (cache->used == 0) {
2379 ASSERT(list_empty(&cache->bg_list));
2380 if (btrfs_test_opt(info, DISCARD_ASYNC))
2381 btrfs_discard_queue_work(&info->discard_ctl, cache);
2382 else
2383 btrfs_mark_bg_unused(cache);
2384 }
2385 } else {
2386 inc_block_group_ro(cache, 1);
2387 }
2388
2389 return 0;
2390error:
2391 btrfs_put_block_group(cache);
2392 return ret;
2393}
2394
2395static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
2396{
2397 struct rb_node *node;
2398 int ret = 0;
2399
2400 for (node = rb_first_cached(&fs_info->mapping_tree); node; node = rb_next(node)) {
2401 struct btrfs_chunk_map *map;
2402 struct btrfs_block_group *bg;
2403
2404 map = rb_entry(node, struct btrfs_chunk_map, rb_node);
2405 bg = btrfs_create_block_group_cache(fs_info, map->start);
2406 if (!bg) {
2407 ret = -ENOMEM;
2408 break;
2409 }
2410
2411 /* Fill dummy cache as FULL */
2412 bg->length = map->chunk_len;
2413 bg->flags = map->type;
2414 bg->cached = BTRFS_CACHE_FINISHED;
2415 bg->used = map->chunk_len;
2416 bg->flags = map->type;
2417 ret = btrfs_add_block_group_cache(fs_info, bg);
2418 /*
2419 * We may have some valid block group cache added already, in
2420 * that case we skip to the next one.
2421 */
2422 if (ret == -EEXIST) {
2423 ret = 0;
2424 btrfs_put_block_group(bg);
2425 continue;
2426 }
2427
2428 if (ret) {
2429 btrfs_remove_free_space_cache(bg);
2430 btrfs_put_block_group(bg);
2431 break;
2432 }
2433
2434 btrfs_add_bg_to_space_info(fs_info, bg);
2435
2436 set_avail_alloc_bits(fs_info, bg->flags);
2437 }
2438 if (!ret)
2439 btrfs_init_global_block_rsv(fs_info);
2440 return ret;
2441}
2442
2443int btrfs_read_block_groups(struct btrfs_fs_info *info)
2444{
2445 struct btrfs_root *root = btrfs_block_group_root(info);
2446 struct btrfs_path *path;
2447 int ret;
2448 struct btrfs_block_group *cache;
2449 struct btrfs_space_info *space_info;
2450 struct btrfs_key key;
2451 int need_clear = 0;
2452 u64 cache_gen;
2453
2454 /*
2455 * Either no extent root (with ibadroots rescue option) or we have
2456 * unsupported RO options. The fs can never be mounted read-write, so no
2457 * need to waste time searching block group items.
2458 *
2459 * This also allows new extent tree related changes to be RO compat,
2460 * no need for a full incompat flag.
2461 */
2462 if (!root || (btrfs_super_compat_ro_flags(info->super_copy) &
2463 ~BTRFS_FEATURE_COMPAT_RO_SUPP))
2464 return fill_dummy_bgs(info);
2465
2466 key.objectid = 0;
2467 key.offset = 0;
2468 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2469 path = btrfs_alloc_path();
2470 if (!path)
2471 return -ENOMEM;
2472
2473 cache_gen = btrfs_super_cache_generation(info->super_copy);
2474 if (btrfs_test_opt(info, SPACE_CACHE) &&
2475 btrfs_super_generation(info->super_copy) != cache_gen)
2476 need_clear = 1;
2477 if (btrfs_test_opt(info, CLEAR_CACHE))
2478 need_clear = 1;
2479
2480 while (1) {
2481 struct btrfs_block_group_item bgi;
2482 struct extent_buffer *leaf;
2483 int slot;
2484
2485 ret = find_first_block_group(info, path, &key);
2486 if (ret > 0)
2487 break;
2488 if (ret != 0)
2489 goto error;
2490
2491 leaf = path->nodes[0];
2492 slot = path->slots[0];
2493
2494 read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2495 sizeof(bgi));
2496
2497 btrfs_item_key_to_cpu(leaf, &key, slot);
2498 btrfs_release_path(path);
2499 ret = read_one_block_group(info, &bgi, &key, need_clear);
2500 if (ret < 0)
2501 goto error;
2502 key.objectid += key.offset;
2503 key.offset = 0;
2504 }
2505 btrfs_release_path(path);
2506
2507 list_for_each_entry(space_info, &info->space_info, list) {
2508 int i;
2509
2510 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2511 if (list_empty(&space_info->block_groups[i]))
2512 continue;
2513 cache = list_first_entry(&space_info->block_groups[i],
2514 struct btrfs_block_group,
2515 list);
2516 btrfs_sysfs_add_block_group_type(cache);
2517 }
2518
2519 if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2520 (BTRFS_BLOCK_GROUP_RAID10 |
2521 BTRFS_BLOCK_GROUP_RAID1_MASK |
2522 BTRFS_BLOCK_GROUP_RAID56_MASK |
2523 BTRFS_BLOCK_GROUP_DUP)))
2524 continue;
2525 /*
2526 * Avoid allocating from un-mirrored block group if there are
2527 * mirrored block groups.
2528 */
2529 list_for_each_entry(cache,
2530 &space_info->block_groups[BTRFS_RAID_RAID0],
2531 list)
2532 inc_block_group_ro(cache, 1);
2533 list_for_each_entry(cache,
2534 &space_info->block_groups[BTRFS_RAID_SINGLE],
2535 list)
2536 inc_block_group_ro(cache, 1);
2537 }
2538
2539 btrfs_init_global_block_rsv(info);
2540 ret = check_chunk_block_group_mappings(info);
2541error:
2542 btrfs_free_path(path);
2543 /*
2544 * We've hit some error while reading the extent tree, and have
2545 * rescue=ibadroots mount option.
2546 * Try to fill the tree using dummy block groups so that the user can
2547 * continue to mount and grab their data.
2548 */
2549 if (ret && btrfs_test_opt(info, IGNOREBADROOTS))
2550 ret = fill_dummy_bgs(info);
2551 return ret;
2552}
2553
2554/*
2555 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2556 * allocation.
2557 *
2558 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2559 * phases.
2560 */
2561static int insert_block_group_item(struct btrfs_trans_handle *trans,
2562 struct btrfs_block_group *block_group)
2563{
2564 struct btrfs_fs_info *fs_info = trans->fs_info;
2565 struct btrfs_block_group_item bgi;
2566 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2567 struct btrfs_key key;
2568 u64 old_commit_used;
2569 int ret;
2570
2571 spin_lock(&block_group->lock);
2572 btrfs_set_stack_block_group_used(&bgi, block_group->used);
2573 btrfs_set_stack_block_group_chunk_objectid(&bgi,
2574 block_group->global_root_id);
2575 btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2576 old_commit_used = block_group->commit_used;
2577 block_group->commit_used = block_group->used;
2578 key.objectid = block_group->start;
2579 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2580 key.offset = block_group->length;
2581 spin_unlock(&block_group->lock);
2582
2583 ret = btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2584 if (ret < 0) {
2585 spin_lock(&block_group->lock);
2586 block_group->commit_used = old_commit_used;
2587 spin_unlock(&block_group->lock);
2588 }
2589
2590 return ret;
2591}
2592
2593static int insert_dev_extent(struct btrfs_trans_handle *trans,
2594 struct btrfs_device *device, u64 chunk_offset,
2595 u64 start, u64 num_bytes)
2596{
2597 struct btrfs_fs_info *fs_info = device->fs_info;
2598 struct btrfs_root *root = fs_info->dev_root;
2599 struct btrfs_path *path;
2600 struct btrfs_dev_extent *extent;
2601 struct extent_buffer *leaf;
2602 struct btrfs_key key;
2603 int ret;
2604
2605 WARN_ON(!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state));
2606 WARN_ON(test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state));
2607 path = btrfs_alloc_path();
2608 if (!path)
2609 return -ENOMEM;
2610
2611 key.objectid = device->devid;
2612 key.type = BTRFS_DEV_EXTENT_KEY;
2613 key.offset = start;
2614 ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*extent));
2615 if (ret)
2616 goto out;
2617
2618 leaf = path->nodes[0];
2619 extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
2620 btrfs_set_dev_extent_chunk_tree(leaf, extent, BTRFS_CHUNK_TREE_OBJECTID);
2621 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
2622 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2623 btrfs_set_dev_extent_chunk_offset(leaf, extent, chunk_offset);
2624
2625 btrfs_set_dev_extent_length(leaf, extent, num_bytes);
2626 btrfs_mark_buffer_dirty(trans, leaf);
2627out:
2628 btrfs_free_path(path);
2629 return ret;
2630}
2631
2632/*
2633 * This function belongs to phase 2.
2634 *
2635 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2636 * phases.
2637 */
2638static int insert_dev_extents(struct btrfs_trans_handle *trans,
2639 u64 chunk_offset, u64 chunk_size)
2640{
2641 struct btrfs_fs_info *fs_info = trans->fs_info;
2642 struct btrfs_device *device;
2643 struct btrfs_chunk_map *map;
2644 u64 dev_offset;
2645 int i;
2646 int ret = 0;
2647
2648 map = btrfs_get_chunk_map(fs_info, chunk_offset, chunk_size);
2649 if (IS_ERR(map))
2650 return PTR_ERR(map);
2651
2652 /*
2653 * Take the device list mutex to prevent races with the final phase of
2654 * a device replace operation that replaces the device object associated
2655 * with the map's stripes, because the device object's id can change
2656 * at any time during that final phase of the device replace operation
2657 * (dev-replace.c:btrfs_dev_replace_finishing()), so we could grab the
2658 * replaced device and then see it with an ID of BTRFS_DEV_REPLACE_DEVID,
2659 * resulting in persisting a device extent item with such ID.
2660 */
2661 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2662 for (i = 0; i < map->num_stripes; i++) {
2663 device = map->stripes[i].dev;
2664 dev_offset = map->stripes[i].physical;
2665
2666 ret = insert_dev_extent(trans, device, chunk_offset, dev_offset,
2667 map->stripe_size);
2668 if (ret)
2669 break;
2670 }
2671 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2672
2673 btrfs_free_chunk_map(map);
2674 return ret;
2675}
2676
2677/*
2678 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2679 * chunk allocation.
2680 *
2681 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2682 * phases.
2683 */
2684void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2685{
2686 struct btrfs_fs_info *fs_info = trans->fs_info;
2687 struct btrfs_block_group *block_group;
2688 int ret = 0;
2689
2690 while (!list_empty(&trans->new_bgs)) {
2691 int index;
2692
2693 block_group = list_first_entry(&trans->new_bgs,
2694 struct btrfs_block_group,
2695 bg_list);
2696 if (ret)
2697 goto next;
2698
2699 index = btrfs_bg_flags_to_raid_index(block_group->flags);
2700
2701 ret = insert_block_group_item(trans, block_group);
2702 if (ret)
2703 btrfs_abort_transaction(trans, ret);
2704 if (!test_bit(BLOCK_GROUP_FLAG_CHUNK_ITEM_INSERTED,
2705 &block_group->runtime_flags)) {
2706 mutex_lock(&fs_info->chunk_mutex);
2707 ret = btrfs_chunk_alloc_add_chunk_item(trans, block_group);
2708 mutex_unlock(&fs_info->chunk_mutex);
2709 if (ret)
2710 btrfs_abort_transaction(trans, ret);
2711 }
2712 ret = insert_dev_extents(trans, block_group->start,
2713 block_group->length);
2714 if (ret)
2715 btrfs_abort_transaction(trans, ret);
2716 add_block_group_free_space(trans, block_group);
2717
2718 /*
2719 * If we restriped during balance, we may have added a new raid
2720 * type, so now add the sysfs entries when it is safe to do so.
2721 * We don't have to worry about locking here as it's handled in
2722 * btrfs_sysfs_add_block_group_type.
2723 */
2724 if (block_group->space_info->block_group_kobjs[index] == NULL)
2725 btrfs_sysfs_add_block_group_type(block_group);
2726
2727 /* Already aborted the transaction if it failed. */
2728next:
2729 btrfs_dec_delayed_refs_rsv_bg_inserts(fs_info);
2730 list_del_init(&block_group->bg_list);
2731 clear_bit(BLOCK_GROUP_FLAG_NEW, &block_group->runtime_flags);
2732
2733 /*
2734 * If the block group is still unused, add it to the list of
2735 * unused block groups. The block group may have been created in
2736 * order to satisfy a space reservation, in which case the
2737 * extent allocation only happens later. But often we don't
2738 * actually need to allocate space that we previously reserved,
2739 * so the block group may become unused for a long time. For
2740 * example for metadata we generally reserve space for a worst
2741 * possible scenario, but then don't end up allocating all that
2742 * space or none at all (due to no need to COW, extent buffers
2743 * were already COWed in the current transaction and still
2744 * unwritten, tree heights lower than the maximum possible
2745 * height, etc). For data we generally reserve the axact amount
2746 * of space we are going to allocate later, the exception is
2747 * when using compression, as we must reserve space based on the
2748 * uncompressed data size, because the compression is only done
2749 * when writeback triggered and we don't know how much space we
2750 * are actually going to need, so we reserve the uncompressed
2751 * size because the data may be uncompressible in the worst case.
2752 */
2753 if (ret == 0) {
2754 bool used;
2755
2756 spin_lock(&block_group->lock);
2757 used = btrfs_is_block_group_used(block_group);
2758 spin_unlock(&block_group->lock);
2759
2760 if (!used)
2761 btrfs_mark_bg_unused(block_group);
2762 }
2763 }
2764 btrfs_trans_release_chunk_metadata(trans);
2765}
2766
2767/*
2768 * For extent tree v2 we use the block_group_item->chunk_offset to point at our
2769 * global root id. For v1 it's always set to BTRFS_FIRST_CHUNK_TREE_OBJECTID.
2770 */
2771static u64 calculate_global_root_id(struct btrfs_fs_info *fs_info, u64 offset)
2772{
2773 u64 div = SZ_1G;
2774 u64 index;
2775
2776 if (!btrfs_fs_incompat(fs_info, EXTENT_TREE_V2))
2777 return BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2778
2779 /* If we have a smaller fs index based on 128MiB. */
2780 if (btrfs_super_total_bytes(fs_info->super_copy) <= (SZ_1G * 10ULL))
2781 div = SZ_128M;
2782
2783 offset = div64_u64(offset, div);
2784 div64_u64_rem(offset, fs_info->nr_global_roots, &index);
2785 return index;
2786}
2787
2788struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
2789 u64 type,
2790 u64 chunk_offset, u64 size)
2791{
2792 struct btrfs_fs_info *fs_info = trans->fs_info;
2793 struct btrfs_block_group *cache;
2794 int ret;
2795
2796 btrfs_set_log_full_commit(trans);
2797
2798 cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2799 if (!cache)
2800 return ERR_PTR(-ENOMEM);
2801
2802 /*
2803 * Mark it as new before adding it to the rbtree of block groups or any
2804 * list, so that no other task finds it and calls btrfs_mark_bg_unused()
2805 * before the new flag is set.
2806 */
2807 set_bit(BLOCK_GROUP_FLAG_NEW, &cache->runtime_flags);
2808
2809 cache->length = size;
2810 set_free_space_tree_thresholds(cache);
2811 cache->flags = type;
2812 cache->cached = BTRFS_CACHE_FINISHED;
2813 cache->global_root_id = calculate_global_root_id(fs_info, cache->start);
2814
2815 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2816 set_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE, &cache->runtime_flags);
2817
2818 ret = btrfs_load_block_group_zone_info(cache, true);
2819 if (ret) {
2820 btrfs_put_block_group(cache);
2821 return ERR_PTR(ret);
2822 }
2823
2824 ret = exclude_super_stripes(cache);
2825 if (ret) {
2826 /* We may have excluded something, so call this just in case */
2827 btrfs_free_excluded_extents(cache);
2828 btrfs_put_block_group(cache);
2829 return ERR_PTR(ret);
2830 }
2831
2832 ret = btrfs_add_new_free_space(cache, chunk_offset, chunk_offset + size, NULL);
2833 btrfs_free_excluded_extents(cache);
2834 if (ret) {
2835 btrfs_put_block_group(cache);
2836 return ERR_PTR(ret);
2837 }
2838
2839 /*
2840 * Ensure the corresponding space_info object is created and
2841 * assigned to our block group. We want our bg to be added to the rbtree
2842 * with its ->space_info set.
2843 */
2844 cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2845 ASSERT(cache->space_info);
2846
2847 ret = btrfs_add_block_group_cache(fs_info, cache);
2848 if (ret) {
2849 btrfs_remove_free_space_cache(cache);
2850 btrfs_put_block_group(cache);
2851 return ERR_PTR(ret);
2852 }
2853
2854 /*
2855 * Now that our block group has its ->space_info set and is inserted in
2856 * the rbtree, update the space info's counters.
2857 */
2858 trace_btrfs_add_block_group(fs_info, cache, 1);
2859 btrfs_add_bg_to_space_info(fs_info, cache);
2860 btrfs_update_global_block_rsv(fs_info);
2861
2862#ifdef CONFIG_BTRFS_DEBUG
2863 if (btrfs_should_fragment_free_space(cache)) {
2864 cache->space_info->bytes_used += size >> 1;
2865 fragment_free_space(cache);
2866 }
2867#endif
2868
2869 list_add_tail(&cache->bg_list, &trans->new_bgs);
2870 btrfs_inc_delayed_refs_rsv_bg_inserts(fs_info);
2871
2872 set_avail_alloc_bits(fs_info, type);
2873 return cache;
2874}
2875
2876/*
2877 * Mark one block group RO, can be called several times for the same block
2878 * group.
2879 *
2880 * @cache: the destination block group
2881 * @do_chunk_alloc: whether need to do chunk pre-allocation, this is to
2882 * ensure we still have some free space after marking this
2883 * block group RO.
2884 */
2885int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2886 bool do_chunk_alloc)
2887{
2888 struct btrfs_fs_info *fs_info = cache->fs_info;
2889 struct btrfs_trans_handle *trans;
2890 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2891 u64 alloc_flags;
2892 int ret;
2893 bool dirty_bg_running;
2894
2895 /*
2896 * This can only happen when we are doing read-only scrub on read-only
2897 * mount.
2898 * In that case we should not start a new transaction on read-only fs.
2899 * Thus here we skip all chunk allocations.
2900 */
2901 if (sb_rdonly(fs_info->sb)) {
2902 mutex_lock(&fs_info->ro_block_group_mutex);
2903 ret = inc_block_group_ro(cache, 0);
2904 mutex_unlock(&fs_info->ro_block_group_mutex);
2905 return ret;
2906 }
2907
2908 do {
2909 trans = btrfs_join_transaction(root);
2910 if (IS_ERR(trans))
2911 return PTR_ERR(trans);
2912
2913 dirty_bg_running = false;
2914
2915 /*
2916 * We're not allowed to set block groups readonly after the dirty
2917 * block group cache has started writing. If it already started,
2918 * back off and let this transaction commit.
2919 */
2920 mutex_lock(&fs_info->ro_block_group_mutex);
2921 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2922 u64 transid = trans->transid;
2923
2924 mutex_unlock(&fs_info->ro_block_group_mutex);
2925 btrfs_end_transaction(trans);
2926
2927 ret = btrfs_wait_for_commit(fs_info, transid);
2928 if (ret)
2929 return ret;
2930 dirty_bg_running = true;
2931 }
2932 } while (dirty_bg_running);
2933
2934 if (do_chunk_alloc) {
2935 /*
2936 * If we are changing raid levels, try to allocate a
2937 * corresponding block group with the new raid level.
2938 */
2939 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2940 if (alloc_flags != cache->flags) {
2941 ret = btrfs_chunk_alloc(trans, alloc_flags,
2942 CHUNK_ALLOC_FORCE);
2943 /*
2944 * ENOSPC is allowed here, we may have enough space
2945 * already allocated at the new raid level to carry on
2946 */
2947 if (ret == -ENOSPC)
2948 ret = 0;
2949 if (ret < 0)
2950 goto out;
2951 }
2952 }
2953
2954 ret = inc_block_group_ro(cache, 0);
2955 if (!ret)
2956 goto out;
2957 if (ret == -ETXTBSY)
2958 goto unlock_out;
2959
2960 /*
2961 * Skip chunk allocation if the bg is SYSTEM, this is to avoid system
2962 * chunk allocation storm to exhaust the system chunk array. Otherwise
2963 * we still want to try our best to mark the block group read-only.
2964 */
2965 if (!do_chunk_alloc && ret == -ENOSPC &&
2966 (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM))
2967 goto unlock_out;
2968
2969 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2970 ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2971 if (ret < 0)
2972 goto out;
2973 /*
2974 * We have allocated a new chunk. We also need to activate that chunk to
2975 * grant metadata tickets for zoned filesystem.
2976 */
2977 ret = btrfs_zoned_activate_one_bg(fs_info, cache->space_info, true);
2978 if (ret < 0)
2979 goto out;
2980
2981 ret = inc_block_group_ro(cache, 0);
2982 if (ret == -ETXTBSY)
2983 goto unlock_out;
2984out:
2985 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2986 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2987 mutex_lock(&fs_info->chunk_mutex);
2988 check_system_chunk(trans, alloc_flags);
2989 mutex_unlock(&fs_info->chunk_mutex);
2990 }
2991unlock_out:
2992 mutex_unlock(&fs_info->ro_block_group_mutex);
2993
2994 btrfs_end_transaction(trans);
2995 return ret;
2996}
2997
2998void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
2999{
3000 struct btrfs_space_info *sinfo = cache->space_info;
3001 u64 num_bytes;
3002
3003 BUG_ON(!cache->ro);
3004
3005 spin_lock(&sinfo->lock);
3006 spin_lock(&cache->lock);
3007 if (!--cache->ro) {
3008 if (btrfs_is_zoned(cache->fs_info)) {
3009 /* Migrate zone_unusable bytes back */
3010 cache->zone_unusable =
3011 (cache->alloc_offset - cache->used) +
3012 (cache->length - cache->zone_capacity);
3013 sinfo->bytes_zone_unusable += cache->zone_unusable;
3014 sinfo->bytes_readonly -= cache->zone_unusable;
3015 }
3016 num_bytes = cache->length - cache->reserved -
3017 cache->pinned - cache->bytes_super -
3018 cache->zone_unusable - cache->used;
3019 sinfo->bytes_readonly -= num_bytes;
3020 list_del_init(&cache->ro_list);
3021 }
3022 spin_unlock(&cache->lock);
3023 spin_unlock(&sinfo->lock);
3024}
3025
3026static int update_block_group_item(struct btrfs_trans_handle *trans,
3027 struct btrfs_path *path,
3028 struct btrfs_block_group *cache)
3029{
3030 struct btrfs_fs_info *fs_info = trans->fs_info;
3031 int ret;
3032 struct btrfs_root *root = btrfs_block_group_root(fs_info);
3033 unsigned long bi;
3034 struct extent_buffer *leaf;
3035 struct btrfs_block_group_item bgi;
3036 struct btrfs_key key;
3037 u64 old_commit_used;
3038 u64 used;
3039
3040 /*
3041 * Block group items update can be triggered out of commit transaction
3042 * critical section, thus we need a consistent view of used bytes.
3043 * We cannot use cache->used directly outside of the spin lock, as it
3044 * may be changed.
3045 */
3046 spin_lock(&cache->lock);
3047 old_commit_used = cache->commit_used;
3048 used = cache->used;
3049 /* No change in used bytes, can safely skip it. */
3050 if (cache->commit_used == used) {
3051 spin_unlock(&cache->lock);
3052 return 0;
3053 }
3054 cache->commit_used = used;
3055 spin_unlock(&cache->lock);
3056
3057 key.objectid = cache->start;
3058 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
3059 key.offset = cache->length;
3060
3061 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3062 if (ret) {
3063 if (ret > 0)
3064 ret = -ENOENT;
3065 goto fail;
3066 }
3067
3068 leaf = path->nodes[0];
3069 bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3070 btrfs_set_stack_block_group_used(&bgi, used);
3071 btrfs_set_stack_block_group_chunk_objectid(&bgi,
3072 cache->global_root_id);
3073 btrfs_set_stack_block_group_flags(&bgi, cache->flags);
3074 write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
3075 btrfs_mark_buffer_dirty(trans, leaf);
3076fail:
3077 btrfs_release_path(path);
3078 /*
3079 * We didn't update the block group item, need to revert commit_used
3080 * unless the block group item didn't exist yet - this is to prevent a
3081 * race with a concurrent insertion of the block group item, with
3082 * insert_block_group_item(), that happened just after we attempted to
3083 * update. In that case we would reset commit_used to 0 just after the
3084 * insertion set it to a value greater than 0 - if the block group later
3085 * becomes with 0 used bytes, we would incorrectly skip its update.
3086 */
3087 if (ret < 0 && ret != -ENOENT) {
3088 spin_lock(&cache->lock);
3089 cache->commit_used = old_commit_used;
3090 spin_unlock(&cache->lock);
3091 }
3092 return ret;
3093
3094}
3095
3096static int cache_save_setup(struct btrfs_block_group *block_group,
3097 struct btrfs_trans_handle *trans,
3098 struct btrfs_path *path)
3099{
3100 struct btrfs_fs_info *fs_info = block_group->fs_info;
3101 struct inode *inode = NULL;
3102 struct extent_changeset *data_reserved = NULL;
3103 u64 alloc_hint = 0;
3104 int dcs = BTRFS_DC_ERROR;
3105 u64 cache_size = 0;
3106 int retries = 0;
3107 int ret = 0;
3108
3109 if (!btrfs_test_opt(fs_info, SPACE_CACHE))
3110 return 0;
3111
3112 /*
3113 * If this block group is smaller than 100 megs don't bother caching the
3114 * block group.
3115 */
3116 if (block_group->length < (100 * SZ_1M)) {
3117 spin_lock(&block_group->lock);
3118 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3119 spin_unlock(&block_group->lock);
3120 return 0;
3121 }
3122
3123 if (TRANS_ABORTED(trans))
3124 return 0;
3125again:
3126 inode = lookup_free_space_inode(block_group, path);
3127 if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3128 ret = PTR_ERR(inode);
3129 btrfs_release_path(path);
3130 goto out;
3131 }
3132
3133 if (IS_ERR(inode)) {
3134 BUG_ON(retries);
3135 retries++;
3136
3137 if (block_group->ro)
3138 goto out_free;
3139
3140 ret = create_free_space_inode(trans, block_group, path);
3141 if (ret)
3142 goto out_free;
3143 goto again;
3144 }
3145
3146 /*
3147 * We want to set the generation to 0, that way if anything goes wrong
3148 * from here on out we know not to trust this cache when we load up next
3149 * time.
3150 */
3151 BTRFS_I(inode)->generation = 0;
3152 ret = btrfs_update_inode(trans, BTRFS_I(inode));
3153 if (ret) {
3154 /*
3155 * So theoretically we could recover from this, simply set the
3156 * super cache generation to 0 so we know to invalidate the
3157 * cache, but then we'd have to keep track of the block groups
3158 * that fail this way so we know we _have_ to reset this cache
3159 * before the next commit or risk reading stale cache. So to
3160 * limit our exposure to horrible edge cases lets just abort the
3161 * transaction, this only happens in really bad situations
3162 * anyway.
3163 */
3164 btrfs_abort_transaction(trans, ret);
3165 goto out_put;
3166 }
3167 WARN_ON(ret);
3168
3169 /* We've already setup this transaction, go ahead and exit */
3170 if (block_group->cache_generation == trans->transid &&
3171 i_size_read(inode)) {
3172 dcs = BTRFS_DC_SETUP;
3173 goto out_put;
3174 }
3175
3176 if (i_size_read(inode) > 0) {
3177 ret = btrfs_check_trunc_cache_free_space(fs_info,
3178 &fs_info->global_block_rsv);
3179 if (ret)
3180 goto out_put;
3181
3182 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3183 if (ret)
3184 goto out_put;
3185 }
3186
3187 spin_lock(&block_group->lock);
3188 if (block_group->cached != BTRFS_CACHE_FINISHED ||
3189 !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3190 /*
3191 * don't bother trying to write stuff out _if_
3192 * a) we're not cached,
3193 * b) we're with nospace_cache mount option,
3194 * c) we're with v2 space_cache (FREE_SPACE_TREE).
3195 */
3196 dcs = BTRFS_DC_WRITTEN;
3197 spin_unlock(&block_group->lock);
3198 goto out_put;
3199 }
3200 spin_unlock(&block_group->lock);
3201
3202 /*
3203 * We hit an ENOSPC when setting up the cache in this transaction, just
3204 * skip doing the setup, we've already cleared the cache so we're safe.
3205 */
3206 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3207 ret = -ENOSPC;
3208 goto out_put;
3209 }
3210
3211 /*
3212 * Try to preallocate enough space based on how big the block group is.
3213 * Keep in mind this has to include any pinned space which could end up
3214 * taking up quite a bit since it's not folded into the other space
3215 * cache.
3216 */
3217 cache_size = div_u64(block_group->length, SZ_256M);
3218 if (!cache_size)
3219 cache_size = 1;
3220
3221 cache_size *= 16;
3222 cache_size *= fs_info->sectorsize;
3223
3224 ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
3225 cache_size, false);
3226 if (ret)
3227 goto out_put;
3228
3229 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, cache_size,
3230 cache_size, cache_size,
3231 &alloc_hint);
3232 /*
3233 * Our cache requires contiguous chunks so that we don't modify a bunch
3234 * of metadata or split extents when writing the cache out, which means
3235 * we can enospc if we are heavily fragmented in addition to just normal
3236 * out of space conditions. So if we hit this just skip setting up any
3237 * other block groups for this transaction, maybe we'll unpin enough
3238 * space the next time around.
3239 */
3240 if (!ret)
3241 dcs = BTRFS_DC_SETUP;
3242 else if (ret == -ENOSPC)
3243 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3244
3245out_put:
3246 iput(inode);
3247out_free:
3248 btrfs_release_path(path);
3249out:
3250 spin_lock(&block_group->lock);
3251 if (!ret && dcs == BTRFS_DC_SETUP)
3252 block_group->cache_generation = trans->transid;
3253 block_group->disk_cache_state = dcs;
3254 spin_unlock(&block_group->lock);
3255
3256 extent_changeset_free(data_reserved);
3257 return ret;
3258}
3259
3260int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
3261{
3262 struct btrfs_fs_info *fs_info = trans->fs_info;
3263 struct btrfs_block_group *cache, *tmp;
3264 struct btrfs_transaction *cur_trans = trans->transaction;
3265 struct btrfs_path *path;
3266
3267 if (list_empty(&cur_trans->dirty_bgs) ||
3268 !btrfs_test_opt(fs_info, SPACE_CACHE))
3269 return 0;
3270
3271 path = btrfs_alloc_path();
3272 if (!path)
3273 return -ENOMEM;
3274
3275 /* Could add new block groups, use _safe just in case */
3276 list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3277 dirty_list) {
3278 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3279 cache_save_setup(cache, trans, path);
3280 }
3281
3282 btrfs_free_path(path);
3283 return 0;
3284}
3285
3286/*
3287 * Transaction commit does final block group cache writeback during a critical
3288 * section where nothing is allowed to change the FS. This is required in
3289 * order for the cache to actually match the block group, but can introduce a
3290 * lot of latency into the commit.
3291 *
3292 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
3293 * There's a chance we'll have to redo some of it if the block group changes
3294 * again during the commit, but it greatly reduces the commit latency by
3295 * getting rid of the easy block groups while we're still allowing others to
3296 * join the commit.
3297 */
3298int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3299{
3300 struct btrfs_fs_info *fs_info = trans->fs_info;
3301 struct btrfs_block_group *cache;
3302 struct btrfs_transaction *cur_trans = trans->transaction;
3303 int ret = 0;
3304 int should_put;
3305 struct btrfs_path *path = NULL;
3306 LIST_HEAD(dirty);
3307 struct list_head *io = &cur_trans->io_bgs;
3308 int loops = 0;
3309
3310 spin_lock(&cur_trans->dirty_bgs_lock);
3311 if (list_empty(&cur_trans->dirty_bgs)) {
3312 spin_unlock(&cur_trans->dirty_bgs_lock);
3313 return 0;
3314 }
3315 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3316 spin_unlock(&cur_trans->dirty_bgs_lock);
3317
3318again:
3319 /* Make sure all the block groups on our dirty list actually exist */
3320 btrfs_create_pending_block_groups(trans);
3321
3322 if (!path) {
3323 path = btrfs_alloc_path();
3324 if (!path) {
3325 ret = -ENOMEM;
3326 goto out;
3327 }
3328 }
3329
3330 /*
3331 * cache_write_mutex is here only to save us from balance or automatic
3332 * removal of empty block groups deleting this block group while we are
3333 * writing out the cache
3334 */
3335 mutex_lock(&trans->transaction->cache_write_mutex);
3336 while (!list_empty(&dirty)) {
3337 bool drop_reserve = true;
3338
3339 cache = list_first_entry(&dirty, struct btrfs_block_group,
3340 dirty_list);
3341 /*
3342 * This can happen if something re-dirties a block group that
3343 * is already under IO. Just wait for it to finish and then do
3344 * it all again
3345 */
3346 if (!list_empty(&cache->io_list)) {
3347 list_del_init(&cache->io_list);
3348 btrfs_wait_cache_io(trans, cache, path);
3349 btrfs_put_block_group(cache);
3350 }
3351
3352
3353 /*
3354 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
3355 * it should update the cache_state. Don't delete until after
3356 * we wait.
3357 *
3358 * Since we're not running in the commit critical section
3359 * we need the dirty_bgs_lock to protect from update_block_group
3360 */
3361 spin_lock(&cur_trans->dirty_bgs_lock);
3362 list_del_init(&cache->dirty_list);
3363 spin_unlock(&cur_trans->dirty_bgs_lock);
3364
3365 should_put = 1;
3366
3367 cache_save_setup(cache, trans, path);
3368
3369 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3370 cache->io_ctl.inode = NULL;
3371 ret = btrfs_write_out_cache(trans, cache, path);
3372 if (ret == 0 && cache->io_ctl.inode) {
3373 should_put = 0;
3374
3375 /*
3376 * The cache_write_mutex is protecting the
3377 * io_list, also refer to the definition of
3378 * btrfs_transaction::io_bgs for more details
3379 */
3380 list_add_tail(&cache->io_list, io);
3381 } else {
3382 /*
3383 * If we failed to write the cache, the
3384 * generation will be bad and life goes on
3385 */
3386 ret = 0;
3387 }
3388 }
3389 if (!ret) {
3390 ret = update_block_group_item(trans, path, cache);
3391 /*
3392 * Our block group might still be attached to the list
3393 * of new block groups in the transaction handle of some
3394 * other task (struct btrfs_trans_handle->new_bgs). This
3395 * means its block group item isn't yet in the extent
3396 * tree. If this happens ignore the error, as we will
3397 * try again later in the critical section of the
3398 * transaction commit.
3399 */
3400 if (ret == -ENOENT) {
3401 ret = 0;
3402 spin_lock(&cur_trans->dirty_bgs_lock);
3403 if (list_empty(&cache->dirty_list)) {
3404 list_add_tail(&cache->dirty_list,
3405 &cur_trans->dirty_bgs);
3406 btrfs_get_block_group(cache);
3407 drop_reserve = false;
3408 }
3409 spin_unlock(&cur_trans->dirty_bgs_lock);
3410 } else if (ret) {
3411 btrfs_abort_transaction(trans, ret);
3412 }
3413 }
3414
3415 /* If it's not on the io list, we need to put the block group */
3416 if (should_put)
3417 btrfs_put_block_group(cache);
3418 if (drop_reserve)
3419 btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
3420 /*
3421 * Avoid blocking other tasks for too long. It might even save
3422 * us from writing caches for block groups that are going to be
3423 * removed.
3424 */
3425 mutex_unlock(&trans->transaction->cache_write_mutex);
3426 if (ret)
3427 goto out;
3428 mutex_lock(&trans->transaction->cache_write_mutex);
3429 }
3430 mutex_unlock(&trans->transaction->cache_write_mutex);
3431
3432 /*
3433 * Go through delayed refs for all the stuff we've just kicked off
3434 * and then loop back (just once)
3435 */
3436 if (!ret)
3437 ret = btrfs_run_delayed_refs(trans, 0);
3438 if (!ret && loops == 0) {
3439 loops++;
3440 spin_lock(&cur_trans->dirty_bgs_lock);
3441 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3442 /*
3443 * dirty_bgs_lock protects us from concurrent block group
3444 * deletes too (not just cache_write_mutex).
3445 */
3446 if (!list_empty(&dirty)) {
3447 spin_unlock(&cur_trans->dirty_bgs_lock);
3448 goto again;
3449 }
3450 spin_unlock(&cur_trans->dirty_bgs_lock);
3451 }
3452out:
3453 if (ret < 0) {
3454 spin_lock(&cur_trans->dirty_bgs_lock);
3455 list_splice_init(&dirty, &cur_trans->dirty_bgs);
3456 spin_unlock(&cur_trans->dirty_bgs_lock);
3457 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3458 }
3459
3460 btrfs_free_path(path);
3461 return ret;
3462}
3463
3464int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3465{
3466 struct btrfs_fs_info *fs_info = trans->fs_info;
3467 struct btrfs_block_group *cache;
3468 struct btrfs_transaction *cur_trans = trans->transaction;
3469 int ret = 0;
3470 int should_put;
3471 struct btrfs_path *path;
3472 struct list_head *io = &cur_trans->io_bgs;
3473
3474 path = btrfs_alloc_path();
3475 if (!path)
3476 return -ENOMEM;
3477
3478 /*
3479 * Even though we are in the critical section of the transaction commit,
3480 * we can still have concurrent tasks adding elements to this
3481 * transaction's list of dirty block groups. These tasks correspond to
3482 * endio free space workers started when writeback finishes for a
3483 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3484 * allocate new block groups as a result of COWing nodes of the root
3485 * tree when updating the free space inode. The writeback for the space
3486 * caches is triggered by an earlier call to
3487 * btrfs_start_dirty_block_groups() and iterations of the following
3488 * loop.
3489 * Also we want to do the cache_save_setup first and then run the
3490 * delayed refs to make sure we have the best chance at doing this all
3491 * in one shot.
3492 */
3493 spin_lock(&cur_trans->dirty_bgs_lock);
3494 while (!list_empty(&cur_trans->dirty_bgs)) {
3495 cache = list_first_entry(&cur_trans->dirty_bgs,
3496 struct btrfs_block_group,
3497 dirty_list);
3498
3499 /*
3500 * This can happen if cache_save_setup re-dirties a block group
3501 * that is already under IO. Just wait for it to finish and
3502 * then do it all again
3503 */
3504 if (!list_empty(&cache->io_list)) {
3505 spin_unlock(&cur_trans->dirty_bgs_lock);
3506 list_del_init(&cache->io_list);
3507 btrfs_wait_cache_io(trans, cache, path);
3508 btrfs_put_block_group(cache);
3509 spin_lock(&cur_trans->dirty_bgs_lock);
3510 }
3511
3512 /*
3513 * Don't remove from the dirty list until after we've waited on
3514 * any pending IO
3515 */
3516 list_del_init(&cache->dirty_list);
3517 spin_unlock(&cur_trans->dirty_bgs_lock);
3518 should_put = 1;
3519
3520 cache_save_setup(cache, trans, path);
3521
3522 if (!ret)
3523 ret = btrfs_run_delayed_refs(trans, U64_MAX);
3524
3525 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3526 cache->io_ctl.inode = NULL;
3527 ret = btrfs_write_out_cache(trans, cache, path);
3528 if (ret == 0 && cache->io_ctl.inode) {
3529 should_put = 0;
3530 list_add_tail(&cache->io_list, io);
3531 } else {
3532 /*
3533 * If we failed to write the cache, the
3534 * generation will be bad and life goes on
3535 */
3536 ret = 0;
3537 }
3538 }
3539 if (!ret) {
3540 ret = update_block_group_item(trans, path, cache);
3541 /*
3542 * One of the free space endio workers might have
3543 * created a new block group while updating a free space
3544 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3545 * and hasn't released its transaction handle yet, in
3546 * which case the new block group is still attached to
3547 * its transaction handle and its creation has not
3548 * finished yet (no block group item in the extent tree
3549 * yet, etc). If this is the case, wait for all free
3550 * space endio workers to finish and retry. This is a
3551 * very rare case so no need for a more efficient and
3552 * complex approach.
3553 */
3554 if (ret == -ENOENT) {
3555 wait_event(cur_trans->writer_wait,
3556 atomic_read(&cur_trans->num_writers) == 1);
3557 ret = update_block_group_item(trans, path, cache);
3558 }
3559 if (ret)
3560 btrfs_abort_transaction(trans, ret);
3561 }
3562
3563 /* If its not on the io list, we need to put the block group */
3564 if (should_put)
3565 btrfs_put_block_group(cache);
3566 btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
3567 spin_lock(&cur_trans->dirty_bgs_lock);
3568 }
3569 spin_unlock(&cur_trans->dirty_bgs_lock);
3570
3571 /*
3572 * Refer to the definition of io_bgs member for details why it's safe
3573 * to use it without any locking
3574 */
3575 while (!list_empty(io)) {
3576 cache = list_first_entry(io, struct btrfs_block_group,
3577 io_list);
3578 list_del_init(&cache->io_list);
3579 btrfs_wait_cache_io(trans, cache, path);
3580 btrfs_put_block_group(cache);
3581 }
3582
3583 btrfs_free_path(path);
3584 return ret;
3585}
3586
3587int btrfs_update_block_group(struct btrfs_trans_handle *trans,
3588 u64 bytenr, u64 num_bytes, bool alloc)
3589{
3590 struct btrfs_fs_info *info = trans->fs_info;
3591 struct btrfs_space_info *space_info;
3592 struct btrfs_block_group *cache;
3593 u64 old_val;
3594 bool reclaim = false;
3595 bool bg_already_dirty = true;
3596 int factor;
3597
3598 /* Block accounting for super block */
3599 spin_lock(&info->delalloc_root_lock);
3600 old_val = btrfs_super_bytes_used(info->super_copy);
3601 if (alloc)
3602 old_val += num_bytes;
3603 else
3604 old_val -= num_bytes;
3605 btrfs_set_super_bytes_used(info->super_copy, old_val);
3606 spin_unlock(&info->delalloc_root_lock);
3607
3608 cache = btrfs_lookup_block_group(info, bytenr);
3609 if (!cache)
3610 return -ENOENT;
3611
3612 /* An extent can not span multiple block groups. */
3613 ASSERT(bytenr + num_bytes <= cache->start + cache->length);
3614
3615 space_info = cache->space_info;
3616 factor = btrfs_bg_type_to_factor(cache->flags);
3617
3618 /*
3619 * If this block group has free space cache written out, we need to make
3620 * sure to load it if we are removing space. This is because we need
3621 * the unpinning stage to actually add the space back to the block group,
3622 * otherwise we will leak space.
3623 */
3624 if (!alloc && !btrfs_block_group_done(cache))
3625 btrfs_cache_block_group(cache, true);
3626
3627 spin_lock(&space_info->lock);
3628 spin_lock(&cache->lock);
3629
3630 if (btrfs_test_opt(info, SPACE_CACHE) &&
3631 cache->disk_cache_state < BTRFS_DC_CLEAR)
3632 cache->disk_cache_state = BTRFS_DC_CLEAR;
3633
3634 old_val = cache->used;
3635 if (alloc) {
3636 old_val += num_bytes;
3637 cache->used = old_val;
3638 cache->reserved -= num_bytes;
3639 space_info->bytes_reserved -= num_bytes;
3640 space_info->bytes_used += num_bytes;
3641 space_info->disk_used += num_bytes * factor;
3642 spin_unlock(&cache->lock);
3643 spin_unlock(&space_info->lock);
3644 } else {
3645 old_val -= num_bytes;
3646 cache->used = old_val;
3647 cache->pinned += num_bytes;
3648 btrfs_space_info_update_bytes_pinned(info, space_info, num_bytes);
3649 space_info->bytes_used -= num_bytes;
3650 space_info->disk_used -= num_bytes * factor;
3651
3652 reclaim = should_reclaim_block_group(cache, num_bytes);
3653
3654 spin_unlock(&cache->lock);
3655 spin_unlock(&space_info->lock);
3656
3657 set_extent_bit(&trans->transaction->pinned_extents, bytenr,
3658 bytenr + num_bytes - 1, EXTENT_DIRTY, NULL);
3659 }
3660
3661 spin_lock(&trans->transaction->dirty_bgs_lock);
3662 if (list_empty(&cache->dirty_list)) {
3663 list_add_tail(&cache->dirty_list, &trans->transaction->dirty_bgs);
3664 bg_already_dirty = false;
3665 btrfs_get_block_group(cache);
3666 }
3667 spin_unlock(&trans->transaction->dirty_bgs_lock);
3668
3669 /*
3670 * No longer have used bytes in this block group, queue it for deletion.
3671 * We do this after adding the block group to the dirty list to avoid
3672 * races between cleaner kthread and space cache writeout.
3673 */
3674 if (!alloc && old_val == 0) {
3675 if (!btrfs_test_opt(info, DISCARD_ASYNC))
3676 btrfs_mark_bg_unused(cache);
3677 } else if (!alloc && reclaim) {
3678 btrfs_mark_bg_to_reclaim(cache);
3679 }
3680
3681 btrfs_put_block_group(cache);
3682
3683 /* Modified block groups are accounted for in the delayed_refs_rsv. */
3684 if (!bg_already_dirty)
3685 btrfs_inc_delayed_refs_rsv_bg_updates(info);
3686
3687 return 0;
3688}
3689
3690/*
3691 * Update the block_group and space info counters.
3692 *
3693 * @cache: The cache we are manipulating
3694 * @ram_bytes: The number of bytes of file content, and will be same to
3695 * @num_bytes except for the compress path.
3696 * @num_bytes: The number of bytes in question
3697 * @delalloc: The blocks are allocated for the delalloc write
3698 *
3699 * This is called by the allocator when it reserves space. If this is a
3700 * reservation and the block group has become read only we cannot make the
3701 * reservation and return -EAGAIN, otherwise this function always succeeds.
3702 */
3703int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3704 u64 ram_bytes, u64 num_bytes, int delalloc,
3705 bool force_wrong_size_class)
3706{
3707 struct btrfs_space_info *space_info = cache->space_info;
3708 enum btrfs_block_group_size_class size_class;
3709 int ret = 0;
3710
3711 spin_lock(&space_info->lock);
3712 spin_lock(&cache->lock);
3713 if (cache->ro) {
3714 ret = -EAGAIN;
3715 goto out;
3716 }
3717
3718 if (btrfs_block_group_should_use_size_class(cache)) {
3719 size_class = btrfs_calc_block_group_size_class(num_bytes);
3720 ret = btrfs_use_block_group_size_class(cache, size_class, force_wrong_size_class);
3721 if (ret)
3722 goto out;
3723 }
3724 cache->reserved += num_bytes;
3725 space_info->bytes_reserved += num_bytes;
3726 trace_btrfs_space_reservation(cache->fs_info, "space_info",
3727 space_info->flags, num_bytes, 1);
3728 btrfs_space_info_update_bytes_may_use(cache->fs_info,
3729 space_info, -ram_bytes);
3730 if (delalloc)
3731 cache->delalloc_bytes += num_bytes;
3732
3733 /*
3734 * Compression can use less space than we reserved, so wake tickets if
3735 * that happens.
3736 */
3737 if (num_bytes < ram_bytes)
3738 btrfs_try_granting_tickets(cache->fs_info, space_info);
3739out:
3740 spin_unlock(&cache->lock);
3741 spin_unlock(&space_info->lock);
3742 return ret;
3743}
3744
3745/*
3746 * Update the block_group and space info counters.
3747 *
3748 * @cache: The cache we are manipulating
3749 * @num_bytes: The number of bytes in question
3750 * @delalloc: The blocks are allocated for the delalloc write
3751 *
3752 * This is called by somebody who is freeing space that was never actually used
3753 * on disk. For example if you reserve some space for a new leaf in transaction
3754 * A and before transaction A commits you free that leaf, you call this with
3755 * reserve set to 0 in order to clear the reservation.
3756 */
3757void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
3758 u64 num_bytes, int delalloc)
3759{
3760 struct btrfs_space_info *space_info = cache->space_info;
3761
3762 spin_lock(&space_info->lock);
3763 spin_lock(&cache->lock);
3764 if (cache->ro)
3765 space_info->bytes_readonly += num_bytes;
3766 cache->reserved -= num_bytes;
3767 space_info->bytes_reserved -= num_bytes;
3768 space_info->max_extent_size = 0;
3769
3770 if (delalloc)
3771 cache->delalloc_bytes -= num_bytes;
3772 spin_unlock(&cache->lock);
3773
3774 btrfs_try_granting_tickets(cache->fs_info, space_info);
3775 spin_unlock(&space_info->lock);
3776}
3777
3778static void force_metadata_allocation(struct btrfs_fs_info *info)
3779{
3780 struct list_head *head = &info->space_info;
3781 struct btrfs_space_info *found;
3782
3783 list_for_each_entry(found, head, list) {
3784 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3785 found->force_alloc = CHUNK_ALLOC_FORCE;
3786 }
3787}
3788
3789static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3790 struct btrfs_space_info *sinfo, int force)
3791{
3792 u64 bytes_used = btrfs_space_info_used(sinfo, false);
3793 u64 thresh;
3794
3795 if (force == CHUNK_ALLOC_FORCE)
3796 return 1;
3797
3798 /*
3799 * in limited mode, we want to have some free space up to
3800 * about 1% of the FS size.
3801 */
3802 if (force == CHUNK_ALLOC_LIMITED) {
3803 thresh = btrfs_super_total_bytes(fs_info->super_copy);
3804 thresh = max_t(u64, SZ_64M, mult_perc(thresh, 1));
3805
3806 if (sinfo->total_bytes - bytes_used < thresh)
3807 return 1;
3808 }
3809
3810 if (bytes_used + SZ_2M < mult_perc(sinfo->total_bytes, 80))
3811 return 0;
3812 return 1;
3813}
3814
3815int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3816{
3817 u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3818
3819 return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3820}
3821
3822static struct btrfs_block_group *do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags)
3823{
3824 struct btrfs_block_group *bg;
3825 int ret;
3826
3827 /*
3828 * Check if we have enough space in the system space info because we
3829 * will need to update device items in the chunk btree and insert a new
3830 * chunk item in the chunk btree as well. This will allocate a new
3831 * system block group if needed.
3832 */
3833 check_system_chunk(trans, flags);
3834
3835 bg = btrfs_create_chunk(trans, flags);
3836 if (IS_ERR(bg)) {
3837 ret = PTR_ERR(bg);
3838 goto out;
3839 }
3840
3841 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3842 /*
3843 * Normally we are not expected to fail with -ENOSPC here, since we have
3844 * previously reserved space in the system space_info and allocated one
3845 * new system chunk if necessary. However there are three exceptions:
3846 *
3847 * 1) We may have enough free space in the system space_info but all the
3848 * existing system block groups have a profile which can not be used
3849 * for extent allocation.
3850 *
3851 * This happens when mounting in degraded mode. For example we have a
3852 * RAID1 filesystem with 2 devices, lose one device and mount the fs
3853 * using the other device in degraded mode. If we then allocate a chunk,
3854 * we may have enough free space in the existing system space_info, but
3855 * none of the block groups can be used for extent allocation since they
3856 * have a RAID1 profile, and because we are in degraded mode with a
3857 * single device, we are forced to allocate a new system chunk with a
3858 * SINGLE profile. Making check_system_chunk() iterate over all system
3859 * block groups and check if they have a usable profile and enough space
3860 * can be slow on very large filesystems, so we tolerate the -ENOSPC and
3861 * try again after forcing allocation of a new system chunk. Like this
3862 * we avoid paying the cost of that search in normal circumstances, when
3863 * we were not mounted in degraded mode;
3864 *
3865 * 2) We had enough free space info the system space_info, and one suitable
3866 * block group to allocate from when we called check_system_chunk()
3867 * above. However right after we called it, the only system block group
3868 * with enough free space got turned into RO mode by a running scrub,
3869 * and in this case we have to allocate a new one and retry. We only
3870 * need do this allocate and retry once, since we have a transaction
3871 * handle and scrub uses the commit root to search for block groups;
3872 *
3873 * 3) We had one system block group with enough free space when we called
3874 * check_system_chunk(), but after that, right before we tried to
3875 * allocate the last extent buffer we needed, a discard operation came
3876 * in and it temporarily removed the last free space entry from the
3877 * block group (discard removes a free space entry, discards it, and
3878 * then adds back the entry to the block group cache).
3879 */
3880 if (ret == -ENOSPC) {
3881 const u64 sys_flags = btrfs_system_alloc_profile(trans->fs_info);
3882 struct btrfs_block_group *sys_bg;
3883
3884 sys_bg = btrfs_create_chunk(trans, sys_flags);
3885 if (IS_ERR(sys_bg)) {
3886 ret = PTR_ERR(sys_bg);
3887 btrfs_abort_transaction(trans, ret);
3888 goto out;
3889 }
3890
3891 ret = btrfs_chunk_alloc_add_chunk_item(trans, sys_bg);
3892 if (ret) {
3893 btrfs_abort_transaction(trans, ret);
3894 goto out;
3895 }
3896
3897 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3898 if (ret) {
3899 btrfs_abort_transaction(trans, ret);
3900 goto out;
3901 }
3902 } else if (ret) {
3903 btrfs_abort_transaction(trans, ret);
3904 goto out;
3905 }
3906out:
3907 btrfs_trans_release_chunk_metadata(trans);
3908
3909 if (ret)
3910 return ERR_PTR(ret);
3911
3912 btrfs_get_block_group(bg);
3913 return bg;
3914}
3915
3916/*
3917 * Chunk allocation is done in 2 phases:
3918 *
3919 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
3920 * the chunk, the chunk mapping, create its block group and add the items
3921 * that belong in the chunk btree to it - more specifically, we need to
3922 * update device items in the chunk btree and add a new chunk item to it.
3923 *
3924 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
3925 * group item to the extent btree and the device extent items to the devices
3926 * btree.
3927 *
3928 * This is done to prevent deadlocks. For example when COWing a node from the
3929 * extent btree we are holding a write lock on the node's parent and if we
3930 * trigger chunk allocation and attempted to insert the new block group item
3931 * in the extent btree right way, we could deadlock because the path for the
3932 * insertion can include that parent node. At first glance it seems impossible
3933 * to trigger chunk allocation after starting a transaction since tasks should
3934 * reserve enough transaction units (metadata space), however while that is true
3935 * most of the time, chunk allocation may still be triggered for several reasons:
3936 *
3937 * 1) When reserving metadata, we check if there is enough free space in the
3938 * metadata space_info and therefore don't trigger allocation of a new chunk.
3939 * However later when the task actually tries to COW an extent buffer from
3940 * the extent btree or from the device btree for example, it is forced to
3941 * allocate a new block group (chunk) because the only one that had enough
3942 * free space was just turned to RO mode by a running scrub for example (or
3943 * device replace, block group reclaim thread, etc), so we can not use it
3944 * for allocating an extent and end up being forced to allocate a new one;
3945 *
3946 * 2) Because we only check that the metadata space_info has enough free bytes,
3947 * we end up not allocating a new metadata chunk in that case. However if
3948 * the filesystem was mounted in degraded mode, none of the existing block
3949 * groups might be suitable for extent allocation due to their incompatible
3950 * profile (for e.g. mounting a 2 devices filesystem, where all block groups
3951 * use a RAID1 profile, in degraded mode using a single device). In this case
3952 * when the task attempts to COW some extent buffer of the extent btree for
3953 * example, it will trigger allocation of a new metadata block group with a
3954 * suitable profile (SINGLE profile in the example of the degraded mount of
3955 * the RAID1 filesystem);
3956 *
3957 * 3) The task has reserved enough transaction units / metadata space, but when
3958 * it attempts to COW an extent buffer from the extent or device btree for
3959 * example, it does not find any free extent in any metadata block group,
3960 * therefore forced to try to allocate a new metadata block group.
3961 * This is because some other task allocated all available extents in the
3962 * meanwhile - this typically happens with tasks that don't reserve space
3963 * properly, either intentionally or as a bug. One example where this is
3964 * done intentionally is fsync, as it does not reserve any transaction units
3965 * and ends up allocating a variable number of metadata extents for log
3966 * tree extent buffers;
3967 *
3968 * 4) The task has reserved enough transaction units / metadata space, but right
3969 * before it tries to allocate the last extent buffer it needs, a discard
3970 * operation comes in and, temporarily, removes the last free space entry from
3971 * the only metadata block group that had free space (discard starts by
3972 * removing a free space entry from a block group, then does the discard
3973 * operation and, once it's done, it adds back the free space entry to the
3974 * block group).
3975 *
3976 * We also need this 2 phases setup when adding a device to a filesystem with
3977 * a seed device - we must create new metadata and system chunks without adding
3978 * any of the block group items to the chunk, extent and device btrees. If we
3979 * did not do it this way, we would get ENOSPC when attempting to update those
3980 * btrees, since all the chunks from the seed device are read-only.
3981 *
3982 * Phase 1 does the updates and insertions to the chunk btree because if we had
3983 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
3984 * parallel, we risk having too many system chunks allocated by many tasks if
3985 * many tasks reach phase 1 without the previous ones completing phase 2. In the
3986 * extreme case this leads to exhaustion of the system chunk array in the
3987 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
3988 * and with RAID filesystems (so we have more device items in the chunk btree).
3989 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
3990 * the system chunk array due to concurrent allocations") provides more details.
3991 *
3992 * Allocation of system chunks does not happen through this function. A task that
3993 * needs to update the chunk btree (the only btree that uses system chunks), must
3994 * preallocate chunk space by calling either check_system_chunk() or
3995 * btrfs_reserve_chunk_metadata() - the former is used when allocating a data or
3996 * metadata chunk or when removing a chunk, while the later is used before doing
3997 * a modification to the chunk btree - use cases for the later are adding,
3998 * removing and resizing a device as well as relocation of a system chunk.
3999 * See the comment below for more details.
4000 *
4001 * The reservation of system space, done through check_system_chunk(), as well
4002 * as all the updates and insertions into the chunk btree must be done while
4003 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
4004 * an extent buffer from the chunks btree we never trigger allocation of a new
4005 * system chunk, which would result in a deadlock (trying to lock twice an
4006 * extent buffer of the chunk btree, first time before triggering the chunk
4007 * allocation and the second time during chunk allocation while attempting to
4008 * update the chunks btree). The system chunk array is also updated while holding
4009 * that mutex. The same logic applies to removing chunks - we must reserve system
4010 * space, update the chunk btree and the system chunk array in the superblock
4011 * while holding fs_info->chunk_mutex.
4012 *
4013 * This function, btrfs_chunk_alloc(), belongs to phase 1.
4014 *
4015 * If @force is CHUNK_ALLOC_FORCE:
4016 * - return 1 if it successfully allocates a chunk,
4017 * - return errors including -ENOSPC otherwise.
4018 * If @force is NOT CHUNK_ALLOC_FORCE:
4019 * - return 0 if it doesn't need to allocate a new chunk,
4020 * - return 1 if it successfully allocates a chunk,
4021 * - return errors including -ENOSPC otherwise.
4022 */
4023int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4024 enum btrfs_chunk_alloc_enum force)
4025{
4026 struct btrfs_fs_info *fs_info = trans->fs_info;
4027 struct btrfs_space_info *space_info;
4028 struct btrfs_block_group *ret_bg;
4029 bool wait_for_alloc = false;
4030 bool should_alloc = false;
4031 bool from_extent_allocation = false;
4032 int ret = 0;
4033
4034 if (force == CHUNK_ALLOC_FORCE_FOR_EXTENT) {
4035 from_extent_allocation = true;
4036 force = CHUNK_ALLOC_FORCE;
4037 }
4038
4039 /* Don't re-enter if we're already allocating a chunk */
4040 if (trans->allocating_chunk)
4041 return -ENOSPC;
4042 /*
4043 * Allocation of system chunks can not happen through this path, as we
4044 * could end up in a deadlock if we are allocating a data or metadata
4045 * chunk and there is another task modifying the chunk btree.
4046 *
4047 * This is because while we are holding the chunk mutex, we will attempt
4048 * to add the new chunk item to the chunk btree or update an existing
4049 * device item in the chunk btree, while the other task that is modifying
4050 * the chunk btree is attempting to COW an extent buffer while holding a
4051 * lock on it and on its parent - if the COW operation triggers a system
4052 * chunk allocation, then we can deadlock because we are holding the
4053 * chunk mutex and we may need to access that extent buffer or its parent
4054 * in order to add the chunk item or update a device item.
4055 *
4056 * Tasks that want to modify the chunk tree should reserve system space
4057 * before updating the chunk btree, by calling either
4058 * btrfs_reserve_chunk_metadata() or check_system_chunk().
4059 * It's possible that after a task reserves the space, it still ends up
4060 * here - this happens in the cases described above at do_chunk_alloc().
4061 * The task will have to either retry or fail.
4062 */
4063 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4064 return -ENOSPC;
4065
4066 space_info = btrfs_find_space_info(fs_info, flags);
4067 ASSERT(space_info);
4068
4069 do {
4070 spin_lock(&space_info->lock);
4071 if (force < space_info->force_alloc)
4072 force = space_info->force_alloc;
4073 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4074 if (space_info->full) {
4075 /* No more free physical space */
4076 if (should_alloc)
4077 ret = -ENOSPC;
4078 else
4079 ret = 0;
4080 spin_unlock(&space_info->lock);
4081 return ret;
4082 } else if (!should_alloc) {
4083 spin_unlock(&space_info->lock);
4084 return 0;
4085 } else if (space_info->chunk_alloc) {
4086 /*
4087 * Someone is already allocating, so we need to block
4088 * until this someone is finished and then loop to
4089 * recheck if we should continue with our allocation
4090 * attempt.
4091 */
4092 wait_for_alloc = true;
4093 force = CHUNK_ALLOC_NO_FORCE;
4094 spin_unlock(&space_info->lock);
4095 mutex_lock(&fs_info->chunk_mutex);
4096 mutex_unlock(&fs_info->chunk_mutex);
4097 } else {
4098 /* Proceed with allocation */
4099 space_info->chunk_alloc = 1;
4100 wait_for_alloc = false;
4101 spin_unlock(&space_info->lock);
4102 }
4103
4104 cond_resched();
4105 } while (wait_for_alloc);
4106
4107 mutex_lock(&fs_info->chunk_mutex);
4108 trans->allocating_chunk = true;
4109
4110 /*
4111 * If we have mixed data/metadata chunks we want to make sure we keep
4112 * allocating mixed chunks instead of individual chunks.
4113 */
4114 if (btrfs_mixed_space_info(space_info))
4115 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4116
4117 /*
4118 * if we're doing a data chunk, go ahead and make sure that
4119 * we keep a reasonable number of metadata chunks allocated in the
4120 * FS as well.
4121 */
4122 if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4123 fs_info->data_chunk_allocations++;
4124 if (!(fs_info->data_chunk_allocations %
4125 fs_info->metadata_ratio))
4126 force_metadata_allocation(fs_info);
4127 }
4128
4129 ret_bg = do_chunk_alloc(trans, flags);
4130 trans->allocating_chunk = false;
4131
4132 if (IS_ERR(ret_bg)) {
4133 ret = PTR_ERR(ret_bg);
4134 } else if (from_extent_allocation && (flags & BTRFS_BLOCK_GROUP_DATA)) {
4135 /*
4136 * New block group is likely to be used soon. Try to activate
4137 * it now. Failure is OK for now.
4138 */
4139 btrfs_zone_activate(ret_bg);
4140 }
4141
4142 if (!ret)
4143 btrfs_put_block_group(ret_bg);
4144
4145 spin_lock(&space_info->lock);
4146 if (ret < 0) {
4147 if (ret == -ENOSPC)
4148 space_info->full = 1;
4149 else
4150 goto out;
4151 } else {
4152 ret = 1;
4153 space_info->max_extent_size = 0;
4154 }
4155
4156 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4157out:
4158 space_info->chunk_alloc = 0;
4159 spin_unlock(&space_info->lock);
4160 mutex_unlock(&fs_info->chunk_mutex);
4161
4162 return ret;
4163}
4164
4165static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4166{
4167 u64 num_dev;
4168
4169 num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
4170 if (!num_dev)
4171 num_dev = fs_info->fs_devices->rw_devices;
4172
4173 return num_dev;
4174}
4175
4176static void reserve_chunk_space(struct btrfs_trans_handle *trans,
4177 u64 bytes,
4178 u64 type)
4179{
4180 struct btrfs_fs_info *fs_info = trans->fs_info;
4181 struct btrfs_space_info *info;
4182 u64 left;
4183 int ret = 0;
4184
4185 /*
4186 * Needed because we can end up allocating a system chunk and for an
4187 * atomic and race free space reservation in the chunk block reserve.
4188 */
4189 lockdep_assert_held(&fs_info->chunk_mutex);
4190
4191 info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4192 spin_lock(&info->lock);
4193 left = info->total_bytes - btrfs_space_info_used(info, true);
4194 spin_unlock(&info->lock);
4195
4196 if (left < bytes && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4197 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4198 left, bytes, type);
4199 btrfs_dump_space_info(fs_info, info, 0, 0);
4200 }
4201
4202 if (left < bytes) {
4203 u64 flags = btrfs_system_alloc_profile(fs_info);
4204 struct btrfs_block_group *bg;
4205
4206 /*
4207 * Ignore failure to create system chunk. We might end up not
4208 * needing it, as we might not need to COW all nodes/leafs from
4209 * the paths we visit in the chunk tree (they were already COWed
4210 * or created in the current transaction for example).
4211 */
4212 bg = btrfs_create_chunk(trans, flags);
4213 if (IS_ERR(bg)) {
4214 ret = PTR_ERR(bg);
4215 } else {
4216 /*
4217 * We have a new chunk. We also need to activate it for
4218 * zoned filesystem.
4219 */
4220 ret = btrfs_zoned_activate_one_bg(fs_info, info, true);
4221 if (ret < 0)
4222 return;
4223
4224 /*
4225 * If we fail to add the chunk item here, we end up
4226 * trying again at phase 2 of chunk allocation, at
4227 * btrfs_create_pending_block_groups(). So ignore
4228 * any error here. An ENOSPC here could happen, due to
4229 * the cases described at do_chunk_alloc() - the system
4230 * block group we just created was just turned into RO
4231 * mode by a scrub for example, or a running discard
4232 * temporarily removed its free space entries, etc.
4233 */
4234 btrfs_chunk_alloc_add_chunk_item(trans, bg);
4235 }
4236 }
4237
4238 if (!ret) {
4239 ret = btrfs_block_rsv_add(fs_info,
4240 &fs_info->chunk_block_rsv,
4241 bytes, BTRFS_RESERVE_NO_FLUSH);
4242 if (!ret)
4243 trans->chunk_bytes_reserved += bytes;
4244 }
4245}
4246
4247/*
4248 * Reserve space in the system space for allocating or removing a chunk.
4249 * The caller must be holding fs_info->chunk_mutex.
4250 */
4251void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4252{
4253 struct btrfs_fs_info *fs_info = trans->fs_info;
4254 const u64 num_devs = get_profile_num_devs(fs_info, type);
4255 u64 bytes;
4256
4257 /* num_devs device items to update and 1 chunk item to add or remove. */
4258 bytes = btrfs_calc_metadata_size(fs_info, num_devs) +
4259 btrfs_calc_insert_metadata_size(fs_info, 1);
4260
4261 reserve_chunk_space(trans, bytes, type);
4262}
4263
4264/*
4265 * Reserve space in the system space, if needed, for doing a modification to the
4266 * chunk btree.
4267 *
4268 * @trans: A transaction handle.
4269 * @is_item_insertion: Indicate if the modification is for inserting a new item
4270 * in the chunk btree or if it's for the deletion or update
4271 * of an existing item.
4272 *
4273 * This is used in a context where we need to update the chunk btree outside
4274 * block group allocation and removal, to avoid a deadlock with a concurrent
4275 * task that is allocating a metadata or data block group and therefore needs to
4276 * update the chunk btree while holding the chunk mutex. After the update to the
4277 * chunk btree is done, btrfs_trans_release_chunk_metadata() should be called.
4278 *
4279 */
4280void btrfs_reserve_chunk_metadata(struct btrfs_trans_handle *trans,
4281 bool is_item_insertion)
4282{
4283 struct btrfs_fs_info *fs_info = trans->fs_info;
4284 u64 bytes;
4285
4286 if (is_item_insertion)
4287 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
4288 else
4289 bytes = btrfs_calc_metadata_size(fs_info, 1);
4290
4291 mutex_lock(&fs_info->chunk_mutex);
4292 reserve_chunk_space(trans, bytes, BTRFS_BLOCK_GROUP_SYSTEM);
4293 mutex_unlock(&fs_info->chunk_mutex);
4294}
4295
4296void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
4297{
4298 struct btrfs_block_group *block_group;
4299
4300 block_group = btrfs_lookup_first_block_group(info, 0);
4301 while (block_group) {
4302 btrfs_wait_block_group_cache_done(block_group);
4303 spin_lock(&block_group->lock);
4304 if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF,
4305 &block_group->runtime_flags)) {
4306 struct inode *inode = block_group->inode;
4307
4308 block_group->inode = NULL;
4309 spin_unlock(&block_group->lock);
4310
4311 ASSERT(block_group->io_ctl.inode == NULL);
4312 iput(inode);
4313 } else {
4314 spin_unlock(&block_group->lock);
4315 }
4316 block_group = btrfs_next_block_group(block_group);
4317 }
4318}
4319
4320/*
4321 * Must be called only after stopping all workers, since we could have block
4322 * group caching kthreads running, and therefore they could race with us if we
4323 * freed the block groups before stopping them.
4324 */
4325int btrfs_free_block_groups(struct btrfs_fs_info *info)
4326{
4327 struct btrfs_block_group *block_group;
4328 struct btrfs_space_info *space_info;
4329 struct btrfs_caching_control *caching_ctl;
4330 struct rb_node *n;
4331
4332 if (btrfs_is_zoned(info)) {
4333 if (info->active_meta_bg) {
4334 btrfs_put_block_group(info->active_meta_bg);
4335 info->active_meta_bg = NULL;
4336 }
4337 if (info->active_system_bg) {
4338 btrfs_put_block_group(info->active_system_bg);
4339 info->active_system_bg = NULL;
4340 }
4341 }
4342
4343 write_lock(&info->block_group_cache_lock);
4344 while (!list_empty(&info->caching_block_groups)) {
4345 caching_ctl = list_entry(info->caching_block_groups.next,
4346 struct btrfs_caching_control, list);
4347 list_del(&caching_ctl->list);
4348 btrfs_put_caching_control(caching_ctl);
4349 }
4350 write_unlock(&info->block_group_cache_lock);
4351
4352 spin_lock(&info->unused_bgs_lock);
4353 while (!list_empty(&info->unused_bgs)) {
4354 block_group = list_first_entry(&info->unused_bgs,
4355 struct btrfs_block_group,
4356 bg_list);
4357 list_del_init(&block_group->bg_list);
4358 btrfs_put_block_group(block_group);
4359 }
4360
4361 while (!list_empty(&info->reclaim_bgs)) {
4362 block_group = list_first_entry(&info->reclaim_bgs,
4363 struct btrfs_block_group,
4364 bg_list);
4365 list_del_init(&block_group->bg_list);
4366 btrfs_put_block_group(block_group);
4367 }
4368 spin_unlock(&info->unused_bgs_lock);
4369
4370 spin_lock(&info->zone_active_bgs_lock);
4371 while (!list_empty(&info->zone_active_bgs)) {
4372 block_group = list_first_entry(&info->zone_active_bgs,
4373 struct btrfs_block_group,
4374 active_bg_list);
4375 list_del_init(&block_group->active_bg_list);
4376 btrfs_put_block_group(block_group);
4377 }
4378 spin_unlock(&info->zone_active_bgs_lock);
4379
4380 write_lock(&info->block_group_cache_lock);
4381 while ((n = rb_last(&info->block_group_cache_tree.rb_root)) != NULL) {
4382 block_group = rb_entry(n, struct btrfs_block_group,
4383 cache_node);
4384 rb_erase_cached(&block_group->cache_node,
4385 &info->block_group_cache_tree);
4386 RB_CLEAR_NODE(&block_group->cache_node);
4387 write_unlock(&info->block_group_cache_lock);
4388
4389 down_write(&block_group->space_info->groups_sem);
4390 list_del(&block_group->list);
4391 up_write(&block_group->space_info->groups_sem);
4392
4393 /*
4394 * We haven't cached this block group, which means we could
4395 * possibly have excluded extents on this block group.
4396 */
4397 if (block_group->cached == BTRFS_CACHE_NO ||
4398 block_group->cached == BTRFS_CACHE_ERROR)
4399 btrfs_free_excluded_extents(block_group);
4400
4401 btrfs_remove_free_space_cache(block_group);
4402 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
4403 ASSERT(list_empty(&block_group->dirty_list));
4404 ASSERT(list_empty(&block_group->io_list));
4405 ASSERT(list_empty(&block_group->bg_list));
4406 ASSERT(refcount_read(&block_group->refs) == 1);
4407 ASSERT(block_group->swap_extents == 0);
4408 btrfs_put_block_group(block_group);
4409
4410 write_lock(&info->block_group_cache_lock);
4411 }
4412 write_unlock(&info->block_group_cache_lock);
4413
4414 btrfs_release_global_block_rsv(info);
4415
4416 while (!list_empty(&info->space_info)) {
4417 space_info = list_entry(info->space_info.next,
4418 struct btrfs_space_info,
4419 list);
4420
4421 /*
4422 * Do not hide this behind enospc_debug, this is actually
4423 * important and indicates a real bug if this happens.
4424 */
4425 if (WARN_ON(space_info->bytes_pinned > 0 ||
4426 space_info->bytes_may_use > 0))
4427 btrfs_dump_space_info(info, space_info, 0, 0);
4428
4429 /*
4430 * If there was a failure to cleanup a log tree, very likely due
4431 * to an IO failure on a writeback attempt of one or more of its
4432 * extent buffers, we could not do proper (and cheap) unaccounting
4433 * of their reserved space, so don't warn on bytes_reserved > 0 in
4434 * that case.
4435 */
4436 if (!(space_info->flags & BTRFS_BLOCK_GROUP_METADATA) ||
4437 !BTRFS_FS_LOG_CLEANUP_ERROR(info)) {
4438 if (WARN_ON(space_info->bytes_reserved > 0))
4439 btrfs_dump_space_info(info, space_info, 0, 0);
4440 }
4441
4442 WARN_ON(space_info->reclaim_size > 0);
4443 list_del(&space_info->list);
4444 btrfs_sysfs_remove_space_info(space_info);
4445 }
4446 return 0;
4447}
4448
4449void btrfs_freeze_block_group(struct btrfs_block_group *cache)
4450{
4451 atomic_inc(&cache->frozen);
4452}
4453
4454void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
4455{
4456 struct btrfs_fs_info *fs_info = block_group->fs_info;
4457 bool cleanup;
4458
4459 spin_lock(&block_group->lock);
4460 cleanup = (atomic_dec_and_test(&block_group->frozen) &&
4461 test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags));
4462 spin_unlock(&block_group->lock);
4463
4464 if (cleanup) {
4465 struct btrfs_chunk_map *map;
4466
4467 map = btrfs_find_chunk_map(fs_info, block_group->start, 1);
4468 /* Logic error, can't happen. */
4469 ASSERT(map);
4470
4471 btrfs_remove_chunk_map(fs_info, map);
4472
4473 /* Once for our lookup reference. */
4474 btrfs_free_chunk_map(map);
4475
4476 /*
4477 * We may have left one free space entry and other possible
4478 * tasks trimming this block group have left 1 entry each one.
4479 * Free them if any.
4480 */
4481 btrfs_remove_free_space_cache(block_group);
4482 }
4483}
4484
4485bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
4486{
4487 bool ret = true;
4488
4489 spin_lock(&bg->lock);
4490 if (bg->ro)
4491 ret = false;
4492 else
4493 bg->swap_extents++;
4494 spin_unlock(&bg->lock);
4495
4496 return ret;
4497}
4498
4499void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
4500{
4501 spin_lock(&bg->lock);
4502 ASSERT(!bg->ro);
4503 ASSERT(bg->swap_extents >= amount);
4504 bg->swap_extents -= amount;
4505 spin_unlock(&bg->lock);
4506}
4507
4508enum btrfs_block_group_size_class btrfs_calc_block_group_size_class(u64 size)
4509{
4510 if (size <= SZ_128K)
4511 return BTRFS_BG_SZ_SMALL;
4512 if (size <= SZ_8M)
4513 return BTRFS_BG_SZ_MEDIUM;
4514 return BTRFS_BG_SZ_LARGE;
4515}
4516
4517/*
4518 * Handle a block group allocating an extent in a size class
4519 *
4520 * @bg: The block group we allocated in.
4521 * @size_class: The size class of the allocation.
4522 * @force_wrong_size_class: Whether we are desperate enough to allow
4523 * mismatched size classes.
4524 *
4525 * Returns: 0 if the size class was valid for this block_group, -EAGAIN in the
4526 * case of a race that leads to the wrong size class without
4527 * force_wrong_size_class set.
4528 *
4529 * find_free_extent will skip block groups with a mismatched size class until
4530 * it really needs to avoid ENOSPC. In that case it will set
4531 * force_wrong_size_class. However, if a block group is newly allocated and
4532 * doesn't yet have a size class, then it is possible for two allocations of
4533 * different sizes to race and both try to use it. The loser is caught here and
4534 * has to retry.
4535 */
4536int btrfs_use_block_group_size_class(struct btrfs_block_group *bg,
4537 enum btrfs_block_group_size_class size_class,
4538 bool force_wrong_size_class)
4539{
4540 ASSERT(size_class != BTRFS_BG_SZ_NONE);
4541
4542 /* The new allocation is in the right size class, do nothing */
4543 if (bg->size_class == size_class)
4544 return 0;
4545 /*
4546 * The new allocation is in a mismatched size class.
4547 * This means one of two things:
4548 *
4549 * 1. Two tasks in find_free_extent for different size_classes raced
4550 * and hit the same empty block_group. Make the loser try again.
4551 * 2. A call to find_free_extent got desperate enough to set
4552 * 'force_wrong_slab'. Don't change the size_class, but allow the
4553 * allocation.
4554 */
4555 if (bg->size_class != BTRFS_BG_SZ_NONE) {
4556 if (force_wrong_size_class)
4557 return 0;
4558 return -EAGAIN;
4559 }
4560 /*
4561 * The happy new block group case: the new allocation is the first
4562 * one in the block_group so we set size_class.
4563 */
4564 bg->size_class = size_class;
4565
4566 return 0;
4567}
4568
4569bool btrfs_block_group_should_use_size_class(struct btrfs_block_group *bg)
4570{
4571 if (btrfs_is_zoned(bg->fs_info))
4572 return false;
4573 if (!btrfs_is_block_group_data_only(bg))
4574 return false;
4575 return true;
4576}
1// SPDX-License-Identifier: GPL-2.0
2
3#include <linux/list_sort.h>
4#include "misc.h"
5#include "ctree.h"
6#include "block-group.h"
7#include "space-info.h"
8#include "disk-io.h"
9#include "free-space-cache.h"
10#include "free-space-tree.h"
11#include "volumes.h"
12#include "transaction.h"
13#include "ref-verify.h"
14#include "sysfs.h"
15#include "tree-log.h"
16#include "delalloc-space.h"
17#include "discard.h"
18#include "raid56.h"
19#include "zoned.h"
20#include "fs.h"
21#include "accessors.h"
22#include "extent-tree.h"
23
24#ifdef CONFIG_BTRFS_DEBUG
25int btrfs_should_fragment_free_space(struct btrfs_block_group *block_group)
26{
27 struct btrfs_fs_info *fs_info = block_group->fs_info;
28
29 return (btrfs_test_opt(fs_info, FRAGMENT_METADATA) &&
30 block_group->flags & BTRFS_BLOCK_GROUP_METADATA) ||
31 (btrfs_test_opt(fs_info, FRAGMENT_DATA) &&
32 block_group->flags & BTRFS_BLOCK_GROUP_DATA);
33}
34#endif
35
36/*
37 * Return target flags in extended format or 0 if restripe for this chunk_type
38 * is not in progress
39 *
40 * Should be called with balance_lock held
41 */
42static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
43{
44 struct btrfs_balance_control *bctl = fs_info->balance_ctl;
45 u64 target = 0;
46
47 if (!bctl)
48 return 0;
49
50 if (flags & BTRFS_BLOCK_GROUP_DATA &&
51 bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
52 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
53 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
54 bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
55 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
56 } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
57 bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
58 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
59 }
60
61 return target;
62}
63
64/*
65 * @flags: available profiles in extended format (see ctree.h)
66 *
67 * Return reduced profile in chunk format. If profile changing is in progress
68 * (either running or paused) picks the target profile (if it's already
69 * available), otherwise falls back to plain reducing.
70 */
71static u64 btrfs_reduce_alloc_profile(struct btrfs_fs_info *fs_info, u64 flags)
72{
73 u64 num_devices = fs_info->fs_devices->rw_devices;
74 u64 target;
75 u64 raid_type;
76 u64 allowed = 0;
77
78 /*
79 * See if restripe for this chunk_type is in progress, if so try to
80 * reduce to the target profile
81 */
82 spin_lock(&fs_info->balance_lock);
83 target = get_restripe_target(fs_info, flags);
84 if (target) {
85 spin_unlock(&fs_info->balance_lock);
86 return extended_to_chunk(target);
87 }
88 spin_unlock(&fs_info->balance_lock);
89
90 /* First, mask out the RAID levels which aren't possible */
91 for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
92 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
93 allowed |= btrfs_raid_array[raid_type].bg_flag;
94 }
95 allowed &= flags;
96
97 if (allowed & BTRFS_BLOCK_GROUP_RAID6)
98 allowed = BTRFS_BLOCK_GROUP_RAID6;
99 else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
100 allowed = BTRFS_BLOCK_GROUP_RAID5;
101 else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
102 allowed = BTRFS_BLOCK_GROUP_RAID10;
103 else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
104 allowed = BTRFS_BLOCK_GROUP_RAID1;
105 else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
106 allowed = BTRFS_BLOCK_GROUP_RAID0;
107
108 flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
109
110 return extended_to_chunk(flags | allowed);
111}
112
113u64 btrfs_get_alloc_profile(struct btrfs_fs_info *fs_info, u64 orig_flags)
114{
115 unsigned seq;
116 u64 flags;
117
118 do {
119 flags = orig_flags;
120 seq = read_seqbegin(&fs_info->profiles_lock);
121
122 if (flags & BTRFS_BLOCK_GROUP_DATA)
123 flags |= fs_info->avail_data_alloc_bits;
124 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
125 flags |= fs_info->avail_system_alloc_bits;
126 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
127 flags |= fs_info->avail_metadata_alloc_bits;
128 } while (read_seqretry(&fs_info->profiles_lock, seq));
129
130 return btrfs_reduce_alloc_profile(fs_info, flags);
131}
132
133void btrfs_get_block_group(struct btrfs_block_group *cache)
134{
135 refcount_inc(&cache->refs);
136}
137
138void btrfs_put_block_group(struct btrfs_block_group *cache)
139{
140 if (refcount_dec_and_test(&cache->refs)) {
141 WARN_ON(cache->pinned > 0);
142 /*
143 * If there was a failure to cleanup a log tree, very likely due
144 * to an IO failure on a writeback attempt of one or more of its
145 * extent buffers, we could not do proper (and cheap) unaccounting
146 * of their reserved space, so don't warn on reserved > 0 in that
147 * case.
148 */
149 if (!(cache->flags & BTRFS_BLOCK_GROUP_METADATA) ||
150 !BTRFS_FS_LOG_CLEANUP_ERROR(cache->fs_info))
151 WARN_ON(cache->reserved > 0);
152
153 /*
154 * A block_group shouldn't be on the discard_list anymore.
155 * Remove the block_group from the discard_list to prevent us
156 * from causing a panic due to NULL pointer dereference.
157 */
158 if (WARN_ON(!list_empty(&cache->discard_list)))
159 btrfs_discard_cancel_work(&cache->fs_info->discard_ctl,
160 cache);
161
162 /*
163 * If not empty, someone is still holding mutex of
164 * full_stripe_lock, which can only be released by caller.
165 * And it will definitely cause use-after-free when caller
166 * tries to release full stripe lock.
167 *
168 * No better way to resolve, but only to warn.
169 */
170 WARN_ON(!RB_EMPTY_ROOT(&cache->full_stripe_locks_root.root));
171 kfree(cache->free_space_ctl);
172 kfree(cache->physical_map);
173 kfree(cache);
174 }
175}
176
177/*
178 * This adds the block group to the fs_info rb tree for the block group cache
179 */
180static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
181 struct btrfs_block_group *block_group)
182{
183 struct rb_node **p;
184 struct rb_node *parent = NULL;
185 struct btrfs_block_group *cache;
186 bool leftmost = true;
187
188 ASSERT(block_group->length != 0);
189
190 write_lock(&info->block_group_cache_lock);
191 p = &info->block_group_cache_tree.rb_root.rb_node;
192
193 while (*p) {
194 parent = *p;
195 cache = rb_entry(parent, struct btrfs_block_group, cache_node);
196 if (block_group->start < cache->start) {
197 p = &(*p)->rb_left;
198 } else if (block_group->start > cache->start) {
199 p = &(*p)->rb_right;
200 leftmost = false;
201 } else {
202 write_unlock(&info->block_group_cache_lock);
203 return -EEXIST;
204 }
205 }
206
207 rb_link_node(&block_group->cache_node, parent, p);
208 rb_insert_color_cached(&block_group->cache_node,
209 &info->block_group_cache_tree, leftmost);
210
211 write_unlock(&info->block_group_cache_lock);
212
213 return 0;
214}
215
216/*
217 * This will return the block group at or after bytenr if contains is 0, else
218 * it will return the block group that contains the bytenr
219 */
220static struct btrfs_block_group *block_group_cache_tree_search(
221 struct btrfs_fs_info *info, u64 bytenr, int contains)
222{
223 struct btrfs_block_group *cache, *ret = NULL;
224 struct rb_node *n;
225 u64 end, start;
226
227 read_lock(&info->block_group_cache_lock);
228 n = info->block_group_cache_tree.rb_root.rb_node;
229
230 while (n) {
231 cache = rb_entry(n, struct btrfs_block_group, cache_node);
232 end = cache->start + cache->length - 1;
233 start = cache->start;
234
235 if (bytenr < start) {
236 if (!contains && (!ret || start < ret->start))
237 ret = cache;
238 n = n->rb_left;
239 } else if (bytenr > start) {
240 if (contains && bytenr <= end) {
241 ret = cache;
242 break;
243 }
244 n = n->rb_right;
245 } else {
246 ret = cache;
247 break;
248 }
249 }
250 if (ret)
251 btrfs_get_block_group(ret);
252 read_unlock(&info->block_group_cache_lock);
253
254 return ret;
255}
256
257/*
258 * Return the block group that starts at or after bytenr
259 */
260struct btrfs_block_group *btrfs_lookup_first_block_group(
261 struct btrfs_fs_info *info, u64 bytenr)
262{
263 return block_group_cache_tree_search(info, bytenr, 0);
264}
265
266/*
267 * Return the block group that contains the given bytenr
268 */
269struct btrfs_block_group *btrfs_lookup_block_group(
270 struct btrfs_fs_info *info, u64 bytenr)
271{
272 return block_group_cache_tree_search(info, bytenr, 1);
273}
274
275struct btrfs_block_group *btrfs_next_block_group(
276 struct btrfs_block_group *cache)
277{
278 struct btrfs_fs_info *fs_info = cache->fs_info;
279 struct rb_node *node;
280
281 read_lock(&fs_info->block_group_cache_lock);
282
283 /* If our block group was removed, we need a full search. */
284 if (RB_EMPTY_NODE(&cache->cache_node)) {
285 const u64 next_bytenr = cache->start + cache->length;
286
287 read_unlock(&fs_info->block_group_cache_lock);
288 btrfs_put_block_group(cache);
289 return btrfs_lookup_first_block_group(fs_info, next_bytenr);
290 }
291 node = rb_next(&cache->cache_node);
292 btrfs_put_block_group(cache);
293 if (node) {
294 cache = rb_entry(node, struct btrfs_block_group, cache_node);
295 btrfs_get_block_group(cache);
296 } else
297 cache = NULL;
298 read_unlock(&fs_info->block_group_cache_lock);
299 return cache;
300}
301
302/*
303 * Check if we can do a NOCOW write for a given extent.
304 *
305 * @fs_info: The filesystem information object.
306 * @bytenr: Logical start address of the extent.
307 *
308 * Check if we can do a NOCOW write for the given extent, and increments the
309 * number of NOCOW writers in the block group that contains the extent, as long
310 * as the block group exists and it's currently not in read-only mode.
311 *
312 * Returns: A non-NULL block group pointer if we can do a NOCOW write, the caller
313 * is responsible for calling btrfs_dec_nocow_writers() later.
314 *
315 * Or NULL if we can not do a NOCOW write
316 */
317struct btrfs_block_group *btrfs_inc_nocow_writers(struct btrfs_fs_info *fs_info,
318 u64 bytenr)
319{
320 struct btrfs_block_group *bg;
321 bool can_nocow = true;
322
323 bg = btrfs_lookup_block_group(fs_info, bytenr);
324 if (!bg)
325 return NULL;
326
327 spin_lock(&bg->lock);
328 if (bg->ro)
329 can_nocow = false;
330 else
331 atomic_inc(&bg->nocow_writers);
332 spin_unlock(&bg->lock);
333
334 if (!can_nocow) {
335 btrfs_put_block_group(bg);
336 return NULL;
337 }
338
339 /* No put on block group, done by btrfs_dec_nocow_writers(). */
340 return bg;
341}
342
343/*
344 * Decrement the number of NOCOW writers in a block group.
345 *
346 * This is meant to be called after a previous call to btrfs_inc_nocow_writers(),
347 * and on the block group returned by that call. Typically this is called after
348 * creating an ordered extent for a NOCOW write, to prevent races with scrub and
349 * relocation.
350 *
351 * After this call, the caller should not use the block group anymore. It it wants
352 * to use it, then it should get a reference on it before calling this function.
353 */
354void btrfs_dec_nocow_writers(struct btrfs_block_group *bg)
355{
356 if (atomic_dec_and_test(&bg->nocow_writers))
357 wake_up_var(&bg->nocow_writers);
358
359 /* For the lookup done by a previous call to btrfs_inc_nocow_writers(). */
360 btrfs_put_block_group(bg);
361}
362
363void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
364{
365 wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
366}
367
368void btrfs_dec_block_group_reservations(struct btrfs_fs_info *fs_info,
369 const u64 start)
370{
371 struct btrfs_block_group *bg;
372
373 bg = btrfs_lookup_block_group(fs_info, start);
374 ASSERT(bg);
375 if (atomic_dec_and_test(&bg->reservations))
376 wake_up_var(&bg->reservations);
377 btrfs_put_block_group(bg);
378}
379
380void btrfs_wait_block_group_reservations(struct btrfs_block_group *bg)
381{
382 struct btrfs_space_info *space_info = bg->space_info;
383
384 ASSERT(bg->ro);
385
386 if (!(bg->flags & BTRFS_BLOCK_GROUP_DATA))
387 return;
388
389 /*
390 * Our block group is read only but before we set it to read only,
391 * some task might have had allocated an extent from it already, but it
392 * has not yet created a respective ordered extent (and added it to a
393 * root's list of ordered extents).
394 * Therefore wait for any task currently allocating extents, since the
395 * block group's reservations counter is incremented while a read lock
396 * on the groups' semaphore is held and decremented after releasing
397 * the read access on that semaphore and creating the ordered extent.
398 */
399 down_write(&space_info->groups_sem);
400 up_write(&space_info->groups_sem);
401
402 wait_var_event(&bg->reservations, !atomic_read(&bg->reservations));
403}
404
405struct btrfs_caching_control *btrfs_get_caching_control(
406 struct btrfs_block_group *cache)
407{
408 struct btrfs_caching_control *ctl;
409
410 spin_lock(&cache->lock);
411 if (!cache->caching_ctl) {
412 spin_unlock(&cache->lock);
413 return NULL;
414 }
415
416 ctl = cache->caching_ctl;
417 refcount_inc(&ctl->count);
418 spin_unlock(&cache->lock);
419 return ctl;
420}
421
422void btrfs_put_caching_control(struct btrfs_caching_control *ctl)
423{
424 if (refcount_dec_and_test(&ctl->count))
425 kfree(ctl);
426}
427
428/*
429 * When we wait for progress in the block group caching, its because our
430 * allocation attempt failed at least once. So, we must sleep and let some
431 * progress happen before we try again.
432 *
433 * This function will sleep at least once waiting for new free space to show
434 * up, and then it will check the block group free space numbers for our min
435 * num_bytes. Another option is to have it go ahead and look in the rbtree for
436 * a free extent of a given size, but this is a good start.
437 *
438 * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
439 * any of the information in this block group.
440 */
441void btrfs_wait_block_group_cache_progress(struct btrfs_block_group *cache,
442 u64 num_bytes)
443{
444 struct btrfs_caching_control *caching_ctl;
445
446 caching_ctl = btrfs_get_caching_control(cache);
447 if (!caching_ctl)
448 return;
449
450 wait_event(caching_ctl->wait, btrfs_block_group_done(cache) ||
451 (cache->free_space_ctl->free_space >= num_bytes));
452
453 btrfs_put_caching_control(caching_ctl);
454}
455
456static int btrfs_caching_ctl_wait_done(struct btrfs_block_group *cache,
457 struct btrfs_caching_control *caching_ctl)
458{
459 wait_event(caching_ctl->wait, btrfs_block_group_done(cache));
460 return cache->cached == BTRFS_CACHE_ERROR ? -EIO : 0;
461}
462
463static int btrfs_wait_block_group_cache_done(struct btrfs_block_group *cache)
464{
465 struct btrfs_caching_control *caching_ctl;
466 int ret;
467
468 caching_ctl = btrfs_get_caching_control(cache);
469 if (!caching_ctl)
470 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
471 ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
472 btrfs_put_caching_control(caching_ctl);
473 return ret;
474}
475
476#ifdef CONFIG_BTRFS_DEBUG
477static void fragment_free_space(struct btrfs_block_group *block_group)
478{
479 struct btrfs_fs_info *fs_info = block_group->fs_info;
480 u64 start = block_group->start;
481 u64 len = block_group->length;
482 u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
483 fs_info->nodesize : fs_info->sectorsize;
484 u64 step = chunk << 1;
485
486 while (len > chunk) {
487 btrfs_remove_free_space(block_group, start, chunk);
488 start += step;
489 if (len < step)
490 len = 0;
491 else
492 len -= step;
493 }
494}
495#endif
496
497/*
498 * This is only called by btrfs_cache_block_group, since we could have freed
499 * extents we need to check the pinned_extents for any extents that can't be
500 * used yet since their free space will be released as soon as the transaction
501 * commits.
502 */
503u64 add_new_free_space(struct btrfs_block_group *block_group, u64 start, u64 end)
504{
505 struct btrfs_fs_info *info = block_group->fs_info;
506 u64 extent_start, extent_end, size, total_added = 0;
507 int ret;
508
509 while (start < end) {
510 ret = find_first_extent_bit(&info->excluded_extents, start,
511 &extent_start, &extent_end,
512 EXTENT_DIRTY | EXTENT_UPTODATE,
513 NULL);
514 if (ret)
515 break;
516
517 if (extent_start <= start) {
518 start = extent_end + 1;
519 } else if (extent_start > start && extent_start < end) {
520 size = extent_start - start;
521 total_added += size;
522 ret = btrfs_add_free_space_async_trimmed(block_group,
523 start, size);
524 BUG_ON(ret); /* -ENOMEM or logic error */
525 start = extent_end + 1;
526 } else {
527 break;
528 }
529 }
530
531 if (start < end) {
532 size = end - start;
533 total_added += size;
534 ret = btrfs_add_free_space_async_trimmed(block_group, start,
535 size);
536 BUG_ON(ret); /* -ENOMEM or logic error */
537 }
538
539 return total_added;
540}
541
542static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
543{
544 struct btrfs_block_group *block_group = caching_ctl->block_group;
545 struct btrfs_fs_info *fs_info = block_group->fs_info;
546 struct btrfs_root *extent_root;
547 struct btrfs_path *path;
548 struct extent_buffer *leaf;
549 struct btrfs_key key;
550 u64 total_found = 0;
551 u64 last = 0;
552 u32 nritems;
553 int ret;
554 bool wakeup = true;
555
556 path = btrfs_alloc_path();
557 if (!path)
558 return -ENOMEM;
559
560 last = max_t(u64, block_group->start, BTRFS_SUPER_INFO_OFFSET);
561 extent_root = btrfs_extent_root(fs_info, last);
562
563#ifdef CONFIG_BTRFS_DEBUG
564 /*
565 * If we're fragmenting we don't want to make anybody think we can
566 * allocate from this block group until we've had a chance to fragment
567 * the free space.
568 */
569 if (btrfs_should_fragment_free_space(block_group))
570 wakeup = false;
571#endif
572 /*
573 * We don't want to deadlock with somebody trying to allocate a new
574 * extent for the extent root while also trying to search the extent
575 * root to add free space. So we skip locking and search the commit
576 * root, since its read-only
577 */
578 path->skip_locking = 1;
579 path->search_commit_root = 1;
580 path->reada = READA_FORWARD;
581
582 key.objectid = last;
583 key.offset = 0;
584 key.type = BTRFS_EXTENT_ITEM_KEY;
585
586next:
587 ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
588 if (ret < 0)
589 goto out;
590
591 leaf = path->nodes[0];
592 nritems = btrfs_header_nritems(leaf);
593
594 while (1) {
595 if (btrfs_fs_closing(fs_info) > 1) {
596 last = (u64)-1;
597 break;
598 }
599
600 if (path->slots[0] < nritems) {
601 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
602 } else {
603 ret = btrfs_find_next_key(extent_root, path, &key, 0, 0);
604 if (ret)
605 break;
606
607 if (need_resched() ||
608 rwsem_is_contended(&fs_info->commit_root_sem)) {
609 btrfs_release_path(path);
610 up_read(&fs_info->commit_root_sem);
611 mutex_unlock(&caching_ctl->mutex);
612 cond_resched();
613 mutex_lock(&caching_ctl->mutex);
614 down_read(&fs_info->commit_root_sem);
615 goto next;
616 }
617
618 ret = btrfs_next_leaf(extent_root, path);
619 if (ret < 0)
620 goto out;
621 if (ret)
622 break;
623 leaf = path->nodes[0];
624 nritems = btrfs_header_nritems(leaf);
625 continue;
626 }
627
628 if (key.objectid < last) {
629 key.objectid = last;
630 key.offset = 0;
631 key.type = BTRFS_EXTENT_ITEM_KEY;
632 btrfs_release_path(path);
633 goto next;
634 }
635
636 if (key.objectid < block_group->start) {
637 path->slots[0]++;
638 continue;
639 }
640
641 if (key.objectid >= block_group->start + block_group->length)
642 break;
643
644 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
645 key.type == BTRFS_METADATA_ITEM_KEY) {
646 total_found += add_new_free_space(block_group, last,
647 key.objectid);
648 if (key.type == BTRFS_METADATA_ITEM_KEY)
649 last = key.objectid +
650 fs_info->nodesize;
651 else
652 last = key.objectid + key.offset;
653
654 if (total_found > CACHING_CTL_WAKE_UP) {
655 total_found = 0;
656 if (wakeup)
657 wake_up(&caching_ctl->wait);
658 }
659 }
660 path->slots[0]++;
661 }
662 ret = 0;
663
664 total_found += add_new_free_space(block_group, last,
665 block_group->start + block_group->length);
666
667out:
668 btrfs_free_path(path);
669 return ret;
670}
671
672static noinline void caching_thread(struct btrfs_work *work)
673{
674 struct btrfs_block_group *block_group;
675 struct btrfs_fs_info *fs_info;
676 struct btrfs_caching_control *caching_ctl;
677 int ret;
678
679 caching_ctl = container_of(work, struct btrfs_caching_control, work);
680 block_group = caching_ctl->block_group;
681 fs_info = block_group->fs_info;
682
683 mutex_lock(&caching_ctl->mutex);
684 down_read(&fs_info->commit_root_sem);
685
686 if (btrfs_test_opt(fs_info, SPACE_CACHE)) {
687 ret = load_free_space_cache(block_group);
688 if (ret == 1) {
689 ret = 0;
690 goto done;
691 }
692
693 /*
694 * We failed to load the space cache, set ourselves to
695 * CACHE_STARTED and carry on.
696 */
697 spin_lock(&block_group->lock);
698 block_group->cached = BTRFS_CACHE_STARTED;
699 spin_unlock(&block_group->lock);
700 wake_up(&caching_ctl->wait);
701 }
702
703 /*
704 * If we are in the transaction that populated the free space tree we
705 * can't actually cache from the free space tree as our commit root and
706 * real root are the same, so we could change the contents of the blocks
707 * while caching. Instead do the slow caching in this case, and after
708 * the transaction has committed we will be safe.
709 */
710 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE) &&
711 !(test_bit(BTRFS_FS_FREE_SPACE_TREE_UNTRUSTED, &fs_info->flags)))
712 ret = load_free_space_tree(caching_ctl);
713 else
714 ret = load_extent_tree_free(caching_ctl);
715done:
716 spin_lock(&block_group->lock);
717 block_group->caching_ctl = NULL;
718 block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
719 spin_unlock(&block_group->lock);
720
721#ifdef CONFIG_BTRFS_DEBUG
722 if (btrfs_should_fragment_free_space(block_group)) {
723 u64 bytes_used;
724
725 spin_lock(&block_group->space_info->lock);
726 spin_lock(&block_group->lock);
727 bytes_used = block_group->length - block_group->used;
728 block_group->space_info->bytes_used += bytes_used >> 1;
729 spin_unlock(&block_group->lock);
730 spin_unlock(&block_group->space_info->lock);
731 fragment_free_space(block_group);
732 }
733#endif
734
735 up_read(&fs_info->commit_root_sem);
736 btrfs_free_excluded_extents(block_group);
737 mutex_unlock(&caching_ctl->mutex);
738
739 wake_up(&caching_ctl->wait);
740
741 btrfs_put_caching_control(caching_ctl);
742 btrfs_put_block_group(block_group);
743}
744
745int btrfs_cache_block_group(struct btrfs_block_group *cache, bool wait)
746{
747 struct btrfs_fs_info *fs_info = cache->fs_info;
748 struct btrfs_caching_control *caching_ctl = NULL;
749 int ret = 0;
750
751 /* Allocator for zoned filesystems does not use the cache at all */
752 if (btrfs_is_zoned(fs_info))
753 return 0;
754
755 caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
756 if (!caching_ctl)
757 return -ENOMEM;
758
759 INIT_LIST_HEAD(&caching_ctl->list);
760 mutex_init(&caching_ctl->mutex);
761 init_waitqueue_head(&caching_ctl->wait);
762 caching_ctl->block_group = cache;
763 refcount_set(&caching_ctl->count, 2);
764 btrfs_init_work(&caching_ctl->work, caching_thread, NULL, NULL);
765
766 spin_lock(&cache->lock);
767 if (cache->cached != BTRFS_CACHE_NO) {
768 kfree(caching_ctl);
769
770 caching_ctl = cache->caching_ctl;
771 if (caching_ctl)
772 refcount_inc(&caching_ctl->count);
773 spin_unlock(&cache->lock);
774 goto out;
775 }
776 WARN_ON(cache->caching_ctl);
777 cache->caching_ctl = caching_ctl;
778 cache->cached = BTRFS_CACHE_STARTED;
779 spin_unlock(&cache->lock);
780
781 write_lock(&fs_info->block_group_cache_lock);
782 refcount_inc(&caching_ctl->count);
783 list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
784 write_unlock(&fs_info->block_group_cache_lock);
785
786 btrfs_get_block_group(cache);
787
788 btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
789out:
790 if (wait && caching_ctl)
791 ret = btrfs_caching_ctl_wait_done(cache, caching_ctl);
792 if (caching_ctl)
793 btrfs_put_caching_control(caching_ctl);
794
795 return ret;
796}
797
798static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
799{
800 u64 extra_flags = chunk_to_extended(flags) &
801 BTRFS_EXTENDED_PROFILE_MASK;
802
803 write_seqlock(&fs_info->profiles_lock);
804 if (flags & BTRFS_BLOCK_GROUP_DATA)
805 fs_info->avail_data_alloc_bits &= ~extra_flags;
806 if (flags & BTRFS_BLOCK_GROUP_METADATA)
807 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
808 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
809 fs_info->avail_system_alloc_bits &= ~extra_flags;
810 write_sequnlock(&fs_info->profiles_lock);
811}
812
813/*
814 * Clear incompat bits for the following feature(s):
815 *
816 * - RAID56 - in case there's neither RAID5 nor RAID6 profile block group
817 * in the whole filesystem
818 *
819 * - RAID1C34 - same as above for RAID1C3 and RAID1C4 block groups
820 */
821static void clear_incompat_bg_bits(struct btrfs_fs_info *fs_info, u64 flags)
822{
823 bool found_raid56 = false;
824 bool found_raid1c34 = false;
825
826 if ((flags & BTRFS_BLOCK_GROUP_RAID56_MASK) ||
827 (flags & BTRFS_BLOCK_GROUP_RAID1C3) ||
828 (flags & BTRFS_BLOCK_GROUP_RAID1C4)) {
829 struct list_head *head = &fs_info->space_info;
830 struct btrfs_space_info *sinfo;
831
832 list_for_each_entry_rcu(sinfo, head, list) {
833 down_read(&sinfo->groups_sem);
834 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID5]))
835 found_raid56 = true;
836 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID6]))
837 found_raid56 = true;
838 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C3]))
839 found_raid1c34 = true;
840 if (!list_empty(&sinfo->block_groups[BTRFS_RAID_RAID1C4]))
841 found_raid1c34 = true;
842 up_read(&sinfo->groups_sem);
843 }
844 if (!found_raid56)
845 btrfs_clear_fs_incompat(fs_info, RAID56);
846 if (!found_raid1c34)
847 btrfs_clear_fs_incompat(fs_info, RAID1C34);
848 }
849}
850
851static int remove_block_group_item(struct btrfs_trans_handle *trans,
852 struct btrfs_path *path,
853 struct btrfs_block_group *block_group)
854{
855 struct btrfs_fs_info *fs_info = trans->fs_info;
856 struct btrfs_root *root;
857 struct btrfs_key key;
858 int ret;
859
860 root = btrfs_block_group_root(fs_info);
861 key.objectid = block_group->start;
862 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
863 key.offset = block_group->length;
864
865 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
866 if (ret > 0)
867 ret = -ENOENT;
868 if (ret < 0)
869 return ret;
870
871 ret = btrfs_del_item(trans, root, path);
872 return ret;
873}
874
875int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
876 u64 group_start, struct extent_map *em)
877{
878 struct btrfs_fs_info *fs_info = trans->fs_info;
879 struct btrfs_path *path;
880 struct btrfs_block_group *block_group;
881 struct btrfs_free_cluster *cluster;
882 struct inode *inode;
883 struct kobject *kobj = NULL;
884 int ret;
885 int index;
886 int factor;
887 struct btrfs_caching_control *caching_ctl = NULL;
888 bool remove_em;
889 bool remove_rsv = false;
890
891 block_group = btrfs_lookup_block_group(fs_info, group_start);
892 BUG_ON(!block_group);
893 BUG_ON(!block_group->ro);
894
895 trace_btrfs_remove_block_group(block_group);
896 /*
897 * Free the reserved super bytes from this block group before
898 * remove it.
899 */
900 btrfs_free_excluded_extents(block_group);
901 btrfs_free_ref_tree_range(fs_info, block_group->start,
902 block_group->length);
903
904 index = btrfs_bg_flags_to_raid_index(block_group->flags);
905 factor = btrfs_bg_type_to_factor(block_group->flags);
906
907 /* make sure this block group isn't part of an allocation cluster */
908 cluster = &fs_info->data_alloc_cluster;
909 spin_lock(&cluster->refill_lock);
910 btrfs_return_cluster_to_free_space(block_group, cluster);
911 spin_unlock(&cluster->refill_lock);
912
913 /*
914 * make sure this block group isn't part of a metadata
915 * allocation cluster
916 */
917 cluster = &fs_info->meta_alloc_cluster;
918 spin_lock(&cluster->refill_lock);
919 btrfs_return_cluster_to_free_space(block_group, cluster);
920 spin_unlock(&cluster->refill_lock);
921
922 btrfs_clear_treelog_bg(block_group);
923 btrfs_clear_data_reloc_bg(block_group);
924
925 path = btrfs_alloc_path();
926 if (!path) {
927 ret = -ENOMEM;
928 goto out;
929 }
930
931 /*
932 * get the inode first so any iput calls done for the io_list
933 * aren't the final iput (no unlinks allowed now)
934 */
935 inode = lookup_free_space_inode(block_group, path);
936
937 mutex_lock(&trans->transaction->cache_write_mutex);
938 /*
939 * Make sure our free space cache IO is done before removing the
940 * free space inode
941 */
942 spin_lock(&trans->transaction->dirty_bgs_lock);
943 if (!list_empty(&block_group->io_list)) {
944 list_del_init(&block_group->io_list);
945
946 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
947
948 spin_unlock(&trans->transaction->dirty_bgs_lock);
949 btrfs_wait_cache_io(trans, block_group, path);
950 btrfs_put_block_group(block_group);
951 spin_lock(&trans->transaction->dirty_bgs_lock);
952 }
953
954 if (!list_empty(&block_group->dirty_list)) {
955 list_del_init(&block_group->dirty_list);
956 remove_rsv = true;
957 btrfs_put_block_group(block_group);
958 }
959 spin_unlock(&trans->transaction->dirty_bgs_lock);
960 mutex_unlock(&trans->transaction->cache_write_mutex);
961
962 ret = btrfs_remove_free_space_inode(trans, inode, block_group);
963 if (ret)
964 goto out;
965
966 write_lock(&fs_info->block_group_cache_lock);
967 rb_erase_cached(&block_group->cache_node,
968 &fs_info->block_group_cache_tree);
969 RB_CLEAR_NODE(&block_group->cache_node);
970
971 /* Once for the block groups rbtree */
972 btrfs_put_block_group(block_group);
973
974 write_unlock(&fs_info->block_group_cache_lock);
975
976 down_write(&block_group->space_info->groups_sem);
977 /*
978 * we must use list_del_init so people can check to see if they
979 * are still on the list after taking the semaphore
980 */
981 list_del_init(&block_group->list);
982 if (list_empty(&block_group->space_info->block_groups[index])) {
983 kobj = block_group->space_info->block_group_kobjs[index];
984 block_group->space_info->block_group_kobjs[index] = NULL;
985 clear_avail_alloc_bits(fs_info, block_group->flags);
986 }
987 up_write(&block_group->space_info->groups_sem);
988 clear_incompat_bg_bits(fs_info, block_group->flags);
989 if (kobj) {
990 kobject_del(kobj);
991 kobject_put(kobj);
992 }
993
994 if (block_group->cached == BTRFS_CACHE_STARTED)
995 btrfs_wait_block_group_cache_done(block_group);
996
997 write_lock(&fs_info->block_group_cache_lock);
998 caching_ctl = btrfs_get_caching_control(block_group);
999 if (!caching_ctl) {
1000 struct btrfs_caching_control *ctl;
1001
1002 list_for_each_entry(ctl, &fs_info->caching_block_groups, list) {
1003 if (ctl->block_group == block_group) {
1004 caching_ctl = ctl;
1005 refcount_inc(&caching_ctl->count);
1006 break;
1007 }
1008 }
1009 }
1010 if (caching_ctl)
1011 list_del_init(&caching_ctl->list);
1012 write_unlock(&fs_info->block_group_cache_lock);
1013
1014 if (caching_ctl) {
1015 /* Once for the caching bgs list and once for us. */
1016 btrfs_put_caching_control(caching_ctl);
1017 btrfs_put_caching_control(caching_ctl);
1018 }
1019
1020 spin_lock(&trans->transaction->dirty_bgs_lock);
1021 WARN_ON(!list_empty(&block_group->dirty_list));
1022 WARN_ON(!list_empty(&block_group->io_list));
1023 spin_unlock(&trans->transaction->dirty_bgs_lock);
1024
1025 btrfs_remove_free_space_cache(block_group);
1026
1027 spin_lock(&block_group->space_info->lock);
1028 list_del_init(&block_group->ro_list);
1029
1030 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1031 WARN_ON(block_group->space_info->total_bytes
1032 < block_group->length);
1033 WARN_ON(block_group->space_info->bytes_readonly
1034 < block_group->length - block_group->zone_unusable);
1035 WARN_ON(block_group->space_info->bytes_zone_unusable
1036 < block_group->zone_unusable);
1037 WARN_ON(block_group->space_info->disk_total
1038 < block_group->length * factor);
1039 WARN_ON(test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
1040 &block_group->runtime_flags) &&
1041 block_group->space_info->active_total_bytes
1042 < block_group->length);
1043 }
1044 block_group->space_info->total_bytes -= block_group->length;
1045 if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags))
1046 block_group->space_info->active_total_bytes -= block_group->length;
1047 block_group->space_info->bytes_readonly -=
1048 (block_group->length - block_group->zone_unusable);
1049 block_group->space_info->bytes_zone_unusable -=
1050 block_group->zone_unusable;
1051 block_group->space_info->disk_total -= block_group->length * factor;
1052
1053 spin_unlock(&block_group->space_info->lock);
1054
1055 /*
1056 * Remove the free space for the block group from the free space tree
1057 * and the block group's item from the extent tree before marking the
1058 * block group as removed. This is to prevent races with tasks that
1059 * freeze and unfreeze a block group, this task and another task
1060 * allocating a new block group - the unfreeze task ends up removing
1061 * the block group's extent map before the task calling this function
1062 * deletes the block group item from the extent tree, allowing for
1063 * another task to attempt to create another block group with the same
1064 * item key (and failing with -EEXIST and a transaction abort).
1065 */
1066 ret = remove_block_group_free_space(trans, block_group);
1067 if (ret)
1068 goto out;
1069
1070 ret = remove_block_group_item(trans, path, block_group);
1071 if (ret < 0)
1072 goto out;
1073
1074 spin_lock(&block_group->lock);
1075 set_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags);
1076
1077 /*
1078 * At this point trimming or scrub can't start on this block group,
1079 * because we removed the block group from the rbtree
1080 * fs_info->block_group_cache_tree so no one can't find it anymore and
1081 * even if someone already got this block group before we removed it
1082 * from the rbtree, they have already incremented block_group->frozen -
1083 * if they didn't, for the trimming case they won't find any free space
1084 * entries because we already removed them all when we called
1085 * btrfs_remove_free_space_cache().
1086 *
1087 * And we must not remove the extent map from the fs_info->mapping_tree
1088 * to prevent the same logical address range and physical device space
1089 * ranges from being reused for a new block group. This is needed to
1090 * avoid races with trimming and scrub.
1091 *
1092 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1093 * completely transactionless, so while it is trimming a range the
1094 * currently running transaction might finish and a new one start,
1095 * allowing for new block groups to be created that can reuse the same
1096 * physical device locations unless we take this special care.
1097 *
1098 * There may also be an implicit trim operation if the file system
1099 * is mounted with -odiscard. The same protections must remain
1100 * in place until the extents have been discarded completely when
1101 * the transaction commit has completed.
1102 */
1103 remove_em = (atomic_read(&block_group->frozen) == 0);
1104 spin_unlock(&block_group->lock);
1105
1106 if (remove_em) {
1107 struct extent_map_tree *em_tree;
1108
1109 em_tree = &fs_info->mapping_tree;
1110 write_lock(&em_tree->lock);
1111 remove_extent_mapping(em_tree, em);
1112 write_unlock(&em_tree->lock);
1113 /* once for the tree */
1114 free_extent_map(em);
1115 }
1116
1117out:
1118 /* Once for the lookup reference */
1119 btrfs_put_block_group(block_group);
1120 if (remove_rsv)
1121 btrfs_delayed_refs_rsv_release(fs_info, 1);
1122 btrfs_free_path(path);
1123 return ret;
1124}
1125
1126struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1127 struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1128{
1129 struct btrfs_root *root = btrfs_block_group_root(fs_info);
1130 struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1131 struct extent_map *em;
1132 struct map_lookup *map;
1133 unsigned int num_items;
1134
1135 read_lock(&em_tree->lock);
1136 em = lookup_extent_mapping(em_tree, chunk_offset, 1);
1137 read_unlock(&em_tree->lock);
1138 ASSERT(em && em->start == chunk_offset);
1139
1140 /*
1141 * We need to reserve 3 + N units from the metadata space info in order
1142 * to remove a block group (done at btrfs_remove_chunk() and at
1143 * btrfs_remove_block_group()), which are used for:
1144 *
1145 * 1 unit for adding the free space inode's orphan (located in the tree
1146 * of tree roots).
1147 * 1 unit for deleting the block group item (located in the extent
1148 * tree).
1149 * 1 unit for deleting the free space item (located in tree of tree
1150 * roots).
1151 * N units for deleting N device extent items corresponding to each
1152 * stripe (located in the device tree).
1153 *
1154 * In order to remove a block group we also need to reserve units in the
1155 * system space info in order to update the chunk tree (update one or
1156 * more device items and remove one chunk item), but this is done at
1157 * btrfs_remove_chunk() through a call to check_system_chunk().
1158 */
1159 map = em->map_lookup;
1160 num_items = 3 + map->num_stripes;
1161 free_extent_map(em);
1162
1163 return btrfs_start_transaction_fallback_global_rsv(root, num_items);
1164}
1165
1166/*
1167 * Mark block group @cache read-only, so later write won't happen to block
1168 * group @cache.
1169 *
1170 * If @force is not set, this function will only mark the block group readonly
1171 * if we have enough free space (1M) in other metadata/system block groups.
1172 * If @force is not set, this function will mark the block group readonly
1173 * without checking free space.
1174 *
1175 * NOTE: This function doesn't care if other block groups can contain all the
1176 * data in this block group. That check should be done by relocation routine,
1177 * not this function.
1178 */
1179static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1180{
1181 struct btrfs_space_info *sinfo = cache->space_info;
1182 u64 num_bytes;
1183 int ret = -ENOSPC;
1184
1185 spin_lock(&sinfo->lock);
1186 spin_lock(&cache->lock);
1187
1188 if (cache->swap_extents) {
1189 ret = -ETXTBSY;
1190 goto out;
1191 }
1192
1193 if (cache->ro) {
1194 cache->ro++;
1195 ret = 0;
1196 goto out;
1197 }
1198
1199 num_bytes = cache->length - cache->reserved - cache->pinned -
1200 cache->bytes_super - cache->zone_unusable - cache->used;
1201
1202 /*
1203 * Data never overcommits, even in mixed mode, so do just the straight
1204 * check of left over space in how much we have allocated.
1205 */
1206 if (force) {
1207 ret = 0;
1208 } else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1209 u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1210
1211 /*
1212 * Here we make sure if we mark this bg RO, we still have enough
1213 * free space as buffer.
1214 */
1215 if (sinfo_used + num_bytes <= sinfo->total_bytes)
1216 ret = 0;
1217 } else {
1218 /*
1219 * We overcommit metadata, so we need to do the
1220 * btrfs_can_overcommit check here, and we need to pass in
1221 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1222 * leeway to allow us to mark this block group as read only.
1223 */
1224 if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1225 BTRFS_RESERVE_NO_FLUSH))
1226 ret = 0;
1227 }
1228
1229 if (!ret) {
1230 sinfo->bytes_readonly += num_bytes;
1231 if (btrfs_is_zoned(cache->fs_info)) {
1232 /* Migrate zone_unusable bytes to readonly */
1233 sinfo->bytes_readonly += cache->zone_unusable;
1234 sinfo->bytes_zone_unusable -= cache->zone_unusable;
1235 cache->zone_unusable = 0;
1236 }
1237 cache->ro++;
1238 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1239 }
1240out:
1241 spin_unlock(&cache->lock);
1242 spin_unlock(&sinfo->lock);
1243 if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1244 btrfs_info(cache->fs_info,
1245 "unable to make block group %llu ro", cache->start);
1246 btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1247 }
1248 return ret;
1249}
1250
1251static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1252 struct btrfs_block_group *bg)
1253{
1254 struct btrfs_fs_info *fs_info = bg->fs_info;
1255 struct btrfs_transaction *prev_trans = NULL;
1256 const u64 start = bg->start;
1257 const u64 end = start + bg->length - 1;
1258 int ret;
1259
1260 spin_lock(&fs_info->trans_lock);
1261 if (trans->transaction->list.prev != &fs_info->trans_list) {
1262 prev_trans = list_last_entry(&trans->transaction->list,
1263 struct btrfs_transaction, list);
1264 refcount_inc(&prev_trans->use_count);
1265 }
1266 spin_unlock(&fs_info->trans_lock);
1267
1268 /*
1269 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1270 * btrfs_finish_extent_commit(). If we are at transaction N, another
1271 * task might be running finish_extent_commit() for the previous
1272 * transaction N - 1, and have seen a range belonging to the block
1273 * group in pinned_extents before we were able to clear the whole block
1274 * group range from pinned_extents. This means that task can lookup for
1275 * the block group after we unpinned it from pinned_extents and removed
1276 * it, leading to a BUG_ON() at unpin_extent_range().
1277 */
1278 mutex_lock(&fs_info->unused_bg_unpin_mutex);
1279 if (prev_trans) {
1280 ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1281 EXTENT_DIRTY);
1282 if (ret)
1283 goto out;
1284 }
1285
1286 ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1287 EXTENT_DIRTY);
1288out:
1289 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1290 if (prev_trans)
1291 btrfs_put_transaction(prev_trans);
1292
1293 return ret == 0;
1294}
1295
1296/*
1297 * Process the unused_bgs list and remove any that don't have any allocated
1298 * space inside of them.
1299 */
1300void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1301{
1302 struct btrfs_block_group *block_group;
1303 struct btrfs_space_info *space_info;
1304 struct btrfs_trans_handle *trans;
1305 const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1306 int ret = 0;
1307
1308 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1309 return;
1310
1311 if (btrfs_fs_closing(fs_info))
1312 return;
1313
1314 /*
1315 * Long running balances can keep us blocked here for eternity, so
1316 * simply skip deletion if we're unable to get the mutex.
1317 */
1318 if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1319 return;
1320
1321 spin_lock(&fs_info->unused_bgs_lock);
1322 while (!list_empty(&fs_info->unused_bgs)) {
1323 int trimming;
1324
1325 block_group = list_first_entry(&fs_info->unused_bgs,
1326 struct btrfs_block_group,
1327 bg_list);
1328 list_del_init(&block_group->bg_list);
1329
1330 space_info = block_group->space_info;
1331
1332 if (ret || btrfs_mixed_space_info(space_info)) {
1333 btrfs_put_block_group(block_group);
1334 continue;
1335 }
1336 spin_unlock(&fs_info->unused_bgs_lock);
1337
1338 btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1339
1340 /* Don't want to race with allocators so take the groups_sem */
1341 down_write(&space_info->groups_sem);
1342
1343 /*
1344 * Async discard moves the final block group discard to be prior
1345 * to the unused_bgs code path. Therefore, if it's not fully
1346 * trimmed, punt it back to the async discard lists.
1347 */
1348 if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1349 !btrfs_is_free_space_trimmed(block_group)) {
1350 trace_btrfs_skip_unused_block_group(block_group);
1351 up_write(&space_info->groups_sem);
1352 /* Requeue if we failed because of async discard */
1353 btrfs_discard_queue_work(&fs_info->discard_ctl,
1354 block_group);
1355 goto next;
1356 }
1357
1358 spin_lock(&block_group->lock);
1359 if (block_group->reserved || block_group->pinned ||
1360 block_group->used || block_group->ro ||
1361 list_is_singular(&block_group->list)) {
1362 /*
1363 * We want to bail if we made new allocations or have
1364 * outstanding allocations in this block group. We do
1365 * the ro check in case balance is currently acting on
1366 * this block group.
1367 */
1368 trace_btrfs_skip_unused_block_group(block_group);
1369 spin_unlock(&block_group->lock);
1370 up_write(&space_info->groups_sem);
1371 goto next;
1372 }
1373 spin_unlock(&block_group->lock);
1374
1375 /* We don't want to force the issue, only flip if it's ok. */
1376 ret = inc_block_group_ro(block_group, 0);
1377 up_write(&space_info->groups_sem);
1378 if (ret < 0) {
1379 ret = 0;
1380 goto next;
1381 }
1382
1383 ret = btrfs_zone_finish(block_group);
1384 if (ret < 0) {
1385 btrfs_dec_block_group_ro(block_group);
1386 if (ret == -EAGAIN)
1387 ret = 0;
1388 goto next;
1389 }
1390
1391 /*
1392 * Want to do this before we do anything else so we can recover
1393 * properly if we fail to join the transaction.
1394 */
1395 trans = btrfs_start_trans_remove_block_group(fs_info,
1396 block_group->start);
1397 if (IS_ERR(trans)) {
1398 btrfs_dec_block_group_ro(block_group);
1399 ret = PTR_ERR(trans);
1400 goto next;
1401 }
1402
1403 /*
1404 * We could have pending pinned extents for this block group,
1405 * just delete them, we don't care about them anymore.
1406 */
1407 if (!clean_pinned_extents(trans, block_group)) {
1408 btrfs_dec_block_group_ro(block_group);
1409 goto end_trans;
1410 }
1411
1412 /*
1413 * At this point, the block_group is read only and should fail
1414 * new allocations. However, btrfs_finish_extent_commit() can
1415 * cause this block_group to be placed back on the discard
1416 * lists because now the block_group isn't fully discarded.
1417 * Bail here and try again later after discarding everything.
1418 */
1419 spin_lock(&fs_info->discard_ctl.lock);
1420 if (!list_empty(&block_group->discard_list)) {
1421 spin_unlock(&fs_info->discard_ctl.lock);
1422 btrfs_dec_block_group_ro(block_group);
1423 btrfs_discard_queue_work(&fs_info->discard_ctl,
1424 block_group);
1425 goto end_trans;
1426 }
1427 spin_unlock(&fs_info->discard_ctl.lock);
1428
1429 /* Reset pinned so btrfs_put_block_group doesn't complain */
1430 spin_lock(&space_info->lock);
1431 spin_lock(&block_group->lock);
1432
1433 btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1434 -block_group->pinned);
1435 space_info->bytes_readonly += block_group->pinned;
1436 block_group->pinned = 0;
1437
1438 spin_unlock(&block_group->lock);
1439 spin_unlock(&space_info->lock);
1440
1441 /*
1442 * The normal path here is an unused block group is passed here,
1443 * then trimming is handled in the transaction commit path.
1444 * Async discard interposes before this to do the trimming
1445 * before coming down the unused block group path as trimming
1446 * will no longer be done later in the transaction commit path.
1447 */
1448 if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1449 goto flip_async;
1450
1451 /*
1452 * DISCARD can flip during remount. On zoned filesystems, we
1453 * need to reset sequential-required zones.
1454 */
1455 trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1456 btrfs_is_zoned(fs_info);
1457
1458 /* Implicit trim during transaction commit. */
1459 if (trimming)
1460 btrfs_freeze_block_group(block_group);
1461
1462 /*
1463 * Btrfs_remove_chunk will abort the transaction if things go
1464 * horribly wrong.
1465 */
1466 ret = btrfs_remove_chunk(trans, block_group->start);
1467
1468 if (ret) {
1469 if (trimming)
1470 btrfs_unfreeze_block_group(block_group);
1471 goto end_trans;
1472 }
1473
1474 /*
1475 * If we're not mounted with -odiscard, we can just forget
1476 * about this block group. Otherwise we'll need to wait
1477 * until transaction commit to do the actual discard.
1478 */
1479 if (trimming) {
1480 spin_lock(&fs_info->unused_bgs_lock);
1481 /*
1482 * A concurrent scrub might have added us to the list
1483 * fs_info->unused_bgs, so use a list_move operation
1484 * to add the block group to the deleted_bgs list.
1485 */
1486 list_move(&block_group->bg_list,
1487 &trans->transaction->deleted_bgs);
1488 spin_unlock(&fs_info->unused_bgs_lock);
1489 btrfs_get_block_group(block_group);
1490 }
1491end_trans:
1492 btrfs_end_transaction(trans);
1493next:
1494 btrfs_put_block_group(block_group);
1495 spin_lock(&fs_info->unused_bgs_lock);
1496 }
1497 spin_unlock(&fs_info->unused_bgs_lock);
1498 mutex_unlock(&fs_info->reclaim_bgs_lock);
1499 return;
1500
1501flip_async:
1502 btrfs_end_transaction(trans);
1503 mutex_unlock(&fs_info->reclaim_bgs_lock);
1504 btrfs_put_block_group(block_group);
1505 btrfs_discard_punt_unused_bgs_list(fs_info);
1506}
1507
1508void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1509{
1510 struct btrfs_fs_info *fs_info = bg->fs_info;
1511
1512 spin_lock(&fs_info->unused_bgs_lock);
1513 if (list_empty(&bg->bg_list)) {
1514 btrfs_get_block_group(bg);
1515 trace_btrfs_add_unused_block_group(bg);
1516 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1517 }
1518 spin_unlock(&fs_info->unused_bgs_lock);
1519}
1520
1521/*
1522 * We want block groups with a low number of used bytes to be in the beginning
1523 * of the list, so they will get reclaimed first.
1524 */
1525static int reclaim_bgs_cmp(void *unused, const struct list_head *a,
1526 const struct list_head *b)
1527{
1528 const struct btrfs_block_group *bg1, *bg2;
1529
1530 bg1 = list_entry(a, struct btrfs_block_group, bg_list);
1531 bg2 = list_entry(b, struct btrfs_block_group, bg_list);
1532
1533 return bg1->used > bg2->used;
1534}
1535
1536static inline bool btrfs_should_reclaim(struct btrfs_fs_info *fs_info)
1537{
1538 if (btrfs_is_zoned(fs_info))
1539 return btrfs_zoned_should_reclaim(fs_info);
1540 return true;
1541}
1542
1543static bool should_reclaim_block_group(struct btrfs_block_group *bg, u64 bytes_freed)
1544{
1545 const struct btrfs_space_info *space_info = bg->space_info;
1546 const int reclaim_thresh = READ_ONCE(space_info->bg_reclaim_threshold);
1547 const u64 new_val = bg->used;
1548 const u64 old_val = new_val + bytes_freed;
1549 u64 thresh;
1550
1551 if (reclaim_thresh == 0)
1552 return false;
1553
1554 thresh = mult_perc(bg->length, reclaim_thresh);
1555
1556 /*
1557 * If we were below the threshold before don't reclaim, we are likely a
1558 * brand new block group and we don't want to relocate new block groups.
1559 */
1560 if (old_val < thresh)
1561 return false;
1562 if (new_val >= thresh)
1563 return false;
1564 return true;
1565}
1566
1567void btrfs_reclaim_bgs_work(struct work_struct *work)
1568{
1569 struct btrfs_fs_info *fs_info =
1570 container_of(work, struct btrfs_fs_info, reclaim_bgs_work);
1571 struct btrfs_block_group *bg;
1572 struct btrfs_space_info *space_info;
1573
1574 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1575 return;
1576
1577 if (btrfs_fs_closing(fs_info))
1578 return;
1579
1580 if (!btrfs_should_reclaim(fs_info))
1581 return;
1582
1583 sb_start_write(fs_info->sb);
1584
1585 if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_BALANCE)) {
1586 sb_end_write(fs_info->sb);
1587 return;
1588 }
1589
1590 /*
1591 * Long running balances can keep us blocked here for eternity, so
1592 * simply skip reclaim if we're unable to get the mutex.
1593 */
1594 if (!mutex_trylock(&fs_info->reclaim_bgs_lock)) {
1595 btrfs_exclop_finish(fs_info);
1596 sb_end_write(fs_info->sb);
1597 return;
1598 }
1599
1600 spin_lock(&fs_info->unused_bgs_lock);
1601 /*
1602 * Sort happens under lock because we can't simply splice it and sort.
1603 * The block groups might still be in use and reachable via bg_list,
1604 * and their presence in the reclaim_bgs list must be preserved.
1605 */
1606 list_sort(NULL, &fs_info->reclaim_bgs, reclaim_bgs_cmp);
1607 while (!list_empty(&fs_info->reclaim_bgs)) {
1608 u64 zone_unusable;
1609 int ret = 0;
1610
1611 bg = list_first_entry(&fs_info->reclaim_bgs,
1612 struct btrfs_block_group,
1613 bg_list);
1614 list_del_init(&bg->bg_list);
1615
1616 space_info = bg->space_info;
1617 spin_unlock(&fs_info->unused_bgs_lock);
1618
1619 /* Don't race with allocators so take the groups_sem */
1620 down_write(&space_info->groups_sem);
1621
1622 spin_lock(&bg->lock);
1623 if (bg->reserved || bg->pinned || bg->ro) {
1624 /*
1625 * We want to bail if we made new allocations or have
1626 * outstanding allocations in this block group. We do
1627 * the ro check in case balance is currently acting on
1628 * this block group.
1629 */
1630 spin_unlock(&bg->lock);
1631 up_write(&space_info->groups_sem);
1632 goto next;
1633 }
1634 if (bg->used == 0) {
1635 /*
1636 * It is possible that we trigger relocation on a block
1637 * group as its extents are deleted and it first goes
1638 * below the threshold, then shortly after goes empty.
1639 *
1640 * In this case, relocating it does delete it, but has
1641 * some overhead in relocation specific metadata, looking
1642 * for the non-existent extents and running some extra
1643 * transactions, which we can avoid by using one of the
1644 * other mechanisms for dealing with empty block groups.
1645 */
1646 if (!btrfs_test_opt(fs_info, DISCARD_ASYNC))
1647 btrfs_mark_bg_unused(bg);
1648 spin_unlock(&bg->lock);
1649 up_write(&space_info->groups_sem);
1650 goto next;
1651
1652 }
1653 /*
1654 * The block group might no longer meet the reclaim condition by
1655 * the time we get around to reclaiming it, so to avoid
1656 * reclaiming overly full block_groups, skip reclaiming them.
1657 *
1658 * Since the decision making process also depends on the amount
1659 * being freed, pass in a fake giant value to skip that extra
1660 * check, which is more meaningful when adding to the list in
1661 * the first place.
1662 */
1663 if (!should_reclaim_block_group(bg, bg->length)) {
1664 spin_unlock(&bg->lock);
1665 up_write(&space_info->groups_sem);
1666 goto next;
1667 }
1668 spin_unlock(&bg->lock);
1669
1670 /* Get out fast, in case we're unmounting the filesystem */
1671 if (btrfs_fs_closing(fs_info)) {
1672 up_write(&space_info->groups_sem);
1673 goto next;
1674 }
1675
1676 /*
1677 * Cache the zone_unusable value before turning the block group
1678 * to read only. As soon as the blog group is read only it's
1679 * zone_unusable value gets moved to the block group's read-only
1680 * bytes and isn't available for calculations anymore.
1681 */
1682 zone_unusable = bg->zone_unusable;
1683 ret = inc_block_group_ro(bg, 0);
1684 up_write(&space_info->groups_sem);
1685 if (ret < 0)
1686 goto next;
1687
1688 btrfs_info(fs_info,
1689 "reclaiming chunk %llu with %llu%% used %llu%% unusable",
1690 bg->start, div_u64(bg->used * 100, bg->length),
1691 div64_u64(zone_unusable * 100, bg->length));
1692 trace_btrfs_reclaim_block_group(bg);
1693 ret = btrfs_relocate_chunk(fs_info, bg->start);
1694 if (ret) {
1695 btrfs_dec_block_group_ro(bg);
1696 btrfs_err(fs_info, "error relocating chunk %llu",
1697 bg->start);
1698 }
1699
1700next:
1701 btrfs_put_block_group(bg);
1702 spin_lock(&fs_info->unused_bgs_lock);
1703 }
1704 spin_unlock(&fs_info->unused_bgs_lock);
1705 mutex_unlock(&fs_info->reclaim_bgs_lock);
1706 btrfs_exclop_finish(fs_info);
1707 sb_end_write(fs_info->sb);
1708}
1709
1710void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info)
1711{
1712 spin_lock(&fs_info->unused_bgs_lock);
1713 if (!list_empty(&fs_info->reclaim_bgs))
1714 queue_work(system_unbound_wq, &fs_info->reclaim_bgs_work);
1715 spin_unlock(&fs_info->unused_bgs_lock);
1716}
1717
1718void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg)
1719{
1720 struct btrfs_fs_info *fs_info = bg->fs_info;
1721
1722 spin_lock(&fs_info->unused_bgs_lock);
1723 if (list_empty(&bg->bg_list)) {
1724 btrfs_get_block_group(bg);
1725 trace_btrfs_add_reclaim_block_group(bg);
1726 list_add_tail(&bg->bg_list, &fs_info->reclaim_bgs);
1727 }
1728 spin_unlock(&fs_info->unused_bgs_lock);
1729}
1730
1731static int read_bg_from_eb(struct btrfs_fs_info *fs_info, struct btrfs_key *key,
1732 struct btrfs_path *path)
1733{
1734 struct extent_map_tree *em_tree;
1735 struct extent_map *em;
1736 struct btrfs_block_group_item bg;
1737 struct extent_buffer *leaf;
1738 int slot;
1739 u64 flags;
1740 int ret = 0;
1741
1742 slot = path->slots[0];
1743 leaf = path->nodes[0];
1744
1745 em_tree = &fs_info->mapping_tree;
1746 read_lock(&em_tree->lock);
1747 em = lookup_extent_mapping(em_tree, key->objectid, key->offset);
1748 read_unlock(&em_tree->lock);
1749 if (!em) {
1750 btrfs_err(fs_info,
1751 "logical %llu len %llu found bg but no related chunk",
1752 key->objectid, key->offset);
1753 return -ENOENT;
1754 }
1755
1756 if (em->start != key->objectid || em->len != key->offset) {
1757 btrfs_err(fs_info,
1758 "block group %llu len %llu mismatch with chunk %llu len %llu",
1759 key->objectid, key->offset, em->start, em->len);
1760 ret = -EUCLEAN;
1761 goto out_free_em;
1762 }
1763
1764 read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
1765 sizeof(bg));
1766 flags = btrfs_stack_block_group_flags(&bg) &
1767 BTRFS_BLOCK_GROUP_TYPE_MASK;
1768
1769 if (flags != (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
1770 btrfs_err(fs_info,
1771"block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
1772 key->objectid, key->offset, flags,
1773 (BTRFS_BLOCK_GROUP_TYPE_MASK & em->map_lookup->type));
1774 ret = -EUCLEAN;
1775 }
1776
1777out_free_em:
1778 free_extent_map(em);
1779 return ret;
1780}
1781
1782static int find_first_block_group(struct btrfs_fs_info *fs_info,
1783 struct btrfs_path *path,
1784 struct btrfs_key *key)
1785{
1786 struct btrfs_root *root = btrfs_block_group_root(fs_info);
1787 int ret;
1788 struct btrfs_key found_key;
1789
1790 btrfs_for_each_slot(root, key, &found_key, path, ret) {
1791 if (found_key.objectid >= key->objectid &&
1792 found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
1793 return read_bg_from_eb(fs_info, &found_key, path);
1794 }
1795 }
1796 return ret;
1797}
1798
1799static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
1800{
1801 u64 extra_flags = chunk_to_extended(flags) &
1802 BTRFS_EXTENDED_PROFILE_MASK;
1803
1804 write_seqlock(&fs_info->profiles_lock);
1805 if (flags & BTRFS_BLOCK_GROUP_DATA)
1806 fs_info->avail_data_alloc_bits |= extra_flags;
1807 if (flags & BTRFS_BLOCK_GROUP_METADATA)
1808 fs_info->avail_metadata_alloc_bits |= extra_flags;
1809 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
1810 fs_info->avail_system_alloc_bits |= extra_flags;
1811 write_sequnlock(&fs_info->profiles_lock);
1812}
1813
1814/*
1815 * Map a physical disk address to a list of logical addresses.
1816 *
1817 * @fs_info: the filesystem
1818 * @chunk_start: logical address of block group
1819 * @bdev: physical device to resolve, can be NULL to indicate any device
1820 * @physical: physical address to map to logical addresses
1821 * @logical: return array of logical addresses which map to @physical
1822 * @naddrs: length of @logical
1823 * @stripe_len: size of IO stripe for the given block group
1824 *
1825 * Maps a particular @physical disk address to a list of @logical addresses.
1826 * Used primarily to exclude those portions of a block group that contain super
1827 * block copies.
1828 */
1829int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
1830 struct block_device *bdev, u64 physical, u64 **logical,
1831 int *naddrs, int *stripe_len)
1832{
1833 struct extent_map *em;
1834 struct map_lookup *map;
1835 u64 *buf;
1836 u64 bytenr;
1837 u64 data_stripe_length;
1838 u64 io_stripe_size;
1839 int i, nr = 0;
1840 int ret = 0;
1841
1842 em = btrfs_get_chunk_map(fs_info, chunk_start, 1);
1843 if (IS_ERR(em))
1844 return -EIO;
1845
1846 map = em->map_lookup;
1847 data_stripe_length = em->orig_block_len;
1848 io_stripe_size = map->stripe_len;
1849 chunk_start = em->start;
1850
1851 /* For RAID5/6 adjust to a full IO stripe length */
1852 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
1853 io_stripe_size = map->stripe_len * nr_data_stripes(map);
1854
1855 buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
1856 if (!buf) {
1857 ret = -ENOMEM;
1858 goto out;
1859 }
1860
1861 for (i = 0; i < map->num_stripes; i++) {
1862 bool already_inserted = false;
1863 u64 stripe_nr;
1864 u64 offset;
1865 int j;
1866
1867 if (!in_range(physical, map->stripes[i].physical,
1868 data_stripe_length))
1869 continue;
1870
1871 if (bdev && map->stripes[i].dev->bdev != bdev)
1872 continue;
1873
1874 stripe_nr = physical - map->stripes[i].physical;
1875 stripe_nr = div64_u64_rem(stripe_nr, map->stripe_len, &offset);
1876
1877 if (map->type & (BTRFS_BLOCK_GROUP_RAID0 |
1878 BTRFS_BLOCK_GROUP_RAID10)) {
1879 stripe_nr = stripe_nr * map->num_stripes + i;
1880 stripe_nr = div_u64(stripe_nr, map->sub_stripes);
1881 }
1882 /*
1883 * The remaining case would be for RAID56, multiply by
1884 * nr_data_stripes(). Alternatively, just use rmap_len below
1885 * instead of map->stripe_len
1886 */
1887
1888 bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
1889
1890 /* Ensure we don't add duplicate addresses */
1891 for (j = 0; j < nr; j++) {
1892 if (buf[j] == bytenr) {
1893 already_inserted = true;
1894 break;
1895 }
1896 }
1897
1898 if (!already_inserted)
1899 buf[nr++] = bytenr;
1900 }
1901
1902 *logical = buf;
1903 *naddrs = nr;
1904 *stripe_len = io_stripe_size;
1905out:
1906 free_extent_map(em);
1907 return ret;
1908}
1909
1910static int exclude_super_stripes(struct btrfs_block_group *cache)
1911{
1912 struct btrfs_fs_info *fs_info = cache->fs_info;
1913 const bool zoned = btrfs_is_zoned(fs_info);
1914 u64 bytenr;
1915 u64 *logical;
1916 int stripe_len;
1917 int i, nr, ret;
1918
1919 if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
1920 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
1921 cache->bytes_super += stripe_len;
1922 ret = btrfs_add_excluded_extent(fs_info, cache->start,
1923 stripe_len);
1924 if (ret)
1925 return ret;
1926 }
1927
1928 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
1929 bytenr = btrfs_sb_offset(i);
1930 ret = btrfs_rmap_block(fs_info, cache->start, NULL,
1931 bytenr, &logical, &nr, &stripe_len);
1932 if (ret)
1933 return ret;
1934
1935 /* Shouldn't have super stripes in sequential zones */
1936 if (zoned && nr) {
1937 btrfs_err(fs_info,
1938 "zoned: block group %llu must not contain super block",
1939 cache->start);
1940 return -EUCLEAN;
1941 }
1942
1943 while (nr--) {
1944 u64 len = min_t(u64, stripe_len,
1945 cache->start + cache->length - logical[nr]);
1946
1947 cache->bytes_super += len;
1948 ret = btrfs_add_excluded_extent(fs_info, logical[nr],
1949 len);
1950 if (ret) {
1951 kfree(logical);
1952 return ret;
1953 }
1954 }
1955
1956 kfree(logical);
1957 }
1958 return 0;
1959}
1960
1961static struct btrfs_block_group *btrfs_create_block_group_cache(
1962 struct btrfs_fs_info *fs_info, u64 start)
1963{
1964 struct btrfs_block_group *cache;
1965
1966 cache = kzalloc(sizeof(*cache), GFP_NOFS);
1967 if (!cache)
1968 return NULL;
1969
1970 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
1971 GFP_NOFS);
1972 if (!cache->free_space_ctl) {
1973 kfree(cache);
1974 return NULL;
1975 }
1976
1977 cache->start = start;
1978
1979 cache->fs_info = fs_info;
1980 cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
1981
1982 cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
1983
1984 refcount_set(&cache->refs, 1);
1985 spin_lock_init(&cache->lock);
1986 init_rwsem(&cache->data_rwsem);
1987 INIT_LIST_HEAD(&cache->list);
1988 INIT_LIST_HEAD(&cache->cluster_list);
1989 INIT_LIST_HEAD(&cache->bg_list);
1990 INIT_LIST_HEAD(&cache->ro_list);
1991 INIT_LIST_HEAD(&cache->discard_list);
1992 INIT_LIST_HEAD(&cache->dirty_list);
1993 INIT_LIST_HEAD(&cache->io_list);
1994 INIT_LIST_HEAD(&cache->active_bg_list);
1995 btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
1996 atomic_set(&cache->frozen, 0);
1997 mutex_init(&cache->free_space_lock);
1998 cache->full_stripe_locks_root.root = RB_ROOT;
1999 mutex_init(&cache->full_stripe_locks_root.lock);
2000
2001 return cache;
2002}
2003
2004/*
2005 * Iterate all chunks and verify that each of them has the corresponding block
2006 * group
2007 */
2008static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
2009{
2010 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
2011 struct extent_map *em;
2012 struct btrfs_block_group *bg;
2013 u64 start = 0;
2014 int ret = 0;
2015
2016 while (1) {
2017 read_lock(&map_tree->lock);
2018 /*
2019 * lookup_extent_mapping will return the first extent map
2020 * intersecting the range, so setting @len to 1 is enough to
2021 * get the first chunk.
2022 */
2023 em = lookup_extent_mapping(map_tree, start, 1);
2024 read_unlock(&map_tree->lock);
2025 if (!em)
2026 break;
2027
2028 bg = btrfs_lookup_block_group(fs_info, em->start);
2029 if (!bg) {
2030 btrfs_err(fs_info,
2031 "chunk start=%llu len=%llu doesn't have corresponding block group",
2032 em->start, em->len);
2033 ret = -EUCLEAN;
2034 free_extent_map(em);
2035 break;
2036 }
2037 if (bg->start != em->start || bg->length != em->len ||
2038 (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
2039 (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2040 btrfs_err(fs_info,
2041"chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
2042 em->start, em->len,
2043 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
2044 bg->start, bg->length,
2045 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
2046 ret = -EUCLEAN;
2047 free_extent_map(em);
2048 btrfs_put_block_group(bg);
2049 break;
2050 }
2051 start = em->start + em->len;
2052 free_extent_map(em);
2053 btrfs_put_block_group(bg);
2054 }
2055 return ret;
2056}
2057
2058static int read_one_block_group(struct btrfs_fs_info *info,
2059 struct btrfs_block_group_item *bgi,
2060 const struct btrfs_key *key,
2061 int need_clear)
2062{
2063 struct btrfs_block_group *cache;
2064 const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
2065 int ret;
2066
2067 ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
2068
2069 cache = btrfs_create_block_group_cache(info, key->objectid);
2070 if (!cache)
2071 return -ENOMEM;
2072
2073 cache->length = key->offset;
2074 cache->used = btrfs_stack_block_group_used(bgi);
2075 cache->commit_used = cache->used;
2076 cache->flags = btrfs_stack_block_group_flags(bgi);
2077 cache->global_root_id = btrfs_stack_block_group_chunk_objectid(bgi);
2078
2079 set_free_space_tree_thresholds(cache);
2080
2081 if (need_clear) {
2082 /*
2083 * When we mount with old space cache, we need to
2084 * set BTRFS_DC_CLEAR and set dirty flag.
2085 *
2086 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
2087 * truncate the old free space cache inode and
2088 * setup a new one.
2089 * b) Setting 'dirty flag' makes sure that we flush
2090 * the new space cache info onto disk.
2091 */
2092 if (btrfs_test_opt(info, SPACE_CACHE))
2093 cache->disk_cache_state = BTRFS_DC_CLEAR;
2094 }
2095 if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
2096 (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
2097 btrfs_err(info,
2098"bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2099 cache->start);
2100 ret = -EINVAL;
2101 goto error;
2102 }
2103
2104 ret = btrfs_load_block_group_zone_info(cache, false);
2105 if (ret) {
2106 btrfs_err(info, "zoned: failed to load zone info of bg %llu",
2107 cache->start);
2108 goto error;
2109 }
2110
2111 /*
2112 * We need to exclude the super stripes now so that the space info has
2113 * super bytes accounted for, otherwise we'll think we have more space
2114 * than we actually do.
2115 */
2116 ret = exclude_super_stripes(cache);
2117 if (ret) {
2118 /* We may have excluded something, so call this just in case. */
2119 btrfs_free_excluded_extents(cache);
2120 goto error;
2121 }
2122
2123 /*
2124 * For zoned filesystem, space after the allocation offset is the only
2125 * free space for a block group. So, we don't need any caching work.
2126 * btrfs_calc_zone_unusable() will set the amount of free space and
2127 * zone_unusable space.
2128 *
2129 * For regular filesystem, check for two cases, either we are full, and
2130 * therefore don't need to bother with the caching work since we won't
2131 * find any space, or we are empty, and we can just add all the space
2132 * in and be done with it. This saves us _a_lot_ of time, particularly
2133 * in the full case.
2134 */
2135 if (btrfs_is_zoned(info)) {
2136 btrfs_calc_zone_unusable(cache);
2137 /* Should not have any excluded extents. Just in case, though. */
2138 btrfs_free_excluded_extents(cache);
2139 } else if (cache->length == cache->used) {
2140 cache->cached = BTRFS_CACHE_FINISHED;
2141 btrfs_free_excluded_extents(cache);
2142 } else if (cache->used == 0) {
2143 cache->cached = BTRFS_CACHE_FINISHED;
2144 add_new_free_space(cache, cache->start,
2145 cache->start + cache->length);
2146 btrfs_free_excluded_extents(cache);
2147 }
2148
2149 ret = btrfs_add_block_group_cache(info, cache);
2150 if (ret) {
2151 btrfs_remove_free_space_cache(cache);
2152 goto error;
2153 }
2154 trace_btrfs_add_block_group(info, cache, 0);
2155 btrfs_add_bg_to_space_info(info, cache);
2156
2157 set_avail_alloc_bits(info, cache->flags);
2158 if (btrfs_chunk_writeable(info, cache->start)) {
2159 if (cache->used == 0) {
2160 ASSERT(list_empty(&cache->bg_list));
2161 if (btrfs_test_opt(info, DISCARD_ASYNC))
2162 btrfs_discard_queue_work(&info->discard_ctl, cache);
2163 else
2164 btrfs_mark_bg_unused(cache);
2165 }
2166 } else {
2167 inc_block_group_ro(cache, 1);
2168 }
2169
2170 return 0;
2171error:
2172 btrfs_put_block_group(cache);
2173 return ret;
2174}
2175
2176static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
2177{
2178 struct extent_map_tree *em_tree = &fs_info->mapping_tree;
2179 struct rb_node *node;
2180 int ret = 0;
2181
2182 for (node = rb_first_cached(&em_tree->map); node; node = rb_next(node)) {
2183 struct extent_map *em;
2184 struct map_lookup *map;
2185 struct btrfs_block_group *bg;
2186
2187 em = rb_entry(node, struct extent_map, rb_node);
2188 map = em->map_lookup;
2189 bg = btrfs_create_block_group_cache(fs_info, em->start);
2190 if (!bg) {
2191 ret = -ENOMEM;
2192 break;
2193 }
2194
2195 /* Fill dummy cache as FULL */
2196 bg->length = em->len;
2197 bg->flags = map->type;
2198 bg->cached = BTRFS_CACHE_FINISHED;
2199 bg->used = em->len;
2200 bg->flags = map->type;
2201 ret = btrfs_add_block_group_cache(fs_info, bg);
2202 /*
2203 * We may have some valid block group cache added already, in
2204 * that case we skip to the next one.
2205 */
2206 if (ret == -EEXIST) {
2207 ret = 0;
2208 btrfs_put_block_group(bg);
2209 continue;
2210 }
2211
2212 if (ret) {
2213 btrfs_remove_free_space_cache(bg);
2214 btrfs_put_block_group(bg);
2215 break;
2216 }
2217
2218 btrfs_add_bg_to_space_info(fs_info, bg);
2219
2220 set_avail_alloc_bits(fs_info, bg->flags);
2221 }
2222 if (!ret)
2223 btrfs_init_global_block_rsv(fs_info);
2224 return ret;
2225}
2226
2227int btrfs_read_block_groups(struct btrfs_fs_info *info)
2228{
2229 struct btrfs_root *root = btrfs_block_group_root(info);
2230 struct btrfs_path *path;
2231 int ret;
2232 struct btrfs_block_group *cache;
2233 struct btrfs_space_info *space_info;
2234 struct btrfs_key key;
2235 int need_clear = 0;
2236 u64 cache_gen;
2237
2238 /*
2239 * Either no extent root (with ibadroots rescue option) or we have
2240 * unsupported RO options. The fs can never be mounted read-write, so no
2241 * need to waste time searching block group items.
2242 *
2243 * This also allows new extent tree related changes to be RO compat,
2244 * no need for a full incompat flag.
2245 */
2246 if (!root || (btrfs_super_compat_ro_flags(info->super_copy) &
2247 ~BTRFS_FEATURE_COMPAT_RO_SUPP))
2248 return fill_dummy_bgs(info);
2249
2250 key.objectid = 0;
2251 key.offset = 0;
2252 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2253 path = btrfs_alloc_path();
2254 if (!path)
2255 return -ENOMEM;
2256
2257 cache_gen = btrfs_super_cache_generation(info->super_copy);
2258 if (btrfs_test_opt(info, SPACE_CACHE) &&
2259 btrfs_super_generation(info->super_copy) != cache_gen)
2260 need_clear = 1;
2261 if (btrfs_test_opt(info, CLEAR_CACHE))
2262 need_clear = 1;
2263
2264 while (1) {
2265 struct btrfs_block_group_item bgi;
2266 struct extent_buffer *leaf;
2267 int slot;
2268
2269 ret = find_first_block_group(info, path, &key);
2270 if (ret > 0)
2271 break;
2272 if (ret != 0)
2273 goto error;
2274
2275 leaf = path->nodes[0];
2276 slot = path->slots[0];
2277
2278 read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2279 sizeof(bgi));
2280
2281 btrfs_item_key_to_cpu(leaf, &key, slot);
2282 btrfs_release_path(path);
2283 ret = read_one_block_group(info, &bgi, &key, need_clear);
2284 if (ret < 0)
2285 goto error;
2286 key.objectid += key.offset;
2287 key.offset = 0;
2288 }
2289 btrfs_release_path(path);
2290
2291 list_for_each_entry(space_info, &info->space_info, list) {
2292 int i;
2293
2294 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2295 if (list_empty(&space_info->block_groups[i]))
2296 continue;
2297 cache = list_first_entry(&space_info->block_groups[i],
2298 struct btrfs_block_group,
2299 list);
2300 btrfs_sysfs_add_block_group_type(cache);
2301 }
2302
2303 if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2304 (BTRFS_BLOCK_GROUP_RAID10 |
2305 BTRFS_BLOCK_GROUP_RAID1_MASK |
2306 BTRFS_BLOCK_GROUP_RAID56_MASK |
2307 BTRFS_BLOCK_GROUP_DUP)))
2308 continue;
2309 /*
2310 * Avoid allocating from un-mirrored block group if there are
2311 * mirrored block groups.
2312 */
2313 list_for_each_entry(cache,
2314 &space_info->block_groups[BTRFS_RAID_RAID0],
2315 list)
2316 inc_block_group_ro(cache, 1);
2317 list_for_each_entry(cache,
2318 &space_info->block_groups[BTRFS_RAID_SINGLE],
2319 list)
2320 inc_block_group_ro(cache, 1);
2321 }
2322
2323 btrfs_init_global_block_rsv(info);
2324 ret = check_chunk_block_group_mappings(info);
2325error:
2326 btrfs_free_path(path);
2327 /*
2328 * We've hit some error while reading the extent tree, and have
2329 * rescue=ibadroots mount option.
2330 * Try to fill the tree using dummy block groups so that the user can
2331 * continue to mount and grab their data.
2332 */
2333 if (ret && btrfs_test_opt(info, IGNOREBADROOTS))
2334 ret = fill_dummy_bgs(info);
2335 return ret;
2336}
2337
2338/*
2339 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2340 * allocation.
2341 *
2342 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2343 * phases.
2344 */
2345static int insert_block_group_item(struct btrfs_trans_handle *trans,
2346 struct btrfs_block_group *block_group)
2347{
2348 struct btrfs_fs_info *fs_info = trans->fs_info;
2349 struct btrfs_block_group_item bgi;
2350 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2351 struct btrfs_key key;
2352
2353 spin_lock(&block_group->lock);
2354 btrfs_set_stack_block_group_used(&bgi, block_group->used);
2355 btrfs_set_stack_block_group_chunk_objectid(&bgi,
2356 block_group->global_root_id);
2357 btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2358 key.objectid = block_group->start;
2359 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2360 key.offset = block_group->length;
2361 spin_unlock(&block_group->lock);
2362
2363 return btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2364}
2365
2366static int insert_dev_extent(struct btrfs_trans_handle *trans,
2367 struct btrfs_device *device, u64 chunk_offset,
2368 u64 start, u64 num_bytes)
2369{
2370 struct btrfs_fs_info *fs_info = device->fs_info;
2371 struct btrfs_root *root = fs_info->dev_root;
2372 struct btrfs_path *path;
2373 struct btrfs_dev_extent *extent;
2374 struct extent_buffer *leaf;
2375 struct btrfs_key key;
2376 int ret;
2377
2378 WARN_ON(!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state));
2379 WARN_ON(test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state));
2380 path = btrfs_alloc_path();
2381 if (!path)
2382 return -ENOMEM;
2383
2384 key.objectid = device->devid;
2385 key.type = BTRFS_DEV_EXTENT_KEY;
2386 key.offset = start;
2387 ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*extent));
2388 if (ret)
2389 goto out;
2390
2391 leaf = path->nodes[0];
2392 extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
2393 btrfs_set_dev_extent_chunk_tree(leaf, extent, BTRFS_CHUNK_TREE_OBJECTID);
2394 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
2395 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2396 btrfs_set_dev_extent_chunk_offset(leaf, extent, chunk_offset);
2397
2398 btrfs_set_dev_extent_length(leaf, extent, num_bytes);
2399 btrfs_mark_buffer_dirty(leaf);
2400out:
2401 btrfs_free_path(path);
2402 return ret;
2403}
2404
2405/*
2406 * This function belongs to phase 2.
2407 *
2408 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2409 * phases.
2410 */
2411static int insert_dev_extents(struct btrfs_trans_handle *trans,
2412 u64 chunk_offset, u64 chunk_size)
2413{
2414 struct btrfs_fs_info *fs_info = trans->fs_info;
2415 struct btrfs_device *device;
2416 struct extent_map *em;
2417 struct map_lookup *map;
2418 u64 dev_offset;
2419 u64 stripe_size;
2420 int i;
2421 int ret = 0;
2422
2423 em = btrfs_get_chunk_map(fs_info, chunk_offset, chunk_size);
2424 if (IS_ERR(em))
2425 return PTR_ERR(em);
2426
2427 map = em->map_lookup;
2428 stripe_size = em->orig_block_len;
2429
2430 /*
2431 * Take the device list mutex to prevent races with the final phase of
2432 * a device replace operation that replaces the device object associated
2433 * with the map's stripes, because the device object's id can change
2434 * at any time during that final phase of the device replace operation
2435 * (dev-replace.c:btrfs_dev_replace_finishing()), so we could grab the
2436 * replaced device and then see it with an ID of BTRFS_DEV_REPLACE_DEVID,
2437 * resulting in persisting a device extent item with such ID.
2438 */
2439 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2440 for (i = 0; i < map->num_stripes; i++) {
2441 device = map->stripes[i].dev;
2442 dev_offset = map->stripes[i].physical;
2443
2444 ret = insert_dev_extent(trans, device, chunk_offset, dev_offset,
2445 stripe_size);
2446 if (ret)
2447 break;
2448 }
2449 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2450
2451 free_extent_map(em);
2452 return ret;
2453}
2454
2455/*
2456 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2457 * chunk allocation.
2458 *
2459 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2460 * phases.
2461 */
2462void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2463{
2464 struct btrfs_fs_info *fs_info = trans->fs_info;
2465 struct btrfs_block_group *block_group;
2466 int ret = 0;
2467
2468 while (!list_empty(&trans->new_bgs)) {
2469 int index;
2470
2471 block_group = list_first_entry(&trans->new_bgs,
2472 struct btrfs_block_group,
2473 bg_list);
2474 if (ret)
2475 goto next;
2476
2477 index = btrfs_bg_flags_to_raid_index(block_group->flags);
2478
2479 ret = insert_block_group_item(trans, block_group);
2480 if (ret)
2481 btrfs_abort_transaction(trans, ret);
2482 if (!test_bit(BLOCK_GROUP_FLAG_CHUNK_ITEM_INSERTED,
2483 &block_group->runtime_flags)) {
2484 mutex_lock(&fs_info->chunk_mutex);
2485 ret = btrfs_chunk_alloc_add_chunk_item(trans, block_group);
2486 mutex_unlock(&fs_info->chunk_mutex);
2487 if (ret)
2488 btrfs_abort_transaction(trans, ret);
2489 }
2490 ret = insert_dev_extents(trans, block_group->start,
2491 block_group->length);
2492 if (ret)
2493 btrfs_abort_transaction(trans, ret);
2494 add_block_group_free_space(trans, block_group);
2495
2496 /*
2497 * If we restriped during balance, we may have added a new raid
2498 * type, so now add the sysfs entries when it is safe to do so.
2499 * We don't have to worry about locking here as it's handled in
2500 * btrfs_sysfs_add_block_group_type.
2501 */
2502 if (block_group->space_info->block_group_kobjs[index] == NULL)
2503 btrfs_sysfs_add_block_group_type(block_group);
2504
2505 /* Already aborted the transaction if it failed. */
2506next:
2507 btrfs_delayed_refs_rsv_release(fs_info, 1);
2508 list_del_init(&block_group->bg_list);
2509 }
2510 btrfs_trans_release_chunk_metadata(trans);
2511}
2512
2513/*
2514 * For extent tree v2 we use the block_group_item->chunk_offset to point at our
2515 * global root id. For v1 it's always set to BTRFS_FIRST_CHUNK_TREE_OBJECTID.
2516 */
2517static u64 calculate_global_root_id(struct btrfs_fs_info *fs_info, u64 offset)
2518{
2519 u64 div = SZ_1G;
2520 u64 index;
2521
2522 if (!btrfs_fs_incompat(fs_info, EXTENT_TREE_V2))
2523 return BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2524
2525 /* If we have a smaller fs index based on 128MiB. */
2526 if (btrfs_super_total_bytes(fs_info->super_copy) <= (SZ_1G * 10ULL))
2527 div = SZ_128M;
2528
2529 offset = div64_u64(offset, div);
2530 div64_u64_rem(offset, fs_info->nr_global_roots, &index);
2531 return index;
2532}
2533
2534struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
2535 u64 bytes_used, u64 type,
2536 u64 chunk_offset, u64 size)
2537{
2538 struct btrfs_fs_info *fs_info = trans->fs_info;
2539 struct btrfs_block_group *cache;
2540 int ret;
2541
2542 btrfs_set_log_full_commit(trans);
2543
2544 cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2545 if (!cache)
2546 return ERR_PTR(-ENOMEM);
2547
2548 cache->length = size;
2549 set_free_space_tree_thresholds(cache);
2550 cache->used = bytes_used;
2551 cache->flags = type;
2552 cache->cached = BTRFS_CACHE_FINISHED;
2553 cache->global_root_id = calculate_global_root_id(fs_info, cache->start);
2554
2555 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2556 set_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE, &cache->runtime_flags);
2557
2558 ret = btrfs_load_block_group_zone_info(cache, true);
2559 if (ret) {
2560 btrfs_put_block_group(cache);
2561 return ERR_PTR(ret);
2562 }
2563
2564 ret = exclude_super_stripes(cache);
2565 if (ret) {
2566 /* We may have excluded something, so call this just in case */
2567 btrfs_free_excluded_extents(cache);
2568 btrfs_put_block_group(cache);
2569 return ERR_PTR(ret);
2570 }
2571
2572 add_new_free_space(cache, chunk_offset, chunk_offset + size);
2573
2574 btrfs_free_excluded_extents(cache);
2575
2576 /*
2577 * Ensure the corresponding space_info object is created and
2578 * assigned to our block group. We want our bg to be added to the rbtree
2579 * with its ->space_info set.
2580 */
2581 cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2582 ASSERT(cache->space_info);
2583
2584 ret = btrfs_add_block_group_cache(fs_info, cache);
2585 if (ret) {
2586 btrfs_remove_free_space_cache(cache);
2587 btrfs_put_block_group(cache);
2588 return ERR_PTR(ret);
2589 }
2590
2591 /*
2592 * Now that our block group has its ->space_info set and is inserted in
2593 * the rbtree, update the space info's counters.
2594 */
2595 trace_btrfs_add_block_group(fs_info, cache, 1);
2596 btrfs_add_bg_to_space_info(fs_info, cache);
2597 btrfs_update_global_block_rsv(fs_info);
2598
2599#ifdef CONFIG_BTRFS_DEBUG
2600 if (btrfs_should_fragment_free_space(cache)) {
2601 u64 new_bytes_used = size - bytes_used;
2602
2603 cache->space_info->bytes_used += new_bytes_used >> 1;
2604 fragment_free_space(cache);
2605 }
2606#endif
2607
2608 list_add_tail(&cache->bg_list, &trans->new_bgs);
2609 trans->delayed_ref_updates++;
2610 btrfs_update_delayed_refs_rsv(trans);
2611
2612 set_avail_alloc_bits(fs_info, type);
2613 return cache;
2614}
2615
2616/*
2617 * Mark one block group RO, can be called several times for the same block
2618 * group.
2619 *
2620 * @cache: the destination block group
2621 * @do_chunk_alloc: whether need to do chunk pre-allocation, this is to
2622 * ensure we still have some free space after marking this
2623 * block group RO.
2624 */
2625int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2626 bool do_chunk_alloc)
2627{
2628 struct btrfs_fs_info *fs_info = cache->fs_info;
2629 struct btrfs_trans_handle *trans;
2630 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2631 u64 alloc_flags;
2632 int ret;
2633 bool dirty_bg_running;
2634
2635 /*
2636 * This can only happen when we are doing read-only scrub on read-only
2637 * mount.
2638 * In that case we should not start a new transaction on read-only fs.
2639 * Thus here we skip all chunk allocations.
2640 */
2641 if (sb_rdonly(fs_info->sb)) {
2642 mutex_lock(&fs_info->ro_block_group_mutex);
2643 ret = inc_block_group_ro(cache, 0);
2644 mutex_unlock(&fs_info->ro_block_group_mutex);
2645 return ret;
2646 }
2647
2648 do {
2649 trans = btrfs_join_transaction(root);
2650 if (IS_ERR(trans))
2651 return PTR_ERR(trans);
2652
2653 dirty_bg_running = false;
2654
2655 /*
2656 * We're not allowed to set block groups readonly after the dirty
2657 * block group cache has started writing. If it already started,
2658 * back off and let this transaction commit.
2659 */
2660 mutex_lock(&fs_info->ro_block_group_mutex);
2661 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2662 u64 transid = trans->transid;
2663
2664 mutex_unlock(&fs_info->ro_block_group_mutex);
2665 btrfs_end_transaction(trans);
2666
2667 ret = btrfs_wait_for_commit(fs_info, transid);
2668 if (ret)
2669 return ret;
2670 dirty_bg_running = true;
2671 }
2672 } while (dirty_bg_running);
2673
2674 if (do_chunk_alloc) {
2675 /*
2676 * If we are changing raid levels, try to allocate a
2677 * corresponding block group with the new raid level.
2678 */
2679 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2680 if (alloc_flags != cache->flags) {
2681 ret = btrfs_chunk_alloc(trans, alloc_flags,
2682 CHUNK_ALLOC_FORCE);
2683 /*
2684 * ENOSPC is allowed here, we may have enough space
2685 * already allocated at the new raid level to carry on
2686 */
2687 if (ret == -ENOSPC)
2688 ret = 0;
2689 if (ret < 0)
2690 goto out;
2691 }
2692 }
2693
2694 ret = inc_block_group_ro(cache, 0);
2695 if (!do_chunk_alloc || ret == -ETXTBSY)
2696 goto unlock_out;
2697 if (!ret)
2698 goto out;
2699 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2700 ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2701 if (ret < 0)
2702 goto out;
2703 /*
2704 * We have allocated a new chunk. We also need to activate that chunk to
2705 * grant metadata tickets for zoned filesystem.
2706 */
2707 ret = btrfs_zoned_activate_one_bg(fs_info, cache->space_info, true);
2708 if (ret < 0)
2709 goto out;
2710
2711 ret = inc_block_group_ro(cache, 0);
2712 if (ret == -ETXTBSY)
2713 goto unlock_out;
2714out:
2715 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2716 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2717 mutex_lock(&fs_info->chunk_mutex);
2718 check_system_chunk(trans, alloc_flags);
2719 mutex_unlock(&fs_info->chunk_mutex);
2720 }
2721unlock_out:
2722 mutex_unlock(&fs_info->ro_block_group_mutex);
2723
2724 btrfs_end_transaction(trans);
2725 return ret;
2726}
2727
2728void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
2729{
2730 struct btrfs_space_info *sinfo = cache->space_info;
2731 u64 num_bytes;
2732
2733 BUG_ON(!cache->ro);
2734
2735 spin_lock(&sinfo->lock);
2736 spin_lock(&cache->lock);
2737 if (!--cache->ro) {
2738 if (btrfs_is_zoned(cache->fs_info)) {
2739 /* Migrate zone_unusable bytes back */
2740 cache->zone_unusable =
2741 (cache->alloc_offset - cache->used) +
2742 (cache->length - cache->zone_capacity);
2743 sinfo->bytes_zone_unusable += cache->zone_unusable;
2744 sinfo->bytes_readonly -= cache->zone_unusable;
2745 }
2746 num_bytes = cache->length - cache->reserved -
2747 cache->pinned - cache->bytes_super -
2748 cache->zone_unusable - cache->used;
2749 sinfo->bytes_readonly -= num_bytes;
2750 list_del_init(&cache->ro_list);
2751 }
2752 spin_unlock(&cache->lock);
2753 spin_unlock(&sinfo->lock);
2754}
2755
2756static int update_block_group_item(struct btrfs_trans_handle *trans,
2757 struct btrfs_path *path,
2758 struct btrfs_block_group *cache)
2759{
2760 struct btrfs_fs_info *fs_info = trans->fs_info;
2761 int ret;
2762 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2763 unsigned long bi;
2764 struct extent_buffer *leaf;
2765 struct btrfs_block_group_item bgi;
2766 struct btrfs_key key;
2767 u64 old_commit_used;
2768 u64 used;
2769
2770 /*
2771 * Block group items update can be triggered out of commit transaction
2772 * critical section, thus we need a consistent view of used bytes.
2773 * We cannot use cache->used directly outside of the spin lock, as it
2774 * may be changed.
2775 */
2776 spin_lock(&cache->lock);
2777 old_commit_used = cache->commit_used;
2778 used = cache->used;
2779 /* No change in used bytes, can safely skip it. */
2780 if (cache->commit_used == used) {
2781 spin_unlock(&cache->lock);
2782 return 0;
2783 }
2784 cache->commit_used = used;
2785 spin_unlock(&cache->lock);
2786
2787 key.objectid = cache->start;
2788 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2789 key.offset = cache->length;
2790
2791 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2792 if (ret) {
2793 if (ret > 0)
2794 ret = -ENOENT;
2795 goto fail;
2796 }
2797
2798 leaf = path->nodes[0];
2799 bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2800 btrfs_set_stack_block_group_used(&bgi, used);
2801 btrfs_set_stack_block_group_chunk_objectid(&bgi,
2802 cache->global_root_id);
2803 btrfs_set_stack_block_group_flags(&bgi, cache->flags);
2804 write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
2805 btrfs_mark_buffer_dirty(leaf);
2806fail:
2807 btrfs_release_path(path);
2808 /* We didn't update the block group item, need to revert @commit_used. */
2809 if (ret < 0) {
2810 spin_lock(&cache->lock);
2811 cache->commit_used = old_commit_used;
2812 spin_unlock(&cache->lock);
2813 }
2814 return ret;
2815
2816}
2817
2818static int cache_save_setup(struct btrfs_block_group *block_group,
2819 struct btrfs_trans_handle *trans,
2820 struct btrfs_path *path)
2821{
2822 struct btrfs_fs_info *fs_info = block_group->fs_info;
2823 struct btrfs_root *root = fs_info->tree_root;
2824 struct inode *inode = NULL;
2825 struct extent_changeset *data_reserved = NULL;
2826 u64 alloc_hint = 0;
2827 int dcs = BTRFS_DC_ERROR;
2828 u64 cache_size = 0;
2829 int retries = 0;
2830 int ret = 0;
2831
2832 if (!btrfs_test_opt(fs_info, SPACE_CACHE))
2833 return 0;
2834
2835 /*
2836 * If this block group is smaller than 100 megs don't bother caching the
2837 * block group.
2838 */
2839 if (block_group->length < (100 * SZ_1M)) {
2840 spin_lock(&block_group->lock);
2841 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2842 spin_unlock(&block_group->lock);
2843 return 0;
2844 }
2845
2846 if (TRANS_ABORTED(trans))
2847 return 0;
2848again:
2849 inode = lookup_free_space_inode(block_group, path);
2850 if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2851 ret = PTR_ERR(inode);
2852 btrfs_release_path(path);
2853 goto out;
2854 }
2855
2856 if (IS_ERR(inode)) {
2857 BUG_ON(retries);
2858 retries++;
2859
2860 if (block_group->ro)
2861 goto out_free;
2862
2863 ret = create_free_space_inode(trans, block_group, path);
2864 if (ret)
2865 goto out_free;
2866 goto again;
2867 }
2868
2869 /*
2870 * We want to set the generation to 0, that way if anything goes wrong
2871 * from here on out we know not to trust this cache when we load up next
2872 * time.
2873 */
2874 BTRFS_I(inode)->generation = 0;
2875 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2876 if (ret) {
2877 /*
2878 * So theoretically we could recover from this, simply set the
2879 * super cache generation to 0 so we know to invalidate the
2880 * cache, but then we'd have to keep track of the block groups
2881 * that fail this way so we know we _have_ to reset this cache
2882 * before the next commit or risk reading stale cache. So to
2883 * limit our exposure to horrible edge cases lets just abort the
2884 * transaction, this only happens in really bad situations
2885 * anyway.
2886 */
2887 btrfs_abort_transaction(trans, ret);
2888 goto out_put;
2889 }
2890 WARN_ON(ret);
2891
2892 /* We've already setup this transaction, go ahead and exit */
2893 if (block_group->cache_generation == trans->transid &&
2894 i_size_read(inode)) {
2895 dcs = BTRFS_DC_SETUP;
2896 goto out_put;
2897 }
2898
2899 if (i_size_read(inode) > 0) {
2900 ret = btrfs_check_trunc_cache_free_space(fs_info,
2901 &fs_info->global_block_rsv);
2902 if (ret)
2903 goto out_put;
2904
2905 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
2906 if (ret)
2907 goto out_put;
2908 }
2909
2910 spin_lock(&block_group->lock);
2911 if (block_group->cached != BTRFS_CACHE_FINISHED ||
2912 !btrfs_test_opt(fs_info, SPACE_CACHE)) {
2913 /*
2914 * don't bother trying to write stuff out _if_
2915 * a) we're not cached,
2916 * b) we're with nospace_cache mount option,
2917 * c) we're with v2 space_cache (FREE_SPACE_TREE).
2918 */
2919 dcs = BTRFS_DC_WRITTEN;
2920 spin_unlock(&block_group->lock);
2921 goto out_put;
2922 }
2923 spin_unlock(&block_group->lock);
2924
2925 /*
2926 * We hit an ENOSPC when setting up the cache in this transaction, just
2927 * skip doing the setup, we've already cleared the cache so we're safe.
2928 */
2929 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
2930 ret = -ENOSPC;
2931 goto out_put;
2932 }
2933
2934 /*
2935 * Try to preallocate enough space based on how big the block group is.
2936 * Keep in mind this has to include any pinned space which could end up
2937 * taking up quite a bit since it's not folded into the other space
2938 * cache.
2939 */
2940 cache_size = div_u64(block_group->length, SZ_256M);
2941 if (!cache_size)
2942 cache_size = 1;
2943
2944 cache_size *= 16;
2945 cache_size *= fs_info->sectorsize;
2946
2947 ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
2948 cache_size, false);
2949 if (ret)
2950 goto out_put;
2951
2952 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, cache_size,
2953 cache_size, cache_size,
2954 &alloc_hint);
2955 /*
2956 * Our cache requires contiguous chunks so that we don't modify a bunch
2957 * of metadata or split extents when writing the cache out, which means
2958 * we can enospc if we are heavily fragmented in addition to just normal
2959 * out of space conditions. So if we hit this just skip setting up any
2960 * other block groups for this transaction, maybe we'll unpin enough
2961 * space the next time around.
2962 */
2963 if (!ret)
2964 dcs = BTRFS_DC_SETUP;
2965 else if (ret == -ENOSPC)
2966 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
2967
2968out_put:
2969 iput(inode);
2970out_free:
2971 btrfs_release_path(path);
2972out:
2973 spin_lock(&block_group->lock);
2974 if (!ret && dcs == BTRFS_DC_SETUP)
2975 block_group->cache_generation = trans->transid;
2976 block_group->disk_cache_state = dcs;
2977 spin_unlock(&block_group->lock);
2978
2979 extent_changeset_free(data_reserved);
2980 return ret;
2981}
2982
2983int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
2984{
2985 struct btrfs_fs_info *fs_info = trans->fs_info;
2986 struct btrfs_block_group *cache, *tmp;
2987 struct btrfs_transaction *cur_trans = trans->transaction;
2988 struct btrfs_path *path;
2989
2990 if (list_empty(&cur_trans->dirty_bgs) ||
2991 !btrfs_test_opt(fs_info, SPACE_CACHE))
2992 return 0;
2993
2994 path = btrfs_alloc_path();
2995 if (!path)
2996 return -ENOMEM;
2997
2998 /* Could add new block groups, use _safe just in case */
2999 list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3000 dirty_list) {
3001 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3002 cache_save_setup(cache, trans, path);
3003 }
3004
3005 btrfs_free_path(path);
3006 return 0;
3007}
3008
3009/*
3010 * Transaction commit does final block group cache writeback during a critical
3011 * section where nothing is allowed to change the FS. This is required in
3012 * order for the cache to actually match the block group, but can introduce a
3013 * lot of latency into the commit.
3014 *
3015 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
3016 * There's a chance we'll have to redo some of it if the block group changes
3017 * again during the commit, but it greatly reduces the commit latency by
3018 * getting rid of the easy block groups while we're still allowing others to
3019 * join the commit.
3020 */
3021int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3022{
3023 struct btrfs_fs_info *fs_info = trans->fs_info;
3024 struct btrfs_block_group *cache;
3025 struct btrfs_transaction *cur_trans = trans->transaction;
3026 int ret = 0;
3027 int should_put;
3028 struct btrfs_path *path = NULL;
3029 LIST_HEAD(dirty);
3030 struct list_head *io = &cur_trans->io_bgs;
3031 int loops = 0;
3032
3033 spin_lock(&cur_trans->dirty_bgs_lock);
3034 if (list_empty(&cur_trans->dirty_bgs)) {
3035 spin_unlock(&cur_trans->dirty_bgs_lock);
3036 return 0;
3037 }
3038 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3039 spin_unlock(&cur_trans->dirty_bgs_lock);
3040
3041again:
3042 /* Make sure all the block groups on our dirty list actually exist */
3043 btrfs_create_pending_block_groups(trans);
3044
3045 if (!path) {
3046 path = btrfs_alloc_path();
3047 if (!path) {
3048 ret = -ENOMEM;
3049 goto out;
3050 }
3051 }
3052
3053 /*
3054 * cache_write_mutex is here only to save us from balance or automatic
3055 * removal of empty block groups deleting this block group while we are
3056 * writing out the cache
3057 */
3058 mutex_lock(&trans->transaction->cache_write_mutex);
3059 while (!list_empty(&dirty)) {
3060 bool drop_reserve = true;
3061
3062 cache = list_first_entry(&dirty, struct btrfs_block_group,
3063 dirty_list);
3064 /*
3065 * This can happen if something re-dirties a block group that
3066 * is already under IO. Just wait for it to finish and then do
3067 * it all again
3068 */
3069 if (!list_empty(&cache->io_list)) {
3070 list_del_init(&cache->io_list);
3071 btrfs_wait_cache_io(trans, cache, path);
3072 btrfs_put_block_group(cache);
3073 }
3074
3075
3076 /*
3077 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
3078 * it should update the cache_state. Don't delete until after
3079 * we wait.
3080 *
3081 * Since we're not running in the commit critical section
3082 * we need the dirty_bgs_lock to protect from update_block_group
3083 */
3084 spin_lock(&cur_trans->dirty_bgs_lock);
3085 list_del_init(&cache->dirty_list);
3086 spin_unlock(&cur_trans->dirty_bgs_lock);
3087
3088 should_put = 1;
3089
3090 cache_save_setup(cache, trans, path);
3091
3092 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3093 cache->io_ctl.inode = NULL;
3094 ret = btrfs_write_out_cache(trans, cache, path);
3095 if (ret == 0 && cache->io_ctl.inode) {
3096 should_put = 0;
3097
3098 /*
3099 * The cache_write_mutex is protecting the
3100 * io_list, also refer to the definition of
3101 * btrfs_transaction::io_bgs for more details
3102 */
3103 list_add_tail(&cache->io_list, io);
3104 } else {
3105 /*
3106 * If we failed to write the cache, the
3107 * generation will be bad and life goes on
3108 */
3109 ret = 0;
3110 }
3111 }
3112 if (!ret) {
3113 ret = update_block_group_item(trans, path, cache);
3114 /*
3115 * Our block group might still be attached to the list
3116 * of new block groups in the transaction handle of some
3117 * other task (struct btrfs_trans_handle->new_bgs). This
3118 * means its block group item isn't yet in the extent
3119 * tree. If this happens ignore the error, as we will
3120 * try again later in the critical section of the
3121 * transaction commit.
3122 */
3123 if (ret == -ENOENT) {
3124 ret = 0;
3125 spin_lock(&cur_trans->dirty_bgs_lock);
3126 if (list_empty(&cache->dirty_list)) {
3127 list_add_tail(&cache->dirty_list,
3128 &cur_trans->dirty_bgs);
3129 btrfs_get_block_group(cache);
3130 drop_reserve = false;
3131 }
3132 spin_unlock(&cur_trans->dirty_bgs_lock);
3133 } else if (ret) {
3134 btrfs_abort_transaction(trans, ret);
3135 }
3136 }
3137
3138 /* If it's not on the io list, we need to put the block group */
3139 if (should_put)
3140 btrfs_put_block_group(cache);
3141 if (drop_reserve)
3142 btrfs_delayed_refs_rsv_release(fs_info, 1);
3143 /*
3144 * Avoid blocking other tasks for too long. It might even save
3145 * us from writing caches for block groups that are going to be
3146 * removed.
3147 */
3148 mutex_unlock(&trans->transaction->cache_write_mutex);
3149 if (ret)
3150 goto out;
3151 mutex_lock(&trans->transaction->cache_write_mutex);
3152 }
3153 mutex_unlock(&trans->transaction->cache_write_mutex);
3154
3155 /*
3156 * Go through delayed refs for all the stuff we've just kicked off
3157 * and then loop back (just once)
3158 */
3159 if (!ret)
3160 ret = btrfs_run_delayed_refs(trans, 0);
3161 if (!ret && loops == 0) {
3162 loops++;
3163 spin_lock(&cur_trans->dirty_bgs_lock);
3164 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3165 /*
3166 * dirty_bgs_lock protects us from concurrent block group
3167 * deletes too (not just cache_write_mutex).
3168 */
3169 if (!list_empty(&dirty)) {
3170 spin_unlock(&cur_trans->dirty_bgs_lock);
3171 goto again;
3172 }
3173 spin_unlock(&cur_trans->dirty_bgs_lock);
3174 }
3175out:
3176 if (ret < 0) {
3177 spin_lock(&cur_trans->dirty_bgs_lock);
3178 list_splice_init(&dirty, &cur_trans->dirty_bgs);
3179 spin_unlock(&cur_trans->dirty_bgs_lock);
3180 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3181 }
3182
3183 btrfs_free_path(path);
3184 return ret;
3185}
3186
3187int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3188{
3189 struct btrfs_fs_info *fs_info = trans->fs_info;
3190 struct btrfs_block_group *cache;
3191 struct btrfs_transaction *cur_trans = trans->transaction;
3192 int ret = 0;
3193 int should_put;
3194 struct btrfs_path *path;
3195 struct list_head *io = &cur_trans->io_bgs;
3196
3197 path = btrfs_alloc_path();
3198 if (!path)
3199 return -ENOMEM;
3200
3201 /*
3202 * Even though we are in the critical section of the transaction commit,
3203 * we can still have concurrent tasks adding elements to this
3204 * transaction's list of dirty block groups. These tasks correspond to
3205 * endio free space workers started when writeback finishes for a
3206 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3207 * allocate new block groups as a result of COWing nodes of the root
3208 * tree when updating the free space inode. The writeback for the space
3209 * caches is triggered by an earlier call to
3210 * btrfs_start_dirty_block_groups() and iterations of the following
3211 * loop.
3212 * Also we want to do the cache_save_setup first and then run the
3213 * delayed refs to make sure we have the best chance at doing this all
3214 * in one shot.
3215 */
3216 spin_lock(&cur_trans->dirty_bgs_lock);
3217 while (!list_empty(&cur_trans->dirty_bgs)) {
3218 cache = list_first_entry(&cur_trans->dirty_bgs,
3219 struct btrfs_block_group,
3220 dirty_list);
3221
3222 /*
3223 * This can happen if cache_save_setup re-dirties a block group
3224 * that is already under IO. Just wait for it to finish and
3225 * then do it all again
3226 */
3227 if (!list_empty(&cache->io_list)) {
3228 spin_unlock(&cur_trans->dirty_bgs_lock);
3229 list_del_init(&cache->io_list);
3230 btrfs_wait_cache_io(trans, cache, path);
3231 btrfs_put_block_group(cache);
3232 spin_lock(&cur_trans->dirty_bgs_lock);
3233 }
3234
3235 /*
3236 * Don't remove from the dirty list until after we've waited on
3237 * any pending IO
3238 */
3239 list_del_init(&cache->dirty_list);
3240 spin_unlock(&cur_trans->dirty_bgs_lock);
3241 should_put = 1;
3242
3243 cache_save_setup(cache, trans, path);
3244
3245 if (!ret)
3246 ret = btrfs_run_delayed_refs(trans,
3247 (unsigned long) -1);
3248
3249 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3250 cache->io_ctl.inode = NULL;
3251 ret = btrfs_write_out_cache(trans, cache, path);
3252 if (ret == 0 && cache->io_ctl.inode) {
3253 should_put = 0;
3254 list_add_tail(&cache->io_list, io);
3255 } else {
3256 /*
3257 * If we failed to write the cache, the
3258 * generation will be bad and life goes on
3259 */
3260 ret = 0;
3261 }
3262 }
3263 if (!ret) {
3264 ret = update_block_group_item(trans, path, cache);
3265 /*
3266 * One of the free space endio workers might have
3267 * created a new block group while updating a free space
3268 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3269 * and hasn't released its transaction handle yet, in
3270 * which case the new block group is still attached to
3271 * its transaction handle and its creation has not
3272 * finished yet (no block group item in the extent tree
3273 * yet, etc). If this is the case, wait for all free
3274 * space endio workers to finish and retry. This is a
3275 * very rare case so no need for a more efficient and
3276 * complex approach.
3277 */
3278 if (ret == -ENOENT) {
3279 wait_event(cur_trans->writer_wait,
3280 atomic_read(&cur_trans->num_writers) == 1);
3281 ret = update_block_group_item(trans, path, cache);
3282 }
3283 if (ret)
3284 btrfs_abort_transaction(trans, ret);
3285 }
3286
3287 /* If its not on the io list, we need to put the block group */
3288 if (should_put)
3289 btrfs_put_block_group(cache);
3290 btrfs_delayed_refs_rsv_release(fs_info, 1);
3291 spin_lock(&cur_trans->dirty_bgs_lock);
3292 }
3293 spin_unlock(&cur_trans->dirty_bgs_lock);
3294
3295 /*
3296 * Refer to the definition of io_bgs member for details why it's safe
3297 * to use it without any locking
3298 */
3299 while (!list_empty(io)) {
3300 cache = list_first_entry(io, struct btrfs_block_group,
3301 io_list);
3302 list_del_init(&cache->io_list);
3303 btrfs_wait_cache_io(trans, cache, path);
3304 btrfs_put_block_group(cache);
3305 }
3306
3307 btrfs_free_path(path);
3308 return ret;
3309}
3310
3311int btrfs_update_block_group(struct btrfs_trans_handle *trans,
3312 u64 bytenr, u64 num_bytes, bool alloc)
3313{
3314 struct btrfs_fs_info *info = trans->fs_info;
3315 struct btrfs_block_group *cache = NULL;
3316 u64 total = num_bytes;
3317 u64 old_val;
3318 u64 byte_in_group;
3319 int factor;
3320 int ret = 0;
3321
3322 /* Block accounting for super block */
3323 spin_lock(&info->delalloc_root_lock);
3324 old_val = btrfs_super_bytes_used(info->super_copy);
3325 if (alloc)
3326 old_val += num_bytes;
3327 else
3328 old_val -= num_bytes;
3329 btrfs_set_super_bytes_used(info->super_copy, old_val);
3330 spin_unlock(&info->delalloc_root_lock);
3331
3332 while (total) {
3333 bool reclaim;
3334
3335 cache = btrfs_lookup_block_group(info, bytenr);
3336 if (!cache) {
3337 ret = -ENOENT;
3338 break;
3339 }
3340 factor = btrfs_bg_type_to_factor(cache->flags);
3341
3342 /*
3343 * If this block group has free space cache written out, we
3344 * need to make sure to load it if we are removing space. This
3345 * is because we need the unpinning stage to actually add the
3346 * space back to the block group, otherwise we will leak space.
3347 */
3348 if (!alloc && !btrfs_block_group_done(cache))
3349 btrfs_cache_block_group(cache, true);
3350
3351 byte_in_group = bytenr - cache->start;
3352 WARN_ON(byte_in_group > cache->length);
3353
3354 spin_lock(&cache->space_info->lock);
3355 spin_lock(&cache->lock);
3356
3357 if (btrfs_test_opt(info, SPACE_CACHE) &&
3358 cache->disk_cache_state < BTRFS_DC_CLEAR)
3359 cache->disk_cache_state = BTRFS_DC_CLEAR;
3360
3361 old_val = cache->used;
3362 num_bytes = min(total, cache->length - byte_in_group);
3363 if (alloc) {
3364 old_val += num_bytes;
3365 cache->used = old_val;
3366 cache->reserved -= num_bytes;
3367 cache->space_info->bytes_reserved -= num_bytes;
3368 cache->space_info->bytes_used += num_bytes;
3369 cache->space_info->disk_used += num_bytes * factor;
3370 spin_unlock(&cache->lock);
3371 spin_unlock(&cache->space_info->lock);
3372 } else {
3373 old_val -= num_bytes;
3374 cache->used = old_val;
3375 cache->pinned += num_bytes;
3376 btrfs_space_info_update_bytes_pinned(info,
3377 cache->space_info, num_bytes);
3378 cache->space_info->bytes_used -= num_bytes;
3379 cache->space_info->disk_used -= num_bytes * factor;
3380
3381 reclaim = should_reclaim_block_group(cache, num_bytes);
3382 spin_unlock(&cache->lock);
3383 spin_unlock(&cache->space_info->lock);
3384
3385 set_extent_dirty(&trans->transaction->pinned_extents,
3386 bytenr, bytenr + num_bytes - 1,
3387 GFP_NOFS | __GFP_NOFAIL);
3388 }
3389
3390 spin_lock(&trans->transaction->dirty_bgs_lock);
3391 if (list_empty(&cache->dirty_list)) {
3392 list_add_tail(&cache->dirty_list,
3393 &trans->transaction->dirty_bgs);
3394 trans->delayed_ref_updates++;
3395 btrfs_get_block_group(cache);
3396 }
3397 spin_unlock(&trans->transaction->dirty_bgs_lock);
3398
3399 /*
3400 * No longer have used bytes in this block group, queue it for
3401 * deletion. We do this after adding the block group to the
3402 * dirty list to avoid races between cleaner kthread and space
3403 * cache writeout.
3404 */
3405 if (!alloc && old_val == 0) {
3406 if (!btrfs_test_opt(info, DISCARD_ASYNC))
3407 btrfs_mark_bg_unused(cache);
3408 } else if (!alloc && reclaim) {
3409 btrfs_mark_bg_to_reclaim(cache);
3410 }
3411
3412 btrfs_put_block_group(cache);
3413 total -= num_bytes;
3414 bytenr += num_bytes;
3415 }
3416
3417 /* Modified block groups are accounted for in the delayed_refs_rsv. */
3418 btrfs_update_delayed_refs_rsv(trans);
3419 return ret;
3420}
3421
3422/*
3423 * Update the block_group and space info counters.
3424 *
3425 * @cache: The cache we are manipulating
3426 * @ram_bytes: The number of bytes of file content, and will be same to
3427 * @num_bytes except for the compress path.
3428 * @num_bytes: The number of bytes in question
3429 * @delalloc: The blocks are allocated for the delalloc write
3430 *
3431 * This is called by the allocator when it reserves space. If this is a
3432 * reservation and the block group has become read only we cannot make the
3433 * reservation and return -EAGAIN, otherwise this function always succeeds.
3434 */
3435int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3436 u64 ram_bytes, u64 num_bytes, int delalloc)
3437{
3438 struct btrfs_space_info *space_info = cache->space_info;
3439 int ret = 0;
3440
3441 spin_lock(&space_info->lock);
3442 spin_lock(&cache->lock);
3443 if (cache->ro) {
3444 ret = -EAGAIN;
3445 } else {
3446 cache->reserved += num_bytes;
3447 space_info->bytes_reserved += num_bytes;
3448 trace_btrfs_space_reservation(cache->fs_info, "space_info",
3449 space_info->flags, num_bytes, 1);
3450 btrfs_space_info_update_bytes_may_use(cache->fs_info,
3451 space_info, -ram_bytes);
3452 if (delalloc)
3453 cache->delalloc_bytes += num_bytes;
3454
3455 /*
3456 * Compression can use less space than we reserved, so wake
3457 * tickets if that happens
3458 */
3459 if (num_bytes < ram_bytes)
3460 btrfs_try_granting_tickets(cache->fs_info, space_info);
3461 }
3462 spin_unlock(&cache->lock);
3463 spin_unlock(&space_info->lock);
3464 return ret;
3465}
3466
3467/*
3468 * Update the block_group and space info counters.
3469 *
3470 * @cache: The cache we are manipulating
3471 * @num_bytes: The number of bytes in question
3472 * @delalloc: The blocks are allocated for the delalloc write
3473 *
3474 * This is called by somebody who is freeing space that was never actually used
3475 * on disk. For example if you reserve some space for a new leaf in transaction
3476 * A and before transaction A commits you free that leaf, you call this with
3477 * reserve set to 0 in order to clear the reservation.
3478 */
3479void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
3480 u64 num_bytes, int delalloc)
3481{
3482 struct btrfs_space_info *space_info = cache->space_info;
3483
3484 spin_lock(&space_info->lock);
3485 spin_lock(&cache->lock);
3486 if (cache->ro)
3487 space_info->bytes_readonly += num_bytes;
3488 cache->reserved -= num_bytes;
3489 space_info->bytes_reserved -= num_bytes;
3490 space_info->max_extent_size = 0;
3491
3492 if (delalloc)
3493 cache->delalloc_bytes -= num_bytes;
3494 spin_unlock(&cache->lock);
3495
3496 btrfs_try_granting_tickets(cache->fs_info, space_info);
3497 spin_unlock(&space_info->lock);
3498}
3499
3500static void force_metadata_allocation(struct btrfs_fs_info *info)
3501{
3502 struct list_head *head = &info->space_info;
3503 struct btrfs_space_info *found;
3504
3505 list_for_each_entry(found, head, list) {
3506 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3507 found->force_alloc = CHUNK_ALLOC_FORCE;
3508 }
3509}
3510
3511static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3512 struct btrfs_space_info *sinfo, int force)
3513{
3514 u64 bytes_used = btrfs_space_info_used(sinfo, false);
3515 u64 thresh;
3516
3517 if (force == CHUNK_ALLOC_FORCE)
3518 return 1;
3519
3520 /*
3521 * in limited mode, we want to have some free space up to
3522 * about 1% of the FS size.
3523 */
3524 if (force == CHUNK_ALLOC_LIMITED) {
3525 thresh = btrfs_super_total_bytes(fs_info->super_copy);
3526 thresh = max_t(u64, SZ_64M, mult_perc(thresh, 1));
3527
3528 if (sinfo->total_bytes - bytes_used < thresh)
3529 return 1;
3530 }
3531
3532 if (bytes_used + SZ_2M < mult_perc(sinfo->total_bytes, 80))
3533 return 0;
3534 return 1;
3535}
3536
3537int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3538{
3539 u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3540
3541 return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3542}
3543
3544static struct btrfs_block_group *do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags)
3545{
3546 struct btrfs_block_group *bg;
3547 int ret;
3548
3549 /*
3550 * Check if we have enough space in the system space info because we
3551 * will need to update device items in the chunk btree and insert a new
3552 * chunk item in the chunk btree as well. This will allocate a new
3553 * system block group if needed.
3554 */
3555 check_system_chunk(trans, flags);
3556
3557 bg = btrfs_create_chunk(trans, flags);
3558 if (IS_ERR(bg)) {
3559 ret = PTR_ERR(bg);
3560 goto out;
3561 }
3562
3563 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3564 /*
3565 * Normally we are not expected to fail with -ENOSPC here, since we have
3566 * previously reserved space in the system space_info and allocated one
3567 * new system chunk if necessary. However there are three exceptions:
3568 *
3569 * 1) We may have enough free space in the system space_info but all the
3570 * existing system block groups have a profile which can not be used
3571 * for extent allocation.
3572 *
3573 * This happens when mounting in degraded mode. For example we have a
3574 * RAID1 filesystem with 2 devices, lose one device and mount the fs
3575 * using the other device in degraded mode. If we then allocate a chunk,
3576 * we may have enough free space in the existing system space_info, but
3577 * none of the block groups can be used for extent allocation since they
3578 * have a RAID1 profile, and because we are in degraded mode with a
3579 * single device, we are forced to allocate a new system chunk with a
3580 * SINGLE profile. Making check_system_chunk() iterate over all system
3581 * block groups and check if they have a usable profile and enough space
3582 * can be slow on very large filesystems, so we tolerate the -ENOSPC and
3583 * try again after forcing allocation of a new system chunk. Like this
3584 * we avoid paying the cost of that search in normal circumstances, when
3585 * we were not mounted in degraded mode;
3586 *
3587 * 2) We had enough free space info the system space_info, and one suitable
3588 * block group to allocate from when we called check_system_chunk()
3589 * above. However right after we called it, the only system block group
3590 * with enough free space got turned into RO mode by a running scrub,
3591 * and in this case we have to allocate a new one and retry. We only
3592 * need do this allocate and retry once, since we have a transaction
3593 * handle and scrub uses the commit root to search for block groups;
3594 *
3595 * 3) We had one system block group with enough free space when we called
3596 * check_system_chunk(), but after that, right before we tried to
3597 * allocate the last extent buffer we needed, a discard operation came
3598 * in and it temporarily removed the last free space entry from the
3599 * block group (discard removes a free space entry, discards it, and
3600 * then adds back the entry to the block group cache).
3601 */
3602 if (ret == -ENOSPC) {
3603 const u64 sys_flags = btrfs_system_alloc_profile(trans->fs_info);
3604 struct btrfs_block_group *sys_bg;
3605
3606 sys_bg = btrfs_create_chunk(trans, sys_flags);
3607 if (IS_ERR(sys_bg)) {
3608 ret = PTR_ERR(sys_bg);
3609 btrfs_abort_transaction(trans, ret);
3610 goto out;
3611 }
3612
3613 ret = btrfs_chunk_alloc_add_chunk_item(trans, sys_bg);
3614 if (ret) {
3615 btrfs_abort_transaction(trans, ret);
3616 goto out;
3617 }
3618
3619 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3620 if (ret) {
3621 btrfs_abort_transaction(trans, ret);
3622 goto out;
3623 }
3624 } else if (ret) {
3625 btrfs_abort_transaction(trans, ret);
3626 goto out;
3627 }
3628out:
3629 btrfs_trans_release_chunk_metadata(trans);
3630
3631 if (ret)
3632 return ERR_PTR(ret);
3633
3634 btrfs_get_block_group(bg);
3635 return bg;
3636}
3637
3638/*
3639 * Chunk allocation is done in 2 phases:
3640 *
3641 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
3642 * the chunk, the chunk mapping, create its block group and add the items
3643 * that belong in the chunk btree to it - more specifically, we need to
3644 * update device items in the chunk btree and add a new chunk item to it.
3645 *
3646 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
3647 * group item to the extent btree and the device extent items to the devices
3648 * btree.
3649 *
3650 * This is done to prevent deadlocks. For example when COWing a node from the
3651 * extent btree we are holding a write lock on the node's parent and if we
3652 * trigger chunk allocation and attempted to insert the new block group item
3653 * in the extent btree right way, we could deadlock because the path for the
3654 * insertion can include that parent node. At first glance it seems impossible
3655 * to trigger chunk allocation after starting a transaction since tasks should
3656 * reserve enough transaction units (metadata space), however while that is true
3657 * most of the time, chunk allocation may still be triggered for several reasons:
3658 *
3659 * 1) When reserving metadata, we check if there is enough free space in the
3660 * metadata space_info and therefore don't trigger allocation of a new chunk.
3661 * However later when the task actually tries to COW an extent buffer from
3662 * the extent btree or from the device btree for example, it is forced to
3663 * allocate a new block group (chunk) because the only one that had enough
3664 * free space was just turned to RO mode by a running scrub for example (or
3665 * device replace, block group reclaim thread, etc), so we can not use it
3666 * for allocating an extent and end up being forced to allocate a new one;
3667 *
3668 * 2) Because we only check that the metadata space_info has enough free bytes,
3669 * we end up not allocating a new metadata chunk in that case. However if
3670 * the filesystem was mounted in degraded mode, none of the existing block
3671 * groups might be suitable for extent allocation due to their incompatible
3672 * profile (for e.g. mounting a 2 devices filesystem, where all block groups
3673 * use a RAID1 profile, in degraded mode using a single device). In this case
3674 * when the task attempts to COW some extent buffer of the extent btree for
3675 * example, it will trigger allocation of a new metadata block group with a
3676 * suitable profile (SINGLE profile in the example of the degraded mount of
3677 * the RAID1 filesystem);
3678 *
3679 * 3) The task has reserved enough transaction units / metadata space, but when
3680 * it attempts to COW an extent buffer from the extent or device btree for
3681 * example, it does not find any free extent in any metadata block group,
3682 * therefore forced to try to allocate a new metadata block group.
3683 * This is because some other task allocated all available extents in the
3684 * meanwhile - this typically happens with tasks that don't reserve space
3685 * properly, either intentionally or as a bug. One example where this is
3686 * done intentionally is fsync, as it does not reserve any transaction units
3687 * and ends up allocating a variable number of metadata extents for log
3688 * tree extent buffers;
3689 *
3690 * 4) The task has reserved enough transaction units / metadata space, but right
3691 * before it tries to allocate the last extent buffer it needs, a discard
3692 * operation comes in and, temporarily, removes the last free space entry from
3693 * the only metadata block group that had free space (discard starts by
3694 * removing a free space entry from a block group, then does the discard
3695 * operation and, once it's done, it adds back the free space entry to the
3696 * block group).
3697 *
3698 * We also need this 2 phases setup when adding a device to a filesystem with
3699 * a seed device - we must create new metadata and system chunks without adding
3700 * any of the block group items to the chunk, extent and device btrees. If we
3701 * did not do it this way, we would get ENOSPC when attempting to update those
3702 * btrees, since all the chunks from the seed device are read-only.
3703 *
3704 * Phase 1 does the updates and insertions to the chunk btree because if we had
3705 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
3706 * parallel, we risk having too many system chunks allocated by many tasks if
3707 * many tasks reach phase 1 without the previous ones completing phase 2. In the
3708 * extreme case this leads to exhaustion of the system chunk array in the
3709 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
3710 * and with RAID filesystems (so we have more device items in the chunk btree).
3711 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
3712 * the system chunk array due to concurrent allocations") provides more details.
3713 *
3714 * Allocation of system chunks does not happen through this function. A task that
3715 * needs to update the chunk btree (the only btree that uses system chunks), must
3716 * preallocate chunk space by calling either check_system_chunk() or
3717 * btrfs_reserve_chunk_metadata() - the former is used when allocating a data or
3718 * metadata chunk or when removing a chunk, while the later is used before doing
3719 * a modification to the chunk btree - use cases for the later are adding,
3720 * removing and resizing a device as well as relocation of a system chunk.
3721 * See the comment below for more details.
3722 *
3723 * The reservation of system space, done through check_system_chunk(), as well
3724 * as all the updates and insertions into the chunk btree must be done while
3725 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
3726 * an extent buffer from the chunks btree we never trigger allocation of a new
3727 * system chunk, which would result in a deadlock (trying to lock twice an
3728 * extent buffer of the chunk btree, first time before triggering the chunk
3729 * allocation and the second time during chunk allocation while attempting to
3730 * update the chunks btree). The system chunk array is also updated while holding
3731 * that mutex. The same logic applies to removing chunks - we must reserve system
3732 * space, update the chunk btree and the system chunk array in the superblock
3733 * while holding fs_info->chunk_mutex.
3734 *
3735 * This function, btrfs_chunk_alloc(), belongs to phase 1.
3736 *
3737 * If @force is CHUNK_ALLOC_FORCE:
3738 * - return 1 if it successfully allocates a chunk,
3739 * - return errors including -ENOSPC otherwise.
3740 * If @force is NOT CHUNK_ALLOC_FORCE:
3741 * - return 0 if it doesn't need to allocate a new chunk,
3742 * - return 1 if it successfully allocates a chunk,
3743 * - return errors including -ENOSPC otherwise.
3744 */
3745int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
3746 enum btrfs_chunk_alloc_enum force)
3747{
3748 struct btrfs_fs_info *fs_info = trans->fs_info;
3749 struct btrfs_space_info *space_info;
3750 struct btrfs_block_group *ret_bg;
3751 bool wait_for_alloc = false;
3752 bool should_alloc = false;
3753 bool from_extent_allocation = false;
3754 int ret = 0;
3755
3756 if (force == CHUNK_ALLOC_FORCE_FOR_EXTENT) {
3757 from_extent_allocation = true;
3758 force = CHUNK_ALLOC_FORCE;
3759 }
3760
3761 /* Don't re-enter if we're already allocating a chunk */
3762 if (trans->allocating_chunk)
3763 return -ENOSPC;
3764 /*
3765 * Allocation of system chunks can not happen through this path, as we
3766 * could end up in a deadlock if we are allocating a data or metadata
3767 * chunk and there is another task modifying the chunk btree.
3768 *
3769 * This is because while we are holding the chunk mutex, we will attempt
3770 * to add the new chunk item to the chunk btree or update an existing
3771 * device item in the chunk btree, while the other task that is modifying
3772 * the chunk btree is attempting to COW an extent buffer while holding a
3773 * lock on it and on its parent - if the COW operation triggers a system
3774 * chunk allocation, then we can deadlock because we are holding the
3775 * chunk mutex and we may need to access that extent buffer or its parent
3776 * in order to add the chunk item or update a device item.
3777 *
3778 * Tasks that want to modify the chunk tree should reserve system space
3779 * before updating the chunk btree, by calling either
3780 * btrfs_reserve_chunk_metadata() or check_system_chunk().
3781 * It's possible that after a task reserves the space, it still ends up
3782 * here - this happens in the cases described above at do_chunk_alloc().
3783 * The task will have to either retry or fail.
3784 */
3785 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3786 return -ENOSPC;
3787
3788 space_info = btrfs_find_space_info(fs_info, flags);
3789 ASSERT(space_info);
3790
3791 do {
3792 spin_lock(&space_info->lock);
3793 if (force < space_info->force_alloc)
3794 force = space_info->force_alloc;
3795 should_alloc = should_alloc_chunk(fs_info, space_info, force);
3796 if (space_info->full) {
3797 /* No more free physical space */
3798 if (should_alloc)
3799 ret = -ENOSPC;
3800 else
3801 ret = 0;
3802 spin_unlock(&space_info->lock);
3803 return ret;
3804 } else if (!should_alloc) {
3805 spin_unlock(&space_info->lock);
3806 return 0;
3807 } else if (space_info->chunk_alloc) {
3808 /*
3809 * Someone is already allocating, so we need to block
3810 * until this someone is finished and then loop to
3811 * recheck if we should continue with our allocation
3812 * attempt.
3813 */
3814 wait_for_alloc = true;
3815 force = CHUNK_ALLOC_NO_FORCE;
3816 spin_unlock(&space_info->lock);
3817 mutex_lock(&fs_info->chunk_mutex);
3818 mutex_unlock(&fs_info->chunk_mutex);
3819 } else {
3820 /* Proceed with allocation */
3821 space_info->chunk_alloc = 1;
3822 wait_for_alloc = false;
3823 spin_unlock(&space_info->lock);
3824 }
3825
3826 cond_resched();
3827 } while (wait_for_alloc);
3828
3829 mutex_lock(&fs_info->chunk_mutex);
3830 trans->allocating_chunk = true;
3831
3832 /*
3833 * If we have mixed data/metadata chunks we want to make sure we keep
3834 * allocating mixed chunks instead of individual chunks.
3835 */
3836 if (btrfs_mixed_space_info(space_info))
3837 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3838
3839 /*
3840 * if we're doing a data chunk, go ahead and make sure that
3841 * we keep a reasonable number of metadata chunks allocated in the
3842 * FS as well.
3843 */
3844 if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3845 fs_info->data_chunk_allocations++;
3846 if (!(fs_info->data_chunk_allocations %
3847 fs_info->metadata_ratio))
3848 force_metadata_allocation(fs_info);
3849 }
3850
3851 ret_bg = do_chunk_alloc(trans, flags);
3852 trans->allocating_chunk = false;
3853
3854 if (IS_ERR(ret_bg)) {
3855 ret = PTR_ERR(ret_bg);
3856 } else if (from_extent_allocation) {
3857 /*
3858 * New block group is likely to be used soon. Try to activate
3859 * it now. Failure is OK for now.
3860 */
3861 btrfs_zone_activate(ret_bg);
3862 }
3863
3864 if (!ret)
3865 btrfs_put_block_group(ret_bg);
3866
3867 spin_lock(&space_info->lock);
3868 if (ret < 0) {
3869 if (ret == -ENOSPC)
3870 space_info->full = 1;
3871 else
3872 goto out;
3873 } else {
3874 ret = 1;
3875 space_info->max_extent_size = 0;
3876 }
3877
3878 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3879out:
3880 space_info->chunk_alloc = 0;
3881 spin_unlock(&space_info->lock);
3882 mutex_unlock(&fs_info->chunk_mutex);
3883
3884 return ret;
3885}
3886
3887static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
3888{
3889 u64 num_dev;
3890
3891 num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
3892 if (!num_dev)
3893 num_dev = fs_info->fs_devices->rw_devices;
3894
3895 return num_dev;
3896}
3897
3898static void reserve_chunk_space(struct btrfs_trans_handle *trans,
3899 u64 bytes,
3900 u64 type)
3901{
3902 struct btrfs_fs_info *fs_info = trans->fs_info;
3903 struct btrfs_space_info *info;
3904 u64 left;
3905 int ret = 0;
3906
3907 /*
3908 * Needed because we can end up allocating a system chunk and for an
3909 * atomic and race free space reservation in the chunk block reserve.
3910 */
3911 lockdep_assert_held(&fs_info->chunk_mutex);
3912
3913 info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3914 spin_lock(&info->lock);
3915 left = info->total_bytes - btrfs_space_info_used(info, true);
3916 spin_unlock(&info->lock);
3917
3918 if (left < bytes && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
3919 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
3920 left, bytes, type);
3921 btrfs_dump_space_info(fs_info, info, 0, 0);
3922 }
3923
3924 if (left < bytes) {
3925 u64 flags = btrfs_system_alloc_profile(fs_info);
3926 struct btrfs_block_group *bg;
3927
3928 /*
3929 * Ignore failure to create system chunk. We might end up not
3930 * needing it, as we might not need to COW all nodes/leafs from
3931 * the paths we visit in the chunk tree (they were already COWed
3932 * or created in the current transaction for example).
3933 */
3934 bg = btrfs_create_chunk(trans, flags);
3935 if (IS_ERR(bg)) {
3936 ret = PTR_ERR(bg);
3937 } else {
3938 /*
3939 * We have a new chunk. We also need to activate it for
3940 * zoned filesystem.
3941 */
3942 ret = btrfs_zoned_activate_one_bg(fs_info, info, true);
3943 if (ret < 0)
3944 return;
3945
3946 /*
3947 * If we fail to add the chunk item here, we end up
3948 * trying again at phase 2 of chunk allocation, at
3949 * btrfs_create_pending_block_groups(). So ignore
3950 * any error here. An ENOSPC here could happen, due to
3951 * the cases described at do_chunk_alloc() - the system
3952 * block group we just created was just turned into RO
3953 * mode by a scrub for example, or a running discard
3954 * temporarily removed its free space entries, etc.
3955 */
3956 btrfs_chunk_alloc_add_chunk_item(trans, bg);
3957 }
3958 }
3959
3960 if (!ret) {
3961 ret = btrfs_block_rsv_add(fs_info,
3962 &fs_info->chunk_block_rsv,
3963 bytes, BTRFS_RESERVE_NO_FLUSH);
3964 if (!ret)
3965 trans->chunk_bytes_reserved += bytes;
3966 }
3967}
3968
3969/*
3970 * Reserve space in the system space for allocating or removing a chunk.
3971 * The caller must be holding fs_info->chunk_mutex.
3972 */
3973void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
3974{
3975 struct btrfs_fs_info *fs_info = trans->fs_info;
3976 const u64 num_devs = get_profile_num_devs(fs_info, type);
3977 u64 bytes;
3978
3979 /* num_devs device items to update and 1 chunk item to add or remove. */
3980 bytes = btrfs_calc_metadata_size(fs_info, num_devs) +
3981 btrfs_calc_insert_metadata_size(fs_info, 1);
3982
3983 reserve_chunk_space(trans, bytes, type);
3984}
3985
3986/*
3987 * Reserve space in the system space, if needed, for doing a modification to the
3988 * chunk btree.
3989 *
3990 * @trans: A transaction handle.
3991 * @is_item_insertion: Indicate if the modification is for inserting a new item
3992 * in the chunk btree or if it's for the deletion or update
3993 * of an existing item.
3994 *
3995 * This is used in a context where we need to update the chunk btree outside
3996 * block group allocation and removal, to avoid a deadlock with a concurrent
3997 * task that is allocating a metadata or data block group and therefore needs to
3998 * update the chunk btree while holding the chunk mutex. After the update to the
3999 * chunk btree is done, btrfs_trans_release_chunk_metadata() should be called.
4000 *
4001 */
4002void btrfs_reserve_chunk_metadata(struct btrfs_trans_handle *trans,
4003 bool is_item_insertion)
4004{
4005 struct btrfs_fs_info *fs_info = trans->fs_info;
4006 u64 bytes;
4007
4008 if (is_item_insertion)
4009 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
4010 else
4011 bytes = btrfs_calc_metadata_size(fs_info, 1);
4012
4013 mutex_lock(&fs_info->chunk_mutex);
4014 reserve_chunk_space(trans, bytes, BTRFS_BLOCK_GROUP_SYSTEM);
4015 mutex_unlock(&fs_info->chunk_mutex);
4016}
4017
4018void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
4019{
4020 struct btrfs_block_group *block_group;
4021
4022 block_group = btrfs_lookup_first_block_group(info, 0);
4023 while (block_group) {
4024 btrfs_wait_block_group_cache_done(block_group);
4025 spin_lock(&block_group->lock);
4026 if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF,
4027 &block_group->runtime_flags)) {
4028 struct inode *inode = block_group->inode;
4029
4030 block_group->inode = NULL;
4031 spin_unlock(&block_group->lock);
4032
4033 ASSERT(block_group->io_ctl.inode == NULL);
4034 iput(inode);
4035 } else {
4036 spin_unlock(&block_group->lock);
4037 }
4038 block_group = btrfs_next_block_group(block_group);
4039 }
4040}
4041
4042/*
4043 * Must be called only after stopping all workers, since we could have block
4044 * group caching kthreads running, and therefore they could race with us if we
4045 * freed the block groups before stopping them.
4046 */
4047int btrfs_free_block_groups(struct btrfs_fs_info *info)
4048{
4049 struct btrfs_block_group *block_group;
4050 struct btrfs_space_info *space_info;
4051 struct btrfs_caching_control *caching_ctl;
4052 struct rb_node *n;
4053
4054 write_lock(&info->block_group_cache_lock);
4055 while (!list_empty(&info->caching_block_groups)) {
4056 caching_ctl = list_entry(info->caching_block_groups.next,
4057 struct btrfs_caching_control, list);
4058 list_del(&caching_ctl->list);
4059 btrfs_put_caching_control(caching_ctl);
4060 }
4061 write_unlock(&info->block_group_cache_lock);
4062
4063 spin_lock(&info->unused_bgs_lock);
4064 while (!list_empty(&info->unused_bgs)) {
4065 block_group = list_first_entry(&info->unused_bgs,
4066 struct btrfs_block_group,
4067 bg_list);
4068 list_del_init(&block_group->bg_list);
4069 btrfs_put_block_group(block_group);
4070 }
4071
4072 while (!list_empty(&info->reclaim_bgs)) {
4073 block_group = list_first_entry(&info->reclaim_bgs,
4074 struct btrfs_block_group,
4075 bg_list);
4076 list_del_init(&block_group->bg_list);
4077 btrfs_put_block_group(block_group);
4078 }
4079 spin_unlock(&info->unused_bgs_lock);
4080
4081 spin_lock(&info->zone_active_bgs_lock);
4082 while (!list_empty(&info->zone_active_bgs)) {
4083 block_group = list_first_entry(&info->zone_active_bgs,
4084 struct btrfs_block_group,
4085 active_bg_list);
4086 list_del_init(&block_group->active_bg_list);
4087 btrfs_put_block_group(block_group);
4088 }
4089 spin_unlock(&info->zone_active_bgs_lock);
4090
4091 write_lock(&info->block_group_cache_lock);
4092 while ((n = rb_last(&info->block_group_cache_tree.rb_root)) != NULL) {
4093 block_group = rb_entry(n, struct btrfs_block_group,
4094 cache_node);
4095 rb_erase_cached(&block_group->cache_node,
4096 &info->block_group_cache_tree);
4097 RB_CLEAR_NODE(&block_group->cache_node);
4098 write_unlock(&info->block_group_cache_lock);
4099
4100 down_write(&block_group->space_info->groups_sem);
4101 list_del(&block_group->list);
4102 up_write(&block_group->space_info->groups_sem);
4103
4104 /*
4105 * We haven't cached this block group, which means we could
4106 * possibly have excluded extents on this block group.
4107 */
4108 if (block_group->cached == BTRFS_CACHE_NO ||
4109 block_group->cached == BTRFS_CACHE_ERROR)
4110 btrfs_free_excluded_extents(block_group);
4111
4112 btrfs_remove_free_space_cache(block_group);
4113 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
4114 ASSERT(list_empty(&block_group->dirty_list));
4115 ASSERT(list_empty(&block_group->io_list));
4116 ASSERT(list_empty(&block_group->bg_list));
4117 ASSERT(refcount_read(&block_group->refs) == 1);
4118 ASSERT(block_group->swap_extents == 0);
4119 btrfs_put_block_group(block_group);
4120
4121 write_lock(&info->block_group_cache_lock);
4122 }
4123 write_unlock(&info->block_group_cache_lock);
4124
4125 btrfs_release_global_block_rsv(info);
4126
4127 while (!list_empty(&info->space_info)) {
4128 space_info = list_entry(info->space_info.next,
4129 struct btrfs_space_info,
4130 list);
4131
4132 /*
4133 * Do not hide this behind enospc_debug, this is actually
4134 * important and indicates a real bug if this happens.
4135 */
4136 if (WARN_ON(space_info->bytes_pinned > 0 ||
4137 space_info->bytes_may_use > 0))
4138 btrfs_dump_space_info(info, space_info, 0, 0);
4139
4140 /*
4141 * If there was a failure to cleanup a log tree, very likely due
4142 * to an IO failure on a writeback attempt of one or more of its
4143 * extent buffers, we could not do proper (and cheap) unaccounting
4144 * of their reserved space, so don't warn on bytes_reserved > 0 in
4145 * that case.
4146 */
4147 if (!(space_info->flags & BTRFS_BLOCK_GROUP_METADATA) ||
4148 !BTRFS_FS_LOG_CLEANUP_ERROR(info)) {
4149 if (WARN_ON(space_info->bytes_reserved > 0))
4150 btrfs_dump_space_info(info, space_info, 0, 0);
4151 }
4152
4153 WARN_ON(space_info->reclaim_size > 0);
4154 list_del(&space_info->list);
4155 btrfs_sysfs_remove_space_info(space_info);
4156 }
4157 return 0;
4158}
4159
4160void btrfs_freeze_block_group(struct btrfs_block_group *cache)
4161{
4162 atomic_inc(&cache->frozen);
4163}
4164
4165void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
4166{
4167 struct btrfs_fs_info *fs_info = block_group->fs_info;
4168 struct extent_map_tree *em_tree;
4169 struct extent_map *em;
4170 bool cleanup;
4171
4172 spin_lock(&block_group->lock);
4173 cleanup = (atomic_dec_and_test(&block_group->frozen) &&
4174 test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags));
4175 spin_unlock(&block_group->lock);
4176
4177 if (cleanup) {
4178 em_tree = &fs_info->mapping_tree;
4179 write_lock(&em_tree->lock);
4180 em = lookup_extent_mapping(em_tree, block_group->start,
4181 1);
4182 BUG_ON(!em); /* logic error, can't happen */
4183 remove_extent_mapping(em_tree, em);
4184 write_unlock(&em_tree->lock);
4185
4186 /* once for us and once for the tree */
4187 free_extent_map(em);
4188 free_extent_map(em);
4189
4190 /*
4191 * We may have left one free space entry and other possible
4192 * tasks trimming this block group have left 1 entry each one.
4193 * Free them if any.
4194 */
4195 btrfs_remove_free_space_cache(block_group);
4196 }
4197}
4198
4199bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
4200{
4201 bool ret = true;
4202
4203 spin_lock(&bg->lock);
4204 if (bg->ro)
4205 ret = false;
4206 else
4207 bg->swap_extents++;
4208 spin_unlock(&bg->lock);
4209
4210 return ret;
4211}
4212
4213void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
4214{
4215 spin_lock(&bg->lock);
4216 ASSERT(!bg->ro);
4217 ASSERT(bg->swap_extents >= amount);
4218 bg->swap_extents -= amount;
4219 spin_unlock(&bg->lock);
4220}