Loading...
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}
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(const 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(const struct btrfs_fs_info *fs_info, u64 flags)
44{
45 const 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
421static void 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 struct btrfs_root *btrfs_block_group_root(struct btrfs_fs_info *fs_info)
1026{
1027 if (btrfs_fs_compat_ro(fs_info, BLOCK_GROUP_TREE))
1028 return fs_info->block_group_root;
1029 return btrfs_extent_root(fs_info, 0);
1030}
1031
1032static int remove_block_group_item(struct btrfs_trans_handle *trans,
1033 struct btrfs_path *path,
1034 struct btrfs_block_group *block_group)
1035{
1036 struct btrfs_fs_info *fs_info = trans->fs_info;
1037 struct btrfs_root *root;
1038 struct btrfs_key key;
1039 int ret;
1040
1041 root = btrfs_block_group_root(fs_info);
1042 key.objectid = block_group->start;
1043 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
1044 key.offset = block_group->length;
1045
1046 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1047 if (ret > 0)
1048 ret = -ENOENT;
1049 if (ret < 0)
1050 return ret;
1051
1052 ret = btrfs_del_item(trans, root, path);
1053 return ret;
1054}
1055
1056int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
1057 struct btrfs_chunk_map *map)
1058{
1059 struct btrfs_fs_info *fs_info = trans->fs_info;
1060 struct btrfs_path *path;
1061 struct btrfs_block_group *block_group;
1062 struct btrfs_free_cluster *cluster;
1063 struct inode *inode;
1064 struct kobject *kobj = NULL;
1065 int ret;
1066 int index;
1067 int factor;
1068 struct btrfs_caching_control *caching_ctl = NULL;
1069 bool remove_map;
1070 bool remove_rsv = false;
1071
1072 block_group = btrfs_lookup_block_group(fs_info, map->start);
1073 if (!block_group)
1074 return -ENOENT;
1075
1076 BUG_ON(!block_group->ro);
1077
1078 trace_btrfs_remove_block_group(block_group);
1079 /*
1080 * Free the reserved super bytes from this block group before
1081 * remove it.
1082 */
1083 btrfs_free_excluded_extents(block_group);
1084 btrfs_free_ref_tree_range(fs_info, block_group->start,
1085 block_group->length);
1086
1087 index = btrfs_bg_flags_to_raid_index(block_group->flags);
1088 factor = btrfs_bg_type_to_factor(block_group->flags);
1089
1090 /* make sure this block group isn't part of an allocation cluster */
1091 cluster = &fs_info->data_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 /*
1097 * make sure this block group isn't part of a metadata
1098 * allocation cluster
1099 */
1100 cluster = &fs_info->meta_alloc_cluster;
1101 spin_lock(&cluster->refill_lock);
1102 btrfs_return_cluster_to_free_space(block_group, cluster);
1103 spin_unlock(&cluster->refill_lock);
1104
1105 btrfs_clear_treelog_bg(block_group);
1106 btrfs_clear_data_reloc_bg(block_group);
1107
1108 path = btrfs_alloc_path();
1109 if (!path) {
1110 ret = -ENOMEM;
1111 goto out;
1112 }
1113
1114 /*
1115 * get the inode first so any iput calls done for the io_list
1116 * aren't the final iput (no unlinks allowed now)
1117 */
1118 inode = lookup_free_space_inode(block_group, path);
1119
1120 mutex_lock(&trans->transaction->cache_write_mutex);
1121 /*
1122 * Make sure our free space cache IO is done before removing the
1123 * free space inode
1124 */
1125 spin_lock(&trans->transaction->dirty_bgs_lock);
1126 if (!list_empty(&block_group->io_list)) {
1127 list_del_init(&block_group->io_list);
1128
1129 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
1130
1131 spin_unlock(&trans->transaction->dirty_bgs_lock);
1132 btrfs_wait_cache_io(trans, block_group, path);
1133 btrfs_put_block_group(block_group);
1134 spin_lock(&trans->transaction->dirty_bgs_lock);
1135 }
1136
1137 if (!list_empty(&block_group->dirty_list)) {
1138 list_del_init(&block_group->dirty_list);
1139 remove_rsv = true;
1140 btrfs_put_block_group(block_group);
1141 }
1142 spin_unlock(&trans->transaction->dirty_bgs_lock);
1143 mutex_unlock(&trans->transaction->cache_write_mutex);
1144
1145 ret = btrfs_remove_free_space_inode(trans, inode, block_group);
1146 if (ret)
1147 goto out;
1148
1149 write_lock(&fs_info->block_group_cache_lock);
1150 rb_erase_cached(&block_group->cache_node,
1151 &fs_info->block_group_cache_tree);
1152 RB_CLEAR_NODE(&block_group->cache_node);
1153
1154 /* Once for the block groups rbtree */
1155 btrfs_put_block_group(block_group);
1156
1157 write_unlock(&fs_info->block_group_cache_lock);
1158
1159 down_write(&block_group->space_info->groups_sem);
1160 /*
1161 * we must use list_del_init so people can check to see if they
1162 * are still on the list after taking the semaphore
1163 */
1164 list_del_init(&block_group->list);
1165 if (list_empty(&block_group->space_info->block_groups[index])) {
1166 kobj = block_group->space_info->block_group_kobjs[index];
1167 block_group->space_info->block_group_kobjs[index] = NULL;
1168 clear_avail_alloc_bits(fs_info, block_group->flags);
1169 }
1170 up_write(&block_group->space_info->groups_sem);
1171 clear_incompat_bg_bits(fs_info, block_group->flags);
1172 if (kobj) {
1173 kobject_del(kobj);
1174 kobject_put(kobj);
1175 }
1176
1177 if (block_group->cached == BTRFS_CACHE_STARTED)
1178 btrfs_wait_block_group_cache_done(block_group);
1179
1180 write_lock(&fs_info->block_group_cache_lock);
1181 caching_ctl = btrfs_get_caching_control(block_group);
1182 if (!caching_ctl) {
1183 struct btrfs_caching_control *ctl;
1184
1185 list_for_each_entry(ctl, &fs_info->caching_block_groups, list) {
1186 if (ctl->block_group == block_group) {
1187 caching_ctl = ctl;
1188 refcount_inc(&caching_ctl->count);
1189 break;
1190 }
1191 }
1192 }
1193 if (caching_ctl)
1194 list_del_init(&caching_ctl->list);
1195 write_unlock(&fs_info->block_group_cache_lock);
1196
1197 if (caching_ctl) {
1198 /* Once for the caching bgs list and once for us. */
1199 btrfs_put_caching_control(caching_ctl);
1200 btrfs_put_caching_control(caching_ctl);
1201 }
1202
1203 spin_lock(&trans->transaction->dirty_bgs_lock);
1204 WARN_ON(!list_empty(&block_group->dirty_list));
1205 WARN_ON(!list_empty(&block_group->io_list));
1206 spin_unlock(&trans->transaction->dirty_bgs_lock);
1207
1208 btrfs_remove_free_space_cache(block_group);
1209
1210 spin_lock(&block_group->space_info->lock);
1211 list_del_init(&block_group->ro_list);
1212
1213 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1214 WARN_ON(block_group->space_info->total_bytes
1215 < block_group->length);
1216 WARN_ON(block_group->space_info->bytes_readonly
1217 < block_group->length - block_group->zone_unusable);
1218 WARN_ON(block_group->space_info->bytes_zone_unusable
1219 < block_group->zone_unusable);
1220 WARN_ON(block_group->space_info->disk_total
1221 < block_group->length * factor);
1222 }
1223 block_group->space_info->total_bytes -= block_group->length;
1224 block_group->space_info->bytes_readonly -=
1225 (block_group->length - block_group->zone_unusable);
1226 btrfs_space_info_update_bytes_zone_unusable(fs_info, block_group->space_info,
1227 -block_group->zone_unusable);
1228 block_group->space_info->disk_total -= block_group->length * factor;
1229
1230 spin_unlock(&block_group->space_info->lock);
1231
1232 /*
1233 * Remove the free space for the block group from the free space tree
1234 * and the block group's item from the extent tree before marking the
1235 * block group as removed. This is to prevent races with tasks that
1236 * freeze and unfreeze a block group, this task and another task
1237 * allocating a new block group - the unfreeze task ends up removing
1238 * the block group's extent map before the task calling this function
1239 * deletes the block group item from the extent tree, allowing for
1240 * another task to attempt to create another block group with the same
1241 * item key (and failing with -EEXIST and a transaction abort).
1242 */
1243 ret = remove_block_group_free_space(trans, block_group);
1244 if (ret)
1245 goto out;
1246
1247 ret = remove_block_group_item(trans, path, block_group);
1248 if (ret < 0)
1249 goto out;
1250
1251 spin_lock(&block_group->lock);
1252 set_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags);
1253
1254 /*
1255 * At this point trimming or scrub can't start on this block group,
1256 * because we removed the block group from the rbtree
1257 * fs_info->block_group_cache_tree so no one can't find it anymore and
1258 * even if someone already got this block group before we removed it
1259 * from the rbtree, they have already incremented block_group->frozen -
1260 * if they didn't, for the trimming case they won't find any free space
1261 * entries because we already removed them all when we called
1262 * btrfs_remove_free_space_cache().
1263 *
1264 * And we must not remove the chunk map from the fs_info->mapping_tree
1265 * to prevent the same logical address range and physical device space
1266 * ranges from being reused for a new block group. This is needed to
1267 * avoid races with trimming and scrub.
1268 *
1269 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1270 * completely transactionless, so while it is trimming a range the
1271 * currently running transaction might finish and a new one start,
1272 * allowing for new block groups to be created that can reuse the same
1273 * physical device locations unless we take this special care.
1274 *
1275 * There may also be an implicit trim operation if the file system
1276 * is mounted with -odiscard. The same protections must remain
1277 * in place until the extents have been discarded completely when
1278 * the transaction commit has completed.
1279 */
1280 remove_map = (atomic_read(&block_group->frozen) == 0);
1281 spin_unlock(&block_group->lock);
1282
1283 if (remove_map)
1284 btrfs_remove_chunk_map(fs_info, map);
1285
1286out:
1287 /* Once for the lookup reference */
1288 btrfs_put_block_group(block_group);
1289 if (remove_rsv)
1290 btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
1291 btrfs_free_path(path);
1292 return ret;
1293}
1294
1295struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1296 struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1297{
1298 struct btrfs_root *root = btrfs_block_group_root(fs_info);
1299 struct btrfs_chunk_map *map;
1300 unsigned int num_items;
1301
1302 map = btrfs_find_chunk_map(fs_info, chunk_offset, 1);
1303 ASSERT(map != NULL);
1304 ASSERT(map->start == chunk_offset);
1305
1306 /*
1307 * We need to reserve 3 + N units from the metadata space info in order
1308 * to remove a block group (done at btrfs_remove_chunk() and at
1309 * btrfs_remove_block_group()), which are used for:
1310 *
1311 * 1 unit for adding the free space inode's orphan (located in the tree
1312 * of tree roots).
1313 * 1 unit for deleting the block group item (located in the extent
1314 * tree).
1315 * 1 unit for deleting the free space item (located in tree of tree
1316 * roots).
1317 * N units for deleting N device extent items corresponding to each
1318 * stripe (located in the device tree).
1319 *
1320 * In order to remove a block group we also need to reserve units in the
1321 * system space info in order to update the chunk tree (update one or
1322 * more device items and remove one chunk item), but this is done at
1323 * btrfs_remove_chunk() through a call to check_system_chunk().
1324 */
1325 num_items = 3 + map->num_stripes;
1326 btrfs_free_chunk_map(map);
1327
1328 return btrfs_start_transaction_fallback_global_rsv(root, num_items);
1329}
1330
1331/*
1332 * Mark block group @cache read-only, so later write won't happen to block
1333 * group @cache.
1334 *
1335 * If @force is not set, this function will only mark the block group readonly
1336 * if we have enough free space (1M) in other metadata/system block groups.
1337 * If @force is not set, this function will mark the block group readonly
1338 * without checking free space.
1339 *
1340 * NOTE: This function doesn't care if other block groups can contain all the
1341 * data in this block group. That check should be done by relocation routine,
1342 * not this function.
1343 */
1344static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1345{
1346 struct btrfs_space_info *sinfo = cache->space_info;
1347 u64 num_bytes;
1348 int ret = -ENOSPC;
1349
1350 spin_lock(&sinfo->lock);
1351 spin_lock(&cache->lock);
1352
1353 if (cache->swap_extents) {
1354 ret = -ETXTBSY;
1355 goto out;
1356 }
1357
1358 if (cache->ro) {
1359 cache->ro++;
1360 ret = 0;
1361 goto out;
1362 }
1363
1364 num_bytes = cache->length - cache->reserved - cache->pinned -
1365 cache->bytes_super - cache->zone_unusable - cache->used;
1366
1367 /*
1368 * Data never overcommits, even in mixed mode, so do just the straight
1369 * check of left over space in how much we have allocated.
1370 */
1371 if (force) {
1372 ret = 0;
1373 } else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1374 u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1375
1376 /*
1377 * Here we make sure if we mark this bg RO, we still have enough
1378 * free space as buffer.
1379 */
1380 if (sinfo_used + num_bytes <= sinfo->total_bytes)
1381 ret = 0;
1382 } else {
1383 /*
1384 * We overcommit metadata, so we need to do the
1385 * btrfs_can_overcommit check here, and we need to pass in
1386 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1387 * leeway to allow us to mark this block group as read only.
1388 */
1389 if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1390 BTRFS_RESERVE_NO_FLUSH))
1391 ret = 0;
1392 }
1393
1394 if (!ret) {
1395 sinfo->bytes_readonly += num_bytes;
1396 if (btrfs_is_zoned(cache->fs_info)) {
1397 /* Migrate zone_unusable bytes to readonly */
1398 sinfo->bytes_readonly += cache->zone_unusable;
1399 btrfs_space_info_update_bytes_zone_unusable(cache->fs_info, sinfo,
1400 -cache->zone_unusable);
1401 cache->zone_unusable = 0;
1402 }
1403 cache->ro++;
1404 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1405 }
1406out:
1407 spin_unlock(&cache->lock);
1408 spin_unlock(&sinfo->lock);
1409 if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1410 btrfs_info(cache->fs_info,
1411 "unable to make block group %llu ro", cache->start);
1412 btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1413 }
1414 return ret;
1415}
1416
1417static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1418 const struct btrfs_block_group *bg)
1419{
1420 struct btrfs_fs_info *fs_info = trans->fs_info;
1421 struct btrfs_transaction *prev_trans = NULL;
1422 const u64 start = bg->start;
1423 const u64 end = start + bg->length - 1;
1424 int ret;
1425
1426 spin_lock(&fs_info->trans_lock);
1427 if (trans->transaction->list.prev != &fs_info->trans_list) {
1428 prev_trans = list_last_entry(&trans->transaction->list,
1429 struct btrfs_transaction, list);
1430 refcount_inc(&prev_trans->use_count);
1431 }
1432 spin_unlock(&fs_info->trans_lock);
1433
1434 /*
1435 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1436 * btrfs_finish_extent_commit(). If we are at transaction N, another
1437 * task might be running finish_extent_commit() for the previous
1438 * transaction N - 1, and have seen a range belonging to the block
1439 * group in pinned_extents before we were able to clear the whole block
1440 * group range from pinned_extents. This means that task can lookup for
1441 * the block group after we unpinned it from pinned_extents and removed
1442 * it, leading to an error at unpin_extent_range().
1443 */
1444 mutex_lock(&fs_info->unused_bg_unpin_mutex);
1445 if (prev_trans) {
1446 ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1447 EXTENT_DIRTY);
1448 if (ret)
1449 goto out;
1450 }
1451
1452 ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1453 EXTENT_DIRTY);
1454out:
1455 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1456 if (prev_trans)
1457 btrfs_put_transaction(prev_trans);
1458
1459 return ret == 0;
1460}
1461
1462/*
1463 * Process the unused_bgs list and remove any that don't have any allocated
1464 * space inside of them.
1465 */
1466void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1467{
1468 LIST_HEAD(retry_list);
1469 struct btrfs_block_group *block_group;
1470 struct btrfs_space_info *space_info;
1471 struct btrfs_trans_handle *trans;
1472 const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1473 int ret = 0;
1474
1475 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1476 return;
1477
1478 if (btrfs_fs_closing(fs_info))
1479 return;
1480
1481 /*
1482 * Long running balances can keep us blocked here for eternity, so
1483 * simply skip deletion if we're unable to get the mutex.
1484 */
1485 if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1486 return;
1487
1488 spin_lock(&fs_info->unused_bgs_lock);
1489 while (!list_empty(&fs_info->unused_bgs)) {
1490 u64 used;
1491 int trimming;
1492
1493 block_group = list_first_entry(&fs_info->unused_bgs,
1494 struct btrfs_block_group,
1495 bg_list);
1496 list_del_init(&block_group->bg_list);
1497
1498 space_info = block_group->space_info;
1499
1500 if (ret || btrfs_mixed_space_info(space_info)) {
1501 btrfs_put_block_group(block_group);
1502 continue;
1503 }
1504 spin_unlock(&fs_info->unused_bgs_lock);
1505
1506 btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1507
1508 /* Don't want to race with allocators so take the groups_sem */
1509 down_write(&space_info->groups_sem);
1510
1511 /*
1512 * Async discard moves the final block group discard to be prior
1513 * to the unused_bgs code path. Therefore, if it's not fully
1514 * trimmed, punt it back to the async discard lists.
1515 */
1516 if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1517 !btrfs_is_free_space_trimmed(block_group)) {
1518 trace_btrfs_skip_unused_block_group(block_group);
1519 up_write(&space_info->groups_sem);
1520 /* Requeue if we failed because of async discard */
1521 btrfs_discard_queue_work(&fs_info->discard_ctl,
1522 block_group);
1523 goto next;
1524 }
1525
1526 spin_lock(&space_info->lock);
1527 spin_lock(&block_group->lock);
1528 if (btrfs_is_block_group_used(block_group) || block_group->ro ||
1529 list_is_singular(&block_group->list)) {
1530 /*
1531 * We want to bail if we made new allocations or have
1532 * outstanding allocations in this block group. We do
1533 * the ro check in case balance is currently acting on
1534 * this block group.
1535 *
1536 * Also bail out if this is the only block group for its
1537 * type, because otherwise we would lose profile
1538 * information from fs_info->avail_*_alloc_bits and the
1539 * next block group of this type would be created with a
1540 * "single" profile (even if we're in a raid fs) because
1541 * fs_info->avail_*_alloc_bits would be 0.
1542 */
1543 trace_btrfs_skip_unused_block_group(block_group);
1544 spin_unlock(&block_group->lock);
1545 spin_unlock(&space_info->lock);
1546 up_write(&space_info->groups_sem);
1547 goto next;
1548 }
1549
1550 /*
1551 * The block group may be unused but there may be space reserved
1552 * accounting with the existence of that block group, that is,
1553 * space_info->bytes_may_use was incremented by a task but no
1554 * space was yet allocated from the block group by the task.
1555 * That space may or may not be allocated, as we are generally
1556 * pessimistic about space reservation for metadata as well as
1557 * for data when using compression (as we reserve space based on
1558 * the worst case, when data can't be compressed, and before
1559 * actually attempting compression, before starting writeback).
1560 *
1561 * So check if the total space of the space_info minus the size
1562 * of this block group is less than the used space of the
1563 * space_info - if that's the case, then it means we have tasks
1564 * that might be relying on the block group in order to allocate
1565 * extents, and add back the block group to the unused list when
1566 * we finish, so that we retry later in case no tasks ended up
1567 * needing to allocate extents from the block group.
1568 */
1569 used = btrfs_space_info_used(space_info, true);
1570 if (space_info->total_bytes - block_group->length < used &&
1571 block_group->zone_unusable < block_group->length) {
1572 /*
1573 * Add a reference for the list, compensate for the ref
1574 * drop under the "next" label for the
1575 * fs_info->unused_bgs list.
1576 */
1577 btrfs_get_block_group(block_group);
1578 list_add_tail(&block_group->bg_list, &retry_list);
1579
1580 trace_btrfs_skip_unused_block_group(block_group);
1581 spin_unlock(&block_group->lock);
1582 spin_unlock(&space_info->lock);
1583 up_write(&space_info->groups_sem);
1584 goto next;
1585 }
1586
1587 spin_unlock(&block_group->lock);
1588 spin_unlock(&space_info->lock);
1589
1590 /* We don't want to force the issue, only flip if it's ok. */
1591 ret = inc_block_group_ro(block_group, 0);
1592 up_write(&space_info->groups_sem);
1593 if (ret < 0) {
1594 ret = 0;
1595 goto next;
1596 }
1597
1598 ret = btrfs_zone_finish(block_group);
1599 if (ret < 0) {
1600 btrfs_dec_block_group_ro(block_group);
1601 if (ret == -EAGAIN)
1602 ret = 0;
1603 goto next;
1604 }
1605
1606 /*
1607 * Want to do this before we do anything else so we can recover
1608 * properly if we fail to join the transaction.
1609 */
1610 trans = btrfs_start_trans_remove_block_group(fs_info,
1611 block_group->start);
1612 if (IS_ERR(trans)) {
1613 btrfs_dec_block_group_ro(block_group);
1614 ret = PTR_ERR(trans);
1615 goto next;
1616 }
1617
1618 /*
1619 * We could have pending pinned extents for this block group,
1620 * just delete them, we don't care about them anymore.
1621 */
1622 if (!clean_pinned_extents(trans, block_group)) {
1623 btrfs_dec_block_group_ro(block_group);
1624 goto end_trans;
1625 }
1626
1627 /*
1628 * At this point, the block_group is read only and should fail
1629 * new allocations. However, btrfs_finish_extent_commit() can
1630 * cause this block_group to be placed back on the discard
1631 * lists because now the block_group isn't fully discarded.
1632 * Bail here and try again later after discarding everything.
1633 */
1634 spin_lock(&fs_info->discard_ctl.lock);
1635 if (!list_empty(&block_group->discard_list)) {
1636 spin_unlock(&fs_info->discard_ctl.lock);
1637 btrfs_dec_block_group_ro(block_group);
1638 btrfs_discard_queue_work(&fs_info->discard_ctl,
1639 block_group);
1640 goto end_trans;
1641 }
1642 spin_unlock(&fs_info->discard_ctl.lock);
1643
1644 /* Reset pinned so btrfs_put_block_group doesn't complain */
1645 spin_lock(&space_info->lock);
1646 spin_lock(&block_group->lock);
1647
1648 btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1649 -block_group->pinned);
1650 space_info->bytes_readonly += block_group->pinned;
1651 block_group->pinned = 0;
1652
1653 spin_unlock(&block_group->lock);
1654 spin_unlock(&space_info->lock);
1655
1656 /*
1657 * The normal path here is an unused block group is passed here,
1658 * then trimming is handled in the transaction commit path.
1659 * Async discard interposes before this to do the trimming
1660 * before coming down the unused block group path as trimming
1661 * will no longer be done later in the transaction commit path.
1662 */
1663 if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1664 goto flip_async;
1665
1666 /*
1667 * DISCARD can flip during remount. On zoned filesystems, we
1668 * need to reset sequential-required zones.
1669 */
1670 trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1671 btrfs_is_zoned(fs_info);
1672
1673 /* Implicit trim during transaction commit. */
1674 if (trimming)
1675 btrfs_freeze_block_group(block_group);
1676
1677 /*
1678 * Btrfs_remove_chunk will abort the transaction if things go
1679 * horribly wrong.
1680 */
1681 ret = btrfs_remove_chunk(trans, block_group->start);
1682
1683 if (ret) {
1684 if (trimming)
1685 btrfs_unfreeze_block_group(block_group);
1686 goto end_trans;
1687 }
1688
1689 /*
1690 * If we're not mounted with -odiscard, we can just forget
1691 * about this block group. Otherwise we'll need to wait
1692 * until transaction commit to do the actual discard.
1693 */
1694 if (trimming) {
1695 spin_lock(&fs_info->unused_bgs_lock);
1696 /*
1697 * A concurrent scrub might have added us to the list
1698 * fs_info->unused_bgs, so use a list_move operation
1699 * to add the block group to the deleted_bgs list.
1700 */
1701 list_move(&block_group->bg_list,
1702 &trans->transaction->deleted_bgs);
1703 spin_unlock(&fs_info->unused_bgs_lock);
1704 btrfs_get_block_group(block_group);
1705 }
1706end_trans:
1707 btrfs_end_transaction(trans);
1708next:
1709 btrfs_put_block_group(block_group);
1710 spin_lock(&fs_info->unused_bgs_lock);
1711 }
1712 list_splice_tail(&retry_list, &fs_info->unused_bgs);
1713 spin_unlock(&fs_info->unused_bgs_lock);
1714 mutex_unlock(&fs_info->reclaim_bgs_lock);
1715 return;
1716
1717flip_async:
1718 btrfs_end_transaction(trans);
1719 spin_lock(&fs_info->unused_bgs_lock);
1720 list_splice_tail(&retry_list, &fs_info->unused_bgs);
1721 spin_unlock(&fs_info->unused_bgs_lock);
1722 mutex_unlock(&fs_info->reclaim_bgs_lock);
1723 btrfs_put_block_group(block_group);
1724 btrfs_discard_punt_unused_bgs_list(fs_info);
1725}
1726
1727void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1728{
1729 struct btrfs_fs_info *fs_info = bg->fs_info;
1730
1731 spin_lock(&fs_info->unused_bgs_lock);
1732 if (list_empty(&bg->bg_list)) {
1733 btrfs_get_block_group(bg);
1734 trace_btrfs_add_unused_block_group(bg);
1735 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1736 } else if (!test_bit(BLOCK_GROUP_FLAG_NEW, &bg->runtime_flags)) {
1737 /* Pull out the block group from the reclaim_bgs list. */
1738 trace_btrfs_add_unused_block_group(bg);
1739 list_move_tail(&bg->bg_list, &fs_info->unused_bgs);
1740 }
1741 spin_unlock(&fs_info->unused_bgs_lock);
1742}
1743
1744/*
1745 * We want block groups with a low number of used bytes to be in the beginning
1746 * of the list, so they will get reclaimed first.
1747 */
1748static int reclaim_bgs_cmp(void *unused, const struct list_head *a,
1749 const struct list_head *b)
1750{
1751 const struct btrfs_block_group *bg1, *bg2;
1752
1753 bg1 = list_entry(a, struct btrfs_block_group, bg_list);
1754 bg2 = list_entry(b, struct btrfs_block_group, bg_list);
1755
1756 return bg1->used > bg2->used;
1757}
1758
1759static inline bool btrfs_should_reclaim(const struct btrfs_fs_info *fs_info)
1760{
1761 if (btrfs_is_zoned(fs_info))
1762 return btrfs_zoned_should_reclaim(fs_info);
1763 return true;
1764}
1765
1766static bool should_reclaim_block_group(const struct btrfs_block_group *bg, u64 bytes_freed)
1767{
1768 const int thresh_pct = btrfs_calc_reclaim_threshold(bg->space_info);
1769 u64 thresh_bytes = mult_perc(bg->length, thresh_pct);
1770 const u64 new_val = bg->used;
1771 const u64 old_val = new_val + bytes_freed;
1772
1773 if (thresh_bytes == 0)
1774 return false;
1775
1776 /*
1777 * If we were below the threshold before don't reclaim, we are likely a
1778 * brand new block group and we don't want to relocate new block groups.
1779 */
1780 if (old_val < thresh_bytes)
1781 return false;
1782 if (new_val >= thresh_bytes)
1783 return false;
1784 return true;
1785}
1786
1787void btrfs_reclaim_bgs_work(struct work_struct *work)
1788{
1789 struct btrfs_fs_info *fs_info =
1790 container_of(work, struct btrfs_fs_info, reclaim_bgs_work);
1791 struct btrfs_block_group *bg;
1792 struct btrfs_space_info *space_info;
1793 LIST_HEAD(retry_list);
1794
1795 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1796 return;
1797
1798 if (btrfs_fs_closing(fs_info))
1799 return;
1800
1801 if (!btrfs_should_reclaim(fs_info))
1802 return;
1803
1804 sb_start_write(fs_info->sb);
1805
1806 if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_BALANCE)) {
1807 sb_end_write(fs_info->sb);
1808 return;
1809 }
1810
1811 /*
1812 * Long running balances can keep us blocked here for eternity, so
1813 * simply skip reclaim if we're unable to get the mutex.
1814 */
1815 if (!mutex_trylock(&fs_info->reclaim_bgs_lock)) {
1816 btrfs_exclop_finish(fs_info);
1817 sb_end_write(fs_info->sb);
1818 return;
1819 }
1820
1821 spin_lock(&fs_info->unused_bgs_lock);
1822 /*
1823 * Sort happens under lock because we can't simply splice it and sort.
1824 * The block groups might still be in use and reachable via bg_list,
1825 * and their presence in the reclaim_bgs list must be preserved.
1826 */
1827 list_sort(NULL, &fs_info->reclaim_bgs, reclaim_bgs_cmp);
1828 while (!list_empty(&fs_info->reclaim_bgs)) {
1829 u64 zone_unusable;
1830 u64 reclaimed;
1831 int ret = 0;
1832
1833 bg = list_first_entry(&fs_info->reclaim_bgs,
1834 struct btrfs_block_group,
1835 bg_list);
1836 list_del_init(&bg->bg_list);
1837
1838 space_info = bg->space_info;
1839 spin_unlock(&fs_info->unused_bgs_lock);
1840
1841 /* Don't race with allocators so take the groups_sem */
1842 down_write(&space_info->groups_sem);
1843
1844 spin_lock(&space_info->lock);
1845 spin_lock(&bg->lock);
1846 if (bg->reserved || bg->pinned || bg->ro) {
1847 /*
1848 * We want to bail if we made new allocations or have
1849 * outstanding allocations in this block group. We do
1850 * the ro check in case balance is currently acting on
1851 * this block group.
1852 */
1853 spin_unlock(&bg->lock);
1854 spin_unlock(&space_info->lock);
1855 up_write(&space_info->groups_sem);
1856 goto next;
1857 }
1858 if (bg->used == 0) {
1859 /*
1860 * It is possible that we trigger relocation on a block
1861 * group as its extents are deleted and it first goes
1862 * below the threshold, then shortly after goes empty.
1863 *
1864 * In this case, relocating it does delete it, but has
1865 * some overhead in relocation specific metadata, looking
1866 * for the non-existent extents and running some extra
1867 * transactions, which we can avoid by using one of the
1868 * other mechanisms for dealing with empty block groups.
1869 */
1870 if (!btrfs_test_opt(fs_info, DISCARD_ASYNC))
1871 btrfs_mark_bg_unused(bg);
1872 spin_unlock(&bg->lock);
1873 spin_unlock(&space_info->lock);
1874 up_write(&space_info->groups_sem);
1875 goto next;
1876
1877 }
1878 /*
1879 * The block group might no longer meet the reclaim condition by
1880 * the time we get around to reclaiming it, so to avoid
1881 * reclaiming overly full block_groups, skip reclaiming them.
1882 *
1883 * Since the decision making process also depends on the amount
1884 * being freed, pass in a fake giant value to skip that extra
1885 * check, which is more meaningful when adding to the list in
1886 * the first place.
1887 */
1888 if (!should_reclaim_block_group(bg, bg->length)) {
1889 spin_unlock(&bg->lock);
1890 spin_unlock(&space_info->lock);
1891 up_write(&space_info->groups_sem);
1892 goto next;
1893 }
1894 spin_unlock(&bg->lock);
1895 spin_unlock(&space_info->lock);
1896
1897 /*
1898 * Get out fast, in case we're read-only or unmounting the
1899 * filesystem. It is OK to drop block groups from the list even
1900 * for the read-only case. As we did sb_start_write(),
1901 * "mount -o remount,ro" won't happen and read-only filesystem
1902 * means it is forced read-only due to a fatal error. So, it
1903 * never gets back to read-write to let us reclaim again.
1904 */
1905 if (btrfs_need_cleaner_sleep(fs_info)) {
1906 up_write(&space_info->groups_sem);
1907 goto next;
1908 }
1909
1910 /*
1911 * Cache the zone_unusable value before turning the block group
1912 * to read only. As soon as the blog group is read only it's
1913 * zone_unusable value gets moved to the block group's read-only
1914 * bytes and isn't available for calculations anymore.
1915 */
1916 zone_unusable = bg->zone_unusable;
1917 ret = inc_block_group_ro(bg, 0);
1918 up_write(&space_info->groups_sem);
1919 if (ret < 0)
1920 goto next;
1921
1922 btrfs_info(fs_info,
1923 "reclaiming chunk %llu with %llu%% used %llu%% unusable",
1924 bg->start,
1925 div64_u64(bg->used * 100, bg->length),
1926 div64_u64(zone_unusable * 100, bg->length));
1927 trace_btrfs_reclaim_block_group(bg);
1928 reclaimed = bg->used;
1929 ret = btrfs_relocate_chunk(fs_info, bg->start);
1930 if (ret) {
1931 btrfs_dec_block_group_ro(bg);
1932 btrfs_err(fs_info, "error relocating chunk %llu",
1933 bg->start);
1934 reclaimed = 0;
1935 spin_lock(&space_info->lock);
1936 space_info->reclaim_errors++;
1937 if (READ_ONCE(space_info->periodic_reclaim))
1938 space_info->periodic_reclaim_ready = false;
1939 spin_unlock(&space_info->lock);
1940 }
1941 spin_lock(&space_info->lock);
1942 space_info->reclaim_count++;
1943 space_info->reclaim_bytes += reclaimed;
1944 spin_unlock(&space_info->lock);
1945
1946next:
1947 if (ret && !READ_ONCE(space_info->periodic_reclaim)) {
1948 /* Refcount held by the reclaim_bgs list after splice. */
1949 spin_lock(&fs_info->unused_bgs_lock);
1950 /*
1951 * This block group might be added to the unused list
1952 * during the above process. Move it back to the
1953 * reclaim list otherwise.
1954 */
1955 if (list_empty(&bg->bg_list)) {
1956 btrfs_get_block_group(bg);
1957 list_add_tail(&bg->bg_list, &retry_list);
1958 }
1959 spin_unlock(&fs_info->unused_bgs_lock);
1960 }
1961 btrfs_put_block_group(bg);
1962
1963 mutex_unlock(&fs_info->reclaim_bgs_lock);
1964 /*
1965 * Reclaiming all the block groups in the list can take really
1966 * long. Prioritize cleaning up unused block groups.
1967 */
1968 btrfs_delete_unused_bgs(fs_info);
1969 /*
1970 * If we are interrupted by a balance, we can just bail out. The
1971 * cleaner thread restart again if necessary.
1972 */
1973 if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1974 goto end;
1975 spin_lock(&fs_info->unused_bgs_lock);
1976 }
1977 spin_unlock(&fs_info->unused_bgs_lock);
1978 mutex_unlock(&fs_info->reclaim_bgs_lock);
1979end:
1980 spin_lock(&fs_info->unused_bgs_lock);
1981 list_splice_tail(&retry_list, &fs_info->reclaim_bgs);
1982 spin_unlock(&fs_info->unused_bgs_lock);
1983 btrfs_exclop_finish(fs_info);
1984 sb_end_write(fs_info->sb);
1985}
1986
1987void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info)
1988{
1989 btrfs_reclaim_sweep(fs_info);
1990 spin_lock(&fs_info->unused_bgs_lock);
1991 if (!list_empty(&fs_info->reclaim_bgs))
1992 queue_work(system_unbound_wq, &fs_info->reclaim_bgs_work);
1993 spin_unlock(&fs_info->unused_bgs_lock);
1994}
1995
1996void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg)
1997{
1998 struct btrfs_fs_info *fs_info = bg->fs_info;
1999
2000 spin_lock(&fs_info->unused_bgs_lock);
2001 if (list_empty(&bg->bg_list)) {
2002 btrfs_get_block_group(bg);
2003 trace_btrfs_add_reclaim_block_group(bg);
2004 list_add_tail(&bg->bg_list, &fs_info->reclaim_bgs);
2005 }
2006 spin_unlock(&fs_info->unused_bgs_lock);
2007}
2008
2009static int read_bg_from_eb(struct btrfs_fs_info *fs_info, const struct btrfs_key *key,
2010 const struct btrfs_path *path)
2011{
2012 struct btrfs_chunk_map *map;
2013 struct btrfs_block_group_item bg;
2014 struct extent_buffer *leaf;
2015 int slot;
2016 u64 flags;
2017 int ret = 0;
2018
2019 slot = path->slots[0];
2020 leaf = path->nodes[0];
2021
2022 map = btrfs_find_chunk_map(fs_info, key->objectid, key->offset);
2023 if (!map) {
2024 btrfs_err(fs_info,
2025 "logical %llu len %llu found bg but no related chunk",
2026 key->objectid, key->offset);
2027 return -ENOENT;
2028 }
2029
2030 if (map->start != key->objectid || map->chunk_len != key->offset) {
2031 btrfs_err(fs_info,
2032 "block group %llu len %llu mismatch with chunk %llu len %llu",
2033 key->objectid, key->offset, map->start, map->chunk_len);
2034 ret = -EUCLEAN;
2035 goto out_free_map;
2036 }
2037
2038 read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
2039 sizeof(bg));
2040 flags = btrfs_stack_block_group_flags(&bg) &
2041 BTRFS_BLOCK_GROUP_TYPE_MASK;
2042
2043 if (flags != (map->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2044 btrfs_err(fs_info,
2045"block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
2046 key->objectid, key->offset, flags,
2047 (BTRFS_BLOCK_GROUP_TYPE_MASK & map->type));
2048 ret = -EUCLEAN;
2049 }
2050
2051out_free_map:
2052 btrfs_free_chunk_map(map);
2053 return ret;
2054}
2055
2056static int find_first_block_group(struct btrfs_fs_info *fs_info,
2057 struct btrfs_path *path,
2058 const struct btrfs_key *key)
2059{
2060 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2061 int ret;
2062 struct btrfs_key found_key;
2063
2064 btrfs_for_each_slot(root, key, &found_key, path, ret) {
2065 if (found_key.objectid >= key->objectid &&
2066 found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
2067 return read_bg_from_eb(fs_info, &found_key, path);
2068 }
2069 }
2070 return ret;
2071}
2072
2073static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2074{
2075 u64 extra_flags = chunk_to_extended(flags) &
2076 BTRFS_EXTENDED_PROFILE_MASK;
2077
2078 write_seqlock(&fs_info->profiles_lock);
2079 if (flags & BTRFS_BLOCK_GROUP_DATA)
2080 fs_info->avail_data_alloc_bits |= extra_flags;
2081 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2082 fs_info->avail_metadata_alloc_bits |= extra_flags;
2083 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2084 fs_info->avail_system_alloc_bits |= extra_flags;
2085 write_sequnlock(&fs_info->profiles_lock);
2086}
2087
2088/*
2089 * Map a physical disk address to a list of logical addresses.
2090 *
2091 * @fs_info: the filesystem
2092 * @chunk_start: logical address of block group
2093 * @physical: physical address to map to logical addresses
2094 * @logical: return array of logical addresses which map to @physical
2095 * @naddrs: length of @logical
2096 * @stripe_len: size of IO stripe for the given block group
2097 *
2098 * Maps a particular @physical disk address to a list of @logical addresses.
2099 * Used primarily to exclude those portions of a block group that contain super
2100 * block copies.
2101 */
2102int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
2103 u64 physical, u64 **logical, int *naddrs, int *stripe_len)
2104{
2105 struct btrfs_chunk_map *map;
2106 u64 *buf;
2107 u64 bytenr;
2108 u64 data_stripe_length;
2109 u64 io_stripe_size;
2110 int i, nr = 0;
2111 int ret = 0;
2112
2113 map = btrfs_get_chunk_map(fs_info, chunk_start, 1);
2114 if (IS_ERR(map))
2115 return -EIO;
2116
2117 data_stripe_length = map->stripe_size;
2118 io_stripe_size = BTRFS_STRIPE_LEN;
2119 chunk_start = map->start;
2120
2121 /* For RAID5/6 adjust to a full IO stripe length */
2122 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2123 io_stripe_size = btrfs_stripe_nr_to_offset(nr_data_stripes(map));
2124
2125 buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
2126 if (!buf) {
2127 ret = -ENOMEM;
2128 goto out;
2129 }
2130
2131 for (i = 0; i < map->num_stripes; i++) {
2132 bool already_inserted = false;
2133 u32 stripe_nr;
2134 u32 offset;
2135 int j;
2136
2137 if (!in_range(physical, map->stripes[i].physical,
2138 data_stripe_length))
2139 continue;
2140
2141 stripe_nr = (physical - map->stripes[i].physical) >>
2142 BTRFS_STRIPE_LEN_SHIFT;
2143 offset = (physical - map->stripes[i].physical) &
2144 BTRFS_STRIPE_LEN_MASK;
2145
2146 if (map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2147 BTRFS_BLOCK_GROUP_RAID10))
2148 stripe_nr = div_u64(stripe_nr * map->num_stripes + i,
2149 map->sub_stripes);
2150 /*
2151 * The remaining case would be for RAID56, multiply by
2152 * nr_data_stripes(). Alternatively, just use rmap_len below
2153 * instead of map->stripe_len
2154 */
2155 bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
2156
2157 /* Ensure we don't add duplicate addresses */
2158 for (j = 0; j < nr; j++) {
2159 if (buf[j] == bytenr) {
2160 already_inserted = true;
2161 break;
2162 }
2163 }
2164
2165 if (!already_inserted)
2166 buf[nr++] = bytenr;
2167 }
2168
2169 *logical = buf;
2170 *naddrs = nr;
2171 *stripe_len = io_stripe_size;
2172out:
2173 btrfs_free_chunk_map(map);
2174 return ret;
2175}
2176
2177static int exclude_super_stripes(struct btrfs_block_group *cache)
2178{
2179 struct btrfs_fs_info *fs_info = cache->fs_info;
2180 const bool zoned = btrfs_is_zoned(fs_info);
2181 u64 bytenr;
2182 u64 *logical;
2183 int stripe_len;
2184 int i, nr, ret;
2185
2186 if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
2187 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
2188 cache->bytes_super += stripe_len;
2189 ret = set_extent_bit(&fs_info->excluded_extents, cache->start,
2190 cache->start + stripe_len - 1,
2191 EXTENT_UPTODATE, NULL);
2192 if (ret)
2193 return ret;
2194 }
2195
2196 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2197 bytenr = btrfs_sb_offset(i);
2198 ret = btrfs_rmap_block(fs_info, cache->start,
2199 bytenr, &logical, &nr, &stripe_len);
2200 if (ret)
2201 return ret;
2202
2203 /* Shouldn't have super stripes in sequential zones */
2204 if (zoned && nr) {
2205 kfree(logical);
2206 btrfs_err(fs_info,
2207 "zoned: block group %llu must not contain super block",
2208 cache->start);
2209 return -EUCLEAN;
2210 }
2211
2212 while (nr--) {
2213 u64 len = min_t(u64, stripe_len,
2214 cache->start + cache->length - logical[nr]);
2215
2216 cache->bytes_super += len;
2217 ret = set_extent_bit(&fs_info->excluded_extents, logical[nr],
2218 logical[nr] + len - 1,
2219 EXTENT_UPTODATE, NULL);
2220 if (ret) {
2221 kfree(logical);
2222 return ret;
2223 }
2224 }
2225
2226 kfree(logical);
2227 }
2228 return 0;
2229}
2230
2231static struct btrfs_block_group *btrfs_create_block_group_cache(
2232 struct btrfs_fs_info *fs_info, u64 start)
2233{
2234 struct btrfs_block_group *cache;
2235
2236 cache = kzalloc(sizeof(*cache), GFP_NOFS);
2237 if (!cache)
2238 return NULL;
2239
2240 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
2241 GFP_NOFS);
2242 if (!cache->free_space_ctl) {
2243 kfree(cache);
2244 return NULL;
2245 }
2246
2247 cache->start = start;
2248
2249 cache->fs_info = fs_info;
2250 cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
2251
2252 cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
2253
2254 refcount_set(&cache->refs, 1);
2255 spin_lock_init(&cache->lock);
2256 init_rwsem(&cache->data_rwsem);
2257 INIT_LIST_HEAD(&cache->list);
2258 INIT_LIST_HEAD(&cache->cluster_list);
2259 INIT_LIST_HEAD(&cache->bg_list);
2260 INIT_LIST_HEAD(&cache->ro_list);
2261 INIT_LIST_HEAD(&cache->discard_list);
2262 INIT_LIST_HEAD(&cache->dirty_list);
2263 INIT_LIST_HEAD(&cache->io_list);
2264 INIT_LIST_HEAD(&cache->active_bg_list);
2265 btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
2266 atomic_set(&cache->frozen, 0);
2267 mutex_init(&cache->free_space_lock);
2268
2269 return cache;
2270}
2271
2272/*
2273 * Iterate all chunks and verify that each of them has the corresponding block
2274 * group
2275 */
2276static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
2277{
2278 u64 start = 0;
2279 int ret = 0;
2280
2281 while (1) {
2282 struct btrfs_chunk_map *map;
2283 struct btrfs_block_group *bg;
2284
2285 /*
2286 * btrfs_find_chunk_map() will return the first chunk map
2287 * intersecting the range, so setting @length to 1 is enough to
2288 * get the first chunk.
2289 */
2290 map = btrfs_find_chunk_map(fs_info, start, 1);
2291 if (!map)
2292 break;
2293
2294 bg = btrfs_lookup_block_group(fs_info, map->start);
2295 if (!bg) {
2296 btrfs_err(fs_info,
2297 "chunk start=%llu len=%llu doesn't have corresponding block group",
2298 map->start, map->chunk_len);
2299 ret = -EUCLEAN;
2300 btrfs_free_chunk_map(map);
2301 break;
2302 }
2303 if (bg->start != map->start || bg->length != map->chunk_len ||
2304 (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
2305 (map->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2306 btrfs_err(fs_info,
2307"chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
2308 map->start, map->chunk_len,
2309 map->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
2310 bg->start, bg->length,
2311 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
2312 ret = -EUCLEAN;
2313 btrfs_free_chunk_map(map);
2314 btrfs_put_block_group(bg);
2315 break;
2316 }
2317 start = map->start + map->chunk_len;
2318 btrfs_free_chunk_map(map);
2319 btrfs_put_block_group(bg);
2320 }
2321 return ret;
2322}
2323
2324static int read_one_block_group(struct btrfs_fs_info *info,
2325 struct btrfs_block_group_item *bgi,
2326 const struct btrfs_key *key,
2327 int need_clear)
2328{
2329 struct btrfs_block_group *cache;
2330 const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
2331 int ret;
2332
2333 ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
2334
2335 cache = btrfs_create_block_group_cache(info, key->objectid);
2336 if (!cache)
2337 return -ENOMEM;
2338
2339 cache->length = key->offset;
2340 cache->used = btrfs_stack_block_group_used(bgi);
2341 cache->commit_used = cache->used;
2342 cache->flags = btrfs_stack_block_group_flags(bgi);
2343 cache->global_root_id = btrfs_stack_block_group_chunk_objectid(bgi);
2344
2345 set_free_space_tree_thresholds(cache);
2346
2347 if (need_clear) {
2348 /*
2349 * When we mount with old space cache, we need to
2350 * set BTRFS_DC_CLEAR and set dirty flag.
2351 *
2352 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
2353 * truncate the old free space cache inode and
2354 * setup a new one.
2355 * b) Setting 'dirty flag' makes sure that we flush
2356 * the new space cache info onto disk.
2357 */
2358 if (btrfs_test_opt(info, SPACE_CACHE))
2359 cache->disk_cache_state = BTRFS_DC_CLEAR;
2360 }
2361 if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
2362 (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
2363 btrfs_err(info,
2364"bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2365 cache->start);
2366 ret = -EINVAL;
2367 goto error;
2368 }
2369
2370 ret = btrfs_load_block_group_zone_info(cache, false);
2371 if (ret) {
2372 btrfs_err(info, "zoned: failed to load zone info of bg %llu",
2373 cache->start);
2374 goto error;
2375 }
2376
2377 /*
2378 * We need to exclude the super stripes now so that the space info has
2379 * super bytes accounted for, otherwise we'll think we have more space
2380 * than we actually do.
2381 */
2382 ret = exclude_super_stripes(cache);
2383 if (ret) {
2384 /* We may have excluded something, so call this just in case. */
2385 btrfs_free_excluded_extents(cache);
2386 goto error;
2387 }
2388
2389 /*
2390 * For zoned filesystem, space after the allocation offset is the only
2391 * free space for a block group. So, we don't need any caching work.
2392 * btrfs_calc_zone_unusable() will set the amount of free space and
2393 * zone_unusable space.
2394 *
2395 * For regular filesystem, check for two cases, either we are full, and
2396 * therefore don't need to bother with the caching work since we won't
2397 * find any space, or we are empty, and we can just add all the space
2398 * in and be done with it. This saves us _a_lot_ of time, particularly
2399 * in the full case.
2400 */
2401 if (btrfs_is_zoned(info)) {
2402 btrfs_calc_zone_unusable(cache);
2403 /* Should not have any excluded extents. Just in case, though. */
2404 btrfs_free_excluded_extents(cache);
2405 } else if (cache->length == cache->used) {
2406 cache->cached = BTRFS_CACHE_FINISHED;
2407 btrfs_free_excluded_extents(cache);
2408 } else if (cache->used == 0) {
2409 cache->cached = BTRFS_CACHE_FINISHED;
2410 ret = btrfs_add_new_free_space(cache, cache->start,
2411 cache->start + cache->length, NULL);
2412 btrfs_free_excluded_extents(cache);
2413 if (ret)
2414 goto error;
2415 }
2416
2417 ret = btrfs_add_block_group_cache(info, cache);
2418 if (ret) {
2419 btrfs_remove_free_space_cache(cache);
2420 goto error;
2421 }
2422 trace_btrfs_add_block_group(info, cache, 0);
2423 btrfs_add_bg_to_space_info(info, cache);
2424
2425 set_avail_alloc_bits(info, cache->flags);
2426 if (btrfs_chunk_writeable(info, cache->start)) {
2427 if (cache->used == 0) {
2428 ASSERT(list_empty(&cache->bg_list));
2429 if (btrfs_test_opt(info, DISCARD_ASYNC))
2430 btrfs_discard_queue_work(&info->discard_ctl, cache);
2431 else
2432 btrfs_mark_bg_unused(cache);
2433 }
2434 } else {
2435 inc_block_group_ro(cache, 1);
2436 }
2437
2438 return 0;
2439error:
2440 btrfs_put_block_group(cache);
2441 return ret;
2442}
2443
2444static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
2445{
2446 struct rb_node *node;
2447 int ret = 0;
2448
2449 for (node = rb_first_cached(&fs_info->mapping_tree); node; node = rb_next(node)) {
2450 struct btrfs_chunk_map *map;
2451 struct btrfs_block_group *bg;
2452
2453 map = rb_entry(node, struct btrfs_chunk_map, rb_node);
2454 bg = btrfs_create_block_group_cache(fs_info, map->start);
2455 if (!bg) {
2456 ret = -ENOMEM;
2457 break;
2458 }
2459
2460 /* Fill dummy cache as FULL */
2461 bg->length = map->chunk_len;
2462 bg->flags = map->type;
2463 bg->cached = BTRFS_CACHE_FINISHED;
2464 bg->used = map->chunk_len;
2465 bg->flags = map->type;
2466 ret = btrfs_add_block_group_cache(fs_info, bg);
2467 /*
2468 * We may have some valid block group cache added already, in
2469 * that case we skip to the next one.
2470 */
2471 if (ret == -EEXIST) {
2472 ret = 0;
2473 btrfs_put_block_group(bg);
2474 continue;
2475 }
2476
2477 if (ret) {
2478 btrfs_remove_free_space_cache(bg);
2479 btrfs_put_block_group(bg);
2480 break;
2481 }
2482
2483 btrfs_add_bg_to_space_info(fs_info, bg);
2484
2485 set_avail_alloc_bits(fs_info, bg->flags);
2486 }
2487 if (!ret)
2488 btrfs_init_global_block_rsv(fs_info);
2489 return ret;
2490}
2491
2492int btrfs_read_block_groups(struct btrfs_fs_info *info)
2493{
2494 struct btrfs_root *root = btrfs_block_group_root(info);
2495 struct btrfs_path *path;
2496 int ret;
2497 struct btrfs_block_group *cache;
2498 struct btrfs_space_info *space_info;
2499 struct btrfs_key key;
2500 int need_clear = 0;
2501 u64 cache_gen;
2502
2503 /*
2504 * Either no extent root (with ibadroots rescue option) or we have
2505 * unsupported RO options. The fs can never be mounted read-write, so no
2506 * need to waste time searching block group items.
2507 *
2508 * This also allows new extent tree related changes to be RO compat,
2509 * no need for a full incompat flag.
2510 */
2511 if (!root || (btrfs_super_compat_ro_flags(info->super_copy) &
2512 ~BTRFS_FEATURE_COMPAT_RO_SUPP))
2513 return fill_dummy_bgs(info);
2514
2515 key.objectid = 0;
2516 key.offset = 0;
2517 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2518 path = btrfs_alloc_path();
2519 if (!path)
2520 return -ENOMEM;
2521
2522 cache_gen = btrfs_super_cache_generation(info->super_copy);
2523 if (btrfs_test_opt(info, SPACE_CACHE) &&
2524 btrfs_super_generation(info->super_copy) != cache_gen)
2525 need_clear = 1;
2526 if (btrfs_test_opt(info, CLEAR_CACHE))
2527 need_clear = 1;
2528
2529 while (1) {
2530 struct btrfs_block_group_item bgi;
2531 struct extent_buffer *leaf;
2532 int slot;
2533
2534 ret = find_first_block_group(info, path, &key);
2535 if (ret > 0)
2536 break;
2537 if (ret != 0)
2538 goto error;
2539
2540 leaf = path->nodes[0];
2541 slot = path->slots[0];
2542
2543 read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2544 sizeof(bgi));
2545
2546 btrfs_item_key_to_cpu(leaf, &key, slot);
2547 btrfs_release_path(path);
2548 ret = read_one_block_group(info, &bgi, &key, need_clear);
2549 if (ret < 0)
2550 goto error;
2551 key.objectid += key.offset;
2552 key.offset = 0;
2553 }
2554 btrfs_release_path(path);
2555
2556 list_for_each_entry(space_info, &info->space_info, list) {
2557 int i;
2558
2559 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2560 if (list_empty(&space_info->block_groups[i]))
2561 continue;
2562 cache = list_first_entry(&space_info->block_groups[i],
2563 struct btrfs_block_group,
2564 list);
2565 btrfs_sysfs_add_block_group_type(cache);
2566 }
2567
2568 if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2569 (BTRFS_BLOCK_GROUP_RAID10 |
2570 BTRFS_BLOCK_GROUP_RAID1_MASK |
2571 BTRFS_BLOCK_GROUP_RAID56_MASK |
2572 BTRFS_BLOCK_GROUP_DUP)))
2573 continue;
2574 /*
2575 * Avoid allocating from un-mirrored block group if there are
2576 * mirrored block groups.
2577 */
2578 list_for_each_entry(cache,
2579 &space_info->block_groups[BTRFS_RAID_RAID0],
2580 list)
2581 inc_block_group_ro(cache, 1);
2582 list_for_each_entry(cache,
2583 &space_info->block_groups[BTRFS_RAID_SINGLE],
2584 list)
2585 inc_block_group_ro(cache, 1);
2586 }
2587
2588 btrfs_init_global_block_rsv(info);
2589 ret = check_chunk_block_group_mappings(info);
2590error:
2591 btrfs_free_path(path);
2592 /*
2593 * We've hit some error while reading the extent tree, and have
2594 * rescue=ibadroots mount option.
2595 * Try to fill the tree using dummy block groups so that the user can
2596 * continue to mount and grab their data.
2597 */
2598 if (ret && btrfs_test_opt(info, IGNOREBADROOTS))
2599 ret = fill_dummy_bgs(info);
2600 return ret;
2601}
2602
2603/*
2604 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2605 * allocation.
2606 *
2607 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2608 * phases.
2609 */
2610static int insert_block_group_item(struct btrfs_trans_handle *trans,
2611 struct btrfs_block_group *block_group)
2612{
2613 struct btrfs_fs_info *fs_info = trans->fs_info;
2614 struct btrfs_block_group_item bgi;
2615 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2616 struct btrfs_key key;
2617 u64 old_commit_used;
2618 int ret;
2619
2620 spin_lock(&block_group->lock);
2621 btrfs_set_stack_block_group_used(&bgi, block_group->used);
2622 btrfs_set_stack_block_group_chunk_objectid(&bgi,
2623 block_group->global_root_id);
2624 btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2625 old_commit_used = block_group->commit_used;
2626 block_group->commit_used = block_group->used;
2627 key.objectid = block_group->start;
2628 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2629 key.offset = block_group->length;
2630 spin_unlock(&block_group->lock);
2631
2632 ret = btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2633 if (ret < 0) {
2634 spin_lock(&block_group->lock);
2635 block_group->commit_used = old_commit_used;
2636 spin_unlock(&block_group->lock);
2637 }
2638
2639 return ret;
2640}
2641
2642static int insert_dev_extent(struct btrfs_trans_handle *trans,
2643 const struct btrfs_device *device, u64 chunk_offset,
2644 u64 start, u64 num_bytes)
2645{
2646 struct btrfs_fs_info *fs_info = device->fs_info;
2647 struct btrfs_root *root = fs_info->dev_root;
2648 struct btrfs_path *path;
2649 struct btrfs_dev_extent *extent;
2650 struct extent_buffer *leaf;
2651 struct btrfs_key key;
2652 int ret;
2653
2654 WARN_ON(!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state));
2655 WARN_ON(test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state));
2656 path = btrfs_alloc_path();
2657 if (!path)
2658 return -ENOMEM;
2659
2660 key.objectid = device->devid;
2661 key.type = BTRFS_DEV_EXTENT_KEY;
2662 key.offset = start;
2663 ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*extent));
2664 if (ret)
2665 goto out;
2666
2667 leaf = path->nodes[0];
2668 extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
2669 btrfs_set_dev_extent_chunk_tree(leaf, extent, BTRFS_CHUNK_TREE_OBJECTID);
2670 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
2671 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2672 btrfs_set_dev_extent_chunk_offset(leaf, extent, chunk_offset);
2673
2674 btrfs_set_dev_extent_length(leaf, extent, num_bytes);
2675 btrfs_mark_buffer_dirty(trans, leaf);
2676out:
2677 btrfs_free_path(path);
2678 return ret;
2679}
2680
2681/*
2682 * This function belongs to phase 2.
2683 *
2684 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2685 * phases.
2686 */
2687static int insert_dev_extents(struct btrfs_trans_handle *trans,
2688 u64 chunk_offset, u64 chunk_size)
2689{
2690 struct btrfs_fs_info *fs_info = trans->fs_info;
2691 struct btrfs_device *device;
2692 struct btrfs_chunk_map *map;
2693 u64 dev_offset;
2694 int i;
2695 int ret = 0;
2696
2697 map = btrfs_get_chunk_map(fs_info, chunk_offset, chunk_size);
2698 if (IS_ERR(map))
2699 return PTR_ERR(map);
2700
2701 /*
2702 * Take the device list mutex to prevent races with the final phase of
2703 * a device replace operation that replaces the device object associated
2704 * with the map's stripes, because the device object's id can change
2705 * at any time during that final phase of the device replace operation
2706 * (dev-replace.c:btrfs_dev_replace_finishing()), so we could grab the
2707 * replaced device and then see it with an ID of BTRFS_DEV_REPLACE_DEVID,
2708 * resulting in persisting a device extent item with such ID.
2709 */
2710 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2711 for (i = 0; i < map->num_stripes; i++) {
2712 device = map->stripes[i].dev;
2713 dev_offset = map->stripes[i].physical;
2714
2715 ret = insert_dev_extent(trans, device, chunk_offset, dev_offset,
2716 map->stripe_size);
2717 if (ret)
2718 break;
2719 }
2720 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2721
2722 btrfs_free_chunk_map(map);
2723 return ret;
2724}
2725
2726/*
2727 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2728 * chunk allocation.
2729 *
2730 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2731 * phases.
2732 */
2733void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2734{
2735 struct btrfs_fs_info *fs_info = trans->fs_info;
2736 struct btrfs_block_group *block_group;
2737 int ret = 0;
2738
2739 while (!list_empty(&trans->new_bgs)) {
2740 int index;
2741
2742 block_group = list_first_entry(&trans->new_bgs,
2743 struct btrfs_block_group,
2744 bg_list);
2745 if (ret)
2746 goto next;
2747
2748 index = btrfs_bg_flags_to_raid_index(block_group->flags);
2749
2750 ret = insert_block_group_item(trans, block_group);
2751 if (ret)
2752 btrfs_abort_transaction(trans, ret);
2753 if (!test_bit(BLOCK_GROUP_FLAG_CHUNK_ITEM_INSERTED,
2754 &block_group->runtime_flags)) {
2755 mutex_lock(&fs_info->chunk_mutex);
2756 ret = btrfs_chunk_alloc_add_chunk_item(trans, block_group);
2757 mutex_unlock(&fs_info->chunk_mutex);
2758 if (ret)
2759 btrfs_abort_transaction(trans, ret);
2760 }
2761 ret = insert_dev_extents(trans, block_group->start,
2762 block_group->length);
2763 if (ret)
2764 btrfs_abort_transaction(trans, ret);
2765 add_block_group_free_space(trans, block_group);
2766
2767 /*
2768 * If we restriped during balance, we may have added a new raid
2769 * type, so now add the sysfs entries when it is safe to do so.
2770 * We don't have to worry about locking here as it's handled in
2771 * btrfs_sysfs_add_block_group_type.
2772 */
2773 if (block_group->space_info->block_group_kobjs[index] == NULL)
2774 btrfs_sysfs_add_block_group_type(block_group);
2775
2776 /* Already aborted the transaction if it failed. */
2777next:
2778 btrfs_dec_delayed_refs_rsv_bg_inserts(fs_info);
2779 list_del_init(&block_group->bg_list);
2780 clear_bit(BLOCK_GROUP_FLAG_NEW, &block_group->runtime_flags);
2781
2782 /*
2783 * If the block group is still unused, add it to the list of
2784 * unused block groups. The block group may have been created in
2785 * order to satisfy a space reservation, in which case the
2786 * extent allocation only happens later. But often we don't
2787 * actually need to allocate space that we previously reserved,
2788 * so the block group may become unused for a long time. For
2789 * example for metadata we generally reserve space for a worst
2790 * possible scenario, but then don't end up allocating all that
2791 * space or none at all (due to no need to COW, extent buffers
2792 * were already COWed in the current transaction and still
2793 * unwritten, tree heights lower than the maximum possible
2794 * height, etc). For data we generally reserve the axact amount
2795 * of space we are going to allocate later, the exception is
2796 * when using compression, as we must reserve space based on the
2797 * uncompressed data size, because the compression is only done
2798 * when writeback triggered and we don't know how much space we
2799 * are actually going to need, so we reserve the uncompressed
2800 * size because the data may be incompressible in the worst case.
2801 */
2802 if (ret == 0) {
2803 bool used;
2804
2805 spin_lock(&block_group->lock);
2806 used = btrfs_is_block_group_used(block_group);
2807 spin_unlock(&block_group->lock);
2808
2809 if (!used)
2810 btrfs_mark_bg_unused(block_group);
2811 }
2812 }
2813 btrfs_trans_release_chunk_metadata(trans);
2814}
2815
2816/*
2817 * For extent tree v2 we use the block_group_item->chunk_offset to point at our
2818 * global root id. For v1 it's always set to BTRFS_FIRST_CHUNK_TREE_OBJECTID.
2819 */
2820static u64 calculate_global_root_id(const struct btrfs_fs_info *fs_info, u64 offset)
2821{
2822 u64 div = SZ_1G;
2823 u64 index;
2824
2825 if (!btrfs_fs_incompat(fs_info, EXTENT_TREE_V2))
2826 return BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2827
2828 /* If we have a smaller fs index based on 128MiB. */
2829 if (btrfs_super_total_bytes(fs_info->super_copy) <= (SZ_1G * 10ULL))
2830 div = SZ_128M;
2831
2832 offset = div64_u64(offset, div);
2833 div64_u64_rem(offset, fs_info->nr_global_roots, &index);
2834 return index;
2835}
2836
2837struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
2838 u64 type,
2839 u64 chunk_offset, u64 size)
2840{
2841 struct btrfs_fs_info *fs_info = trans->fs_info;
2842 struct btrfs_block_group *cache;
2843 int ret;
2844
2845 btrfs_set_log_full_commit(trans);
2846
2847 cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2848 if (!cache)
2849 return ERR_PTR(-ENOMEM);
2850
2851 /*
2852 * Mark it as new before adding it to the rbtree of block groups or any
2853 * list, so that no other task finds it and calls btrfs_mark_bg_unused()
2854 * before the new flag is set.
2855 */
2856 set_bit(BLOCK_GROUP_FLAG_NEW, &cache->runtime_flags);
2857
2858 cache->length = size;
2859 set_free_space_tree_thresholds(cache);
2860 cache->flags = type;
2861 cache->cached = BTRFS_CACHE_FINISHED;
2862 cache->global_root_id = calculate_global_root_id(fs_info, cache->start);
2863
2864 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2865 set_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE, &cache->runtime_flags);
2866
2867 ret = btrfs_load_block_group_zone_info(cache, true);
2868 if (ret) {
2869 btrfs_put_block_group(cache);
2870 return ERR_PTR(ret);
2871 }
2872
2873 ret = exclude_super_stripes(cache);
2874 if (ret) {
2875 /* We may have excluded something, so call this just in case */
2876 btrfs_free_excluded_extents(cache);
2877 btrfs_put_block_group(cache);
2878 return ERR_PTR(ret);
2879 }
2880
2881 ret = btrfs_add_new_free_space(cache, chunk_offset, chunk_offset + size, NULL);
2882 btrfs_free_excluded_extents(cache);
2883 if (ret) {
2884 btrfs_put_block_group(cache);
2885 return ERR_PTR(ret);
2886 }
2887
2888 /*
2889 * Ensure the corresponding space_info object is created and
2890 * assigned to our block group. We want our bg to be added to the rbtree
2891 * with its ->space_info set.
2892 */
2893 cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2894 ASSERT(cache->space_info);
2895
2896 ret = btrfs_add_block_group_cache(fs_info, cache);
2897 if (ret) {
2898 btrfs_remove_free_space_cache(cache);
2899 btrfs_put_block_group(cache);
2900 return ERR_PTR(ret);
2901 }
2902
2903 /*
2904 * Now that our block group has its ->space_info set and is inserted in
2905 * the rbtree, update the space info's counters.
2906 */
2907 trace_btrfs_add_block_group(fs_info, cache, 1);
2908 btrfs_add_bg_to_space_info(fs_info, cache);
2909 btrfs_update_global_block_rsv(fs_info);
2910
2911#ifdef CONFIG_BTRFS_DEBUG
2912 if (btrfs_should_fragment_free_space(cache)) {
2913 cache->space_info->bytes_used += size >> 1;
2914 fragment_free_space(cache);
2915 }
2916#endif
2917
2918 list_add_tail(&cache->bg_list, &trans->new_bgs);
2919 btrfs_inc_delayed_refs_rsv_bg_inserts(fs_info);
2920
2921 set_avail_alloc_bits(fs_info, type);
2922 return cache;
2923}
2924
2925/*
2926 * Mark one block group RO, can be called several times for the same block
2927 * group.
2928 *
2929 * @cache: the destination block group
2930 * @do_chunk_alloc: whether need to do chunk pre-allocation, this is to
2931 * ensure we still have some free space after marking this
2932 * block group RO.
2933 */
2934int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2935 bool do_chunk_alloc)
2936{
2937 struct btrfs_fs_info *fs_info = cache->fs_info;
2938 struct btrfs_trans_handle *trans;
2939 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2940 u64 alloc_flags;
2941 int ret;
2942 bool dirty_bg_running;
2943
2944 /*
2945 * This can only happen when we are doing read-only scrub on read-only
2946 * mount.
2947 * In that case we should not start a new transaction on read-only fs.
2948 * Thus here we skip all chunk allocations.
2949 */
2950 if (sb_rdonly(fs_info->sb)) {
2951 mutex_lock(&fs_info->ro_block_group_mutex);
2952 ret = inc_block_group_ro(cache, 0);
2953 mutex_unlock(&fs_info->ro_block_group_mutex);
2954 return ret;
2955 }
2956
2957 do {
2958 trans = btrfs_join_transaction(root);
2959 if (IS_ERR(trans))
2960 return PTR_ERR(trans);
2961
2962 dirty_bg_running = false;
2963
2964 /*
2965 * We're not allowed to set block groups readonly after the dirty
2966 * block group cache has started writing. If it already started,
2967 * back off and let this transaction commit.
2968 */
2969 mutex_lock(&fs_info->ro_block_group_mutex);
2970 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2971 u64 transid = trans->transid;
2972
2973 mutex_unlock(&fs_info->ro_block_group_mutex);
2974 btrfs_end_transaction(trans);
2975
2976 ret = btrfs_wait_for_commit(fs_info, transid);
2977 if (ret)
2978 return ret;
2979 dirty_bg_running = true;
2980 }
2981 } while (dirty_bg_running);
2982
2983 if (do_chunk_alloc) {
2984 /*
2985 * If we are changing raid levels, try to allocate a
2986 * corresponding block group with the new raid level.
2987 */
2988 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2989 if (alloc_flags != cache->flags) {
2990 ret = btrfs_chunk_alloc(trans, alloc_flags,
2991 CHUNK_ALLOC_FORCE);
2992 /*
2993 * ENOSPC is allowed here, we may have enough space
2994 * already allocated at the new raid level to carry on
2995 */
2996 if (ret == -ENOSPC)
2997 ret = 0;
2998 if (ret < 0)
2999 goto out;
3000 }
3001 }
3002
3003 ret = inc_block_group_ro(cache, 0);
3004 if (!ret)
3005 goto out;
3006 if (ret == -ETXTBSY)
3007 goto unlock_out;
3008
3009 /*
3010 * Skip chunk allocation if the bg is SYSTEM, this is to avoid system
3011 * chunk allocation storm to exhaust the system chunk array. Otherwise
3012 * we still want to try our best to mark the block group read-only.
3013 */
3014 if (!do_chunk_alloc && ret == -ENOSPC &&
3015 (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM))
3016 goto unlock_out;
3017
3018 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
3019 ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3020 if (ret < 0)
3021 goto out;
3022 /*
3023 * We have allocated a new chunk. We also need to activate that chunk to
3024 * grant metadata tickets for zoned filesystem.
3025 */
3026 ret = btrfs_zoned_activate_one_bg(fs_info, cache->space_info, true);
3027 if (ret < 0)
3028 goto out;
3029
3030 ret = inc_block_group_ro(cache, 0);
3031 if (ret == -ETXTBSY)
3032 goto unlock_out;
3033out:
3034 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
3035 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
3036 mutex_lock(&fs_info->chunk_mutex);
3037 check_system_chunk(trans, alloc_flags);
3038 mutex_unlock(&fs_info->chunk_mutex);
3039 }
3040unlock_out:
3041 mutex_unlock(&fs_info->ro_block_group_mutex);
3042
3043 btrfs_end_transaction(trans);
3044 return ret;
3045}
3046
3047void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
3048{
3049 struct btrfs_space_info *sinfo = cache->space_info;
3050 u64 num_bytes;
3051
3052 BUG_ON(!cache->ro);
3053
3054 spin_lock(&sinfo->lock);
3055 spin_lock(&cache->lock);
3056 if (!--cache->ro) {
3057 if (btrfs_is_zoned(cache->fs_info)) {
3058 /* Migrate zone_unusable bytes back */
3059 cache->zone_unusable =
3060 (cache->alloc_offset - cache->used - cache->pinned -
3061 cache->reserved) +
3062 (cache->length - cache->zone_capacity);
3063 btrfs_space_info_update_bytes_zone_unusable(cache->fs_info, sinfo,
3064 cache->zone_unusable);
3065 sinfo->bytes_readonly -= cache->zone_unusable;
3066 }
3067 num_bytes = cache->length - cache->reserved -
3068 cache->pinned - cache->bytes_super -
3069 cache->zone_unusable - cache->used;
3070 sinfo->bytes_readonly -= num_bytes;
3071 list_del_init(&cache->ro_list);
3072 }
3073 spin_unlock(&cache->lock);
3074 spin_unlock(&sinfo->lock);
3075}
3076
3077static int update_block_group_item(struct btrfs_trans_handle *trans,
3078 struct btrfs_path *path,
3079 struct btrfs_block_group *cache)
3080{
3081 struct btrfs_fs_info *fs_info = trans->fs_info;
3082 int ret;
3083 struct btrfs_root *root = btrfs_block_group_root(fs_info);
3084 unsigned long bi;
3085 struct extent_buffer *leaf;
3086 struct btrfs_block_group_item bgi;
3087 struct btrfs_key key;
3088 u64 old_commit_used;
3089 u64 used;
3090
3091 /*
3092 * Block group items update can be triggered out of commit transaction
3093 * critical section, thus we need a consistent view of used bytes.
3094 * We cannot use cache->used directly outside of the spin lock, as it
3095 * may be changed.
3096 */
3097 spin_lock(&cache->lock);
3098 old_commit_used = cache->commit_used;
3099 used = cache->used;
3100 /* No change in used bytes, can safely skip it. */
3101 if (cache->commit_used == used) {
3102 spin_unlock(&cache->lock);
3103 return 0;
3104 }
3105 cache->commit_used = used;
3106 spin_unlock(&cache->lock);
3107
3108 key.objectid = cache->start;
3109 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
3110 key.offset = cache->length;
3111
3112 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3113 if (ret) {
3114 if (ret > 0)
3115 ret = -ENOENT;
3116 goto fail;
3117 }
3118
3119 leaf = path->nodes[0];
3120 bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3121 btrfs_set_stack_block_group_used(&bgi, used);
3122 btrfs_set_stack_block_group_chunk_objectid(&bgi,
3123 cache->global_root_id);
3124 btrfs_set_stack_block_group_flags(&bgi, cache->flags);
3125 write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
3126 btrfs_mark_buffer_dirty(trans, leaf);
3127fail:
3128 btrfs_release_path(path);
3129 /*
3130 * We didn't update the block group item, need to revert commit_used
3131 * unless the block group item didn't exist yet - this is to prevent a
3132 * race with a concurrent insertion of the block group item, with
3133 * insert_block_group_item(), that happened just after we attempted to
3134 * update. In that case we would reset commit_used to 0 just after the
3135 * insertion set it to a value greater than 0 - if the block group later
3136 * becomes with 0 used bytes, we would incorrectly skip its update.
3137 */
3138 if (ret < 0 && ret != -ENOENT) {
3139 spin_lock(&cache->lock);
3140 cache->commit_used = old_commit_used;
3141 spin_unlock(&cache->lock);
3142 }
3143 return ret;
3144
3145}
3146
3147static int cache_save_setup(struct btrfs_block_group *block_group,
3148 struct btrfs_trans_handle *trans,
3149 struct btrfs_path *path)
3150{
3151 struct btrfs_fs_info *fs_info = block_group->fs_info;
3152 struct inode *inode = NULL;
3153 struct extent_changeset *data_reserved = NULL;
3154 u64 alloc_hint = 0;
3155 int dcs = BTRFS_DC_ERROR;
3156 u64 cache_size = 0;
3157 int retries = 0;
3158 int ret = 0;
3159
3160 if (!btrfs_test_opt(fs_info, SPACE_CACHE))
3161 return 0;
3162
3163 /*
3164 * If this block group is smaller than 100 megs don't bother caching the
3165 * block group.
3166 */
3167 if (block_group->length < (100 * SZ_1M)) {
3168 spin_lock(&block_group->lock);
3169 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3170 spin_unlock(&block_group->lock);
3171 return 0;
3172 }
3173
3174 if (TRANS_ABORTED(trans))
3175 return 0;
3176again:
3177 inode = lookup_free_space_inode(block_group, path);
3178 if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3179 ret = PTR_ERR(inode);
3180 btrfs_release_path(path);
3181 goto out;
3182 }
3183
3184 if (IS_ERR(inode)) {
3185 BUG_ON(retries);
3186 retries++;
3187
3188 if (block_group->ro)
3189 goto out_free;
3190
3191 ret = create_free_space_inode(trans, block_group, path);
3192 if (ret)
3193 goto out_free;
3194 goto again;
3195 }
3196
3197 /*
3198 * We want to set the generation to 0, that way if anything goes wrong
3199 * from here on out we know not to trust this cache when we load up next
3200 * time.
3201 */
3202 BTRFS_I(inode)->generation = 0;
3203 ret = btrfs_update_inode(trans, BTRFS_I(inode));
3204 if (ret) {
3205 /*
3206 * So theoretically we could recover from this, simply set the
3207 * super cache generation to 0 so we know to invalidate the
3208 * cache, but then we'd have to keep track of the block groups
3209 * that fail this way so we know we _have_ to reset this cache
3210 * before the next commit or risk reading stale cache. So to
3211 * limit our exposure to horrible edge cases lets just abort the
3212 * transaction, this only happens in really bad situations
3213 * anyway.
3214 */
3215 btrfs_abort_transaction(trans, ret);
3216 goto out_put;
3217 }
3218 WARN_ON(ret);
3219
3220 /* We've already setup this transaction, go ahead and exit */
3221 if (block_group->cache_generation == trans->transid &&
3222 i_size_read(inode)) {
3223 dcs = BTRFS_DC_SETUP;
3224 goto out_put;
3225 }
3226
3227 if (i_size_read(inode) > 0) {
3228 ret = btrfs_check_trunc_cache_free_space(fs_info,
3229 &fs_info->global_block_rsv);
3230 if (ret)
3231 goto out_put;
3232
3233 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3234 if (ret)
3235 goto out_put;
3236 }
3237
3238 spin_lock(&block_group->lock);
3239 if (block_group->cached != BTRFS_CACHE_FINISHED ||
3240 !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3241 /*
3242 * don't bother trying to write stuff out _if_
3243 * a) we're not cached,
3244 * b) we're with nospace_cache mount option,
3245 * c) we're with v2 space_cache (FREE_SPACE_TREE).
3246 */
3247 dcs = BTRFS_DC_WRITTEN;
3248 spin_unlock(&block_group->lock);
3249 goto out_put;
3250 }
3251 spin_unlock(&block_group->lock);
3252
3253 /*
3254 * We hit an ENOSPC when setting up the cache in this transaction, just
3255 * skip doing the setup, we've already cleared the cache so we're safe.
3256 */
3257 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3258 ret = -ENOSPC;
3259 goto out_put;
3260 }
3261
3262 /*
3263 * Try to preallocate enough space based on how big the block group is.
3264 * Keep in mind this has to include any pinned space which could end up
3265 * taking up quite a bit since it's not folded into the other space
3266 * cache.
3267 */
3268 cache_size = div_u64(block_group->length, SZ_256M);
3269 if (!cache_size)
3270 cache_size = 1;
3271
3272 cache_size *= 16;
3273 cache_size *= fs_info->sectorsize;
3274
3275 ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
3276 cache_size, false);
3277 if (ret)
3278 goto out_put;
3279
3280 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, cache_size,
3281 cache_size, cache_size,
3282 &alloc_hint);
3283 /*
3284 * Our cache requires contiguous chunks so that we don't modify a bunch
3285 * of metadata or split extents when writing the cache out, which means
3286 * we can enospc if we are heavily fragmented in addition to just normal
3287 * out of space conditions. So if we hit this just skip setting up any
3288 * other block groups for this transaction, maybe we'll unpin enough
3289 * space the next time around.
3290 */
3291 if (!ret)
3292 dcs = BTRFS_DC_SETUP;
3293 else if (ret == -ENOSPC)
3294 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3295
3296out_put:
3297 iput(inode);
3298out_free:
3299 btrfs_release_path(path);
3300out:
3301 spin_lock(&block_group->lock);
3302 if (!ret && dcs == BTRFS_DC_SETUP)
3303 block_group->cache_generation = trans->transid;
3304 block_group->disk_cache_state = dcs;
3305 spin_unlock(&block_group->lock);
3306
3307 extent_changeset_free(data_reserved);
3308 return ret;
3309}
3310
3311int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
3312{
3313 struct btrfs_fs_info *fs_info = trans->fs_info;
3314 struct btrfs_block_group *cache, *tmp;
3315 struct btrfs_transaction *cur_trans = trans->transaction;
3316 struct btrfs_path *path;
3317
3318 if (list_empty(&cur_trans->dirty_bgs) ||
3319 !btrfs_test_opt(fs_info, SPACE_CACHE))
3320 return 0;
3321
3322 path = btrfs_alloc_path();
3323 if (!path)
3324 return -ENOMEM;
3325
3326 /* Could add new block groups, use _safe just in case */
3327 list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3328 dirty_list) {
3329 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3330 cache_save_setup(cache, trans, path);
3331 }
3332
3333 btrfs_free_path(path);
3334 return 0;
3335}
3336
3337/*
3338 * Transaction commit does final block group cache writeback during a critical
3339 * section where nothing is allowed to change the FS. This is required in
3340 * order for the cache to actually match the block group, but can introduce a
3341 * lot of latency into the commit.
3342 *
3343 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
3344 * There's a chance we'll have to redo some of it if the block group changes
3345 * again during the commit, but it greatly reduces the commit latency by
3346 * getting rid of the easy block groups while we're still allowing others to
3347 * join the commit.
3348 */
3349int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3350{
3351 struct btrfs_fs_info *fs_info = trans->fs_info;
3352 struct btrfs_block_group *cache;
3353 struct btrfs_transaction *cur_trans = trans->transaction;
3354 int ret = 0;
3355 int should_put;
3356 struct btrfs_path *path = NULL;
3357 LIST_HEAD(dirty);
3358 struct list_head *io = &cur_trans->io_bgs;
3359 int loops = 0;
3360
3361 spin_lock(&cur_trans->dirty_bgs_lock);
3362 if (list_empty(&cur_trans->dirty_bgs)) {
3363 spin_unlock(&cur_trans->dirty_bgs_lock);
3364 return 0;
3365 }
3366 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3367 spin_unlock(&cur_trans->dirty_bgs_lock);
3368
3369again:
3370 /* Make sure all the block groups on our dirty list actually exist */
3371 btrfs_create_pending_block_groups(trans);
3372
3373 if (!path) {
3374 path = btrfs_alloc_path();
3375 if (!path) {
3376 ret = -ENOMEM;
3377 goto out;
3378 }
3379 }
3380
3381 /*
3382 * cache_write_mutex is here only to save us from balance or automatic
3383 * removal of empty block groups deleting this block group while we are
3384 * writing out the cache
3385 */
3386 mutex_lock(&trans->transaction->cache_write_mutex);
3387 while (!list_empty(&dirty)) {
3388 bool drop_reserve = true;
3389
3390 cache = list_first_entry(&dirty, struct btrfs_block_group,
3391 dirty_list);
3392 /*
3393 * This can happen if something re-dirties a block group that
3394 * is already under IO. Just wait for it to finish and then do
3395 * it all again
3396 */
3397 if (!list_empty(&cache->io_list)) {
3398 list_del_init(&cache->io_list);
3399 btrfs_wait_cache_io(trans, cache, path);
3400 btrfs_put_block_group(cache);
3401 }
3402
3403
3404 /*
3405 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
3406 * it should update the cache_state. Don't delete until after
3407 * we wait.
3408 *
3409 * Since we're not running in the commit critical section
3410 * we need the dirty_bgs_lock to protect from update_block_group
3411 */
3412 spin_lock(&cur_trans->dirty_bgs_lock);
3413 list_del_init(&cache->dirty_list);
3414 spin_unlock(&cur_trans->dirty_bgs_lock);
3415
3416 should_put = 1;
3417
3418 cache_save_setup(cache, trans, path);
3419
3420 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3421 cache->io_ctl.inode = NULL;
3422 ret = btrfs_write_out_cache(trans, cache, path);
3423 if (ret == 0 && cache->io_ctl.inode) {
3424 should_put = 0;
3425
3426 /*
3427 * The cache_write_mutex is protecting the
3428 * io_list, also refer to the definition of
3429 * btrfs_transaction::io_bgs for more details
3430 */
3431 list_add_tail(&cache->io_list, io);
3432 } else {
3433 /*
3434 * If we failed to write the cache, the
3435 * generation will be bad and life goes on
3436 */
3437 ret = 0;
3438 }
3439 }
3440 if (!ret) {
3441 ret = update_block_group_item(trans, path, cache);
3442 /*
3443 * Our block group might still be attached to the list
3444 * of new block groups in the transaction handle of some
3445 * other task (struct btrfs_trans_handle->new_bgs). This
3446 * means its block group item isn't yet in the extent
3447 * tree. If this happens ignore the error, as we will
3448 * try again later in the critical section of the
3449 * transaction commit.
3450 */
3451 if (ret == -ENOENT) {
3452 ret = 0;
3453 spin_lock(&cur_trans->dirty_bgs_lock);
3454 if (list_empty(&cache->dirty_list)) {
3455 list_add_tail(&cache->dirty_list,
3456 &cur_trans->dirty_bgs);
3457 btrfs_get_block_group(cache);
3458 drop_reserve = false;
3459 }
3460 spin_unlock(&cur_trans->dirty_bgs_lock);
3461 } else if (ret) {
3462 btrfs_abort_transaction(trans, ret);
3463 }
3464 }
3465
3466 /* If it's not on the io list, we need to put the block group */
3467 if (should_put)
3468 btrfs_put_block_group(cache);
3469 if (drop_reserve)
3470 btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
3471 /*
3472 * Avoid blocking other tasks for too long. It might even save
3473 * us from writing caches for block groups that are going to be
3474 * removed.
3475 */
3476 mutex_unlock(&trans->transaction->cache_write_mutex);
3477 if (ret)
3478 goto out;
3479 mutex_lock(&trans->transaction->cache_write_mutex);
3480 }
3481 mutex_unlock(&trans->transaction->cache_write_mutex);
3482
3483 /*
3484 * Go through delayed refs for all the stuff we've just kicked off
3485 * and then loop back (just once)
3486 */
3487 if (!ret)
3488 ret = btrfs_run_delayed_refs(trans, 0);
3489 if (!ret && loops == 0) {
3490 loops++;
3491 spin_lock(&cur_trans->dirty_bgs_lock);
3492 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3493 /*
3494 * dirty_bgs_lock protects us from concurrent block group
3495 * deletes too (not just cache_write_mutex).
3496 */
3497 if (!list_empty(&dirty)) {
3498 spin_unlock(&cur_trans->dirty_bgs_lock);
3499 goto again;
3500 }
3501 spin_unlock(&cur_trans->dirty_bgs_lock);
3502 }
3503out:
3504 if (ret < 0) {
3505 spin_lock(&cur_trans->dirty_bgs_lock);
3506 list_splice_init(&dirty, &cur_trans->dirty_bgs);
3507 spin_unlock(&cur_trans->dirty_bgs_lock);
3508 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3509 }
3510
3511 btrfs_free_path(path);
3512 return ret;
3513}
3514
3515int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3516{
3517 struct btrfs_fs_info *fs_info = trans->fs_info;
3518 struct btrfs_block_group *cache;
3519 struct btrfs_transaction *cur_trans = trans->transaction;
3520 int ret = 0;
3521 int should_put;
3522 struct btrfs_path *path;
3523 struct list_head *io = &cur_trans->io_bgs;
3524
3525 path = btrfs_alloc_path();
3526 if (!path)
3527 return -ENOMEM;
3528
3529 /*
3530 * Even though we are in the critical section of the transaction commit,
3531 * we can still have concurrent tasks adding elements to this
3532 * transaction's list of dirty block groups. These tasks correspond to
3533 * endio free space workers started when writeback finishes for a
3534 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3535 * allocate new block groups as a result of COWing nodes of the root
3536 * tree when updating the free space inode. The writeback for the space
3537 * caches is triggered by an earlier call to
3538 * btrfs_start_dirty_block_groups() and iterations of the following
3539 * loop.
3540 * Also we want to do the cache_save_setup first and then run the
3541 * delayed refs to make sure we have the best chance at doing this all
3542 * in one shot.
3543 */
3544 spin_lock(&cur_trans->dirty_bgs_lock);
3545 while (!list_empty(&cur_trans->dirty_bgs)) {
3546 cache = list_first_entry(&cur_trans->dirty_bgs,
3547 struct btrfs_block_group,
3548 dirty_list);
3549
3550 /*
3551 * This can happen if cache_save_setup re-dirties a block group
3552 * that is already under IO. Just wait for it to finish and
3553 * then do it all again
3554 */
3555 if (!list_empty(&cache->io_list)) {
3556 spin_unlock(&cur_trans->dirty_bgs_lock);
3557 list_del_init(&cache->io_list);
3558 btrfs_wait_cache_io(trans, cache, path);
3559 btrfs_put_block_group(cache);
3560 spin_lock(&cur_trans->dirty_bgs_lock);
3561 }
3562
3563 /*
3564 * Don't remove from the dirty list until after we've waited on
3565 * any pending IO
3566 */
3567 list_del_init(&cache->dirty_list);
3568 spin_unlock(&cur_trans->dirty_bgs_lock);
3569 should_put = 1;
3570
3571 cache_save_setup(cache, trans, path);
3572
3573 if (!ret)
3574 ret = btrfs_run_delayed_refs(trans, U64_MAX);
3575
3576 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3577 cache->io_ctl.inode = NULL;
3578 ret = btrfs_write_out_cache(trans, cache, path);
3579 if (ret == 0 && cache->io_ctl.inode) {
3580 should_put = 0;
3581 list_add_tail(&cache->io_list, io);
3582 } else {
3583 /*
3584 * If we failed to write the cache, the
3585 * generation will be bad and life goes on
3586 */
3587 ret = 0;
3588 }
3589 }
3590 if (!ret) {
3591 ret = update_block_group_item(trans, path, cache);
3592 /*
3593 * One of the free space endio workers might have
3594 * created a new block group while updating a free space
3595 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3596 * and hasn't released its transaction handle yet, in
3597 * which case the new block group is still attached to
3598 * its transaction handle and its creation has not
3599 * finished yet (no block group item in the extent tree
3600 * yet, etc). If this is the case, wait for all free
3601 * space endio workers to finish and retry. This is a
3602 * very rare case so no need for a more efficient and
3603 * complex approach.
3604 */
3605 if (ret == -ENOENT) {
3606 wait_event(cur_trans->writer_wait,
3607 atomic_read(&cur_trans->num_writers) == 1);
3608 ret = update_block_group_item(trans, path, cache);
3609 }
3610 if (ret)
3611 btrfs_abort_transaction(trans, ret);
3612 }
3613
3614 /* If its not on the io list, we need to put the block group */
3615 if (should_put)
3616 btrfs_put_block_group(cache);
3617 btrfs_dec_delayed_refs_rsv_bg_updates(fs_info);
3618 spin_lock(&cur_trans->dirty_bgs_lock);
3619 }
3620 spin_unlock(&cur_trans->dirty_bgs_lock);
3621
3622 /*
3623 * Refer to the definition of io_bgs member for details why it's safe
3624 * to use it without any locking
3625 */
3626 while (!list_empty(io)) {
3627 cache = list_first_entry(io, struct btrfs_block_group,
3628 io_list);
3629 list_del_init(&cache->io_list);
3630 btrfs_wait_cache_io(trans, cache, path);
3631 btrfs_put_block_group(cache);
3632 }
3633
3634 btrfs_free_path(path);
3635 return ret;
3636}
3637
3638int btrfs_update_block_group(struct btrfs_trans_handle *trans,
3639 u64 bytenr, u64 num_bytes, bool alloc)
3640{
3641 struct btrfs_fs_info *info = trans->fs_info;
3642 struct btrfs_space_info *space_info;
3643 struct btrfs_block_group *cache;
3644 u64 old_val;
3645 bool reclaim = false;
3646 bool bg_already_dirty = true;
3647 int factor;
3648
3649 /* Block accounting for super block */
3650 spin_lock(&info->delalloc_root_lock);
3651 old_val = btrfs_super_bytes_used(info->super_copy);
3652 if (alloc)
3653 old_val += num_bytes;
3654 else
3655 old_val -= num_bytes;
3656 btrfs_set_super_bytes_used(info->super_copy, old_val);
3657 spin_unlock(&info->delalloc_root_lock);
3658
3659 cache = btrfs_lookup_block_group(info, bytenr);
3660 if (!cache)
3661 return -ENOENT;
3662
3663 /* An extent can not span multiple block groups. */
3664 ASSERT(bytenr + num_bytes <= cache->start + cache->length);
3665
3666 space_info = cache->space_info;
3667 factor = btrfs_bg_type_to_factor(cache->flags);
3668
3669 /*
3670 * If this block group has free space cache written out, we need to make
3671 * sure to load it if we are removing space. This is because we need
3672 * the unpinning stage to actually add the space back to the block group,
3673 * otherwise we will leak space.
3674 */
3675 if (!alloc && !btrfs_block_group_done(cache))
3676 btrfs_cache_block_group(cache, true);
3677
3678 spin_lock(&space_info->lock);
3679 spin_lock(&cache->lock);
3680
3681 if (btrfs_test_opt(info, SPACE_CACHE) &&
3682 cache->disk_cache_state < BTRFS_DC_CLEAR)
3683 cache->disk_cache_state = BTRFS_DC_CLEAR;
3684
3685 old_val = cache->used;
3686 if (alloc) {
3687 old_val += num_bytes;
3688 cache->used = old_val;
3689 cache->reserved -= num_bytes;
3690 cache->reclaim_mark = 0;
3691 space_info->bytes_reserved -= num_bytes;
3692 space_info->bytes_used += num_bytes;
3693 space_info->disk_used += num_bytes * factor;
3694 if (READ_ONCE(space_info->periodic_reclaim))
3695 btrfs_space_info_update_reclaimable(space_info, -num_bytes);
3696 spin_unlock(&cache->lock);
3697 spin_unlock(&space_info->lock);
3698 } else {
3699 old_val -= num_bytes;
3700 cache->used = old_val;
3701 cache->pinned += num_bytes;
3702 btrfs_space_info_update_bytes_pinned(info, space_info, num_bytes);
3703 space_info->bytes_used -= num_bytes;
3704 space_info->disk_used -= num_bytes * factor;
3705 if (READ_ONCE(space_info->periodic_reclaim))
3706 btrfs_space_info_update_reclaimable(space_info, num_bytes);
3707 else
3708 reclaim = should_reclaim_block_group(cache, num_bytes);
3709
3710 spin_unlock(&cache->lock);
3711 spin_unlock(&space_info->lock);
3712
3713 set_extent_bit(&trans->transaction->pinned_extents, bytenr,
3714 bytenr + num_bytes - 1, EXTENT_DIRTY, NULL);
3715 }
3716
3717 spin_lock(&trans->transaction->dirty_bgs_lock);
3718 if (list_empty(&cache->dirty_list)) {
3719 list_add_tail(&cache->dirty_list, &trans->transaction->dirty_bgs);
3720 bg_already_dirty = false;
3721 btrfs_get_block_group(cache);
3722 }
3723 spin_unlock(&trans->transaction->dirty_bgs_lock);
3724
3725 /*
3726 * No longer have used bytes in this block group, queue it for deletion.
3727 * We do this after adding the block group to the dirty list to avoid
3728 * races between cleaner kthread and space cache writeout.
3729 */
3730 if (!alloc && old_val == 0) {
3731 if (!btrfs_test_opt(info, DISCARD_ASYNC))
3732 btrfs_mark_bg_unused(cache);
3733 } else if (!alloc && reclaim) {
3734 btrfs_mark_bg_to_reclaim(cache);
3735 }
3736
3737 btrfs_put_block_group(cache);
3738
3739 /* Modified block groups are accounted for in the delayed_refs_rsv. */
3740 if (!bg_already_dirty)
3741 btrfs_inc_delayed_refs_rsv_bg_updates(info);
3742
3743 return 0;
3744}
3745
3746/*
3747 * Update the block_group and space info counters.
3748 *
3749 * @cache: The cache we are manipulating
3750 * @ram_bytes: The number of bytes of file content, and will be same to
3751 * @num_bytes except for the compress path.
3752 * @num_bytes: The number of bytes in question
3753 * @delalloc: The blocks are allocated for the delalloc write
3754 *
3755 * This is called by the allocator when it reserves space. If this is a
3756 * reservation and the block group has become read only we cannot make the
3757 * reservation and return -EAGAIN, otherwise this function always succeeds.
3758 */
3759int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3760 u64 ram_bytes, u64 num_bytes, int delalloc,
3761 bool force_wrong_size_class)
3762{
3763 struct btrfs_space_info *space_info = cache->space_info;
3764 enum btrfs_block_group_size_class size_class;
3765 int ret = 0;
3766
3767 spin_lock(&space_info->lock);
3768 spin_lock(&cache->lock);
3769 if (cache->ro) {
3770 ret = -EAGAIN;
3771 goto out;
3772 }
3773
3774 if (btrfs_block_group_should_use_size_class(cache)) {
3775 size_class = btrfs_calc_block_group_size_class(num_bytes);
3776 ret = btrfs_use_block_group_size_class(cache, size_class, force_wrong_size_class);
3777 if (ret)
3778 goto out;
3779 }
3780 cache->reserved += num_bytes;
3781 space_info->bytes_reserved += num_bytes;
3782 trace_btrfs_space_reservation(cache->fs_info, "space_info",
3783 space_info->flags, num_bytes, 1);
3784 btrfs_space_info_update_bytes_may_use(cache->fs_info,
3785 space_info, -ram_bytes);
3786 if (delalloc)
3787 cache->delalloc_bytes += num_bytes;
3788
3789 /*
3790 * Compression can use less space than we reserved, so wake tickets if
3791 * that happens.
3792 */
3793 if (num_bytes < ram_bytes)
3794 btrfs_try_granting_tickets(cache->fs_info, space_info);
3795out:
3796 spin_unlock(&cache->lock);
3797 spin_unlock(&space_info->lock);
3798 return ret;
3799}
3800
3801/*
3802 * Update the block_group and space info counters.
3803 *
3804 * @cache: The cache we are manipulating
3805 * @num_bytes: The number of bytes in question
3806 * @delalloc: The blocks are allocated for the delalloc write
3807 *
3808 * This is called by somebody who is freeing space that was never actually used
3809 * on disk. For example if you reserve some space for a new leaf in transaction
3810 * A and before transaction A commits you free that leaf, you call this with
3811 * reserve set to 0 in order to clear the reservation.
3812 */
3813void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
3814 u64 num_bytes, int delalloc)
3815{
3816 struct btrfs_space_info *space_info = cache->space_info;
3817
3818 spin_lock(&space_info->lock);
3819 spin_lock(&cache->lock);
3820 if (cache->ro)
3821 space_info->bytes_readonly += num_bytes;
3822 else if (btrfs_is_zoned(cache->fs_info))
3823 space_info->bytes_zone_unusable += num_bytes;
3824 cache->reserved -= num_bytes;
3825 space_info->bytes_reserved -= num_bytes;
3826 space_info->max_extent_size = 0;
3827
3828 if (delalloc)
3829 cache->delalloc_bytes -= num_bytes;
3830 spin_unlock(&cache->lock);
3831
3832 btrfs_try_granting_tickets(cache->fs_info, space_info);
3833 spin_unlock(&space_info->lock);
3834}
3835
3836static void force_metadata_allocation(struct btrfs_fs_info *info)
3837{
3838 struct list_head *head = &info->space_info;
3839 struct btrfs_space_info *found;
3840
3841 list_for_each_entry(found, head, list) {
3842 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3843 found->force_alloc = CHUNK_ALLOC_FORCE;
3844 }
3845}
3846
3847static int should_alloc_chunk(const struct btrfs_fs_info *fs_info,
3848 const struct btrfs_space_info *sinfo, int force)
3849{
3850 u64 bytes_used = btrfs_space_info_used(sinfo, false);
3851 u64 thresh;
3852
3853 if (force == CHUNK_ALLOC_FORCE)
3854 return 1;
3855
3856 /*
3857 * in limited mode, we want to have some free space up to
3858 * about 1% of the FS size.
3859 */
3860 if (force == CHUNK_ALLOC_LIMITED) {
3861 thresh = btrfs_super_total_bytes(fs_info->super_copy);
3862 thresh = max_t(u64, SZ_64M, mult_perc(thresh, 1));
3863
3864 if (sinfo->total_bytes - bytes_used < thresh)
3865 return 1;
3866 }
3867
3868 if (bytes_used + SZ_2M < mult_perc(sinfo->total_bytes, 80))
3869 return 0;
3870 return 1;
3871}
3872
3873int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3874{
3875 u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3876
3877 return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3878}
3879
3880static struct btrfs_block_group *do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags)
3881{
3882 struct btrfs_block_group *bg;
3883 int ret;
3884
3885 /*
3886 * Check if we have enough space in the system space info because we
3887 * will need to update device items in the chunk btree and insert a new
3888 * chunk item in the chunk btree as well. This will allocate a new
3889 * system block group if needed.
3890 */
3891 check_system_chunk(trans, flags);
3892
3893 bg = btrfs_create_chunk(trans, flags);
3894 if (IS_ERR(bg)) {
3895 ret = PTR_ERR(bg);
3896 goto out;
3897 }
3898
3899 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3900 /*
3901 * Normally we are not expected to fail with -ENOSPC here, since we have
3902 * previously reserved space in the system space_info and allocated one
3903 * new system chunk if necessary. However there are three exceptions:
3904 *
3905 * 1) We may have enough free space in the system space_info but all the
3906 * existing system block groups have a profile which can not be used
3907 * for extent allocation.
3908 *
3909 * This happens when mounting in degraded mode. For example we have a
3910 * RAID1 filesystem with 2 devices, lose one device and mount the fs
3911 * using the other device in degraded mode. If we then allocate a chunk,
3912 * we may have enough free space in the existing system space_info, but
3913 * none of the block groups can be used for extent allocation since they
3914 * have a RAID1 profile, and because we are in degraded mode with a
3915 * single device, we are forced to allocate a new system chunk with a
3916 * SINGLE profile. Making check_system_chunk() iterate over all system
3917 * block groups and check if they have a usable profile and enough space
3918 * can be slow on very large filesystems, so we tolerate the -ENOSPC and
3919 * try again after forcing allocation of a new system chunk. Like this
3920 * we avoid paying the cost of that search in normal circumstances, when
3921 * we were not mounted in degraded mode;
3922 *
3923 * 2) We had enough free space info the system space_info, and one suitable
3924 * block group to allocate from when we called check_system_chunk()
3925 * above. However right after we called it, the only system block group
3926 * with enough free space got turned into RO mode by a running scrub,
3927 * and in this case we have to allocate a new one and retry. We only
3928 * need do this allocate and retry once, since we have a transaction
3929 * handle and scrub uses the commit root to search for block groups;
3930 *
3931 * 3) We had one system block group with enough free space when we called
3932 * check_system_chunk(), but after that, right before we tried to
3933 * allocate the last extent buffer we needed, a discard operation came
3934 * in and it temporarily removed the last free space entry from the
3935 * block group (discard removes a free space entry, discards it, and
3936 * then adds back the entry to the block group cache).
3937 */
3938 if (ret == -ENOSPC) {
3939 const u64 sys_flags = btrfs_system_alloc_profile(trans->fs_info);
3940 struct btrfs_block_group *sys_bg;
3941
3942 sys_bg = btrfs_create_chunk(trans, sys_flags);
3943 if (IS_ERR(sys_bg)) {
3944 ret = PTR_ERR(sys_bg);
3945 btrfs_abort_transaction(trans, ret);
3946 goto out;
3947 }
3948
3949 ret = btrfs_chunk_alloc_add_chunk_item(trans, sys_bg);
3950 if (ret) {
3951 btrfs_abort_transaction(trans, ret);
3952 goto out;
3953 }
3954
3955 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3956 if (ret) {
3957 btrfs_abort_transaction(trans, ret);
3958 goto out;
3959 }
3960 } else if (ret) {
3961 btrfs_abort_transaction(trans, ret);
3962 goto out;
3963 }
3964out:
3965 btrfs_trans_release_chunk_metadata(trans);
3966
3967 if (ret)
3968 return ERR_PTR(ret);
3969
3970 btrfs_get_block_group(bg);
3971 return bg;
3972}
3973
3974/*
3975 * Chunk allocation is done in 2 phases:
3976 *
3977 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
3978 * the chunk, the chunk mapping, create its block group and add the items
3979 * that belong in the chunk btree to it - more specifically, we need to
3980 * update device items in the chunk btree and add a new chunk item to it.
3981 *
3982 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
3983 * group item to the extent btree and the device extent items to the devices
3984 * btree.
3985 *
3986 * This is done to prevent deadlocks. For example when COWing a node from the
3987 * extent btree we are holding a write lock on the node's parent and if we
3988 * trigger chunk allocation and attempted to insert the new block group item
3989 * in the extent btree right way, we could deadlock because the path for the
3990 * insertion can include that parent node. At first glance it seems impossible
3991 * to trigger chunk allocation after starting a transaction since tasks should
3992 * reserve enough transaction units (metadata space), however while that is true
3993 * most of the time, chunk allocation may still be triggered for several reasons:
3994 *
3995 * 1) When reserving metadata, we check if there is enough free space in the
3996 * metadata space_info and therefore don't trigger allocation of a new chunk.
3997 * However later when the task actually tries to COW an extent buffer from
3998 * the extent btree or from the device btree for example, it is forced to
3999 * allocate a new block group (chunk) because the only one that had enough
4000 * free space was just turned to RO mode by a running scrub for example (or
4001 * device replace, block group reclaim thread, etc), so we can not use it
4002 * for allocating an extent and end up being forced to allocate a new one;
4003 *
4004 * 2) Because we only check that the metadata space_info has enough free bytes,
4005 * we end up not allocating a new metadata chunk in that case. However if
4006 * the filesystem was mounted in degraded mode, none of the existing block
4007 * groups might be suitable for extent allocation due to their incompatible
4008 * profile (for e.g. mounting a 2 devices filesystem, where all block groups
4009 * use a RAID1 profile, in degraded mode using a single device). In this case
4010 * when the task attempts to COW some extent buffer of the extent btree for
4011 * example, it will trigger allocation of a new metadata block group with a
4012 * suitable profile (SINGLE profile in the example of the degraded mount of
4013 * the RAID1 filesystem);
4014 *
4015 * 3) The task has reserved enough transaction units / metadata space, but when
4016 * it attempts to COW an extent buffer from the extent or device btree for
4017 * example, it does not find any free extent in any metadata block group,
4018 * therefore forced to try to allocate a new metadata block group.
4019 * This is because some other task allocated all available extents in the
4020 * meanwhile - this typically happens with tasks that don't reserve space
4021 * properly, either intentionally or as a bug. One example where this is
4022 * done intentionally is fsync, as it does not reserve any transaction units
4023 * and ends up allocating a variable number of metadata extents for log
4024 * tree extent buffers;
4025 *
4026 * 4) The task has reserved enough transaction units / metadata space, but right
4027 * before it tries to allocate the last extent buffer it needs, a discard
4028 * operation comes in and, temporarily, removes the last free space entry from
4029 * the only metadata block group that had free space (discard starts by
4030 * removing a free space entry from a block group, then does the discard
4031 * operation and, once it's done, it adds back the free space entry to the
4032 * block group).
4033 *
4034 * We also need this 2 phases setup when adding a device to a filesystem with
4035 * a seed device - we must create new metadata and system chunks without adding
4036 * any of the block group items to the chunk, extent and device btrees. If we
4037 * did not do it this way, we would get ENOSPC when attempting to update those
4038 * btrees, since all the chunks from the seed device are read-only.
4039 *
4040 * Phase 1 does the updates and insertions to the chunk btree because if we had
4041 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
4042 * parallel, we risk having too many system chunks allocated by many tasks if
4043 * many tasks reach phase 1 without the previous ones completing phase 2. In the
4044 * extreme case this leads to exhaustion of the system chunk array in the
4045 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
4046 * and with RAID filesystems (so we have more device items in the chunk btree).
4047 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
4048 * the system chunk array due to concurrent allocations") provides more details.
4049 *
4050 * Allocation of system chunks does not happen through this function. A task that
4051 * needs to update the chunk btree (the only btree that uses system chunks), must
4052 * preallocate chunk space by calling either check_system_chunk() or
4053 * btrfs_reserve_chunk_metadata() - the former is used when allocating a data or
4054 * metadata chunk or when removing a chunk, while the later is used before doing
4055 * a modification to the chunk btree - use cases for the later are adding,
4056 * removing and resizing a device as well as relocation of a system chunk.
4057 * See the comment below for more details.
4058 *
4059 * The reservation of system space, done through check_system_chunk(), as well
4060 * as all the updates and insertions into the chunk btree must be done while
4061 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
4062 * an extent buffer from the chunks btree we never trigger allocation of a new
4063 * system chunk, which would result in a deadlock (trying to lock twice an
4064 * extent buffer of the chunk btree, first time before triggering the chunk
4065 * allocation and the second time during chunk allocation while attempting to
4066 * update the chunks btree). The system chunk array is also updated while holding
4067 * that mutex. The same logic applies to removing chunks - we must reserve system
4068 * space, update the chunk btree and the system chunk array in the superblock
4069 * while holding fs_info->chunk_mutex.
4070 *
4071 * This function, btrfs_chunk_alloc(), belongs to phase 1.
4072 *
4073 * If @force is CHUNK_ALLOC_FORCE:
4074 * - return 1 if it successfully allocates a chunk,
4075 * - return errors including -ENOSPC otherwise.
4076 * If @force is NOT CHUNK_ALLOC_FORCE:
4077 * - return 0 if it doesn't need to allocate a new chunk,
4078 * - return 1 if it successfully allocates a chunk,
4079 * - return errors including -ENOSPC otherwise.
4080 */
4081int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4082 enum btrfs_chunk_alloc_enum force)
4083{
4084 struct btrfs_fs_info *fs_info = trans->fs_info;
4085 struct btrfs_space_info *space_info;
4086 struct btrfs_block_group *ret_bg;
4087 bool wait_for_alloc = false;
4088 bool should_alloc = false;
4089 bool from_extent_allocation = false;
4090 int ret = 0;
4091
4092 if (force == CHUNK_ALLOC_FORCE_FOR_EXTENT) {
4093 from_extent_allocation = true;
4094 force = CHUNK_ALLOC_FORCE;
4095 }
4096
4097 /* Don't re-enter if we're already allocating a chunk */
4098 if (trans->allocating_chunk)
4099 return -ENOSPC;
4100 /*
4101 * Allocation of system chunks can not happen through this path, as we
4102 * could end up in a deadlock if we are allocating a data or metadata
4103 * chunk and there is another task modifying the chunk btree.
4104 *
4105 * This is because while we are holding the chunk mutex, we will attempt
4106 * to add the new chunk item to the chunk btree or update an existing
4107 * device item in the chunk btree, while the other task that is modifying
4108 * the chunk btree is attempting to COW an extent buffer while holding a
4109 * lock on it and on its parent - if the COW operation triggers a system
4110 * chunk allocation, then we can deadlock because we are holding the
4111 * chunk mutex and we may need to access that extent buffer or its parent
4112 * in order to add the chunk item or update a device item.
4113 *
4114 * Tasks that want to modify the chunk tree should reserve system space
4115 * before updating the chunk btree, by calling either
4116 * btrfs_reserve_chunk_metadata() or check_system_chunk().
4117 * It's possible that after a task reserves the space, it still ends up
4118 * here - this happens in the cases described above at do_chunk_alloc().
4119 * The task will have to either retry or fail.
4120 */
4121 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4122 return -ENOSPC;
4123
4124 space_info = btrfs_find_space_info(fs_info, flags);
4125 ASSERT(space_info);
4126
4127 do {
4128 spin_lock(&space_info->lock);
4129 if (force < space_info->force_alloc)
4130 force = space_info->force_alloc;
4131 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4132 if (space_info->full) {
4133 /* No more free physical space */
4134 if (should_alloc)
4135 ret = -ENOSPC;
4136 else
4137 ret = 0;
4138 spin_unlock(&space_info->lock);
4139 return ret;
4140 } else if (!should_alloc) {
4141 spin_unlock(&space_info->lock);
4142 return 0;
4143 } else if (space_info->chunk_alloc) {
4144 /*
4145 * Someone is already allocating, so we need to block
4146 * until this someone is finished and then loop to
4147 * recheck if we should continue with our allocation
4148 * attempt.
4149 */
4150 wait_for_alloc = true;
4151 force = CHUNK_ALLOC_NO_FORCE;
4152 spin_unlock(&space_info->lock);
4153 mutex_lock(&fs_info->chunk_mutex);
4154 mutex_unlock(&fs_info->chunk_mutex);
4155 } else {
4156 /* Proceed with allocation */
4157 space_info->chunk_alloc = 1;
4158 wait_for_alloc = false;
4159 spin_unlock(&space_info->lock);
4160 }
4161
4162 cond_resched();
4163 } while (wait_for_alloc);
4164
4165 mutex_lock(&fs_info->chunk_mutex);
4166 trans->allocating_chunk = true;
4167
4168 /*
4169 * If we have mixed data/metadata chunks we want to make sure we keep
4170 * allocating mixed chunks instead of individual chunks.
4171 */
4172 if (btrfs_mixed_space_info(space_info))
4173 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4174
4175 /*
4176 * if we're doing a data chunk, go ahead and make sure that
4177 * we keep a reasonable number of metadata chunks allocated in the
4178 * FS as well.
4179 */
4180 if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4181 fs_info->data_chunk_allocations++;
4182 if (!(fs_info->data_chunk_allocations %
4183 fs_info->metadata_ratio))
4184 force_metadata_allocation(fs_info);
4185 }
4186
4187 ret_bg = do_chunk_alloc(trans, flags);
4188 trans->allocating_chunk = false;
4189
4190 if (IS_ERR(ret_bg)) {
4191 ret = PTR_ERR(ret_bg);
4192 } else if (from_extent_allocation && (flags & BTRFS_BLOCK_GROUP_DATA)) {
4193 /*
4194 * New block group is likely to be used soon. Try to activate
4195 * it now. Failure is OK for now.
4196 */
4197 btrfs_zone_activate(ret_bg);
4198 }
4199
4200 if (!ret)
4201 btrfs_put_block_group(ret_bg);
4202
4203 spin_lock(&space_info->lock);
4204 if (ret < 0) {
4205 if (ret == -ENOSPC)
4206 space_info->full = 1;
4207 else
4208 goto out;
4209 } else {
4210 ret = 1;
4211 space_info->max_extent_size = 0;
4212 }
4213
4214 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4215out:
4216 space_info->chunk_alloc = 0;
4217 spin_unlock(&space_info->lock);
4218 mutex_unlock(&fs_info->chunk_mutex);
4219
4220 return ret;
4221}
4222
4223static u64 get_profile_num_devs(const struct btrfs_fs_info *fs_info, u64 type)
4224{
4225 u64 num_dev;
4226
4227 num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
4228 if (!num_dev)
4229 num_dev = fs_info->fs_devices->rw_devices;
4230
4231 return num_dev;
4232}
4233
4234static void reserve_chunk_space(struct btrfs_trans_handle *trans,
4235 u64 bytes,
4236 u64 type)
4237{
4238 struct btrfs_fs_info *fs_info = trans->fs_info;
4239 struct btrfs_space_info *info;
4240 u64 left;
4241 int ret = 0;
4242
4243 /*
4244 * Needed because we can end up allocating a system chunk and for an
4245 * atomic and race free space reservation in the chunk block reserve.
4246 */
4247 lockdep_assert_held(&fs_info->chunk_mutex);
4248
4249 info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4250 spin_lock(&info->lock);
4251 left = info->total_bytes - btrfs_space_info_used(info, true);
4252 spin_unlock(&info->lock);
4253
4254 if (left < bytes && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4255 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4256 left, bytes, type);
4257 btrfs_dump_space_info(fs_info, info, 0, 0);
4258 }
4259
4260 if (left < bytes) {
4261 u64 flags = btrfs_system_alloc_profile(fs_info);
4262 struct btrfs_block_group *bg;
4263
4264 /*
4265 * Ignore failure to create system chunk. We might end up not
4266 * needing it, as we might not need to COW all nodes/leafs from
4267 * the paths we visit in the chunk tree (they were already COWed
4268 * or created in the current transaction for example).
4269 */
4270 bg = btrfs_create_chunk(trans, flags);
4271 if (IS_ERR(bg)) {
4272 ret = PTR_ERR(bg);
4273 } else {
4274 /*
4275 * We have a new chunk. We also need to activate it for
4276 * zoned filesystem.
4277 */
4278 ret = btrfs_zoned_activate_one_bg(fs_info, info, true);
4279 if (ret < 0)
4280 return;
4281
4282 /*
4283 * If we fail to add the chunk item here, we end up
4284 * trying again at phase 2 of chunk allocation, at
4285 * btrfs_create_pending_block_groups(). So ignore
4286 * any error here. An ENOSPC here could happen, due to
4287 * the cases described at do_chunk_alloc() - the system
4288 * block group we just created was just turned into RO
4289 * mode by a scrub for example, or a running discard
4290 * temporarily removed its free space entries, etc.
4291 */
4292 btrfs_chunk_alloc_add_chunk_item(trans, bg);
4293 }
4294 }
4295
4296 if (!ret) {
4297 ret = btrfs_block_rsv_add(fs_info,
4298 &fs_info->chunk_block_rsv,
4299 bytes, BTRFS_RESERVE_NO_FLUSH);
4300 if (!ret)
4301 trans->chunk_bytes_reserved += bytes;
4302 }
4303}
4304
4305/*
4306 * Reserve space in the system space for allocating or removing a chunk.
4307 * The caller must be holding fs_info->chunk_mutex.
4308 */
4309void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4310{
4311 struct btrfs_fs_info *fs_info = trans->fs_info;
4312 const u64 num_devs = get_profile_num_devs(fs_info, type);
4313 u64 bytes;
4314
4315 /* num_devs device items to update and 1 chunk item to add or remove. */
4316 bytes = btrfs_calc_metadata_size(fs_info, num_devs) +
4317 btrfs_calc_insert_metadata_size(fs_info, 1);
4318
4319 reserve_chunk_space(trans, bytes, type);
4320}
4321
4322/*
4323 * Reserve space in the system space, if needed, for doing a modification to the
4324 * chunk btree.
4325 *
4326 * @trans: A transaction handle.
4327 * @is_item_insertion: Indicate if the modification is for inserting a new item
4328 * in the chunk btree or if it's for the deletion or update
4329 * of an existing item.
4330 *
4331 * This is used in a context where we need to update the chunk btree outside
4332 * block group allocation and removal, to avoid a deadlock with a concurrent
4333 * task that is allocating a metadata or data block group and therefore needs to
4334 * update the chunk btree while holding the chunk mutex. After the update to the
4335 * chunk btree is done, btrfs_trans_release_chunk_metadata() should be called.
4336 *
4337 */
4338void btrfs_reserve_chunk_metadata(struct btrfs_trans_handle *trans,
4339 bool is_item_insertion)
4340{
4341 struct btrfs_fs_info *fs_info = trans->fs_info;
4342 u64 bytes;
4343
4344 if (is_item_insertion)
4345 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
4346 else
4347 bytes = btrfs_calc_metadata_size(fs_info, 1);
4348
4349 mutex_lock(&fs_info->chunk_mutex);
4350 reserve_chunk_space(trans, bytes, BTRFS_BLOCK_GROUP_SYSTEM);
4351 mutex_unlock(&fs_info->chunk_mutex);
4352}
4353
4354void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
4355{
4356 struct btrfs_block_group *block_group;
4357
4358 block_group = btrfs_lookup_first_block_group(info, 0);
4359 while (block_group) {
4360 btrfs_wait_block_group_cache_done(block_group);
4361 spin_lock(&block_group->lock);
4362 if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF,
4363 &block_group->runtime_flags)) {
4364 struct btrfs_inode *inode = block_group->inode;
4365
4366 block_group->inode = NULL;
4367 spin_unlock(&block_group->lock);
4368
4369 ASSERT(block_group->io_ctl.inode == NULL);
4370 iput(&inode->vfs_inode);
4371 } else {
4372 spin_unlock(&block_group->lock);
4373 }
4374 block_group = btrfs_next_block_group(block_group);
4375 }
4376}
4377
4378/*
4379 * Must be called only after stopping all workers, since we could have block
4380 * group caching kthreads running, and therefore they could race with us if we
4381 * freed the block groups before stopping them.
4382 */
4383int btrfs_free_block_groups(struct btrfs_fs_info *info)
4384{
4385 struct btrfs_block_group *block_group;
4386 struct btrfs_space_info *space_info;
4387 struct btrfs_caching_control *caching_ctl;
4388 struct rb_node *n;
4389
4390 if (btrfs_is_zoned(info)) {
4391 if (info->active_meta_bg) {
4392 btrfs_put_block_group(info->active_meta_bg);
4393 info->active_meta_bg = NULL;
4394 }
4395 if (info->active_system_bg) {
4396 btrfs_put_block_group(info->active_system_bg);
4397 info->active_system_bg = NULL;
4398 }
4399 }
4400
4401 write_lock(&info->block_group_cache_lock);
4402 while (!list_empty(&info->caching_block_groups)) {
4403 caching_ctl = list_entry(info->caching_block_groups.next,
4404 struct btrfs_caching_control, list);
4405 list_del(&caching_ctl->list);
4406 btrfs_put_caching_control(caching_ctl);
4407 }
4408 write_unlock(&info->block_group_cache_lock);
4409
4410 spin_lock(&info->unused_bgs_lock);
4411 while (!list_empty(&info->unused_bgs)) {
4412 block_group = list_first_entry(&info->unused_bgs,
4413 struct btrfs_block_group,
4414 bg_list);
4415 list_del_init(&block_group->bg_list);
4416 btrfs_put_block_group(block_group);
4417 }
4418
4419 while (!list_empty(&info->reclaim_bgs)) {
4420 block_group = list_first_entry(&info->reclaim_bgs,
4421 struct btrfs_block_group,
4422 bg_list);
4423 list_del_init(&block_group->bg_list);
4424 btrfs_put_block_group(block_group);
4425 }
4426 spin_unlock(&info->unused_bgs_lock);
4427
4428 spin_lock(&info->zone_active_bgs_lock);
4429 while (!list_empty(&info->zone_active_bgs)) {
4430 block_group = list_first_entry(&info->zone_active_bgs,
4431 struct btrfs_block_group,
4432 active_bg_list);
4433 list_del_init(&block_group->active_bg_list);
4434 btrfs_put_block_group(block_group);
4435 }
4436 spin_unlock(&info->zone_active_bgs_lock);
4437
4438 write_lock(&info->block_group_cache_lock);
4439 while ((n = rb_last(&info->block_group_cache_tree.rb_root)) != NULL) {
4440 block_group = rb_entry(n, struct btrfs_block_group,
4441 cache_node);
4442 rb_erase_cached(&block_group->cache_node,
4443 &info->block_group_cache_tree);
4444 RB_CLEAR_NODE(&block_group->cache_node);
4445 write_unlock(&info->block_group_cache_lock);
4446
4447 down_write(&block_group->space_info->groups_sem);
4448 list_del(&block_group->list);
4449 up_write(&block_group->space_info->groups_sem);
4450
4451 /*
4452 * We haven't cached this block group, which means we could
4453 * possibly have excluded extents on this block group.
4454 */
4455 if (block_group->cached == BTRFS_CACHE_NO ||
4456 block_group->cached == BTRFS_CACHE_ERROR)
4457 btrfs_free_excluded_extents(block_group);
4458
4459 btrfs_remove_free_space_cache(block_group);
4460 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
4461 ASSERT(list_empty(&block_group->dirty_list));
4462 ASSERT(list_empty(&block_group->io_list));
4463 ASSERT(list_empty(&block_group->bg_list));
4464 ASSERT(refcount_read(&block_group->refs) == 1);
4465 ASSERT(block_group->swap_extents == 0);
4466 btrfs_put_block_group(block_group);
4467
4468 write_lock(&info->block_group_cache_lock);
4469 }
4470 write_unlock(&info->block_group_cache_lock);
4471
4472 btrfs_release_global_block_rsv(info);
4473
4474 while (!list_empty(&info->space_info)) {
4475 space_info = list_entry(info->space_info.next,
4476 struct btrfs_space_info,
4477 list);
4478
4479 /*
4480 * Do not hide this behind enospc_debug, this is actually
4481 * important and indicates a real bug if this happens.
4482 */
4483 if (WARN_ON(space_info->bytes_pinned > 0 ||
4484 space_info->bytes_may_use > 0))
4485 btrfs_dump_space_info(info, space_info, 0, 0);
4486
4487 /*
4488 * If there was a failure to cleanup a log tree, very likely due
4489 * to an IO failure on a writeback attempt of one or more of its
4490 * extent buffers, we could not do proper (and cheap) unaccounting
4491 * of their reserved space, so don't warn on bytes_reserved > 0 in
4492 * that case.
4493 */
4494 if (!(space_info->flags & BTRFS_BLOCK_GROUP_METADATA) ||
4495 !BTRFS_FS_LOG_CLEANUP_ERROR(info)) {
4496 if (WARN_ON(space_info->bytes_reserved > 0))
4497 btrfs_dump_space_info(info, space_info, 0, 0);
4498 }
4499
4500 WARN_ON(space_info->reclaim_size > 0);
4501 list_del(&space_info->list);
4502 btrfs_sysfs_remove_space_info(space_info);
4503 }
4504 return 0;
4505}
4506
4507void btrfs_freeze_block_group(struct btrfs_block_group *cache)
4508{
4509 atomic_inc(&cache->frozen);
4510}
4511
4512void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
4513{
4514 struct btrfs_fs_info *fs_info = block_group->fs_info;
4515 bool cleanup;
4516
4517 spin_lock(&block_group->lock);
4518 cleanup = (atomic_dec_and_test(&block_group->frozen) &&
4519 test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags));
4520 spin_unlock(&block_group->lock);
4521
4522 if (cleanup) {
4523 struct btrfs_chunk_map *map;
4524
4525 map = btrfs_find_chunk_map(fs_info, block_group->start, 1);
4526 /* Logic error, can't happen. */
4527 ASSERT(map);
4528
4529 btrfs_remove_chunk_map(fs_info, map);
4530
4531 /* Once for our lookup reference. */
4532 btrfs_free_chunk_map(map);
4533
4534 /*
4535 * We may have left one free space entry and other possible
4536 * tasks trimming this block group have left 1 entry each one.
4537 * Free them if any.
4538 */
4539 btrfs_remove_free_space_cache(block_group);
4540 }
4541}
4542
4543bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
4544{
4545 bool ret = true;
4546
4547 spin_lock(&bg->lock);
4548 if (bg->ro)
4549 ret = false;
4550 else
4551 bg->swap_extents++;
4552 spin_unlock(&bg->lock);
4553
4554 return ret;
4555}
4556
4557void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
4558{
4559 spin_lock(&bg->lock);
4560 ASSERT(!bg->ro);
4561 ASSERT(bg->swap_extents >= amount);
4562 bg->swap_extents -= amount;
4563 spin_unlock(&bg->lock);
4564}
4565
4566enum btrfs_block_group_size_class btrfs_calc_block_group_size_class(u64 size)
4567{
4568 if (size <= SZ_128K)
4569 return BTRFS_BG_SZ_SMALL;
4570 if (size <= SZ_8M)
4571 return BTRFS_BG_SZ_MEDIUM;
4572 return BTRFS_BG_SZ_LARGE;
4573}
4574
4575/*
4576 * Handle a block group allocating an extent in a size class
4577 *
4578 * @bg: The block group we allocated in.
4579 * @size_class: The size class of the allocation.
4580 * @force_wrong_size_class: Whether we are desperate enough to allow
4581 * mismatched size classes.
4582 *
4583 * Returns: 0 if the size class was valid for this block_group, -EAGAIN in the
4584 * case of a race that leads to the wrong size class without
4585 * force_wrong_size_class set.
4586 *
4587 * find_free_extent will skip block groups with a mismatched size class until
4588 * it really needs to avoid ENOSPC. In that case it will set
4589 * force_wrong_size_class. However, if a block group is newly allocated and
4590 * doesn't yet have a size class, then it is possible for two allocations of
4591 * different sizes to race and both try to use it. The loser is caught here and
4592 * has to retry.
4593 */
4594int btrfs_use_block_group_size_class(struct btrfs_block_group *bg,
4595 enum btrfs_block_group_size_class size_class,
4596 bool force_wrong_size_class)
4597{
4598 ASSERT(size_class != BTRFS_BG_SZ_NONE);
4599
4600 /* The new allocation is in the right size class, do nothing */
4601 if (bg->size_class == size_class)
4602 return 0;
4603 /*
4604 * The new allocation is in a mismatched size class.
4605 * This means one of two things:
4606 *
4607 * 1. Two tasks in find_free_extent for different size_classes raced
4608 * and hit the same empty block_group. Make the loser try again.
4609 * 2. A call to find_free_extent got desperate enough to set
4610 * 'force_wrong_slab'. Don't change the size_class, but allow the
4611 * allocation.
4612 */
4613 if (bg->size_class != BTRFS_BG_SZ_NONE) {
4614 if (force_wrong_size_class)
4615 return 0;
4616 return -EAGAIN;
4617 }
4618 /*
4619 * The happy new block group case: the new allocation is the first
4620 * one in the block_group so we set size_class.
4621 */
4622 bg->size_class = size_class;
4623
4624 return 0;
4625}
4626
4627bool btrfs_block_group_should_use_size_class(const struct btrfs_block_group *bg)
4628{
4629 if (btrfs_is_zoned(bg->fs_info))
4630 return false;
4631 if (!btrfs_is_block_group_data_only(bg))
4632 return false;
4633 return true;
4634}