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
btrfs_should_fragment_free_space(struct btrfs_block_group * block_group)26 int btrfs_should_fragment_free_space(struct btrfs_block_group *block_group)
27 {
28 struct btrfs_fs_info *fs_info = block_group->fs_info;
29
30 return (btrfs_test_opt(fs_info, FRAGMENT_METADATA) &&
31 block_group->flags & BTRFS_BLOCK_GROUP_METADATA) ||
32 (btrfs_test_opt(fs_info, FRAGMENT_DATA) &&
33 block_group->flags & BTRFS_BLOCK_GROUP_DATA);
34 }
35 #endif
36
37 /*
38 * Return target flags in extended format or 0 if restripe for this chunk_type
39 * is not in progress
40 *
41 * Should be called with balance_lock held
42 */
get_restripe_target(struct btrfs_fs_info * fs_info,u64 flags)43 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
44 {
45 struct btrfs_balance_control *bctl = fs_info->balance_ctl;
46 u64 target = 0;
47
48 if (!bctl)
49 return 0;
50
51 if (flags & BTRFS_BLOCK_GROUP_DATA &&
52 bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
53 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
54 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
55 bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
56 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
57 } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
58 bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
59 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
60 }
61
62 return target;
63 }
64
65 /*
66 * @flags: available profiles in extended format (see ctree.h)
67 *
68 * Return reduced profile in chunk format. If profile changing is in progress
69 * (either running or paused) picks the target profile (if it's already
70 * available), otherwise falls back to plain reducing.
71 */
btrfs_reduce_alloc_profile(struct btrfs_fs_info * fs_info,u64 flags)72 static 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
btrfs_get_alloc_profile(struct btrfs_fs_info * fs_info,u64 orig_flags)121 u64 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
btrfs_get_block_group(struct btrfs_block_group * cache)141 void btrfs_get_block_group(struct btrfs_block_group *cache)
142 {
143 refcount_inc(&cache->refs);
144 }
145
btrfs_put_block_group(struct btrfs_block_group * cache)146 void 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 kfree(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 */
btrfs_add_block_group_cache(struct btrfs_fs_info * info,struct btrfs_block_group * block_group)179 static 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 */
block_group_cache_tree_search(struct btrfs_fs_info * info,u64 bytenr,int contains)219 static 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 */
btrfs_lookup_first_block_group(struct btrfs_fs_info * info,u64 bytenr)259 struct 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 */
btrfs_lookup_block_group(struct btrfs_fs_info * info,u64 bytenr)268 struct 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
btrfs_next_block_group(struct btrfs_block_group * cache)274 struct 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 */
btrfs_inc_nocow_writers(struct btrfs_fs_info * fs_info,u64 bytenr)316 struct 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 */
btrfs_dec_nocow_writers(struct btrfs_block_group * bg)353 void 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
btrfs_wait_nocow_writers(struct btrfs_block_group * bg)362 void btrfs_wait_nocow_writers(struct btrfs_block_group *bg)
363 {
364 wait_var_event(&bg->nocow_writers, !atomic_read(&bg->nocow_writers));
365 }
366
btrfs_dec_block_group_reservations(struct btrfs_fs_info * fs_info,const u64 start)367 void 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
btrfs_wait_block_group_reservations(struct btrfs_block_group * bg)379 void 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
btrfs_get_caching_control(struct btrfs_block_group * cache)404 struct 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
btrfs_put_caching_control(struct btrfs_caching_control * ctl)421 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 */
btrfs_wait_block_group_cache_progress(struct btrfs_block_group * cache,u64 num_bytes)440 void 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
btrfs_caching_ctl_wait_done(struct btrfs_block_group * cache,struct btrfs_caching_control * caching_ctl)465 static 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
btrfs_wait_block_group_cache_done(struct btrfs_block_group * cache)472 static 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
fragment_free_space(struct btrfs_block_group * block_group)486 static 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 */
btrfs_add_new_free_space(struct btrfs_block_group * block_group,u64 start,u64 end,u64 * total_added_ret)519 int 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 */
sample_block_group_extent_item(struct btrfs_caching_control * caching_ctl,struct btrfs_block_group * block_group,int index,int max_index,struct btrfs_key * found_key)580 static 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 */
load_block_group_size_class(struct btrfs_caching_control * caching_ctl,struct btrfs_block_group * block_group)669 static 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 }
698 out:
699 return ret;
700 }
701
load_extent_tree_free(struct btrfs_caching_control * caching_ctl)702 static 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
746 next:
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);
833 out:
834 btrfs_free_path(path);
835 return ret;
836 }
837
btrfs_free_excluded_extents(const struct btrfs_block_group * bg)838 static 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
caching_thread(struct btrfs_work * work)844 static 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);
888 done:
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
btrfs_cache_block_group(struct btrfs_block_group * cache,bool wait)918 int 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, 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);
963 out:
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
clear_avail_alloc_bits(struct btrfs_fs_info * fs_info,u64 flags)972 static 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 */
clear_incompat_bg_bits(struct btrfs_fs_info * fs_info,u64 flags)995 static 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
remove_block_group_item(struct btrfs_trans_handle * trans,struct btrfs_path * path,struct btrfs_block_group * block_group)1025 static int remove_block_group_item(struct btrfs_trans_handle *trans,
1026 struct btrfs_path *path,
1027 struct btrfs_block_group *block_group)
1028 {
1029 struct btrfs_fs_info *fs_info = trans->fs_info;
1030 struct btrfs_root *root;
1031 struct btrfs_key key;
1032 int ret;
1033
1034 root = btrfs_block_group_root(fs_info);
1035 key.objectid = block_group->start;
1036 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
1037 key.offset = block_group->length;
1038
1039 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1040 if (ret > 0)
1041 ret = -ENOENT;
1042 if (ret < 0)
1043 return ret;
1044
1045 ret = btrfs_del_item(trans, root, path);
1046 return ret;
1047 }
1048
btrfs_remove_block_group(struct btrfs_trans_handle * trans,u64 group_start,struct extent_map * em)1049 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
1050 u64 group_start, struct extent_map *em)
1051 {
1052 struct btrfs_fs_info *fs_info = trans->fs_info;
1053 struct btrfs_path *path;
1054 struct btrfs_block_group *block_group;
1055 struct btrfs_free_cluster *cluster;
1056 struct inode *inode;
1057 struct kobject *kobj = NULL;
1058 int ret;
1059 int index;
1060 int factor;
1061 struct btrfs_caching_control *caching_ctl = NULL;
1062 bool remove_em;
1063 bool remove_rsv = false;
1064
1065 block_group = btrfs_lookup_block_group(fs_info, group_start);
1066 BUG_ON(!block_group);
1067 BUG_ON(!block_group->ro);
1068
1069 trace_btrfs_remove_block_group(block_group);
1070 /*
1071 * Free the reserved super bytes from this block group before
1072 * remove it.
1073 */
1074 btrfs_free_excluded_extents(block_group);
1075 btrfs_free_ref_tree_range(fs_info, block_group->start,
1076 block_group->length);
1077
1078 index = btrfs_bg_flags_to_raid_index(block_group->flags);
1079 factor = btrfs_bg_type_to_factor(block_group->flags);
1080
1081 /* make sure this block group isn't part of an allocation cluster */
1082 cluster = &fs_info->data_alloc_cluster;
1083 spin_lock(&cluster->refill_lock);
1084 btrfs_return_cluster_to_free_space(block_group, cluster);
1085 spin_unlock(&cluster->refill_lock);
1086
1087 /*
1088 * make sure this block group isn't part of a metadata
1089 * allocation cluster
1090 */
1091 cluster = &fs_info->meta_alloc_cluster;
1092 spin_lock(&cluster->refill_lock);
1093 btrfs_return_cluster_to_free_space(block_group, cluster);
1094 spin_unlock(&cluster->refill_lock);
1095
1096 btrfs_clear_treelog_bg(block_group);
1097 btrfs_clear_data_reloc_bg(block_group);
1098
1099 path = btrfs_alloc_path();
1100 if (!path) {
1101 ret = -ENOMEM;
1102 goto out;
1103 }
1104
1105 /*
1106 * get the inode first so any iput calls done for the io_list
1107 * aren't the final iput (no unlinks allowed now)
1108 */
1109 inode = lookup_free_space_inode(block_group, path);
1110
1111 mutex_lock(&trans->transaction->cache_write_mutex);
1112 /*
1113 * Make sure our free space cache IO is done before removing the
1114 * free space inode
1115 */
1116 spin_lock(&trans->transaction->dirty_bgs_lock);
1117 if (!list_empty(&block_group->io_list)) {
1118 list_del_init(&block_group->io_list);
1119
1120 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
1121
1122 spin_unlock(&trans->transaction->dirty_bgs_lock);
1123 btrfs_wait_cache_io(trans, block_group, path);
1124 btrfs_put_block_group(block_group);
1125 spin_lock(&trans->transaction->dirty_bgs_lock);
1126 }
1127
1128 if (!list_empty(&block_group->dirty_list)) {
1129 list_del_init(&block_group->dirty_list);
1130 remove_rsv = true;
1131 btrfs_put_block_group(block_group);
1132 }
1133 spin_unlock(&trans->transaction->dirty_bgs_lock);
1134 mutex_unlock(&trans->transaction->cache_write_mutex);
1135
1136 ret = btrfs_remove_free_space_inode(trans, inode, block_group);
1137 if (ret)
1138 goto out;
1139
1140 write_lock(&fs_info->block_group_cache_lock);
1141 rb_erase_cached(&block_group->cache_node,
1142 &fs_info->block_group_cache_tree);
1143 RB_CLEAR_NODE(&block_group->cache_node);
1144
1145 /* Once for the block groups rbtree */
1146 btrfs_put_block_group(block_group);
1147
1148 write_unlock(&fs_info->block_group_cache_lock);
1149
1150 down_write(&block_group->space_info->groups_sem);
1151 /*
1152 * we must use list_del_init so people can check to see if they
1153 * are still on the list after taking the semaphore
1154 */
1155 list_del_init(&block_group->list);
1156 if (list_empty(&block_group->space_info->block_groups[index])) {
1157 kobj = block_group->space_info->block_group_kobjs[index];
1158 block_group->space_info->block_group_kobjs[index] = NULL;
1159 clear_avail_alloc_bits(fs_info, block_group->flags);
1160 }
1161 up_write(&block_group->space_info->groups_sem);
1162 clear_incompat_bg_bits(fs_info, block_group->flags);
1163 if (kobj) {
1164 kobject_del(kobj);
1165 kobject_put(kobj);
1166 }
1167
1168 if (block_group->cached == BTRFS_CACHE_STARTED)
1169 btrfs_wait_block_group_cache_done(block_group);
1170
1171 write_lock(&fs_info->block_group_cache_lock);
1172 caching_ctl = btrfs_get_caching_control(block_group);
1173 if (!caching_ctl) {
1174 struct btrfs_caching_control *ctl;
1175
1176 list_for_each_entry(ctl, &fs_info->caching_block_groups, list) {
1177 if (ctl->block_group == block_group) {
1178 caching_ctl = ctl;
1179 refcount_inc(&caching_ctl->count);
1180 break;
1181 }
1182 }
1183 }
1184 if (caching_ctl)
1185 list_del_init(&caching_ctl->list);
1186 write_unlock(&fs_info->block_group_cache_lock);
1187
1188 if (caching_ctl) {
1189 /* Once for the caching bgs list and once for us. */
1190 btrfs_put_caching_control(caching_ctl);
1191 btrfs_put_caching_control(caching_ctl);
1192 }
1193
1194 spin_lock(&trans->transaction->dirty_bgs_lock);
1195 WARN_ON(!list_empty(&block_group->dirty_list));
1196 WARN_ON(!list_empty(&block_group->io_list));
1197 spin_unlock(&trans->transaction->dirty_bgs_lock);
1198
1199 btrfs_remove_free_space_cache(block_group);
1200
1201 spin_lock(&block_group->space_info->lock);
1202 list_del_init(&block_group->ro_list);
1203
1204 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1205 WARN_ON(block_group->space_info->total_bytes
1206 < block_group->length);
1207 WARN_ON(block_group->space_info->bytes_readonly
1208 < block_group->length - block_group->zone_unusable);
1209 WARN_ON(block_group->space_info->bytes_zone_unusable
1210 < block_group->zone_unusable);
1211 WARN_ON(block_group->space_info->disk_total
1212 < block_group->length * factor);
1213 }
1214 block_group->space_info->total_bytes -= block_group->length;
1215 block_group->space_info->bytes_readonly -=
1216 (block_group->length - block_group->zone_unusable);
1217 block_group->space_info->bytes_zone_unusable -=
1218 block_group->zone_unusable;
1219 block_group->space_info->disk_total -= block_group->length * factor;
1220
1221 spin_unlock(&block_group->space_info->lock);
1222
1223 /*
1224 * Remove the free space for the block group from the free space tree
1225 * and the block group's item from the extent tree before marking the
1226 * block group as removed. This is to prevent races with tasks that
1227 * freeze and unfreeze a block group, this task and another task
1228 * allocating a new block group - the unfreeze task ends up removing
1229 * the block group's extent map before the task calling this function
1230 * deletes the block group item from the extent tree, allowing for
1231 * another task to attempt to create another block group with the same
1232 * item key (and failing with -EEXIST and a transaction abort).
1233 */
1234 ret = remove_block_group_free_space(trans, block_group);
1235 if (ret)
1236 goto out;
1237
1238 ret = remove_block_group_item(trans, path, block_group);
1239 if (ret < 0)
1240 goto out;
1241
1242 spin_lock(&block_group->lock);
1243 set_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags);
1244
1245 /*
1246 * At this point trimming or scrub can't start on this block group,
1247 * because we removed the block group from the rbtree
1248 * fs_info->block_group_cache_tree so no one can't find it anymore and
1249 * even if someone already got this block group before we removed it
1250 * from the rbtree, they have already incremented block_group->frozen -
1251 * if they didn't, for the trimming case they won't find any free space
1252 * entries because we already removed them all when we called
1253 * btrfs_remove_free_space_cache().
1254 *
1255 * And we must not remove the extent map from the fs_info->mapping_tree
1256 * to prevent the same logical address range and physical device space
1257 * ranges from being reused for a new block group. This is needed to
1258 * avoid races with trimming and scrub.
1259 *
1260 * An fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
1261 * completely transactionless, so while it is trimming a range the
1262 * currently running transaction might finish and a new one start,
1263 * allowing for new block groups to be created that can reuse the same
1264 * physical device locations unless we take this special care.
1265 *
1266 * There may also be an implicit trim operation if the file system
1267 * is mounted with -odiscard. The same protections must remain
1268 * in place until the extents have been discarded completely when
1269 * the transaction commit has completed.
1270 */
1271 remove_em = (atomic_read(&block_group->frozen) == 0);
1272 spin_unlock(&block_group->lock);
1273
1274 if (remove_em) {
1275 struct extent_map_tree *em_tree;
1276
1277 em_tree = &fs_info->mapping_tree;
1278 write_lock(&em_tree->lock);
1279 remove_extent_mapping(em_tree, em);
1280 write_unlock(&em_tree->lock);
1281 /* once for the tree */
1282 free_extent_map(em);
1283 }
1284
1285 out:
1286 /* Once for the lookup reference */
1287 btrfs_put_block_group(block_group);
1288 if (remove_rsv)
1289 btrfs_delayed_refs_rsv_release(fs_info, 1);
1290 btrfs_free_path(path);
1291 return ret;
1292 }
1293
btrfs_start_trans_remove_block_group(struct btrfs_fs_info * fs_info,const u64 chunk_offset)1294 struct btrfs_trans_handle *btrfs_start_trans_remove_block_group(
1295 struct btrfs_fs_info *fs_info, const u64 chunk_offset)
1296 {
1297 struct btrfs_root *root = btrfs_block_group_root(fs_info);
1298 struct extent_map_tree *em_tree = &fs_info->mapping_tree;
1299 struct extent_map *em;
1300 struct map_lookup *map;
1301 unsigned int num_items;
1302
1303 read_lock(&em_tree->lock);
1304 em = lookup_extent_mapping(em_tree, chunk_offset, 1);
1305 read_unlock(&em_tree->lock);
1306 ASSERT(em && em->start == chunk_offset);
1307
1308 /*
1309 * We need to reserve 3 + N units from the metadata space info in order
1310 * to remove a block group (done at btrfs_remove_chunk() and at
1311 * btrfs_remove_block_group()), which are used for:
1312 *
1313 * 1 unit for adding the free space inode's orphan (located in the tree
1314 * of tree roots).
1315 * 1 unit for deleting the block group item (located in the extent
1316 * tree).
1317 * 1 unit for deleting the free space item (located in tree of tree
1318 * roots).
1319 * N units for deleting N device extent items corresponding to each
1320 * stripe (located in the device tree).
1321 *
1322 * In order to remove a block group we also need to reserve units in the
1323 * system space info in order to update the chunk tree (update one or
1324 * more device items and remove one chunk item), but this is done at
1325 * btrfs_remove_chunk() through a call to check_system_chunk().
1326 */
1327 map = em->map_lookup;
1328 num_items = 3 + map->num_stripes;
1329 free_extent_map(em);
1330
1331 return btrfs_start_transaction_fallback_global_rsv(root, num_items);
1332 }
1333
1334 /*
1335 * Mark block group @cache read-only, so later write won't happen to block
1336 * group @cache.
1337 *
1338 * If @force is not set, this function will only mark the block group readonly
1339 * if we have enough free space (1M) in other metadata/system block groups.
1340 * If @force is not set, this function will mark the block group readonly
1341 * without checking free space.
1342 *
1343 * NOTE: This function doesn't care if other block groups can contain all the
1344 * data in this block group. That check should be done by relocation routine,
1345 * not this function.
1346 */
inc_block_group_ro(struct btrfs_block_group * cache,int force)1347 static int inc_block_group_ro(struct btrfs_block_group *cache, int force)
1348 {
1349 struct btrfs_space_info *sinfo = cache->space_info;
1350 u64 num_bytes;
1351 int ret = -ENOSPC;
1352
1353 spin_lock(&sinfo->lock);
1354 spin_lock(&cache->lock);
1355
1356 if (cache->swap_extents) {
1357 ret = -ETXTBSY;
1358 goto out;
1359 }
1360
1361 if (cache->ro) {
1362 cache->ro++;
1363 ret = 0;
1364 goto out;
1365 }
1366
1367 num_bytes = cache->length - cache->reserved - cache->pinned -
1368 cache->bytes_super - cache->zone_unusable - cache->used;
1369
1370 /*
1371 * Data never overcommits, even in mixed mode, so do just the straight
1372 * check of left over space in how much we have allocated.
1373 */
1374 if (force) {
1375 ret = 0;
1376 } else if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA) {
1377 u64 sinfo_used = btrfs_space_info_used(sinfo, true);
1378
1379 /*
1380 * Here we make sure if we mark this bg RO, we still have enough
1381 * free space as buffer.
1382 */
1383 if (sinfo_used + num_bytes <= sinfo->total_bytes)
1384 ret = 0;
1385 } else {
1386 /*
1387 * We overcommit metadata, so we need to do the
1388 * btrfs_can_overcommit check here, and we need to pass in
1389 * BTRFS_RESERVE_NO_FLUSH to give ourselves the most amount of
1390 * leeway to allow us to mark this block group as read only.
1391 */
1392 if (btrfs_can_overcommit(cache->fs_info, sinfo, num_bytes,
1393 BTRFS_RESERVE_NO_FLUSH))
1394 ret = 0;
1395 }
1396
1397 if (!ret) {
1398 sinfo->bytes_readonly += num_bytes;
1399 if (btrfs_is_zoned(cache->fs_info)) {
1400 /* Migrate zone_unusable bytes to readonly */
1401 sinfo->bytes_readonly += cache->zone_unusable;
1402 sinfo->bytes_zone_unusable -= cache->zone_unusable;
1403 cache->zone_unusable = 0;
1404 }
1405 cache->ro++;
1406 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
1407 }
1408 out:
1409 spin_unlock(&cache->lock);
1410 spin_unlock(&sinfo->lock);
1411 if (ret == -ENOSPC && btrfs_test_opt(cache->fs_info, ENOSPC_DEBUG)) {
1412 btrfs_info(cache->fs_info,
1413 "unable to make block group %llu ro", cache->start);
1414 btrfs_dump_space_info(cache->fs_info, cache->space_info, 0, 0);
1415 }
1416 return ret;
1417 }
1418
clean_pinned_extents(struct btrfs_trans_handle * trans,struct btrfs_block_group * bg)1419 static bool clean_pinned_extents(struct btrfs_trans_handle *trans,
1420 struct btrfs_block_group *bg)
1421 {
1422 struct btrfs_fs_info *fs_info = bg->fs_info;
1423 struct btrfs_transaction *prev_trans = NULL;
1424 const u64 start = bg->start;
1425 const u64 end = start + bg->length - 1;
1426 int ret;
1427
1428 spin_lock(&fs_info->trans_lock);
1429 if (trans->transaction->list.prev != &fs_info->trans_list) {
1430 prev_trans = list_last_entry(&trans->transaction->list,
1431 struct btrfs_transaction, list);
1432 refcount_inc(&prev_trans->use_count);
1433 }
1434 spin_unlock(&fs_info->trans_lock);
1435
1436 /*
1437 * Hold the unused_bg_unpin_mutex lock to avoid racing with
1438 * btrfs_finish_extent_commit(). If we are at transaction N, another
1439 * task might be running finish_extent_commit() for the previous
1440 * transaction N - 1, and have seen a range belonging to the block
1441 * group in pinned_extents before we were able to clear the whole block
1442 * group range from pinned_extents. This means that task can lookup for
1443 * the block group after we unpinned it from pinned_extents and removed
1444 * it, leading to a BUG_ON() at unpin_extent_range().
1445 */
1446 mutex_lock(&fs_info->unused_bg_unpin_mutex);
1447 if (prev_trans) {
1448 ret = clear_extent_bits(&prev_trans->pinned_extents, start, end,
1449 EXTENT_DIRTY);
1450 if (ret)
1451 goto out;
1452 }
1453
1454 ret = clear_extent_bits(&trans->transaction->pinned_extents, start, end,
1455 EXTENT_DIRTY);
1456 out:
1457 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
1458 if (prev_trans)
1459 btrfs_put_transaction(prev_trans);
1460
1461 return ret == 0;
1462 }
1463
1464 /*
1465 * Process the unused_bgs list and remove any that don't have any allocated
1466 * space inside of them.
1467 */
btrfs_delete_unused_bgs(struct btrfs_fs_info * fs_info)1468 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
1469 {
1470 LIST_HEAD(retry_list);
1471 struct btrfs_block_group *block_group;
1472 struct btrfs_space_info *space_info;
1473 struct btrfs_trans_handle *trans;
1474 const bool async_trim_enabled = btrfs_test_opt(fs_info, DISCARD_ASYNC);
1475 int ret = 0;
1476
1477 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1478 return;
1479
1480 if (btrfs_fs_closing(fs_info))
1481 return;
1482
1483 /*
1484 * Long running balances can keep us blocked here for eternity, so
1485 * simply skip deletion if we're unable to get the mutex.
1486 */
1487 if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1488 return;
1489
1490 spin_lock(&fs_info->unused_bgs_lock);
1491 while (!list_empty(&fs_info->unused_bgs)) {
1492 u64 used;
1493 int trimming;
1494
1495 block_group = list_first_entry(&fs_info->unused_bgs,
1496 struct btrfs_block_group,
1497 bg_list);
1498 list_del_init(&block_group->bg_list);
1499
1500 space_info = block_group->space_info;
1501
1502 if (ret || btrfs_mixed_space_info(space_info)) {
1503 btrfs_put_block_group(block_group);
1504 continue;
1505 }
1506 spin_unlock(&fs_info->unused_bgs_lock);
1507
1508 btrfs_discard_cancel_work(&fs_info->discard_ctl, block_group);
1509
1510 /* Don't want to race with allocators so take the groups_sem */
1511 down_write(&space_info->groups_sem);
1512
1513 /*
1514 * Async discard moves the final block group discard to be prior
1515 * to the unused_bgs code path. Therefore, if it's not fully
1516 * trimmed, punt it back to the async discard lists.
1517 */
1518 if (btrfs_test_opt(fs_info, DISCARD_ASYNC) &&
1519 !btrfs_is_free_space_trimmed(block_group)) {
1520 trace_btrfs_skip_unused_block_group(block_group);
1521 up_write(&space_info->groups_sem);
1522 /* Requeue if we failed because of async discard */
1523 btrfs_discard_queue_work(&fs_info->discard_ctl,
1524 block_group);
1525 goto next;
1526 }
1527
1528 spin_lock(&space_info->lock);
1529 spin_lock(&block_group->lock);
1530 if (btrfs_is_block_group_used(block_group) || block_group->ro ||
1531 list_is_singular(&block_group->list)) {
1532 /*
1533 * We want to bail if we made new allocations or have
1534 * outstanding allocations in this block group. We do
1535 * the ro check in case balance is currently acting on
1536 * this block group.
1537 */
1538 trace_btrfs_skip_unused_block_group(block_group);
1539 spin_unlock(&block_group->lock);
1540 spin_unlock(&space_info->lock);
1541 up_write(&space_info->groups_sem);
1542 goto next;
1543 }
1544
1545 /*
1546 * The block group may be unused but there may be space reserved
1547 * accounting with the existence of that block group, that is,
1548 * space_info->bytes_may_use was incremented by a task but no
1549 * space was yet allocated from the block group by the task.
1550 * That space may or may not be allocated, as we are generally
1551 * pessimistic about space reservation for metadata as well as
1552 * for data when using compression (as we reserve space based on
1553 * the worst case, when data can't be compressed, and before
1554 * actually attempting compression, before starting writeback).
1555 *
1556 * So check if the total space of the space_info minus the size
1557 * of this block group is less than the used space of the
1558 * space_info - if that's the case, then it means we have tasks
1559 * that might be relying on the block group in order to allocate
1560 * extents, and add back the block group to the unused list when
1561 * we finish, so that we retry later in case no tasks ended up
1562 * needing to allocate extents from the block group.
1563 */
1564 used = btrfs_space_info_used(space_info, true);
1565 if (space_info->total_bytes - block_group->length < used &&
1566 block_group->zone_unusable < block_group->length) {
1567 /*
1568 * Add a reference for the list, compensate for the ref
1569 * drop under the "next" label for the
1570 * fs_info->unused_bgs list.
1571 */
1572 btrfs_get_block_group(block_group);
1573 list_add_tail(&block_group->bg_list, &retry_list);
1574
1575 trace_btrfs_skip_unused_block_group(block_group);
1576 spin_unlock(&block_group->lock);
1577 spin_unlock(&space_info->lock);
1578 up_write(&space_info->groups_sem);
1579 goto next;
1580 }
1581
1582 spin_unlock(&block_group->lock);
1583 spin_unlock(&space_info->lock);
1584
1585 /* We don't want to force the issue, only flip if it's ok. */
1586 ret = inc_block_group_ro(block_group, 0);
1587 up_write(&space_info->groups_sem);
1588 if (ret < 0) {
1589 ret = 0;
1590 goto next;
1591 }
1592
1593 ret = btrfs_zone_finish(block_group);
1594 if (ret < 0) {
1595 btrfs_dec_block_group_ro(block_group);
1596 if (ret == -EAGAIN)
1597 ret = 0;
1598 goto next;
1599 }
1600
1601 /*
1602 * Want to do this before we do anything else so we can recover
1603 * properly if we fail to join the transaction.
1604 */
1605 trans = btrfs_start_trans_remove_block_group(fs_info,
1606 block_group->start);
1607 if (IS_ERR(trans)) {
1608 btrfs_dec_block_group_ro(block_group);
1609 ret = PTR_ERR(trans);
1610 goto next;
1611 }
1612
1613 /*
1614 * We could have pending pinned extents for this block group,
1615 * just delete them, we don't care about them anymore.
1616 */
1617 if (!clean_pinned_extents(trans, block_group)) {
1618 btrfs_dec_block_group_ro(block_group);
1619 goto end_trans;
1620 }
1621
1622 /*
1623 * At this point, the block_group is read only and should fail
1624 * new allocations. However, btrfs_finish_extent_commit() can
1625 * cause this block_group to be placed back on the discard
1626 * lists because now the block_group isn't fully discarded.
1627 * Bail here and try again later after discarding everything.
1628 */
1629 spin_lock(&fs_info->discard_ctl.lock);
1630 if (!list_empty(&block_group->discard_list)) {
1631 spin_unlock(&fs_info->discard_ctl.lock);
1632 btrfs_dec_block_group_ro(block_group);
1633 btrfs_discard_queue_work(&fs_info->discard_ctl,
1634 block_group);
1635 goto end_trans;
1636 }
1637 spin_unlock(&fs_info->discard_ctl.lock);
1638
1639 /* Reset pinned so btrfs_put_block_group doesn't complain */
1640 spin_lock(&space_info->lock);
1641 spin_lock(&block_group->lock);
1642
1643 btrfs_space_info_update_bytes_pinned(fs_info, space_info,
1644 -block_group->pinned);
1645 space_info->bytes_readonly += block_group->pinned;
1646 block_group->pinned = 0;
1647
1648 spin_unlock(&block_group->lock);
1649 spin_unlock(&space_info->lock);
1650
1651 /*
1652 * The normal path here is an unused block group is passed here,
1653 * then trimming is handled in the transaction commit path.
1654 * Async discard interposes before this to do the trimming
1655 * before coming down the unused block group path as trimming
1656 * will no longer be done later in the transaction commit path.
1657 */
1658 if (!async_trim_enabled && btrfs_test_opt(fs_info, DISCARD_ASYNC))
1659 goto flip_async;
1660
1661 /*
1662 * DISCARD can flip during remount. On zoned filesystems, we
1663 * need to reset sequential-required zones.
1664 */
1665 trimming = btrfs_test_opt(fs_info, DISCARD_SYNC) ||
1666 btrfs_is_zoned(fs_info);
1667
1668 /* Implicit trim during transaction commit. */
1669 if (trimming)
1670 btrfs_freeze_block_group(block_group);
1671
1672 /*
1673 * Btrfs_remove_chunk will abort the transaction if things go
1674 * horribly wrong.
1675 */
1676 ret = btrfs_remove_chunk(trans, block_group->start);
1677
1678 if (ret) {
1679 if (trimming)
1680 btrfs_unfreeze_block_group(block_group);
1681 goto end_trans;
1682 }
1683
1684 /*
1685 * If we're not mounted with -odiscard, we can just forget
1686 * about this block group. Otherwise we'll need to wait
1687 * until transaction commit to do the actual discard.
1688 */
1689 if (trimming) {
1690 spin_lock(&fs_info->unused_bgs_lock);
1691 /*
1692 * A concurrent scrub might have added us to the list
1693 * fs_info->unused_bgs, so use a list_move operation
1694 * to add the block group to the deleted_bgs list.
1695 */
1696 list_move(&block_group->bg_list,
1697 &trans->transaction->deleted_bgs);
1698 spin_unlock(&fs_info->unused_bgs_lock);
1699 btrfs_get_block_group(block_group);
1700 }
1701 end_trans:
1702 btrfs_end_transaction(trans);
1703 next:
1704 btrfs_put_block_group(block_group);
1705 spin_lock(&fs_info->unused_bgs_lock);
1706 }
1707 list_splice_tail(&retry_list, &fs_info->unused_bgs);
1708 spin_unlock(&fs_info->unused_bgs_lock);
1709 mutex_unlock(&fs_info->reclaim_bgs_lock);
1710 return;
1711
1712 flip_async:
1713 btrfs_end_transaction(trans);
1714 spin_lock(&fs_info->unused_bgs_lock);
1715 list_splice_tail(&retry_list, &fs_info->unused_bgs);
1716 spin_unlock(&fs_info->unused_bgs_lock);
1717 mutex_unlock(&fs_info->reclaim_bgs_lock);
1718 btrfs_put_block_group(block_group);
1719 btrfs_discard_punt_unused_bgs_list(fs_info);
1720 }
1721
btrfs_mark_bg_unused(struct btrfs_block_group * bg)1722 void btrfs_mark_bg_unused(struct btrfs_block_group *bg)
1723 {
1724 struct btrfs_fs_info *fs_info = bg->fs_info;
1725
1726 spin_lock(&fs_info->unused_bgs_lock);
1727 if (list_empty(&bg->bg_list)) {
1728 btrfs_get_block_group(bg);
1729 trace_btrfs_add_unused_block_group(bg);
1730 list_add_tail(&bg->bg_list, &fs_info->unused_bgs);
1731 } else if (!test_bit(BLOCK_GROUP_FLAG_NEW, &bg->runtime_flags)) {
1732 /* Pull out the block group from the reclaim_bgs list. */
1733 trace_btrfs_add_unused_block_group(bg);
1734 list_move_tail(&bg->bg_list, &fs_info->unused_bgs);
1735 }
1736 spin_unlock(&fs_info->unused_bgs_lock);
1737 }
1738
1739 /*
1740 * We want block groups with a low number of used bytes to be in the beginning
1741 * of the list, so they will get reclaimed first.
1742 */
reclaim_bgs_cmp(void * unused,const struct list_head * a,const struct list_head * b)1743 static int reclaim_bgs_cmp(void *unused, const struct list_head *a,
1744 const struct list_head *b)
1745 {
1746 const struct btrfs_block_group *bg1, *bg2;
1747
1748 bg1 = list_entry(a, struct btrfs_block_group, bg_list);
1749 bg2 = list_entry(b, struct btrfs_block_group, bg_list);
1750
1751 return bg1->used > bg2->used;
1752 }
1753
btrfs_should_reclaim(struct btrfs_fs_info * fs_info)1754 static inline bool btrfs_should_reclaim(struct btrfs_fs_info *fs_info)
1755 {
1756 if (btrfs_is_zoned(fs_info))
1757 return btrfs_zoned_should_reclaim(fs_info);
1758 return true;
1759 }
1760
should_reclaim_block_group(struct btrfs_block_group * bg,u64 bytes_freed)1761 static bool should_reclaim_block_group(struct btrfs_block_group *bg, u64 bytes_freed)
1762 {
1763 const struct btrfs_space_info *space_info = bg->space_info;
1764 const int reclaim_thresh = READ_ONCE(space_info->bg_reclaim_threshold);
1765 const u64 new_val = bg->used;
1766 const u64 old_val = new_val + bytes_freed;
1767 u64 thresh;
1768
1769 if (reclaim_thresh == 0)
1770 return false;
1771
1772 thresh = mult_perc(bg->length, reclaim_thresh);
1773
1774 /*
1775 * If we were below the threshold before don't reclaim, we are likely a
1776 * brand new block group and we don't want to relocate new block groups.
1777 */
1778 if (old_val < thresh)
1779 return false;
1780 if (new_val >= thresh)
1781 return false;
1782 return true;
1783 }
1784
btrfs_reclaim_bgs_work(struct work_struct * work)1785 void btrfs_reclaim_bgs_work(struct work_struct *work)
1786 {
1787 struct btrfs_fs_info *fs_info =
1788 container_of(work, struct btrfs_fs_info, reclaim_bgs_work);
1789 struct btrfs_block_group *bg;
1790 struct btrfs_space_info *space_info;
1791
1792 if (!test_bit(BTRFS_FS_OPEN, &fs_info->flags))
1793 return;
1794
1795 if (btrfs_fs_closing(fs_info))
1796 return;
1797
1798 if (!btrfs_should_reclaim(fs_info))
1799 return;
1800
1801 sb_start_write(fs_info->sb);
1802
1803 if (!btrfs_exclop_start(fs_info, BTRFS_EXCLOP_BALANCE)) {
1804 sb_end_write(fs_info->sb);
1805 return;
1806 }
1807
1808 /*
1809 * Long running balances can keep us blocked here for eternity, so
1810 * simply skip reclaim if we're unable to get the mutex.
1811 */
1812 if (!mutex_trylock(&fs_info->reclaim_bgs_lock)) {
1813 btrfs_exclop_finish(fs_info);
1814 sb_end_write(fs_info->sb);
1815 return;
1816 }
1817
1818 spin_lock(&fs_info->unused_bgs_lock);
1819 /*
1820 * Sort happens under lock because we can't simply splice it and sort.
1821 * The block groups might still be in use and reachable via bg_list,
1822 * and their presence in the reclaim_bgs list must be preserved.
1823 */
1824 list_sort(NULL, &fs_info->reclaim_bgs, reclaim_bgs_cmp);
1825 while (!list_empty(&fs_info->reclaim_bgs)) {
1826 u64 zone_unusable;
1827 int ret = 0;
1828
1829 bg = list_first_entry(&fs_info->reclaim_bgs,
1830 struct btrfs_block_group,
1831 bg_list);
1832 list_del_init(&bg->bg_list);
1833
1834 space_info = bg->space_info;
1835 spin_unlock(&fs_info->unused_bgs_lock);
1836
1837 /* Don't race with allocators so take the groups_sem */
1838 down_write(&space_info->groups_sem);
1839
1840 spin_lock(&bg->lock);
1841 if (bg->reserved || bg->pinned || bg->ro) {
1842 /*
1843 * We want to bail if we made new allocations or have
1844 * outstanding allocations in this block group. We do
1845 * the ro check in case balance is currently acting on
1846 * this block group.
1847 */
1848 spin_unlock(&bg->lock);
1849 up_write(&space_info->groups_sem);
1850 goto next;
1851 }
1852 if (bg->used == 0) {
1853 /*
1854 * It is possible that we trigger relocation on a block
1855 * group as its extents are deleted and it first goes
1856 * below the threshold, then shortly after goes empty.
1857 *
1858 * In this case, relocating it does delete it, but has
1859 * some overhead in relocation specific metadata, looking
1860 * for the non-existent extents and running some extra
1861 * transactions, which we can avoid by using one of the
1862 * other mechanisms for dealing with empty block groups.
1863 */
1864 if (!btrfs_test_opt(fs_info, DISCARD_ASYNC))
1865 btrfs_mark_bg_unused(bg);
1866 spin_unlock(&bg->lock);
1867 up_write(&space_info->groups_sem);
1868 goto next;
1869
1870 }
1871 /*
1872 * The block group might no longer meet the reclaim condition by
1873 * the time we get around to reclaiming it, so to avoid
1874 * reclaiming overly full block_groups, skip reclaiming them.
1875 *
1876 * Since the decision making process also depends on the amount
1877 * being freed, pass in a fake giant value to skip that extra
1878 * check, which is more meaningful when adding to the list in
1879 * the first place.
1880 */
1881 if (!should_reclaim_block_group(bg, bg->length)) {
1882 spin_unlock(&bg->lock);
1883 up_write(&space_info->groups_sem);
1884 goto next;
1885 }
1886 spin_unlock(&bg->lock);
1887
1888 /*
1889 * Get out fast, in case we're read-only or unmounting the
1890 * filesystem. It is OK to drop block groups from the list even
1891 * for the read-only case. As we did sb_start_write(),
1892 * "mount -o remount,ro" won't happen and read-only filesystem
1893 * means it is forced read-only due to a fatal error. So, it
1894 * never gets back to read-write to let us reclaim again.
1895 */
1896 if (btrfs_need_cleaner_sleep(fs_info)) {
1897 up_write(&space_info->groups_sem);
1898 goto next;
1899 }
1900
1901 /*
1902 * Cache the zone_unusable value before turning the block group
1903 * to read only. As soon as the blog group is read only it's
1904 * zone_unusable value gets moved to the block group's read-only
1905 * bytes and isn't available for calculations anymore.
1906 */
1907 zone_unusable = bg->zone_unusable;
1908 ret = inc_block_group_ro(bg, 0);
1909 up_write(&space_info->groups_sem);
1910 if (ret < 0)
1911 goto next;
1912
1913 btrfs_info(fs_info,
1914 "reclaiming chunk %llu with %llu%% used %llu%% unusable",
1915 bg->start,
1916 div64_u64(bg->used * 100, bg->length),
1917 div64_u64(zone_unusable * 100, bg->length));
1918 trace_btrfs_reclaim_block_group(bg);
1919 ret = btrfs_relocate_chunk(fs_info, bg->start);
1920 if (ret) {
1921 btrfs_dec_block_group_ro(bg);
1922 btrfs_err(fs_info, "error relocating chunk %llu",
1923 bg->start);
1924 }
1925
1926 next:
1927 if (ret)
1928 btrfs_mark_bg_to_reclaim(bg);
1929 btrfs_put_block_group(bg);
1930
1931 mutex_unlock(&fs_info->reclaim_bgs_lock);
1932 /*
1933 * Reclaiming all the block groups in the list can take really
1934 * long. Prioritize cleaning up unused block groups.
1935 */
1936 btrfs_delete_unused_bgs(fs_info);
1937 /*
1938 * If we are interrupted by a balance, we can just bail out. The
1939 * cleaner thread restart again if necessary.
1940 */
1941 if (!mutex_trylock(&fs_info->reclaim_bgs_lock))
1942 goto end;
1943 spin_lock(&fs_info->unused_bgs_lock);
1944 }
1945 spin_unlock(&fs_info->unused_bgs_lock);
1946 mutex_unlock(&fs_info->reclaim_bgs_lock);
1947 end:
1948 btrfs_exclop_finish(fs_info);
1949 sb_end_write(fs_info->sb);
1950 }
1951
btrfs_reclaim_bgs(struct btrfs_fs_info * fs_info)1952 void btrfs_reclaim_bgs(struct btrfs_fs_info *fs_info)
1953 {
1954 spin_lock(&fs_info->unused_bgs_lock);
1955 if (!list_empty(&fs_info->reclaim_bgs))
1956 queue_work(system_unbound_wq, &fs_info->reclaim_bgs_work);
1957 spin_unlock(&fs_info->unused_bgs_lock);
1958 }
1959
btrfs_mark_bg_to_reclaim(struct btrfs_block_group * bg)1960 void btrfs_mark_bg_to_reclaim(struct btrfs_block_group *bg)
1961 {
1962 struct btrfs_fs_info *fs_info = bg->fs_info;
1963
1964 spin_lock(&fs_info->unused_bgs_lock);
1965 if (list_empty(&bg->bg_list)) {
1966 btrfs_get_block_group(bg);
1967 trace_btrfs_add_reclaim_block_group(bg);
1968 list_add_tail(&bg->bg_list, &fs_info->reclaim_bgs);
1969 }
1970 spin_unlock(&fs_info->unused_bgs_lock);
1971 }
1972
read_bg_from_eb(struct btrfs_fs_info * fs_info,struct btrfs_key * key,struct btrfs_path * path)1973 static int read_bg_from_eb(struct btrfs_fs_info *fs_info, struct btrfs_key *key,
1974 struct btrfs_path *path)
1975 {
1976 struct extent_map_tree *em_tree;
1977 struct extent_map *em;
1978 struct btrfs_block_group_item bg;
1979 struct extent_buffer *leaf;
1980 int slot;
1981 u64 flags;
1982 int ret = 0;
1983
1984 slot = path->slots[0];
1985 leaf = path->nodes[0];
1986
1987 em_tree = &fs_info->mapping_tree;
1988 read_lock(&em_tree->lock);
1989 em = lookup_extent_mapping(em_tree, key->objectid, key->offset);
1990 read_unlock(&em_tree->lock);
1991 if (!em) {
1992 btrfs_err(fs_info,
1993 "logical %llu len %llu found bg but no related chunk",
1994 key->objectid, key->offset);
1995 return -ENOENT;
1996 }
1997
1998 if (em->start != key->objectid || em->len != key->offset) {
1999 btrfs_err(fs_info,
2000 "block group %llu len %llu mismatch with chunk %llu len %llu",
2001 key->objectid, key->offset, em->start, em->len);
2002 ret = -EUCLEAN;
2003 goto out_free_em;
2004 }
2005
2006 read_extent_buffer(leaf, &bg, btrfs_item_ptr_offset(leaf, slot),
2007 sizeof(bg));
2008 flags = btrfs_stack_block_group_flags(&bg) &
2009 BTRFS_BLOCK_GROUP_TYPE_MASK;
2010
2011 if (flags != (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2012 btrfs_err(fs_info,
2013 "block group %llu len %llu type flags 0x%llx mismatch with chunk type flags 0x%llx",
2014 key->objectid, key->offset, flags,
2015 (BTRFS_BLOCK_GROUP_TYPE_MASK & em->map_lookup->type));
2016 ret = -EUCLEAN;
2017 }
2018
2019 out_free_em:
2020 free_extent_map(em);
2021 return ret;
2022 }
2023
find_first_block_group(struct btrfs_fs_info * fs_info,struct btrfs_path * path,struct btrfs_key * key)2024 static int find_first_block_group(struct btrfs_fs_info *fs_info,
2025 struct btrfs_path *path,
2026 struct btrfs_key *key)
2027 {
2028 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2029 int ret;
2030 struct btrfs_key found_key;
2031
2032 btrfs_for_each_slot(root, key, &found_key, path, ret) {
2033 if (found_key.objectid >= key->objectid &&
2034 found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
2035 return read_bg_from_eb(fs_info, &found_key, path);
2036 }
2037 }
2038 return ret;
2039 }
2040
set_avail_alloc_bits(struct btrfs_fs_info * fs_info,u64 flags)2041 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2042 {
2043 u64 extra_flags = chunk_to_extended(flags) &
2044 BTRFS_EXTENDED_PROFILE_MASK;
2045
2046 write_seqlock(&fs_info->profiles_lock);
2047 if (flags & BTRFS_BLOCK_GROUP_DATA)
2048 fs_info->avail_data_alloc_bits |= extra_flags;
2049 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2050 fs_info->avail_metadata_alloc_bits |= extra_flags;
2051 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2052 fs_info->avail_system_alloc_bits |= extra_flags;
2053 write_sequnlock(&fs_info->profiles_lock);
2054 }
2055
2056 /*
2057 * Map a physical disk address to a list of logical addresses.
2058 *
2059 * @fs_info: the filesystem
2060 * @chunk_start: logical address of block group
2061 * @physical: physical address to map to logical addresses
2062 * @logical: return array of logical addresses which map to @physical
2063 * @naddrs: length of @logical
2064 * @stripe_len: size of IO stripe for the given block group
2065 *
2066 * Maps a particular @physical disk address to a list of @logical addresses.
2067 * Used primarily to exclude those portions of a block group that contain super
2068 * block copies.
2069 */
btrfs_rmap_block(struct btrfs_fs_info * fs_info,u64 chunk_start,u64 physical,u64 ** logical,int * naddrs,int * stripe_len)2070 int btrfs_rmap_block(struct btrfs_fs_info *fs_info, u64 chunk_start,
2071 u64 physical, u64 **logical, int *naddrs, int *stripe_len)
2072 {
2073 struct extent_map *em;
2074 struct map_lookup *map;
2075 u64 *buf;
2076 u64 bytenr;
2077 u64 data_stripe_length;
2078 u64 io_stripe_size;
2079 int i, nr = 0;
2080 int ret = 0;
2081
2082 em = btrfs_get_chunk_map(fs_info, chunk_start, 1);
2083 if (IS_ERR(em))
2084 return -EIO;
2085
2086 map = em->map_lookup;
2087 data_stripe_length = em->orig_block_len;
2088 io_stripe_size = BTRFS_STRIPE_LEN;
2089 chunk_start = em->start;
2090
2091 /* For RAID5/6 adjust to a full IO stripe length */
2092 if (map->type & BTRFS_BLOCK_GROUP_RAID56_MASK)
2093 io_stripe_size = btrfs_stripe_nr_to_offset(nr_data_stripes(map));
2094
2095 buf = kcalloc(map->num_stripes, sizeof(u64), GFP_NOFS);
2096 if (!buf) {
2097 ret = -ENOMEM;
2098 goto out;
2099 }
2100
2101 for (i = 0; i < map->num_stripes; i++) {
2102 bool already_inserted = false;
2103 u32 stripe_nr;
2104 u32 offset;
2105 int j;
2106
2107 if (!in_range(physical, map->stripes[i].physical,
2108 data_stripe_length))
2109 continue;
2110
2111 stripe_nr = (physical - map->stripes[i].physical) >>
2112 BTRFS_STRIPE_LEN_SHIFT;
2113 offset = (physical - map->stripes[i].physical) &
2114 BTRFS_STRIPE_LEN_MASK;
2115
2116 if (map->type & (BTRFS_BLOCK_GROUP_RAID0 |
2117 BTRFS_BLOCK_GROUP_RAID10))
2118 stripe_nr = div_u64(stripe_nr * map->num_stripes + i,
2119 map->sub_stripes);
2120 /*
2121 * The remaining case would be for RAID56, multiply by
2122 * nr_data_stripes(). Alternatively, just use rmap_len below
2123 * instead of map->stripe_len
2124 */
2125 bytenr = chunk_start + stripe_nr * io_stripe_size + offset;
2126
2127 /* Ensure we don't add duplicate addresses */
2128 for (j = 0; j < nr; j++) {
2129 if (buf[j] == bytenr) {
2130 already_inserted = true;
2131 break;
2132 }
2133 }
2134
2135 if (!already_inserted)
2136 buf[nr++] = bytenr;
2137 }
2138
2139 *logical = buf;
2140 *naddrs = nr;
2141 *stripe_len = io_stripe_size;
2142 out:
2143 free_extent_map(em);
2144 return ret;
2145 }
2146
exclude_super_stripes(struct btrfs_block_group * cache)2147 static int exclude_super_stripes(struct btrfs_block_group *cache)
2148 {
2149 struct btrfs_fs_info *fs_info = cache->fs_info;
2150 const bool zoned = btrfs_is_zoned(fs_info);
2151 u64 bytenr;
2152 u64 *logical;
2153 int stripe_len;
2154 int i, nr, ret;
2155
2156 if (cache->start < BTRFS_SUPER_INFO_OFFSET) {
2157 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->start;
2158 cache->bytes_super += stripe_len;
2159 ret = set_extent_bit(&fs_info->excluded_extents, cache->start,
2160 cache->start + stripe_len - 1,
2161 EXTENT_UPTODATE, NULL);
2162 if (ret)
2163 return ret;
2164 }
2165
2166 for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
2167 bytenr = btrfs_sb_offset(i);
2168 ret = btrfs_rmap_block(fs_info, cache->start,
2169 bytenr, &logical, &nr, &stripe_len);
2170 if (ret)
2171 return ret;
2172
2173 /* Shouldn't have super stripes in sequential zones */
2174 if (zoned && nr) {
2175 kfree(logical);
2176 btrfs_err(fs_info,
2177 "zoned: block group %llu must not contain super block",
2178 cache->start);
2179 return -EUCLEAN;
2180 }
2181
2182 while (nr--) {
2183 u64 len = min_t(u64, stripe_len,
2184 cache->start + cache->length - logical[nr]);
2185
2186 cache->bytes_super += len;
2187 ret = set_extent_bit(&fs_info->excluded_extents, logical[nr],
2188 logical[nr] + len - 1,
2189 EXTENT_UPTODATE, NULL);
2190 if (ret) {
2191 kfree(logical);
2192 return ret;
2193 }
2194 }
2195
2196 kfree(logical);
2197 }
2198 return 0;
2199 }
2200
btrfs_create_block_group_cache(struct btrfs_fs_info * fs_info,u64 start)2201 static struct btrfs_block_group *btrfs_create_block_group_cache(
2202 struct btrfs_fs_info *fs_info, u64 start)
2203 {
2204 struct btrfs_block_group *cache;
2205
2206 cache = kzalloc(sizeof(*cache), GFP_NOFS);
2207 if (!cache)
2208 return NULL;
2209
2210 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
2211 GFP_NOFS);
2212 if (!cache->free_space_ctl) {
2213 kfree(cache);
2214 return NULL;
2215 }
2216
2217 cache->start = start;
2218
2219 cache->fs_info = fs_info;
2220 cache->full_stripe_len = btrfs_full_stripe_len(fs_info, start);
2221
2222 cache->discard_index = BTRFS_DISCARD_INDEX_UNUSED;
2223
2224 refcount_set(&cache->refs, 1);
2225 spin_lock_init(&cache->lock);
2226 init_rwsem(&cache->data_rwsem);
2227 INIT_LIST_HEAD(&cache->list);
2228 INIT_LIST_HEAD(&cache->cluster_list);
2229 INIT_LIST_HEAD(&cache->bg_list);
2230 INIT_LIST_HEAD(&cache->ro_list);
2231 INIT_LIST_HEAD(&cache->discard_list);
2232 INIT_LIST_HEAD(&cache->dirty_list);
2233 INIT_LIST_HEAD(&cache->io_list);
2234 INIT_LIST_HEAD(&cache->active_bg_list);
2235 btrfs_init_free_space_ctl(cache, cache->free_space_ctl);
2236 atomic_set(&cache->frozen, 0);
2237 mutex_init(&cache->free_space_lock);
2238
2239 return cache;
2240 }
2241
2242 /*
2243 * Iterate all chunks and verify that each of them has the corresponding block
2244 * group
2245 */
check_chunk_block_group_mappings(struct btrfs_fs_info * fs_info)2246 static int check_chunk_block_group_mappings(struct btrfs_fs_info *fs_info)
2247 {
2248 struct extent_map_tree *map_tree = &fs_info->mapping_tree;
2249 struct extent_map *em;
2250 struct btrfs_block_group *bg;
2251 u64 start = 0;
2252 int ret = 0;
2253
2254 while (1) {
2255 read_lock(&map_tree->lock);
2256 /*
2257 * lookup_extent_mapping will return the first extent map
2258 * intersecting the range, so setting @len to 1 is enough to
2259 * get the first chunk.
2260 */
2261 em = lookup_extent_mapping(map_tree, start, 1);
2262 read_unlock(&map_tree->lock);
2263 if (!em)
2264 break;
2265
2266 bg = btrfs_lookup_block_group(fs_info, em->start);
2267 if (!bg) {
2268 btrfs_err(fs_info,
2269 "chunk start=%llu len=%llu doesn't have corresponding block group",
2270 em->start, em->len);
2271 ret = -EUCLEAN;
2272 free_extent_map(em);
2273 break;
2274 }
2275 if (bg->start != em->start || bg->length != em->len ||
2276 (bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK) !=
2277 (em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK)) {
2278 btrfs_err(fs_info,
2279 "chunk start=%llu len=%llu flags=0x%llx doesn't match block group start=%llu len=%llu flags=0x%llx",
2280 em->start, em->len,
2281 em->map_lookup->type & BTRFS_BLOCK_GROUP_TYPE_MASK,
2282 bg->start, bg->length,
2283 bg->flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
2284 ret = -EUCLEAN;
2285 free_extent_map(em);
2286 btrfs_put_block_group(bg);
2287 break;
2288 }
2289 start = em->start + em->len;
2290 free_extent_map(em);
2291 btrfs_put_block_group(bg);
2292 }
2293 return ret;
2294 }
2295
read_one_block_group(struct btrfs_fs_info * info,struct btrfs_block_group_item * bgi,const struct btrfs_key * key,int need_clear)2296 static int read_one_block_group(struct btrfs_fs_info *info,
2297 struct btrfs_block_group_item *bgi,
2298 const struct btrfs_key *key,
2299 int need_clear)
2300 {
2301 struct btrfs_block_group *cache;
2302 const bool mixed = btrfs_fs_incompat(info, MIXED_GROUPS);
2303 int ret;
2304
2305 ASSERT(key->type == BTRFS_BLOCK_GROUP_ITEM_KEY);
2306
2307 cache = btrfs_create_block_group_cache(info, key->objectid);
2308 if (!cache)
2309 return -ENOMEM;
2310
2311 cache->length = key->offset;
2312 cache->used = btrfs_stack_block_group_used(bgi);
2313 cache->commit_used = cache->used;
2314 cache->flags = btrfs_stack_block_group_flags(bgi);
2315 cache->global_root_id = btrfs_stack_block_group_chunk_objectid(bgi);
2316
2317 set_free_space_tree_thresholds(cache);
2318
2319 if (need_clear) {
2320 /*
2321 * When we mount with old space cache, we need to
2322 * set BTRFS_DC_CLEAR and set dirty flag.
2323 *
2324 * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
2325 * truncate the old free space cache inode and
2326 * setup a new one.
2327 * b) Setting 'dirty flag' makes sure that we flush
2328 * the new space cache info onto disk.
2329 */
2330 if (btrfs_test_opt(info, SPACE_CACHE))
2331 cache->disk_cache_state = BTRFS_DC_CLEAR;
2332 }
2333 if (!mixed && ((cache->flags & BTRFS_BLOCK_GROUP_METADATA) &&
2334 (cache->flags & BTRFS_BLOCK_GROUP_DATA))) {
2335 btrfs_err(info,
2336 "bg %llu is a mixed block group but filesystem hasn't enabled mixed block groups",
2337 cache->start);
2338 ret = -EINVAL;
2339 goto error;
2340 }
2341
2342 ret = btrfs_load_block_group_zone_info(cache, false);
2343 if (ret) {
2344 btrfs_err(info, "zoned: failed to load zone info of bg %llu",
2345 cache->start);
2346 goto error;
2347 }
2348
2349 /*
2350 * We need to exclude the super stripes now so that the space info has
2351 * super bytes accounted for, otherwise we'll think we have more space
2352 * than we actually do.
2353 */
2354 ret = exclude_super_stripes(cache);
2355 if (ret) {
2356 /* We may have excluded something, so call this just in case. */
2357 btrfs_free_excluded_extents(cache);
2358 goto error;
2359 }
2360
2361 /*
2362 * For zoned filesystem, space after the allocation offset is the only
2363 * free space for a block group. So, we don't need any caching work.
2364 * btrfs_calc_zone_unusable() will set the amount of free space and
2365 * zone_unusable space.
2366 *
2367 * For regular filesystem, check for two cases, either we are full, and
2368 * therefore don't need to bother with the caching work since we won't
2369 * find any space, or we are empty, and we can just add all the space
2370 * in and be done with it. This saves us _a_lot_ of time, particularly
2371 * in the full case.
2372 */
2373 if (btrfs_is_zoned(info)) {
2374 btrfs_calc_zone_unusable(cache);
2375 /* Should not have any excluded extents. Just in case, though. */
2376 btrfs_free_excluded_extents(cache);
2377 } else if (cache->length == cache->used) {
2378 cache->cached = BTRFS_CACHE_FINISHED;
2379 btrfs_free_excluded_extents(cache);
2380 } else if (cache->used == 0) {
2381 cache->cached = BTRFS_CACHE_FINISHED;
2382 ret = btrfs_add_new_free_space(cache, cache->start,
2383 cache->start + cache->length, NULL);
2384 btrfs_free_excluded_extents(cache);
2385 if (ret)
2386 goto error;
2387 }
2388
2389 ret = btrfs_add_block_group_cache(info, cache);
2390 if (ret) {
2391 btrfs_remove_free_space_cache(cache);
2392 goto error;
2393 }
2394 trace_btrfs_add_block_group(info, cache, 0);
2395 btrfs_add_bg_to_space_info(info, cache);
2396
2397 set_avail_alloc_bits(info, cache->flags);
2398 if (btrfs_chunk_writeable(info, cache->start)) {
2399 if (cache->used == 0) {
2400 ASSERT(list_empty(&cache->bg_list));
2401 if (btrfs_test_opt(info, DISCARD_ASYNC))
2402 btrfs_discard_queue_work(&info->discard_ctl, cache);
2403 else
2404 btrfs_mark_bg_unused(cache);
2405 }
2406 } else {
2407 inc_block_group_ro(cache, 1);
2408 }
2409
2410 return 0;
2411 error:
2412 btrfs_put_block_group(cache);
2413 return ret;
2414 }
2415
fill_dummy_bgs(struct btrfs_fs_info * fs_info)2416 static int fill_dummy_bgs(struct btrfs_fs_info *fs_info)
2417 {
2418 struct extent_map_tree *em_tree = &fs_info->mapping_tree;
2419 struct rb_node *node;
2420 int ret = 0;
2421
2422 for (node = rb_first_cached(&em_tree->map); node; node = rb_next(node)) {
2423 struct extent_map *em;
2424 struct map_lookup *map;
2425 struct btrfs_block_group *bg;
2426
2427 em = rb_entry(node, struct extent_map, rb_node);
2428 map = em->map_lookup;
2429 bg = btrfs_create_block_group_cache(fs_info, em->start);
2430 if (!bg) {
2431 ret = -ENOMEM;
2432 break;
2433 }
2434
2435 /* Fill dummy cache as FULL */
2436 bg->length = em->len;
2437 bg->flags = map->type;
2438 bg->cached = BTRFS_CACHE_FINISHED;
2439 bg->used = em->len;
2440 bg->flags = map->type;
2441 ret = btrfs_add_block_group_cache(fs_info, bg);
2442 /*
2443 * We may have some valid block group cache added already, in
2444 * that case we skip to the next one.
2445 */
2446 if (ret == -EEXIST) {
2447 ret = 0;
2448 btrfs_put_block_group(bg);
2449 continue;
2450 }
2451
2452 if (ret) {
2453 btrfs_remove_free_space_cache(bg);
2454 btrfs_put_block_group(bg);
2455 break;
2456 }
2457
2458 btrfs_add_bg_to_space_info(fs_info, bg);
2459
2460 set_avail_alloc_bits(fs_info, bg->flags);
2461 }
2462 if (!ret)
2463 btrfs_init_global_block_rsv(fs_info);
2464 return ret;
2465 }
2466
btrfs_read_block_groups(struct btrfs_fs_info * info)2467 int btrfs_read_block_groups(struct btrfs_fs_info *info)
2468 {
2469 struct btrfs_root *root = btrfs_block_group_root(info);
2470 struct btrfs_path *path;
2471 int ret;
2472 struct btrfs_block_group *cache;
2473 struct btrfs_space_info *space_info;
2474 struct btrfs_key key;
2475 int need_clear = 0;
2476 u64 cache_gen;
2477
2478 /*
2479 * Either no extent root (with ibadroots rescue option) or we have
2480 * unsupported RO options. The fs can never be mounted read-write, so no
2481 * need to waste time searching block group items.
2482 *
2483 * This also allows new extent tree related changes to be RO compat,
2484 * no need for a full incompat flag.
2485 */
2486 if (!root || (btrfs_super_compat_ro_flags(info->super_copy) &
2487 ~BTRFS_FEATURE_COMPAT_RO_SUPP))
2488 return fill_dummy_bgs(info);
2489
2490 key.objectid = 0;
2491 key.offset = 0;
2492 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2493 path = btrfs_alloc_path();
2494 if (!path)
2495 return -ENOMEM;
2496
2497 cache_gen = btrfs_super_cache_generation(info->super_copy);
2498 if (btrfs_test_opt(info, SPACE_CACHE) &&
2499 btrfs_super_generation(info->super_copy) != cache_gen)
2500 need_clear = 1;
2501 if (btrfs_test_opt(info, CLEAR_CACHE))
2502 need_clear = 1;
2503
2504 while (1) {
2505 struct btrfs_block_group_item bgi;
2506 struct extent_buffer *leaf;
2507 int slot;
2508
2509 ret = find_first_block_group(info, path, &key);
2510 if (ret > 0)
2511 break;
2512 if (ret != 0)
2513 goto error;
2514
2515 leaf = path->nodes[0];
2516 slot = path->slots[0];
2517
2518 read_extent_buffer(leaf, &bgi, btrfs_item_ptr_offset(leaf, slot),
2519 sizeof(bgi));
2520
2521 btrfs_item_key_to_cpu(leaf, &key, slot);
2522 btrfs_release_path(path);
2523 ret = read_one_block_group(info, &bgi, &key, need_clear);
2524 if (ret < 0)
2525 goto error;
2526 key.objectid += key.offset;
2527 key.offset = 0;
2528 }
2529 btrfs_release_path(path);
2530
2531 list_for_each_entry(space_info, &info->space_info, list) {
2532 int i;
2533
2534 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
2535 if (list_empty(&space_info->block_groups[i]))
2536 continue;
2537 cache = list_first_entry(&space_info->block_groups[i],
2538 struct btrfs_block_group,
2539 list);
2540 btrfs_sysfs_add_block_group_type(cache);
2541 }
2542
2543 if (!(btrfs_get_alloc_profile(info, space_info->flags) &
2544 (BTRFS_BLOCK_GROUP_RAID10 |
2545 BTRFS_BLOCK_GROUP_RAID1_MASK |
2546 BTRFS_BLOCK_GROUP_RAID56_MASK |
2547 BTRFS_BLOCK_GROUP_DUP)))
2548 continue;
2549 /*
2550 * Avoid allocating from un-mirrored block group if there are
2551 * mirrored block groups.
2552 */
2553 list_for_each_entry(cache,
2554 &space_info->block_groups[BTRFS_RAID_RAID0],
2555 list)
2556 inc_block_group_ro(cache, 1);
2557 list_for_each_entry(cache,
2558 &space_info->block_groups[BTRFS_RAID_SINGLE],
2559 list)
2560 inc_block_group_ro(cache, 1);
2561 }
2562
2563 btrfs_init_global_block_rsv(info);
2564 ret = check_chunk_block_group_mappings(info);
2565 error:
2566 btrfs_free_path(path);
2567 /*
2568 * We've hit some error while reading the extent tree, and have
2569 * rescue=ibadroots mount option.
2570 * Try to fill the tree using dummy block groups so that the user can
2571 * continue to mount and grab their data.
2572 */
2573 if (ret && btrfs_test_opt(info, IGNOREBADROOTS))
2574 ret = fill_dummy_bgs(info);
2575 return ret;
2576 }
2577
2578 /*
2579 * This function, insert_block_group_item(), belongs to the phase 2 of chunk
2580 * allocation.
2581 *
2582 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2583 * phases.
2584 */
insert_block_group_item(struct btrfs_trans_handle * trans,struct btrfs_block_group * block_group)2585 static int insert_block_group_item(struct btrfs_trans_handle *trans,
2586 struct btrfs_block_group *block_group)
2587 {
2588 struct btrfs_fs_info *fs_info = trans->fs_info;
2589 struct btrfs_block_group_item bgi;
2590 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2591 struct btrfs_key key;
2592 u64 old_commit_used;
2593 int ret;
2594
2595 spin_lock(&block_group->lock);
2596 btrfs_set_stack_block_group_used(&bgi, block_group->used);
2597 btrfs_set_stack_block_group_chunk_objectid(&bgi,
2598 block_group->global_root_id);
2599 btrfs_set_stack_block_group_flags(&bgi, block_group->flags);
2600 old_commit_used = block_group->commit_used;
2601 block_group->commit_used = block_group->used;
2602 key.objectid = block_group->start;
2603 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
2604 key.offset = block_group->length;
2605 spin_unlock(&block_group->lock);
2606
2607 ret = btrfs_insert_item(trans, root, &key, &bgi, sizeof(bgi));
2608 if (ret < 0) {
2609 spin_lock(&block_group->lock);
2610 block_group->commit_used = old_commit_used;
2611 spin_unlock(&block_group->lock);
2612 }
2613
2614 return ret;
2615 }
2616
insert_dev_extent(struct btrfs_trans_handle * trans,struct btrfs_device * device,u64 chunk_offset,u64 start,u64 num_bytes)2617 static int insert_dev_extent(struct btrfs_trans_handle *trans,
2618 struct btrfs_device *device, u64 chunk_offset,
2619 u64 start, u64 num_bytes)
2620 {
2621 struct btrfs_fs_info *fs_info = device->fs_info;
2622 struct btrfs_root *root = fs_info->dev_root;
2623 struct btrfs_path *path;
2624 struct btrfs_dev_extent *extent;
2625 struct extent_buffer *leaf;
2626 struct btrfs_key key;
2627 int ret;
2628
2629 WARN_ON(!test_bit(BTRFS_DEV_STATE_IN_FS_METADATA, &device->dev_state));
2630 WARN_ON(test_bit(BTRFS_DEV_STATE_REPLACE_TGT, &device->dev_state));
2631 path = btrfs_alloc_path();
2632 if (!path)
2633 return -ENOMEM;
2634
2635 key.objectid = device->devid;
2636 key.type = BTRFS_DEV_EXTENT_KEY;
2637 key.offset = start;
2638 ret = btrfs_insert_empty_item(trans, root, path, &key, sizeof(*extent));
2639 if (ret)
2640 goto out;
2641
2642 leaf = path->nodes[0];
2643 extent = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_dev_extent);
2644 btrfs_set_dev_extent_chunk_tree(leaf, extent, BTRFS_CHUNK_TREE_OBJECTID);
2645 btrfs_set_dev_extent_chunk_objectid(leaf, extent,
2646 BTRFS_FIRST_CHUNK_TREE_OBJECTID);
2647 btrfs_set_dev_extent_chunk_offset(leaf, extent, chunk_offset);
2648
2649 btrfs_set_dev_extent_length(leaf, extent, num_bytes);
2650 btrfs_mark_buffer_dirty(trans, leaf);
2651 out:
2652 btrfs_free_path(path);
2653 return ret;
2654 }
2655
2656 /*
2657 * This function belongs to phase 2.
2658 *
2659 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2660 * phases.
2661 */
insert_dev_extents(struct btrfs_trans_handle * trans,u64 chunk_offset,u64 chunk_size)2662 static int insert_dev_extents(struct btrfs_trans_handle *trans,
2663 u64 chunk_offset, u64 chunk_size)
2664 {
2665 struct btrfs_fs_info *fs_info = trans->fs_info;
2666 struct btrfs_device *device;
2667 struct extent_map *em;
2668 struct map_lookup *map;
2669 u64 dev_offset;
2670 u64 stripe_size;
2671 int i;
2672 int ret = 0;
2673
2674 em = btrfs_get_chunk_map(fs_info, chunk_offset, chunk_size);
2675 if (IS_ERR(em))
2676 return PTR_ERR(em);
2677
2678 map = em->map_lookup;
2679 stripe_size = em->orig_block_len;
2680
2681 /*
2682 * Take the device list mutex to prevent races with the final phase of
2683 * a device replace operation that replaces the device object associated
2684 * with the map's stripes, because the device object's id can change
2685 * at any time during that final phase of the device replace operation
2686 * (dev-replace.c:btrfs_dev_replace_finishing()), so we could grab the
2687 * replaced device and then see it with an ID of BTRFS_DEV_REPLACE_DEVID,
2688 * resulting in persisting a device extent item with such ID.
2689 */
2690 mutex_lock(&fs_info->fs_devices->device_list_mutex);
2691 for (i = 0; i < map->num_stripes; i++) {
2692 device = map->stripes[i].dev;
2693 dev_offset = map->stripes[i].physical;
2694
2695 ret = insert_dev_extent(trans, device, chunk_offset, dev_offset,
2696 stripe_size);
2697 if (ret)
2698 break;
2699 }
2700 mutex_unlock(&fs_info->fs_devices->device_list_mutex);
2701
2702 free_extent_map(em);
2703 return ret;
2704 }
2705
2706 /*
2707 * This function, btrfs_create_pending_block_groups(), belongs to the phase 2 of
2708 * chunk allocation.
2709 *
2710 * See the comment at btrfs_chunk_alloc() for details about the chunk allocation
2711 * phases.
2712 */
btrfs_create_pending_block_groups(struct btrfs_trans_handle * trans)2713 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans)
2714 {
2715 struct btrfs_fs_info *fs_info = trans->fs_info;
2716 struct btrfs_block_group *block_group;
2717 int ret = 0;
2718
2719 while (!list_empty(&trans->new_bgs)) {
2720 int index;
2721
2722 block_group = list_first_entry(&trans->new_bgs,
2723 struct btrfs_block_group,
2724 bg_list);
2725 if (ret)
2726 goto next;
2727
2728 index = btrfs_bg_flags_to_raid_index(block_group->flags);
2729
2730 ret = insert_block_group_item(trans, block_group);
2731 if (ret)
2732 btrfs_abort_transaction(trans, ret);
2733 if (!test_bit(BLOCK_GROUP_FLAG_CHUNK_ITEM_INSERTED,
2734 &block_group->runtime_flags)) {
2735 mutex_lock(&fs_info->chunk_mutex);
2736 ret = btrfs_chunk_alloc_add_chunk_item(trans, block_group);
2737 mutex_unlock(&fs_info->chunk_mutex);
2738 if (ret)
2739 btrfs_abort_transaction(trans, ret);
2740 }
2741 ret = insert_dev_extents(trans, block_group->start,
2742 block_group->length);
2743 if (ret)
2744 btrfs_abort_transaction(trans, ret);
2745 add_block_group_free_space(trans, block_group);
2746
2747 /*
2748 * If we restriped during balance, we may have added a new raid
2749 * type, so now add the sysfs entries when it is safe to do so.
2750 * We don't have to worry about locking here as it's handled in
2751 * btrfs_sysfs_add_block_group_type.
2752 */
2753 if (block_group->space_info->block_group_kobjs[index] == NULL)
2754 btrfs_sysfs_add_block_group_type(block_group);
2755
2756 /* Already aborted the transaction if it failed. */
2757 next:
2758 btrfs_delayed_refs_rsv_release(fs_info, 1);
2759 list_del_init(&block_group->bg_list);
2760 clear_bit(BLOCK_GROUP_FLAG_NEW, &block_group->runtime_flags);
2761 }
2762 btrfs_trans_release_chunk_metadata(trans);
2763 }
2764
2765 /*
2766 * For extent tree v2 we use the block_group_item->chunk_offset to point at our
2767 * global root id. For v1 it's always set to BTRFS_FIRST_CHUNK_TREE_OBJECTID.
2768 */
calculate_global_root_id(struct btrfs_fs_info * fs_info,u64 offset)2769 static u64 calculate_global_root_id(struct btrfs_fs_info *fs_info, u64 offset)
2770 {
2771 u64 div = SZ_1G;
2772 u64 index;
2773
2774 if (!btrfs_fs_incompat(fs_info, EXTENT_TREE_V2))
2775 return BTRFS_FIRST_CHUNK_TREE_OBJECTID;
2776
2777 /* If we have a smaller fs index based on 128MiB. */
2778 if (btrfs_super_total_bytes(fs_info->super_copy) <= (SZ_1G * 10ULL))
2779 div = SZ_128M;
2780
2781 offset = div64_u64(offset, div);
2782 div64_u64_rem(offset, fs_info->nr_global_roots, &index);
2783 return index;
2784 }
2785
btrfs_make_block_group(struct btrfs_trans_handle * trans,u64 type,u64 chunk_offset,u64 size)2786 struct btrfs_block_group *btrfs_make_block_group(struct btrfs_trans_handle *trans,
2787 u64 type,
2788 u64 chunk_offset, u64 size)
2789 {
2790 struct btrfs_fs_info *fs_info = trans->fs_info;
2791 struct btrfs_block_group *cache;
2792 int ret;
2793
2794 btrfs_set_log_full_commit(trans);
2795
2796 cache = btrfs_create_block_group_cache(fs_info, chunk_offset);
2797 if (!cache)
2798 return ERR_PTR(-ENOMEM);
2799
2800 /*
2801 * Mark it as new before adding it to the rbtree of block groups or any
2802 * list, so that no other task finds it and calls btrfs_mark_bg_unused()
2803 * before the new flag is set.
2804 */
2805 set_bit(BLOCK_GROUP_FLAG_NEW, &cache->runtime_flags);
2806
2807 cache->length = size;
2808 set_free_space_tree_thresholds(cache);
2809 cache->flags = type;
2810 cache->cached = BTRFS_CACHE_FINISHED;
2811 cache->global_root_id = calculate_global_root_id(fs_info, cache->start);
2812
2813 if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
2814 set_bit(BLOCK_GROUP_FLAG_NEEDS_FREE_SPACE, &cache->runtime_flags);
2815
2816 ret = btrfs_load_block_group_zone_info(cache, true);
2817 if (ret) {
2818 btrfs_put_block_group(cache);
2819 return ERR_PTR(ret);
2820 }
2821
2822 ret = exclude_super_stripes(cache);
2823 if (ret) {
2824 /* We may have excluded something, so call this just in case */
2825 btrfs_free_excluded_extents(cache);
2826 btrfs_put_block_group(cache);
2827 return ERR_PTR(ret);
2828 }
2829
2830 ret = btrfs_add_new_free_space(cache, chunk_offset, chunk_offset + size, NULL);
2831 btrfs_free_excluded_extents(cache);
2832 if (ret) {
2833 btrfs_put_block_group(cache);
2834 return ERR_PTR(ret);
2835 }
2836
2837 /*
2838 * Ensure the corresponding space_info object is created and
2839 * assigned to our block group. We want our bg to be added to the rbtree
2840 * with its ->space_info set.
2841 */
2842 cache->space_info = btrfs_find_space_info(fs_info, cache->flags);
2843 ASSERT(cache->space_info);
2844
2845 ret = btrfs_add_block_group_cache(fs_info, cache);
2846 if (ret) {
2847 btrfs_remove_free_space_cache(cache);
2848 btrfs_put_block_group(cache);
2849 return ERR_PTR(ret);
2850 }
2851
2852 /*
2853 * Now that our block group has its ->space_info set and is inserted in
2854 * the rbtree, update the space info's counters.
2855 */
2856 trace_btrfs_add_block_group(fs_info, cache, 1);
2857 btrfs_add_bg_to_space_info(fs_info, cache);
2858 btrfs_update_global_block_rsv(fs_info);
2859
2860 #ifdef CONFIG_BTRFS_DEBUG
2861 if (btrfs_should_fragment_free_space(cache)) {
2862 cache->space_info->bytes_used += size >> 1;
2863 fragment_free_space(cache);
2864 }
2865 #endif
2866
2867 list_add_tail(&cache->bg_list, &trans->new_bgs);
2868 trans->delayed_ref_updates++;
2869 btrfs_update_delayed_refs_rsv(trans);
2870
2871 set_avail_alloc_bits(fs_info, type);
2872 return cache;
2873 }
2874
2875 /*
2876 * Mark one block group RO, can be called several times for the same block
2877 * group.
2878 *
2879 * @cache: the destination block group
2880 * @do_chunk_alloc: whether need to do chunk pre-allocation, this is to
2881 * ensure we still have some free space after marking this
2882 * block group RO.
2883 */
btrfs_inc_block_group_ro(struct btrfs_block_group * cache,bool do_chunk_alloc)2884 int btrfs_inc_block_group_ro(struct btrfs_block_group *cache,
2885 bool do_chunk_alloc)
2886 {
2887 struct btrfs_fs_info *fs_info = cache->fs_info;
2888 struct btrfs_trans_handle *trans;
2889 struct btrfs_root *root = btrfs_block_group_root(fs_info);
2890 u64 alloc_flags;
2891 int ret;
2892 bool dirty_bg_running;
2893
2894 /*
2895 * This can only happen when we are doing read-only scrub on read-only
2896 * mount.
2897 * In that case we should not start a new transaction on read-only fs.
2898 * Thus here we skip all chunk allocations.
2899 */
2900 if (sb_rdonly(fs_info->sb)) {
2901 mutex_lock(&fs_info->ro_block_group_mutex);
2902 ret = inc_block_group_ro(cache, 0);
2903 mutex_unlock(&fs_info->ro_block_group_mutex);
2904 return ret;
2905 }
2906
2907 do {
2908 trans = btrfs_join_transaction(root);
2909 if (IS_ERR(trans))
2910 return PTR_ERR(trans);
2911
2912 dirty_bg_running = false;
2913
2914 /*
2915 * We're not allowed to set block groups readonly after the dirty
2916 * block group cache has started writing. If it already started,
2917 * back off and let this transaction commit.
2918 */
2919 mutex_lock(&fs_info->ro_block_group_mutex);
2920 if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
2921 u64 transid = trans->transid;
2922
2923 mutex_unlock(&fs_info->ro_block_group_mutex);
2924 btrfs_end_transaction(trans);
2925
2926 ret = btrfs_wait_for_commit(fs_info, transid);
2927 if (ret)
2928 return ret;
2929 dirty_bg_running = true;
2930 }
2931 } while (dirty_bg_running);
2932
2933 if (do_chunk_alloc) {
2934 /*
2935 * If we are changing raid levels, try to allocate a
2936 * corresponding block group with the new raid level.
2937 */
2938 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2939 if (alloc_flags != cache->flags) {
2940 ret = btrfs_chunk_alloc(trans, alloc_flags,
2941 CHUNK_ALLOC_FORCE);
2942 /*
2943 * ENOSPC is allowed here, we may have enough space
2944 * already allocated at the new raid level to carry on
2945 */
2946 if (ret == -ENOSPC)
2947 ret = 0;
2948 if (ret < 0)
2949 goto out;
2950 }
2951 }
2952
2953 ret = inc_block_group_ro(cache, 0);
2954 if (!ret)
2955 goto out;
2956 if (ret == -ETXTBSY)
2957 goto unlock_out;
2958
2959 /*
2960 * Skip chunk alloction if the bg is SYSTEM, this is to avoid system
2961 * chunk allocation storm to exhaust the system chunk array. Otherwise
2962 * we still want to try our best to mark the block group read-only.
2963 */
2964 if (!do_chunk_alloc && ret == -ENOSPC &&
2965 (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM))
2966 goto unlock_out;
2967
2968 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->space_info->flags);
2969 ret = btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
2970 if (ret < 0)
2971 goto out;
2972 /*
2973 * We have allocated a new chunk. We also need to activate that chunk to
2974 * grant metadata tickets for zoned filesystem.
2975 */
2976 ret = btrfs_zoned_activate_one_bg(fs_info, cache->space_info, true);
2977 if (ret < 0)
2978 goto out;
2979
2980 ret = inc_block_group_ro(cache, 0);
2981 if (ret == -ETXTBSY)
2982 goto unlock_out;
2983 out:
2984 if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
2985 alloc_flags = btrfs_get_alloc_profile(fs_info, cache->flags);
2986 mutex_lock(&fs_info->chunk_mutex);
2987 check_system_chunk(trans, alloc_flags);
2988 mutex_unlock(&fs_info->chunk_mutex);
2989 }
2990 unlock_out:
2991 mutex_unlock(&fs_info->ro_block_group_mutex);
2992
2993 btrfs_end_transaction(trans);
2994 return ret;
2995 }
2996
btrfs_dec_block_group_ro(struct btrfs_block_group * cache)2997 void btrfs_dec_block_group_ro(struct btrfs_block_group *cache)
2998 {
2999 struct btrfs_space_info *sinfo = cache->space_info;
3000 u64 num_bytes;
3001
3002 BUG_ON(!cache->ro);
3003
3004 spin_lock(&sinfo->lock);
3005 spin_lock(&cache->lock);
3006 if (!--cache->ro) {
3007 if (btrfs_is_zoned(cache->fs_info)) {
3008 /* Migrate zone_unusable bytes back */
3009 cache->zone_unusable =
3010 (cache->alloc_offset - cache->used) +
3011 (cache->length - cache->zone_capacity);
3012 sinfo->bytes_zone_unusable += cache->zone_unusable;
3013 sinfo->bytes_readonly -= cache->zone_unusable;
3014 }
3015 num_bytes = cache->length - cache->reserved -
3016 cache->pinned - cache->bytes_super -
3017 cache->zone_unusable - cache->used;
3018 sinfo->bytes_readonly -= num_bytes;
3019 list_del_init(&cache->ro_list);
3020 }
3021 spin_unlock(&cache->lock);
3022 spin_unlock(&sinfo->lock);
3023 }
3024
update_block_group_item(struct btrfs_trans_handle * trans,struct btrfs_path * path,struct btrfs_block_group * cache)3025 static int update_block_group_item(struct btrfs_trans_handle *trans,
3026 struct btrfs_path *path,
3027 struct btrfs_block_group *cache)
3028 {
3029 struct btrfs_fs_info *fs_info = trans->fs_info;
3030 int ret;
3031 struct btrfs_root *root = btrfs_block_group_root(fs_info);
3032 unsigned long bi;
3033 struct extent_buffer *leaf;
3034 struct btrfs_block_group_item bgi;
3035 struct btrfs_key key;
3036 u64 old_commit_used;
3037 u64 used;
3038
3039 /*
3040 * Block group items update can be triggered out of commit transaction
3041 * critical section, thus we need a consistent view of used bytes.
3042 * We cannot use cache->used directly outside of the spin lock, as it
3043 * may be changed.
3044 */
3045 spin_lock(&cache->lock);
3046 old_commit_used = cache->commit_used;
3047 used = cache->used;
3048 /* No change in used bytes, can safely skip it. */
3049 if (cache->commit_used == used) {
3050 spin_unlock(&cache->lock);
3051 return 0;
3052 }
3053 cache->commit_used = used;
3054 spin_unlock(&cache->lock);
3055
3056 key.objectid = cache->start;
3057 key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
3058 key.offset = cache->length;
3059
3060 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
3061 if (ret) {
3062 if (ret > 0)
3063 ret = -ENOENT;
3064 goto fail;
3065 }
3066
3067 leaf = path->nodes[0];
3068 bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3069 btrfs_set_stack_block_group_used(&bgi, used);
3070 btrfs_set_stack_block_group_chunk_objectid(&bgi,
3071 cache->global_root_id);
3072 btrfs_set_stack_block_group_flags(&bgi, cache->flags);
3073 write_extent_buffer(leaf, &bgi, bi, sizeof(bgi));
3074 btrfs_mark_buffer_dirty(trans, leaf);
3075 fail:
3076 btrfs_release_path(path);
3077 /*
3078 * We didn't update the block group item, need to revert commit_used
3079 * unless the block group item didn't exist yet - this is to prevent a
3080 * race with a concurrent insertion of the block group item, with
3081 * insert_block_group_item(), that happened just after we attempted to
3082 * update. In that case we would reset commit_used to 0 just after the
3083 * insertion set it to a value greater than 0 - if the block group later
3084 * becomes with 0 used bytes, we would incorrectly skip its update.
3085 */
3086 if (ret < 0 && ret != -ENOENT) {
3087 spin_lock(&cache->lock);
3088 cache->commit_used = old_commit_used;
3089 spin_unlock(&cache->lock);
3090 }
3091 return ret;
3092
3093 }
3094
cache_save_setup(struct btrfs_block_group * block_group,struct btrfs_trans_handle * trans,struct btrfs_path * path)3095 static int cache_save_setup(struct btrfs_block_group *block_group,
3096 struct btrfs_trans_handle *trans,
3097 struct btrfs_path *path)
3098 {
3099 struct btrfs_fs_info *fs_info = block_group->fs_info;
3100 struct btrfs_root *root = fs_info->tree_root;
3101 struct inode *inode = NULL;
3102 struct extent_changeset *data_reserved = NULL;
3103 u64 alloc_hint = 0;
3104 int dcs = BTRFS_DC_ERROR;
3105 u64 cache_size = 0;
3106 int retries = 0;
3107 int ret = 0;
3108
3109 if (!btrfs_test_opt(fs_info, SPACE_CACHE))
3110 return 0;
3111
3112 /*
3113 * If this block group is smaller than 100 megs don't bother caching the
3114 * block group.
3115 */
3116 if (block_group->length < (100 * SZ_1M)) {
3117 spin_lock(&block_group->lock);
3118 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3119 spin_unlock(&block_group->lock);
3120 return 0;
3121 }
3122
3123 if (TRANS_ABORTED(trans))
3124 return 0;
3125 again:
3126 inode = lookup_free_space_inode(block_group, path);
3127 if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3128 ret = PTR_ERR(inode);
3129 btrfs_release_path(path);
3130 goto out;
3131 }
3132
3133 if (IS_ERR(inode)) {
3134 BUG_ON(retries);
3135 retries++;
3136
3137 if (block_group->ro)
3138 goto out_free;
3139
3140 ret = create_free_space_inode(trans, block_group, path);
3141 if (ret)
3142 goto out_free;
3143 goto again;
3144 }
3145
3146 /*
3147 * We want to set the generation to 0, that way if anything goes wrong
3148 * from here on out we know not to trust this cache when we load up next
3149 * time.
3150 */
3151 BTRFS_I(inode)->generation = 0;
3152 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3153 if (ret) {
3154 /*
3155 * So theoretically we could recover from this, simply set the
3156 * super cache generation to 0 so we know to invalidate the
3157 * cache, but then we'd have to keep track of the block groups
3158 * that fail this way so we know we _have_ to reset this cache
3159 * before the next commit or risk reading stale cache. So to
3160 * limit our exposure to horrible edge cases lets just abort the
3161 * transaction, this only happens in really bad situations
3162 * anyway.
3163 */
3164 btrfs_abort_transaction(trans, ret);
3165 goto out_put;
3166 }
3167 WARN_ON(ret);
3168
3169 /* We've already setup this transaction, go ahead and exit */
3170 if (block_group->cache_generation == trans->transid &&
3171 i_size_read(inode)) {
3172 dcs = BTRFS_DC_SETUP;
3173 goto out_put;
3174 }
3175
3176 if (i_size_read(inode) > 0) {
3177 ret = btrfs_check_trunc_cache_free_space(fs_info,
3178 &fs_info->global_block_rsv);
3179 if (ret)
3180 goto out_put;
3181
3182 ret = btrfs_truncate_free_space_cache(trans, NULL, inode);
3183 if (ret)
3184 goto out_put;
3185 }
3186
3187 spin_lock(&block_group->lock);
3188 if (block_group->cached != BTRFS_CACHE_FINISHED ||
3189 !btrfs_test_opt(fs_info, SPACE_CACHE)) {
3190 /*
3191 * don't bother trying to write stuff out _if_
3192 * a) we're not cached,
3193 * b) we're with nospace_cache mount option,
3194 * c) we're with v2 space_cache (FREE_SPACE_TREE).
3195 */
3196 dcs = BTRFS_DC_WRITTEN;
3197 spin_unlock(&block_group->lock);
3198 goto out_put;
3199 }
3200 spin_unlock(&block_group->lock);
3201
3202 /*
3203 * We hit an ENOSPC when setting up the cache in this transaction, just
3204 * skip doing the setup, we've already cleared the cache so we're safe.
3205 */
3206 if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3207 ret = -ENOSPC;
3208 goto out_put;
3209 }
3210
3211 /*
3212 * Try to preallocate enough space based on how big the block group is.
3213 * Keep in mind this has to include any pinned space which could end up
3214 * taking up quite a bit since it's not folded into the other space
3215 * cache.
3216 */
3217 cache_size = div_u64(block_group->length, SZ_256M);
3218 if (!cache_size)
3219 cache_size = 1;
3220
3221 cache_size *= 16;
3222 cache_size *= fs_info->sectorsize;
3223
3224 ret = btrfs_check_data_free_space(BTRFS_I(inode), &data_reserved, 0,
3225 cache_size, false);
3226 if (ret)
3227 goto out_put;
3228
3229 ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, cache_size,
3230 cache_size, cache_size,
3231 &alloc_hint);
3232 /*
3233 * Our cache requires contiguous chunks so that we don't modify a bunch
3234 * of metadata or split extents when writing the cache out, which means
3235 * we can enospc if we are heavily fragmented in addition to just normal
3236 * out of space conditions. So if we hit this just skip setting up any
3237 * other block groups for this transaction, maybe we'll unpin enough
3238 * space the next time around.
3239 */
3240 if (!ret)
3241 dcs = BTRFS_DC_SETUP;
3242 else if (ret == -ENOSPC)
3243 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3244
3245 out_put:
3246 iput(inode);
3247 out_free:
3248 btrfs_release_path(path);
3249 out:
3250 spin_lock(&block_group->lock);
3251 if (!ret && dcs == BTRFS_DC_SETUP)
3252 block_group->cache_generation = trans->transid;
3253 block_group->disk_cache_state = dcs;
3254 spin_unlock(&block_group->lock);
3255
3256 extent_changeset_free(data_reserved);
3257 return ret;
3258 }
3259
btrfs_setup_space_cache(struct btrfs_trans_handle * trans)3260 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans)
3261 {
3262 struct btrfs_fs_info *fs_info = trans->fs_info;
3263 struct btrfs_block_group *cache, *tmp;
3264 struct btrfs_transaction *cur_trans = trans->transaction;
3265 struct btrfs_path *path;
3266
3267 if (list_empty(&cur_trans->dirty_bgs) ||
3268 !btrfs_test_opt(fs_info, SPACE_CACHE))
3269 return 0;
3270
3271 path = btrfs_alloc_path();
3272 if (!path)
3273 return -ENOMEM;
3274
3275 /* Could add new block groups, use _safe just in case */
3276 list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3277 dirty_list) {
3278 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3279 cache_save_setup(cache, trans, path);
3280 }
3281
3282 btrfs_free_path(path);
3283 return 0;
3284 }
3285
3286 /*
3287 * Transaction commit does final block group cache writeback during a critical
3288 * section where nothing is allowed to change the FS. This is required in
3289 * order for the cache to actually match the block group, but can introduce a
3290 * lot of latency into the commit.
3291 *
3292 * So, btrfs_start_dirty_block_groups is here to kick off block group cache IO.
3293 * There's a chance we'll have to redo some of it if the block group changes
3294 * again during the commit, but it greatly reduces the commit latency by
3295 * getting rid of the easy block groups while we're still allowing others to
3296 * join the commit.
3297 */
btrfs_start_dirty_block_groups(struct btrfs_trans_handle * trans)3298 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans)
3299 {
3300 struct btrfs_fs_info *fs_info = trans->fs_info;
3301 struct btrfs_block_group *cache;
3302 struct btrfs_transaction *cur_trans = trans->transaction;
3303 int ret = 0;
3304 int should_put;
3305 struct btrfs_path *path = NULL;
3306 LIST_HEAD(dirty);
3307 struct list_head *io = &cur_trans->io_bgs;
3308 int loops = 0;
3309
3310 spin_lock(&cur_trans->dirty_bgs_lock);
3311 if (list_empty(&cur_trans->dirty_bgs)) {
3312 spin_unlock(&cur_trans->dirty_bgs_lock);
3313 return 0;
3314 }
3315 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3316 spin_unlock(&cur_trans->dirty_bgs_lock);
3317
3318 again:
3319 /* Make sure all the block groups on our dirty list actually exist */
3320 btrfs_create_pending_block_groups(trans);
3321
3322 if (!path) {
3323 path = btrfs_alloc_path();
3324 if (!path) {
3325 ret = -ENOMEM;
3326 goto out;
3327 }
3328 }
3329
3330 /*
3331 * cache_write_mutex is here only to save us from balance or automatic
3332 * removal of empty block groups deleting this block group while we are
3333 * writing out the cache
3334 */
3335 mutex_lock(&trans->transaction->cache_write_mutex);
3336 while (!list_empty(&dirty)) {
3337 bool drop_reserve = true;
3338
3339 cache = list_first_entry(&dirty, struct btrfs_block_group,
3340 dirty_list);
3341 /*
3342 * This can happen if something re-dirties a block group that
3343 * is already under IO. Just wait for it to finish and then do
3344 * it all again
3345 */
3346 if (!list_empty(&cache->io_list)) {
3347 list_del_init(&cache->io_list);
3348 btrfs_wait_cache_io(trans, cache, path);
3349 btrfs_put_block_group(cache);
3350 }
3351
3352
3353 /*
3354 * btrfs_wait_cache_io uses the cache->dirty_list to decide if
3355 * it should update the cache_state. Don't delete until after
3356 * we wait.
3357 *
3358 * Since we're not running in the commit critical section
3359 * we need the dirty_bgs_lock to protect from update_block_group
3360 */
3361 spin_lock(&cur_trans->dirty_bgs_lock);
3362 list_del_init(&cache->dirty_list);
3363 spin_unlock(&cur_trans->dirty_bgs_lock);
3364
3365 should_put = 1;
3366
3367 cache_save_setup(cache, trans, path);
3368
3369 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3370 cache->io_ctl.inode = NULL;
3371 ret = btrfs_write_out_cache(trans, cache, path);
3372 if (ret == 0 && cache->io_ctl.inode) {
3373 should_put = 0;
3374
3375 /*
3376 * The cache_write_mutex is protecting the
3377 * io_list, also refer to the definition of
3378 * btrfs_transaction::io_bgs for more details
3379 */
3380 list_add_tail(&cache->io_list, io);
3381 } else {
3382 /*
3383 * If we failed to write the cache, the
3384 * generation will be bad and life goes on
3385 */
3386 ret = 0;
3387 }
3388 }
3389 if (!ret) {
3390 ret = update_block_group_item(trans, path, cache);
3391 /*
3392 * Our block group might still be attached to the list
3393 * of new block groups in the transaction handle of some
3394 * other task (struct btrfs_trans_handle->new_bgs). This
3395 * means its block group item isn't yet in the extent
3396 * tree. If this happens ignore the error, as we will
3397 * try again later in the critical section of the
3398 * transaction commit.
3399 */
3400 if (ret == -ENOENT) {
3401 ret = 0;
3402 spin_lock(&cur_trans->dirty_bgs_lock);
3403 if (list_empty(&cache->dirty_list)) {
3404 list_add_tail(&cache->dirty_list,
3405 &cur_trans->dirty_bgs);
3406 btrfs_get_block_group(cache);
3407 drop_reserve = false;
3408 }
3409 spin_unlock(&cur_trans->dirty_bgs_lock);
3410 } else if (ret) {
3411 btrfs_abort_transaction(trans, ret);
3412 }
3413 }
3414
3415 /* If it's not on the io list, we need to put the block group */
3416 if (should_put)
3417 btrfs_put_block_group(cache);
3418 if (drop_reserve)
3419 btrfs_delayed_refs_rsv_release(fs_info, 1);
3420 /*
3421 * Avoid blocking other tasks for too long. It might even save
3422 * us from writing caches for block groups that are going to be
3423 * removed.
3424 */
3425 mutex_unlock(&trans->transaction->cache_write_mutex);
3426 if (ret)
3427 goto out;
3428 mutex_lock(&trans->transaction->cache_write_mutex);
3429 }
3430 mutex_unlock(&trans->transaction->cache_write_mutex);
3431
3432 /*
3433 * Go through delayed refs for all the stuff we've just kicked off
3434 * and then loop back (just once)
3435 */
3436 if (!ret)
3437 ret = btrfs_run_delayed_refs(trans, 0);
3438 if (!ret && loops == 0) {
3439 loops++;
3440 spin_lock(&cur_trans->dirty_bgs_lock);
3441 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3442 /*
3443 * dirty_bgs_lock protects us from concurrent block group
3444 * deletes too (not just cache_write_mutex).
3445 */
3446 if (!list_empty(&dirty)) {
3447 spin_unlock(&cur_trans->dirty_bgs_lock);
3448 goto again;
3449 }
3450 spin_unlock(&cur_trans->dirty_bgs_lock);
3451 }
3452 out:
3453 if (ret < 0) {
3454 spin_lock(&cur_trans->dirty_bgs_lock);
3455 list_splice_init(&dirty, &cur_trans->dirty_bgs);
3456 spin_unlock(&cur_trans->dirty_bgs_lock);
3457 btrfs_cleanup_dirty_bgs(cur_trans, fs_info);
3458 }
3459
3460 btrfs_free_path(path);
3461 return ret;
3462 }
3463
btrfs_write_dirty_block_groups(struct btrfs_trans_handle * trans)3464 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans)
3465 {
3466 struct btrfs_fs_info *fs_info = trans->fs_info;
3467 struct btrfs_block_group *cache;
3468 struct btrfs_transaction *cur_trans = trans->transaction;
3469 int ret = 0;
3470 int should_put;
3471 struct btrfs_path *path;
3472 struct list_head *io = &cur_trans->io_bgs;
3473
3474 path = btrfs_alloc_path();
3475 if (!path)
3476 return -ENOMEM;
3477
3478 /*
3479 * Even though we are in the critical section of the transaction commit,
3480 * we can still have concurrent tasks adding elements to this
3481 * transaction's list of dirty block groups. These tasks correspond to
3482 * endio free space workers started when writeback finishes for a
3483 * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3484 * allocate new block groups as a result of COWing nodes of the root
3485 * tree when updating the free space inode. The writeback for the space
3486 * caches is triggered by an earlier call to
3487 * btrfs_start_dirty_block_groups() and iterations of the following
3488 * loop.
3489 * Also we want to do the cache_save_setup first and then run the
3490 * delayed refs to make sure we have the best chance at doing this all
3491 * in one shot.
3492 */
3493 spin_lock(&cur_trans->dirty_bgs_lock);
3494 while (!list_empty(&cur_trans->dirty_bgs)) {
3495 cache = list_first_entry(&cur_trans->dirty_bgs,
3496 struct btrfs_block_group,
3497 dirty_list);
3498
3499 /*
3500 * This can happen if cache_save_setup re-dirties a block group
3501 * that is already under IO. Just wait for it to finish and
3502 * then do it all again
3503 */
3504 if (!list_empty(&cache->io_list)) {
3505 spin_unlock(&cur_trans->dirty_bgs_lock);
3506 list_del_init(&cache->io_list);
3507 btrfs_wait_cache_io(trans, cache, path);
3508 btrfs_put_block_group(cache);
3509 spin_lock(&cur_trans->dirty_bgs_lock);
3510 }
3511
3512 /*
3513 * Don't remove from the dirty list until after we've waited on
3514 * any pending IO
3515 */
3516 list_del_init(&cache->dirty_list);
3517 spin_unlock(&cur_trans->dirty_bgs_lock);
3518 should_put = 1;
3519
3520 cache_save_setup(cache, trans, path);
3521
3522 if (!ret)
3523 ret = btrfs_run_delayed_refs(trans,
3524 (unsigned long) -1);
3525
3526 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3527 cache->io_ctl.inode = NULL;
3528 ret = btrfs_write_out_cache(trans, cache, path);
3529 if (ret == 0 && cache->io_ctl.inode) {
3530 should_put = 0;
3531 list_add_tail(&cache->io_list, io);
3532 } else {
3533 /*
3534 * If we failed to write the cache, the
3535 * generation will be bad and life goes on
3536 */
3537 ret = 0;
3538 }
3539 }
3540 if (!ret) {
3541 ret = update_block_group_item(trans, path, cache);
3542 /*
3543 * One of the free space endio workers might have
3544 * created a new block group while updating a free space
3545 * cache's inode (at inode.c:btrfs_finish_ordered_io())
3546 * and hasn't released its transaction handle yet, in
3547 * which case the new block group is still attached to
3548 * its transaction handle and its creation has not
3549 * finished yet (no block group item in the extent tree
3550 * yet, etc). If this is the case, wait for all free
3551 * space endio workers to finish and retry. This is a
3552 * very rare case so no need for a more efficient and
3553 * complex approach.
3554 */
3555 if (ret == -ENOENT) {
3556 wait_event(cur_trans->writer_wait,
3557 atomic_read(&cur_trans->num_writers) == 1);
3558 ret = update_block_group_item(trans, path, cache);
3559 }
3560 if (ret)
3561 btrfs_abort_transaction(trans, ret);
3562 }
3563
3564 /* If its not on the io list, we need to put the block group */
3565 if (should_put)
3566 btrfs_put_block_group(cache);
3567 btrfs_delayed_refs_rsv_release(fs_info, 1);
3568 spin_lock(&cur_trans->dirty_bgs_lock);
3569 }
3570 spin_unlock(&cur_trans->dirty_bgs_lock);
3571
3572 /*
3573 * Refer to the definition of io_bgs member for details why it's safe
3574 * to use it without any locking
3575 */
3576 while (!list_empty(io)) {
3577 cache = list_first_entry(io, struct btrfs_block_group,
3578 io_list);
3579 list_del_init(&cache->io_list);
3580 btrfs_wait_cache_io(trans, cache, path);
3581 btrfs_put_block_group(cache);
3582 }
3583
3584 btrfs_free_path(path);
3585 return ret;
3586 }
3587
btrfs_update_block_group(struct btrfs_trans_handle * trans,u64 bytenr,u64 num_bytes,bool alloc)3588 int btrfs_update_block_group(struct btrfs_trans_handle *trans,
3589 u64 bytenr, u64 num_bytes, bool alloc)
3590 {
3591 struct btrfs_fs_info *info = trans->fs_info;
3592 struct btrfs_block_group *cache = NULL;
3593 u64 total = num_bytes;
3594 u64 old_val;
3595 u64 byte_in_group;
3596 int factor;
3597 int ret = 0;
3598
3599 /* Block accounting for super block */
3600 spin_lock(&info->delalloc_root_lock);
3601 old_val = btrfs_super_bytes_used(info->super_copy);
3602 if (alloc)
3603 old_val += num_bytes;
3604 else
3605 old_val -= num_bytes;
3606 btrfs_set_super_bytes_used(info->super_copy, old_val);
3607 spin_unlock(&info->delalloc_root_lock);
3608
3609 while (total) {
3610 struct btrfs_space_info *space_info;
3611 bool reclaim = false;
3612
3613 cache = btrfs_lookup_block_group(info, bytenr);
3614 if (!cache) {
3615 ret = -ENOENT;
3616 break;
3617 }
3618 space_info = cache->space_info;
3619 factor = btrfs_bg_type_to_factor(cache->flags);
3620
3621 /*
3622 * If this block group has free space cache written out, we
3623 * need to make sure to load it if we are removing space. This
3624 * is because we need the unpinning stage to actually add the
3625 * space back to the block group, otherwise we will leak space.
3626 */
3627 if (!alloc && !btrfs_block_group_done(cache))
3628 btrfs_cache_block_group(cache, true);
3629
3630 byte_in_group = bytenr - cache->start;
3631 WARN_ON(byte_in_group > cache->length);
3632
3633 spin_lock(&space_info->lock);
3634 spin_lock(&cache->lock);
3635
3636 if (btrfs_test_opt(info, SPACE_CACHE) &&
3637 cache->disk_cache_state < BTRFS_DC_CLEAR)
3638 cache->disk_cache_state = BTRFS_DC_CLEAR;
3639
3640 old_val = cache->used;
3641 num_bytes = min(total, cache->length - byte_in_group);
3642 if (alloc) {
3643 old_val += num_bytes;
3644 cache->used = old_val;
3645 cache->reserved -= num_bytes;
3646 space_info->bytes_reserved -= num_bytes;
3647 space_info->bytes_used += num_bytes;
3648 space_info->disk_used += num_bytes * factor;
3649 spin_unlock(&cache->lock);
3650 spin_unlock(&space_info->lock);
3651 } else {
3652 old_val -= num_bytes;
3653 cache->used = old_val;
3654 cache->pinned += num_bytes;
3655 btrfs_space_info_update_bytes_pinned(info, space_info,
3656 num_bytes);
3657 space_info->bytes_used -= num_bytes;
3658 space_info->disk_used -= num_bytes * factor;
3659
3660 reclaim = should_reclaim_block_group(cache, num_bytes);
3661
3662 spin_unlock(&cache->lock);
3663 spin_unlock(&space_info->lock);
3664
3665 set_extent_bit(&trans->transaction->pinned_extents,
3666 bytenr, bytenr + num_bytes - 1,
3667 EXTENT_DIRTY, NULL);
3668 }
3669
3670 spin_lock(&trans->transaction->dirty_bgs_lock);
3671 if (list_empty(&cache->dirty_list)) {
3672 list_add_tail(&cache->dirty_list,
3673 &trans->transaction->dirty_bgs);
3674 trans->delayed_ref_updates++;
3675 btrfs_get_block_group(cache);
3676 }
3677 spin_unlock(&trans->transaction->dirty_bgs_lock);
3678
3679 /*
3680 * No longer have used bytes in this block group, queue it for
3681 * deletion. We do this after adding the block group to the
3682 * dirty list to avoid races between cleaner kthread and space
3683 * cache writeout.
3684 */
3685 if (!alloc && old_val == 0) {
3686 if (!btrfs_test_opt(info, DISCARD_ASYNC))
3687 btrfs_mark_bg_unused(cache);
3688 } else if (!alloc && reclaim) {
3689 btrfs_mark_bg_to_reclaim(cache);
3690 }
3691
3692 btrfs_put_block_group(cache);
3693 total -= num_bytes;
3694 bytenr += num_bytes;
3695 }
3696
3697 /* Modified block groups are accounted for in the delayed_refs_rsv. */
3698 btrfs_update_delayed_refs_rsv(trans);
3699 return ret;
3700 }
3701
3702 /*
3703 * Update the block_group and space info counters.
3704 *
3705 * @cache: The cache we are manipulating
3706 * @ram_bytes: The number of bytes of file content, and will be same to
3707 * @num_bytes except for the compress path.
3708 * @num_bytes: The number of bytes in question
3709 * @delalloc: The blocks are allocated for the delalloc write
3710 *
3711 * This is called by the allocator when it reserves space. If this is a
3712 * reservation and the block group has become read only we cannot make the
3713 * reservation and return -EAGAIN, otherwise this function always succeeds.
3714 */
btrfs_add_reserved_bytes(struct btrfs_block_group * cache,u64 ram_bytes,u64 num_bytes,int delalloc,bool force_wrong_size_class)3715 int btrfs_add_reserved_bytes(struct btrfs_block_group *cache,
3716 u64 ram_bytes, u64 num_bytes, int delalloc,
3717 bool force_wrong_size_class)
3718 {
3719 struct btrfs_space_info *space_info = cache->space_info;
3720 enum btrfs_block_group_size_class size_class;
3721 int ret = 0;
3722
3723 spin_lock(&space_info->lock);
3724 spin_lock(&cache->lock);
3725 if (cache->ro) {
3726 ret = -EAGAIN;
3727 goto out;
3728 }
3729
3730 if (btrfs_block_group_should_use_size_class(cache)) {
3731 size_class = btrfs_calc_block_group_size_class(num_bytes);
3732 ret = btrfs_use_block_group_size_class(cache, size_class, force_wrong_size_class);
3733 if (ret)
3734 goto out;
3735 }
3736 cache->reserved += num_bytes;
3737 space_info->bytes_reserved += num_bytes;
3738 trace_btrfs_space_reservation(cache->fs_info, "space_info",
3739 space_info->flags, num_bytes, 1);
3740 btrfs_space_info_update_bytes_may_use(cache->fs_info,
3741 space_info, -ram_bytes);
3742 if (delalloc)
3743 cache->delalloc_bytes += num_bytes;
3744
3745 /*
3746 * Compression can use less space than we reserved, so wake tickets if
3747 * that happens.
3748 */
3749 if (num_bytes < ram_bytes)
3750 btrfs_try_granting_tickets(cache->fs_info, space_info);
3751 out:
3752 spin_unlock(&cache->lock);
3753 spin_unlock(&space_info->lock);
3754 return ret;
3755 }
3756
3757 /*
3758 * Update the block_group and space info counters.
3759 *
3760 * @cache: The cache we are manipulating
3761 * @num_bytes: The number of bytes in question
3762 * @delalloc: The blocks are allocated for the delalloc write
3763 *
3764 * This is called by somebody who is freeing space that was never actually used
3765 * on disk. For example if you reserve some space for a new leaf in transaction
3766 * A and before transaction A commits you free that leaf, you call this with
3767 * reserve set to 0 in order to clear the reservation.
3768 */
btrfs_free_reserved_bytes(struct btrfs_block_group * cache,u64 num_bytes,int delalloc)3769 void btrfs_free_reserved_bytes(struct btrfs_block_group *cache,
3770 u64 num_bytes, int delalloc)
3771 {
3772 struct btrfs_space_info *space_info = cache->space_info;
3773
3774 spin_lock(&space_info->lock);
3775 spin_lock(&cache->lock);
3776 if (cache->ro)
3777 space_info->bytes_readonly += num_bytes;
3778 cache->reserved -= num_bytes;
3779 space_info->bytes_reserved -= num_bytes;
3780 space_info->max_extent_size = 0;
3781
3782 if (delalloc)
3783 cache->delalloc_bytes -= num_bytes;
3784 spin_unlock(&cache->lock);
3785
3786 btrfs_try_granting_tickets(cache->fs_info, space_info);
3787 spin_unlock(&space_info->lock);
3788 }
3789
force_metadata_allocation(struct btrfs_fs_info * info)3790 static void force_metadata_allocation(struct btrfs_fs_info *info)
3791 {
3792 struct list_head *head = &info->space_info;
3793 struct btrfs_space_info *found;
3794
3795 list_for_each_entry(found, head, list) {
3796 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3797 found->force_alloc = CHUNK_ALLOC_FORCE;
3798 }
3799 }
3800
should_alloc_chunk(struct btrfs_fs_info * fs_info,struct btrfs_space_info * sinfo,int force)3801 static int should_alloc_chunk(struct btrfs_fs_info *fs_info,
3802 struct btrfs_space_info *sinfo, int force)
3803 {
3804 u64 bytes_used = btrfs_space_info_used(sinfo, false);
3805 u64 thresh;
3806
3807 if (force == CHUNK_ALLOC_FORCE)
3808 return 1;
3809
3810 /*
3811 * in limited mode, we want to have some free space up to
3812 * about 1% of the FS size.
3813 */
3814 if (force == CHUNK_ALLOC_LIMITED) {
3815 thresh = btrfs_super_total_bytes(fs_info->super_copy);
3816 thresh = max_t(u64, SZ_64M, mult_perc(thresh, 1));
3817
3818 if (sinfo->total_bytes - bytes_used < thresh)
3819 return 1;
3820 }
3821
3822 if (bytes_used + SZ_2M < mult_perc(sinfo->total_bytes, 80))
3823 return 0;
3824 return 1;
3825 }
3826
btrfs_force_chunk_alloc(struct btrfs_trans_handle * trans,u64 type)3827 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans, u64 type)
3828 {
3829 u64 alloc_flags = btrfs_get_alloc_profile(trans->fs_info, type);
3830
3831 return btrfs_chunk_alloc(trans, alloc_flags, CHUNK_ALLOC_FORCE);
3832 }
3833
do_chunk_alloc(struct btrfs_trans_handle * trans,u64 flags)3834 static struct btrfs_block_group *do_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags)
3835 {
3836 struct btrfs_block_group *bg;
3837 int ret;
3838
3839 /*
3840 * Check if we have enough space in the system space info because we
3841 * will need to update device items in the chunk btree and insert a new
3842 * chunk item in the chunk btree as well. This will allocate a new
3843 * system block group if needed.
3844 */
3845 check_system_chunk(trans, flags);
3846
3847 bg = btrfs_create_chunk(trans, flags);
3848 if (IS_ERR(bg)) {
3849 ret = PTR_ERR(bg);
3850 goto out;
3851 }
3852
3853 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3854 /*
3855 * Normally we are not expected to fail with -ENOSPC here, since we have
3856 * previously reserved space in the system space_info and allocated one
3857 * new system chunk if necessary. However there are three exceptions:
3858 *
3859 * 1) We may have enough free space in the system space_info but all the
3860 * existing system block groups have a profile which can not be used
3861 * for extent allocation.
3862 *
3863 * This happens when mounting in degraded mode. For example we have a
3864 * RAID1 filesystem with 2 devices, lose one device and mount the fs
3865 * using the other device in degraded mode. If we then allocate a chunk,
3866 * we may have enough free space in the existing system space_info, but
3867 * none of the block groups can be used for extent allocation since they
3868 * have a RAID1 profile, and because we are in degraded mode with a
3869 * single device, we are forced to allocate a new system chunk with a
3870 * SINGLE profile. Making check_system_chunk() iterate over all system
3871 * block groups and check if they have a usable profile and enough space
3872 * can be slow on very large filesystems, so we tolerate the -ENOSPC and
3873 * try again after forcing allocation of a new system chunk. Like this
3874 * we avoid paying the cost of that search in normal circumstances, when
3875 * we were not mounted in degraded mode;
3876 *
3877 * 2) We had enough free space info the system space_info, and one suitable
3878 * block group to allocate from when we called check_system_chunk()
3879 * above. However right after we called it, the only system block group
3880 * with enough free space got turned into RO mode by a running scrub,
3881 * and in this case we have to allocate a new one and retry. We only
3882 * need do this allocate and retry once, since we have a transaction
3883 * handle and scrub uses the commit root to search for block groups;
3884 *
3885 * 3) We had one system block group with enough free space when we called
3886 * check_system_chunk(), but after that, right before we tried to
3887 * allocate the last extent buffer we needed, a discard operation came
3888 * in and it temporarily removed the last free space entry from the
3889 * block group (discard removes a free space entry, discards it, and
3890 * then adds back the entry to the block group cache).
3891 */
3892 if (ret == -ENOSPC) {
3893 const u64 sys_flags = btrfs_system_alloc_profile(trans->fs_info);
3894 struct btrfs_block_group *sys_bg;
3895
3896 sys_bg = btrfs_create_chunk(trans, sys_flags);
3897 if (IS_ERR(sys_bg)) {
3898 ret = PTR_ERR(sys_bg);
3899 btrfs_abort_transaction(trans, ret);
3900 goto out;
3901 }
3902
3903 ret = btrfs_chunk_alloc_add_chunk_item(trans, sys_bg);
3904 if (ret) {
3905 btrfs_abort_transaction(trans, ret);
3906 goto out;
3907 }
3908
3909 ret = btrfs_chunk_alloc_add_chunk_item(trans, bg);
3910 if (ret) {
3911 btrfs_abort_transaction(trans, ret);
3912 goto out;
3913 }
3914 } else if (ret) {
3915 btrfs_abort_transaction(trans, ret);
3916 goto out;
3917 }
3918 out:
3919 btrfs_trans_release_chunk_metadata(trans);
3920
3921 if (ret)
3922 return ERR_PTR(ret);
3923
3924 btrfs_get_block_group(bg);
3925 return bg;
3926 }
3927
3928 /*
3929 * Chunk allocation is done in 2 phases:
3930 *
3931 * 1) Phase 1 - through btrfs_chunk_alloc() we allocate device extents for
3932 * the chunk, the chunk mapping, create its block group and add the items
3933 * that belong in the chunk btree to it - more specifically, we need to
3934 * update device items in the chunk btree and add a new chunk item to it.
3935 *
3936 * 2) Phase 2 - through btrfs_create_pending_block_groups(), we add the block
3937 * group item to the extent btree and the device extent items to the devices
3938 * btree.
3939 *
3940 * This is done to prevent deadlocks. For example when COWing a node from the
3941 * extent btree we are holding a write lock on the node's parent and if we
3942 * trigger chunk allocation and attempted to insert the new block group item
3943 * in the extent btree right way, we could deadlock because the path for the
3944 * insertion can include that parent node. At first glance it seems impossible
3945 * to trigger chunk allocation after starting a transaction since tasks should
3946 * reserve enough transaction units (metadata space), however while that is true
3947 * most of the time, chunk allocation may still be triggered for several reasons:
3948 *
3949 * 1) When reserving metadata, we check if there is enough free space in the
3950 * metadata space_info and therefore don't trigger allocation of a new chunk.
3951 * However later when the task actually tries to COW an extent buffer from
3952 * the extent btree or from the device btree for example, it is forced to
3953 * allocate a new block group (chunk) because the only one that had enough
3954 * free space was just turned to RO mode by a running scrub for example (or
3955 * device replace, block group reclaim thread, etc), so we can not use it
3956 * for allocating an extent and end up being forced to allocate a new one;
3957 *
3958 * 2) Because we only check that the metadata space_info has enough free bytes,
3959 * we end up not allocating a new metadata chunk in that case. However if
3960 * the filesystem was mounted in degraded mode, none of the existing block
3961 * groups might be suitable for extent allocation due to their incompatible
3962 * profile (for e.g. mounting a 2 devices filesystem, where all block groups
3963 * use a RAID1 profile, in degraded mode using a single device). In this case
3964 * when the task attempts to COW some extent buffer of the extent btree for
3965 * example, it will trigger allocation of a new metadata block group with a
3966 * suitable profile (SINGLE profile in the example of the degraded mount of
3967 * the RAID1 filesystem);
3968 *
3969 * 3) The task has reserved enough transaction units / metadata space, but when
3970 * it attempts to COW an extent buffer from the extent or device btree for
3971 * example, it does not find any free extent in any metadata block group,
3972 * therefore forced to try to allocate a new metadata block group.
3973 * This is because some other task allocated all available extents in the
3974 * meanwhile - this typically happens with tasks that don't reserve space
3975 * properly, either intentionally or as a bug. One example where this is
3976 * done intentionally is fsync, as it does not reserve any transaction units
3977 * and ends up allocating a variable number of metadata extents for log
3978 * tree extent buffers;
3979 *
3980 * 4) The task has reserved enough transaction units / metadata space, but right
3981 * before it tries to allocate the last extent buffer it needs, a discard
3982 * operation comes in and, temporarily, removes the last free space entry from
3983 * the only metadata block group that had free space (discard starts by
3984 * removing a free space entry from a block group, then does the discard
3985 * operation and, once it's done, it adds back the free space entry to the
3986 * block group).
3987 *
3988 * We also need this 2 phases setup when adding a device to a filesystem with
3989 * a seed device - we must create new metadata and system chunks without adding
3990 * any of the block group items to the chunk, extent and device btrees. If we
3991 * did not do it this way, we would get ENOSPC when attempting to update those
3992 * btrees, since all the chunks from the seed device are read-only.
3993 *
3994 * Phase 1 does the updates and insertions to the chunk btree because if we had
3995 * it done in phase 2 and have a thundering herd of tasks allocating chunks in
3996 * parallel, we risk having too many system chunks allocated by many tasks if
3997 * many tasks reach phase 1 without the previous ones completing phase 2. In the
3998 * extreme case this leads to exhaustion of the system chunk array in the
3999 * superblock. This is easier to trigger if using a btree node/leaf size of 64K
4000 * and with RAID filesystems (so we have more device items in the chunk btree).
4001 * This has happened before and commit eafa4fd0ad0607 ("btrfs: fix exhaustion of
4002 * the system chunk array due to concurrent allocations") provides more details.
4003 *
4004 * Allocation of system chunks does not happen through this function. A task that
4005 * needs to update the chunk btree (the only btree that uses system chunks), must
4006 * preallocate chunk space by calling either check_system_chunk() or
4007 * btrfs_reserve_chunk_metadata() - the former is used when allocating a data or
4008 * metadata chunk or when removing a chunk, while the later is used before doing
4009 * a modification to the chunk btree - use cases for the later are adding,
4010 * removing and resizing a device as well as relocation of a system chunk.
4011 * See the comment below for more details.
4012 *
4013 * The reservation of system space, done through check_system_chunk(), as well
4014 * as all the updates and insertions into the chunk btree must be done while
4015 * holding fs_info->chunk_mutex. This is important to guarantee that while COWing
4016 * an extent buffer from the chunks btree we never trigger allocation of a new
4017 * system chunk, which would result in a deadlock (trying to lock twice an
4018 * extent buffer of the chunk btree, first time before triggering the chunk
4019 * allocation and the second time during chunk allocation while attempting to
4020 * update the chunks btree). The system chunk array is also updated while holding
4021 * that mutex. The same logic applies to removing chunks - we must reserve system
4022 * space, update the chunk btree and the system chunk array in the superblock
4023 * while holding fs_info->chunk_mutex.
4024 *
4025 * This function, btrfs_chunk_alloc(), belongs to phase 1.
4026 *
4027 * If @force is CHUNK_ALLOC_FORCE:
4028 * - return 1 if it successfully allocates a chunk,
4029 * - return errors including -ENOSPC otherwise.
4030 * If @force is NOT CHUNK_ALLOC_FORCE:
4031 * - return 0 if it doesn't need to allocate a new chunk,
4032 * - return 1 if it successfully allocates a chunk,
4033 * - return errors including -ENOSPC otherwise.
4034 */
btrfs_chunk_alloc(struct btrfs_trans_handle * trans,u64 flags,enum btrfs_chunk_alloc_enum force)4035 int btrfs_chunk_alloc(struct btrfs_trans_handle *trans, u64 flags,
4036 enum btrfs_chunk_alloc_enum force)
4037 {
4038 struct btrfs_fs_info *fs_info = trans->fs_info;
4039 struct btrfs_space_info *space_info;
4040 struct btrfs_block_group *ret_bg;
4041 bool wait_for_alloc = false;
4042 bool should_alloc = false;
4043 bool from_extent_allocation = false;
4044 int ret = 0;
4045
4046 if (force == CHUNK_ALLOC_FORCE_FOR_EXTENT) {
4047 from_extent_allocation = true;
4048 force = CHUNK_ALLOC_FORCE;
4049 }
4050
4051 /* Don't re-enter if we're already allocating a chunk */
4052 if (trans->allocating_chunk)
4053 return -ENOSPC;
4054 /*
4055 * Allocation of system chunks can not happen through this path, as we
4056 * could end up in a deadlock if we are allocating a data or metadata
4057 * chunk and there is another task modifying the chunk btree.
4058 *
4059 * This is because while we are holding the chunk mutex, we will attempt
4060 * to add the new chunk item to the chunk btree or update an existing
4061 * device item in the chunk btree, while the other task that is modifying
4062 * the chunk btree is attempting to COW an extent buffer while holding a
4063 * lock on it and on its parent - if the COW operation triggers a system
4064 * chunk allocation, then we can deadlock because we are holding the
4065 * chunk mutex and we may need to access that extent buffer or its parent
4066 * in order to add the chunk item or update a device item.
4067 *
4068 * Tasks that want to modify the chunk tree should reserve system space
4069 * before updating the chunk btree, by calling either
4070 * btrfs_reserve_chunk_metadata() or check_system_chunk().
4071 * It's possible that after a task reserves the space, it still ends up
4072 * here - this happens in the cases described above at do_chunk_alloc().
4073 * The task will have to either retry or fail.
4074 */
4075 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4076 return -ENOSPC;
4077
4078 space_info = btrfs_find_space_info(fs_info, flags);
4079 ASSERT(space_info);
4080
4081 do {
4082 spin_lock(&space_info->lock);
4083 if (force < space_info->force_alloc)
4084 force = space_info->force_alloc;
4085 should_alloc = should_alloc_chunk(fs_info, space_info, force);
4086 if (space_info->full) {
4087 /* No more free physical space */
4088 if (should_alloc)
4089 ret = -ENOSPC;
4090 else
4091 ret = 0;
4092 spin_unlock(&space_info->lock);
4093 return ret;
4094 } else if (!should_alloc) {
4095 spin_unlock(&space_info->lock);
4096 return 0;
4097 } else if (space_info->chunk_alloc) {
4098 /*
4099 * Someone is already allocating, so we need to block
4100 * until this someone is finished and then loop to
4101 * recheck if we should continue with our allocation
4102 * attempt.
4103 */
4104 wait_for_alloc = true;
4105 force = CHUNK_ALLOC_NO_FORCE;
4106 spin_unlock(&space_info->lock);
4107 mutex_lock(&fs_info->chunk_mutex);
4108 mutex_unlock(&fs_info->chunk_mutex);
4109 } else {
4110 /* Proceed with allocation */
4111 space_info->chunk_alloc = 1;
4112 wait_for_alloc = false;
4113 spin_unlock(&space_info->lock);
4114 }
4115
4116 cond_resched();
4117 } while (wait_for_alloc);
4118
4119 mutex_lock(&fs_info->chunk_mutex);
4120 trans->allocating_chunk = true;
4121
4122 /*
4123 * If we have mixed data/metadata chunks we want to make sure we keep
4124 * allocating mixed chunks instead of individual chunks.
4125 */
4126 if (btrfs_mixed_space_info(space_info))
4127 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4128
4129 /*
4130 * if we're doing a data chunk, go ahead and make sure that
4131 * we keep a reasonable number of metadata chunks allocated in the
4132 * FS as well.
4133 */
4134 if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4135 fs_info->data_chunk_allocations++;
4136 if (!(fs_info->data_chunk_allocations %
4137 fs_info->metadata_ratio))
4138 force_metadata_allocation(fs_info);
4139 }
4140
4141 ret_bg = do_chunk_alloc(trans, flags);
4142 trans->allocating_chunk = false;
4143
4144 if (IS_ERR(ret_bg)) {
4145 ret = PTR_ERR(ret_bg);
4146 } else if (from_extent_allocation && (flags & BTRFS_BLOCK_GROUP_DATA)) {
4147 /*
4148 * New block group is likely to be used soon. Try to activate
4149 * it now. Failure is OK for now.
4150 */
4151 btrfs_zone_activate(ret_bg);
4152 }
4153
4154 if (!ret)
4155 btrfs_put_block_group(ret_bg);
4156
4157 spin_lock(&space_info->lock);
4158 if (ret < 0) {
4159 if (ret == -ENOSPC)
4160 space_info->full = 1;
4161 else
4162 goto out;
4163 } else {
4164 ret = 1;
4165 space_info->max_extent_size = 0;
4166 }
4167
4168 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4169 out:
4170 space_info->chunk_alloc = 0;
4171 spin_unlock(&space_info->lock);
4172 mutex_unlock(&fs_info->chunk_mutex);
4173
4174 return ret;
4175 }
4176
get_profile_num_devs(struct btrfs_fs_info * fs_info,u64 type)4177 static u64 get_profile_num_devs(struct btrfs_fs_info *fs_info, u64 type)
4178 {
4179 u64 num_dev;
4180
4181 num_dev = btrfs_raid_array[btrfs_bg_flags_to_raid_index(type)].devs_max;
4182 if (!num_dev)
4183 num_dev = fs_info->fs_devices->rw_devices;
4184
4185 return num_dev;
4186 }
4187
reserve_chunk_space(struct btrfs_trans_handle * trans,u64 bytes,u64 type)4188 static void reserve_chunk_space(struct btrfs_trans_handle *trans,
4189 u64 bytes,
4190 u64 type)
4191 {
4192 struct btrfs_fs_info *fs_info = trans->fs_info;
4193 struct btrfs_space_info *info;
4194 u64 left;
4195 int ret = 0;
4196
4197 /*
4198 * Needed because we can end up allocating a system chunk and for an
4199 * atomic and race free space reservation in the chunk block reserve.
4200 */
4201 lockdep_assert_held(&fs_info->chunk_mutex);
4202
4203 info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4204 spin_lock(&info->lock);
4205 left = info->total_bytes - btrfs_space_info_used(info, true);
4206 spin_unlock(&info->lock);
4207
4208 if (left < bytes && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
4209 btrfs_info(fs_info, "left=%llu, need=%llu, flags=%llu",
4210 left, bytes, type);
4211 btrfs_dump_space_info(fs_info, info, 0, 0);
4212 }
4213
4214 if (left < bytes) {
4215 u64 flags = btrfs_system_alloc_profile(fs_info);
4216 struct btrfs_block_group *bg;
4217
4218 /*
4219 * Ignore failure to create system chunk. We might end up not
4220 * needing it, as we might not need to COW all nodes/leafs from
4221 * the paths we visit in the chunk tree (they were already COWed
4222 * or created in the current transaction for example).
4223 */
4224 bg = btrfs_create_chunk(trans, flags);
4225 if (IS_ERR(bg)) {
4226 ret = PTR_ERR(bg);
4227 } else {
4228 /*
4229 * We have a new chunk. We also need to activate it for
4230 * zoned filesystem.
4231 */
4232 ret = btrfs_zoned_activate_one_bg(fs_info, info, true);
4233 if (ret < 0)
4234 return;
4235
4236 /*
4237 * If we fail to add the chunk item here, we end up
4238 * trying again at phase 2 of chunk allocation, at
4239 * btrfs_create_pending_block_groups(). So ignore
4240 * any error here. An ENOSPC here could happen, due to
4241 * the cases described at do_chunk_alloc() - the system
4242 * block group we just created was just turned into RO
4243 * mode by a scrub for example, or a running discard
4244 * temporarily removed its free space entries, etc.
4245 */
4246 btrfs_chunk_alloc_add_chunk_item(trans, bg);
4247 }
4248 }
4249
4250 if (!ret) {
4251 ret = btrfs_block_rsv_add(fs_info,
4252 &fs_info->chunk_block_rsv,
4253 bytes, BTRFS_RESERVE_NO_FLUSH);
4254 if (!ret)
4255 trans->chunk_bytes_reserved += bytes;
4256 }
4257 }
4258
4259 /*
4260 * Reserve space in the system space for allocating or removing a chunk.
4261 * The caller must be holding fs_info->chunk_mutex.
4262 */
check_system_chunk(struct btrfs_trans_handle * trans,u64 type)4263 void check_system_chunk(struct btrfs_trans_handle *trans, u64 type)
4264 {
4265 struct btrfs_fs_info *fs_info = trans->fs_info;
4266 const u64 num_devs = get_profile_num_devs(fs_info, type);
4267 u64 bytes;
4268
4269 /* num_devs device items to update and 1 chunk item to add or remove. */
4270 bytes = btrfs_calc_metadata_size(fs_info, num_devs) +
4271 btrfs_calc_insert_metadata_size(fs_info, 1);
4272
4273 reserve_chunk_space(trans, bytes, type);
4274 }
4275
4276 /*
4277 * Reserve space in the system space, if needed, for doing a modification to the
4278 * chunk btree.
4279 *
4280 * @trans: A transaction handle.
4281 * @is_item_insertion: Indicate if the modification is for inserting a new item
4282 * in the chunk btree or if it's for the deletion or update
4283 * of an existing item.
4284 *
4285 * This is used in a context where we need to update the chunk btree outside
4286 * block group allocation and removal, to avoid a deadlock with a concurrent
4287 * task that is allocating a metadata or data block group and therefore needs to
4288 * update the chunk btree while holding the chunk mutex. After the update to the
4289 * chunk btree is done, btrfs_trans_release_chunk_metadata() should be called.
4290 *
4291 */
btrfs_reserve_chunk_metadata(struct btrfs_trans_handle * trans,bool is_item_insertion)4292 void btrfs_reserve_chunk_metadata(struct btrfs_trans_handle *trans,
4293 bool is_item_insertion)
4294 {
4295 struct btrfs_fs_info *fs_info = trans->fs_info;
4296 u64 bytes;
4297
4298 if (is_item_insertion)
4299 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
4300 else
4301 bytes = btrfs_calc_metadata_size(fs_info, 1);
4302
4303 mutex_lock(&fs_info->chunk_mutex);
4304 reserve_chunk_space(trans, bytes, BTRFS_BLOCK_GROUP_SYSTEM);
4305 mutex_unlock(&fs_info->chunk_mutex);
4306 }
4307
btrfs_put_block_group_cache(struct btrfs_fs_info * info)4308 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
4309 {
4310 struct btrfs_block_group *block_group;
4311
4312 block_group = btrfs_lookup_first_block_group(info, 0);
4313 while (block_group) {
4314 btrfs_wait_block_group_cache_done(block_group);
4315 spin_lock(&block_group->lock);
4316 if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF,
4317 &block_group->runtime_flags)) {
4318 struct inode *inode = block_group->inode;
4319
4320 block_group->inode = NULL;
4321 spin_unlock(&block_group->lock);
4322
4323 ASSERT(block_group->io_ctl.inode == NULL);
4324 iput(inode);
4325 } else {
4326 spin_unlock(&block_group->lock);
4327 }
4328 block_group = btrfs_next_block_group(block_group);
4329 }
4330 }
4331
4332 /*
4333 * Must be called only after stopping all workers, since we could have block
4334 * group caching kthreads running, and therefore they could race with us if we
4335 * freed the block groups before stopping them.
4336 */
btrfs_free_block_groups(struct btrfs_fs_info * info)4337 int btrfs_free_block_groups(struct btrfs_fs_info *info)
4338 {
4339 struct btrfs_block_group *block_group;
4340 struct btrfs_space_info *space_info;
4341 struct btrfs_caching_control *caching_ctl;
4342 struct rb_node *n;
4343
4344 if (btrfs_is_zoned(info)) {
4345 if (info->active_meta_bg) {
4346 btrfs_put_block_group(info->active_meta_bg);
4347 info->active_meta_bg = NULL;
4348 }
4349 if (info->active_system_bg) {
4350 btrfs_put_block_group(info->active_system_bg);
4351 info->active_system_bg = NULL;
4352 }
4353 }
4354
4355 write_lock(&info->block_group_cache_lock);
4356 while (!list_empty(&info->caching_block_groups)) {
4357 caching_ctl = list_entry(info->caching_block_groups.next,
4358 struct btrfs_caching_control, list);
4359 list_del(&caching_ctl->list);
4360 btrfs_put_caching_control(caching_ctl);
4361 }
4362 write_unlock(&info->block_group_cache_lock);
4363
4364 spin_lock(&info->unused_bgs_lock);
4365 while (!list_empty(&info->unused_bgs)) {
4366 block_group = list_first_entry(&info->unused_bgs,
4367 struct btrfs_block_group,
4368 bg_list);
4369 list_del_init(&block_group->bg_list);
4370 btrfs_put_block_group(block_group);
4371 }
4372
4373 while (!list_empty(&info->reclaim_bgs)) {
4374 block_group = list_first_entry(&info->reclaim_bgs,
4375 struct btrfs_block_group,
4376 bg_list);
4377 list_del_init(&block_group->bg_list);
4378 btrfs_put_block_group(block_group);
4379 }
4380 spin_unlock(&info->unused_bgs_lock);
4381
4382 spin_lock(&info->zone_active_bgs_lock);
4383 while (!list_empty(&info->zone_active_bgs)) {
4384 block_group = list_first_entry(&info->zone_active_bgs,
4385 struct btrfs_block_group,
4386 active_bg_list);
4387 list_del_init(&block_group->active_bg_list);
4388 btrfs_put_block_group(block_group);
4389 }
4390 spin_unlock(&info->zone_active_bgs_lock);
4391
4392 write_lock(&info->block_group_cache_lock);
4393 while ((n = rb_last(&info->block_group_cache_tree.rb_root)) != NULL) {
4394 block_group = rb_entry(n, struct btrfs_block_group,
4395 cache_node);
4396 rb_erase_cached(&block_group->cache_node,
4397 &info->block_group_cache_tree);
4398 RB_CLEAR_NODE(&block_group->cache_node);
4399 write_unlock(&info->block_group_cache_lock);
4400
4401 down_write(&block_group->space_info->groups_sem);
4402 list_del(&block_group->list);
4403 up_write(&block_group->space_info->groups_sem);
4404
4405 /*
4406 * We haven't cached this block group, which means we could
4407 * possibly have excluded extents on this block group.
4408 */
4409 if (block_group->cached == BTRFS_CACHE_NO ||
4410 block_group->cached == BTRFS_CACHE_ERROR)
4411 btrfs_free_excluded_extents(block_group);
4412
4413 btrfs_remove_free_space_cache(block_group);
4414 ASSERT(block_group->cached != BTRFS_CACHE_STARTED);
4415 ASSERT(list_empty(&block_group->dirty_list));
4416 ASSERT(list_empty(&block_group->io_list));
4417 ASSERT(list_empty(&block_group->bg_list));
4418 ASSERT(refcount_read(&block_group->refs) == 1);
4419 ASSERT(block_group->swap_extents == 0);
4420 btrfs_put_block_group(block_group);
4421
4422 write_lock(&info->block_group_cache_lock);
4423 }
4424 write_unlock(&info->block_group_cache_lock);
4425
4426 btrfs_release_global_block_rsv(info);
4427
4428 while (!list_empty(&info->space_info)) {
4429 space_info = list_entry(info->space_info.next,
4430 struct btrfs_space_info,
4431 list);
4432
4433 /*
4434 * Do not hide this behind enospc_debug, this is actually
4435 * important and indicates a real bug if this happens.
4436 */
4437 if (WARN_ON(space_info->bytes_pinned > 0 ||
4438 space_info->bytes_may_use > 0))
4439 btrfs_dump_space_info(info, space_info, 0, 0);
4440
4441 /*
4442 * If there was a failure to cleanup a log tree, very likely due
4443 * to an IO failure on a writeback attempt of one or more of its
4444 * extent buffers, we could not do proper (and cheap) unaccounting
4445 * of their reserved space, so don't warn on bytes_reserved > 0 in
4446 * that case.
4447 */
4448 if (!(space_info->flags & BTRFS_BLOCK_GROUP_METADATA) ||
4449 !BTRFS_FS_LOG_CLEANUP_ERROR(info)) {
4450 if (WARN_ON(space_info->bytes_reserved > 0))
4451 btrfs_dump_space_info(info, space_info, 0, 0);
4452 }
4453
4454 WARN_ON(space_info->reclaim_size > 0);
4455 list_del(&space_info->list);
4456 btrfs_sysfs_remove_space_info(space_info);
4457 }
4458 return 0;
4459 }
4460
btrfs_freeze_block_group(struct btrfs_block_group * cache)4461 void btrfs_freeze_block_group(struct btrfs_block_group *cache)
4462 {
4463 atomic_inc(&cache->frozen);
4464 }
4465
btrfs_unfreeze_block_group(struct btrfs_block_group * block_group)4466 void btrfs_unfreeze_block_group(struct btrfs_block_group *block_group)
4467 {
4468 struct btrfs_fs_info *fs_info = block_group->fs_info;
4469 struct extent_map_tree *em_tree;
4470 struct extent_map *em;
4471 bool cleanup;
4472
4473 spin_lock(&block_group->lock);
4474 cleanup = (atomic_dec_and_test(&block_group->frozen) &&
4475 test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags));
4476 spin_unlock(&block_group->lock);
4477
4478 if (cleanup) {
4479 em_tree = &fs_info->mapping_tree;
4480 write_lock(&em_tree->lock);
4481 em = lookup_extent_mapping(em_tree, block_group->start,
4482 1);
4483 BUG_ON(!em); /* logic error, can't happen */
4484 remove_extent_mapping(em_tree, em);
4485 write_unlock(&em_tree->lock);
4486
4487 /* once for us and once for the tree */
4488 free_extent_map(em);
4489 free_extent_map(em);
4490
4491 /*
4492 * We may have left one free space entry and other possible
4493 * tasks trimming this block group have left 1 entry each one.
4494 * Free them if any.
4495 */
4496 btrfs_remove_free_space_cache(block_group);
4497 }
4498 }
4499
btrfs_inc_block_group_swap_extents(struct btrfs_block_group * bg)4500 bool btrfs_inc_block_group_swap_extents(struct btrfs_block_group *bg)
4501 {
4502 bool ret = true;
4503
4504 spin_lock(&bg->lock);
4505 if (bg->ro)
4506 ret = false;
4507 else
4508 bg->swap_extents++;
4509 spin_unlock(&bg->lock);
4510
4511 return ret;
4512 }
4513
btrfs_dec_block_group_swap_extents(struct btrfs_block_group * bg,int amount)4514 void btrfs_dec_block_group_swap_extents(struct btrfs_block_group *bg, int amount)
4515 {
4516 spin_lock(&bg->lock);
4517 ASSERT(!bg->ro);
4518 ASSERT(bg->swap_extents >= amount);
4519 bg->swap_extents -= amount;
4520 spin_unlock(&bg->lock);
4521 }
4522
btrfs_calc_block_group_size_class(u64 size)4523 enum btrfs_block_group_size_class btrfs_calc_block_group_size_class(u64 size)
4524 {
4525 if (size <= SZ_128K)
4526 return BTRFS_BG_SZ_SMALL;
4527 if (size <= SZ_8M)
4528 return BTRFS_BG_SZ_MEDIUM;
4529 return BTRFS_BG_SZ_LARGE;
4530 }
4531
4532 /*
4533 * Handle a block group allocating an extent in a size class
4534 *
4535 * @bg: The block group we allocated in.
4536 * @size_class: The size class of the allocation.
4537 * @force_wrong_size_class: Whether we are desperate enough to allow
4538 * mismatched size classes.
4539 *
4540 * Returns: 0 if the size class was valid for this block_group, -EAGAIN in the
4541 * case of a race that leads to the wrong size class without
4542 * force_wrong_size_class set.
4543 *
4544 * find_free_extent will skip block groups with a mismatched size class until
4545 * it really needs to avoid ENOSPC. In that case it will set
4546 * force_wrong_size_class. However, if a block group is newly allocated and
4547 * doesn't yet have a size class, then it is possible for two allocations of
4548 * different sizes to race and both try to use it. The loser is caught here and
4549 * has to retry.
4550 */
btrfs_use_block_group_size_class(struct btrfs_block_group * bg,enum btrfs_block_group_size_class size_class,bool force_wrong_size_class)4551 int btrfs_use_block_group_size_class(struct btrfs_block_group *bg,
4552 enum btrfs_block_group_size_class size_class,
4553 bool force_wrong_size_class)
4554 {
4555 ASSERT(size_class != BTRFS_BG_SZ_NONE);
4556
4557 /* The new allocation is in the right size class, do nothing */
4558 if (bg->size_class == size_class)
4559 return 0;
4560 /*
4561 * The new allocation is in a mismatched size class.
4562 * This means one of two things:
4563 *
4564 * 1. Two tasks in find_free_extent for different size_classes raced
4565 * and hit the same empty block_group. Make the loser try again.
4566 * 2. A call to find_free_extent got desperate enough to set
4567 * 'force_wrong_slab'. Don't change the size_class, but allow the
4568 * allocation.
4569 */
4570 if (bg->size_class != BTRFS_BG_SZ_NONE) {
4571 if (force_wrong_size_class)
4572 return 0;
4573 return -EAGAIN;
4574 }
4575 /*
4576 * The happy new block group case: the new allocation is the first
4577 * one in the block_group so we set size_class.
4578 */
4579 bg->size_class = size_class;
4580
4581 return 0;
4582 }
4583
btrfs_block_group_should_use_size_class(struct btrfs_block_group * bg)4584 bool btrfs_block_group_should_use_size_class(struct btrfs_block_group *bg)
4585 {
4586 if (btrfs_is_zoned(bg->fs_info))
4587 return false;
4588 if (!btrfs_is_block_group_data_only(bg))
4589 return false;
4590 return true;
4591 }
4592