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