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