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