Loading...
1// SPDX-License-Identifier: GPL-2.0
2
3#include "misc.h"
4#include "ctree.h"
5#include "space-info.h"
6#include "sysfs.h"
7#include "volumes.h"
8#include "free-space-cache.h"
9#include "ordered-data.h"
10#include "transaction.h"
11#include "block-group.h"
12
13u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
14 bool may_use_included)
15{
16 ASSERT(s_info);
17 return s_info->bytes_used + s_info->bytes_reserved +
18 s_info->bytes_pinned + s_info->bytes_readonly +
19 (may_use_included ? s_info->bytes_may_use : 0);
20}
21
22/*
23 * after adding space to the filesystem, we need to clear the full flags
24 * on all the space infos.
25 */
26void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
27{
28 struct list_head *head = &info->space_info;
29 struct btrfs_space_info *found;
30
31 rcu_read_lock();
32 list_for_each_entry_rcu(found, head, list)
33 found->full = 0;
34 rcu_read_unlock();
35}
36
37static int create_space_info(struct btrfs_fs_info *info, u64 flags)
38{
39
40 struct btrfs_space_info *space_info;
41 int i;
42 int ret;
43
44 space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
45 if (!space_info)
46 return -ENOMEM;
47
48 ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
49 GFP_KERNEL);
50 if (ret) {
51 kfree(space_info);
52 return ret;
53 }
54
55 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
56 INIT_LIST_HEAD(&space_info->block_groups[i]);
57 init_rwsem(&space_info->groups_sem);
58 spin_lock_init(&space_info->lock);
59 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
60 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
61 init_waitqueue_head(&space_info->wait);
62 INIT_LIST_HEAD(&space_info->ro_bgs);
63 INIT_LIST_HEAD(&space_info->tickets);
64 INIT_LIST_HEAD(&space_info->priority_tickets);
65
66 ret = btrfs_sysfs_add_space_info_type(info, space_info);
67 if (ret)
68 return ret;
69
70 list_add_rcu(&space_info->list, &info->space_info);
71 if (flags & BTRFS_BLOCK_GROUP_DATA)
72 info->data_sinfo = space_info;
73
74 return ret;
75}
76
77int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
78{
79 struct btrfs_super_block *disk_super;
80 u64 features;
81 u64 flags;
82 int mixed = 0;
83 int ret;
84
85 disk_super = fs_info->super_copy;
86 if (!btrfs_super_root(disk_super))
87 return -EINVAL;
88
89 features = btrfs_super_incompat_flags(disk_super);
90 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
91 mixed = 1;
92
93 flags = BTRFS_BLOCK_GROUP_SYSTEM;
94 ret = create_space_info(fs_info, flags);
95 if (ret)
96 goto out;
97
98 if (mixed) {
99 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
100 ret = create_space_info(fs_info, flags);
101 } else {
102 flags = BTRFS_BLOCK_GROUP_METADATA;
103 ret = create_space_info(fs_info, flags);
104 if (ret)
105 goto out;
106
107 flags = BTRFS_BLOCK_GROUP_DATA;
108 ret = create_space_info(fs_info, flags);
109 }
110out:
111 return ret;
112}
113
114void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags,
115 u64 total_bytes, u64 bytes_used,
116 u64 bytes_readonly,
117 struct btrfs_space_info **space_info)
118{
119 struct btrfs_space_info *found;
120 int factor;
121
122 factor = btrfs_bg_type_to_factor(flags);
123
124 found = btrfs_find_space_info(info, flags);
125 ASSERT(found);
126 spin_lock(&found->lock);
127 found->total_bytes += total_bytes;
128 found->disk_total += total_bytes * factor;
129 found->bytes_used += bytes_used;
130 found->disk_used += bytes_used * factor;
131 found->bytes_readonly += bytes_readonly;
132 if (total_bytes > 0)
133 found->full = 0;
134 btrfs_try_granting_tickets(info, found);
135 spin_unlock(&found->lock);
136 *space_info = found;
137}
138
139struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
140 u64 flags)
141{
142 struct list_head *head = &info->space_info;
143 struct btrfs_space_info *found;
144
145 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
146
147 rcu_read_lock();
148 list_for_each_entry_rcu(found, head, list) {
149 if (found->flags & flags) {
150 rcu_read_unlock();
151 return found;
152 }
153 }
154 rcu_read_unlock();
155 return NULL;
156}
157
158static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
159{
160 return (global->size << 1);
161}
162
163static int can_overcommit(struct btrfs_fs_info *fs_info,
164 struct btrfs_space_info *space_info, u64 bytes,
165 enum btrfs_reserve_flush_enum flush,
166 bool system_chunk)
167{
168 u64 profile;
169 u64 avail;
170 u64 used;
171 int factor;
172
173 /* Don't overcommit when in mixed mode. */
174 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
175 return 0;
176
177 if (system_chunk)
178 profile = btrfs_system_alloc_profile(fs_info);
179 else
180 profile = btrfs_metadata_alloc_profile(fs_info);
181
182 used = btrfs_space_info_used(space_info, true);
183 avail = atomic64_read(&fs_info->free_chunk_space);
184
185 /*
186 * If we have dup, raid1 or raid10 then only half of the free
187 * space is actually usable. For raid56, the space info used
188 * doesn't include the parity drive, so we don't have to
189 * change the math
190 */
191 factor = btrfs_bg_type_to_factor(profile);
192 avail = div_u64(avail, factor);
193
194 /*
195 * If we aren't flushing all things, let us overcommit up to
196 * 1/2th of the space. If we can flush, don't let us overcommit
197 * too much, let it overcommit up to 1/8 of the space.
198 */
199 if (flush == BTRFS_RESERVE_FLUSH_ALL)
200 avail >>= 3;
201 else
202 avail >>= 1;
203
204 if (used + bytes < space_info->total_bytes + avail)
205 return 1;
206 return 0;
207}
208
209/*
210 * This is for space we already have accounted in space_info->bytes_may_use, so
211 * basically when we're returning space from block_rsv's.
212 */
213void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
214 struct btrfs_space_info *space_info)
215{
216 struct list_head *head;
217 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
218
219 lockdep_assert_held(&space_info->lock);
220
221 head = &space_info->priority_tickets;
222again:
223 while (!list_empty(head)) {
224 struct reserve_ticket *ticket;
225 u64 used = btrfs_space_info_used(space_info, true);
226
227 ticket = list_first_entry(head, struct reserve_ticket, list);
228
229 /* Check and see if our ticket can be satisified now. */
230 if ((used + ticket->bytes <= space_info->total_bytes) ||
231 can_overcommit(fs_info, space_info, ticket->bytes, flush,
232 false)) {
233 btrfs_space_info_update_bytes_may_use(fs_info,
234 space_info,
235 ticket->bytes);
236 list_del_init(&ticket->list);
237 ticket->bytes = 0;
238 space_info->tickets_id++;
239 wake_up(&ticket->wait);
240 } else {
241 break;
242 }
243 }
244
245 if (head == &space_info->priority_tickets) {
246 head = &space_info->tickets;
247 flush = BTRFS_RESERVE_FLUSH_ALL;
248 goto again;
249 }
250}
251
252#define DUMP_BLOCK_RSV(fs_info, rsv_name) \
253do { \
254 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \
255 spin_lock(&__rsv->lock); \
256 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \
257 __rsv->size, __rsv->reserved); \
258 spin_unlock(&__rsv->lock); \
259} while (0)
260
261static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
262 struct btrfs_space_info *info)
263{
264 lockdep_assert_held(&info->lock);
265
266 btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
267 info->flags,
268 info->total_bytes - btrfs_space_info_used(info, true),
269 info->full ? "" : "not ");
270 btrfs_info(fs_info,
271 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
272 info->total_bytes, info->bytes_used, info->bytes_pinned,
273 info->bytes_reserved, info->bytes_may_use,
274 info->bytes_readonly);
275
276 DUMP_BLOCK_RSV(fs_info, global_block_rsv);
277 DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
278 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
279 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
280 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
281
282}
283
284void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
285 struct btrfs_space_info *info, u64 bytes,
286 int dump_block_groups)
287{
288 struct btrfs_block_group_cache *cache;
289 int index = 0;
290
291 spin_lock(&info->lock);
292 __btrfs_dump_space_info(fs_info, info);
293 spin_unlock(&info->lock);
294
295 if (!dump_block_groups)
296 return;
297
298 down_read(&info->groups_sem);
299again:
300 list_for_each_entry(cache, &info->block_groups[index], list) {
301 spin_lock(&cache->lock);
302 btrfs_info(fs_info,
303 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
304 cache->key.objectid, cache->key.offset,
305 btrfs_block_group_used(&cache->item), cache->pinned,
306 cache->reserved, cache->ro ? "[readonly]" : "");
307 btrfs_dump_free_space(cache, bytes);
308 spin_unlock(&cache->lock);
309 }
310 if (++index < BTRFS_NR_RAID_TYPES)
311 goto again;
312 up_read(&info->groups_sem);
313}
314
315static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
316 unsigned long nr_pages, int nr_items)
317{
318 struct super_block *sb = fs_info->sb;
319
320 if (down_read_trylock(&sb->s_umount)) {
321 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
322 up_read(&sb->s_umount);
323 } else {
324 /*
325 * We needn't worry the filesystem going from r/w to r/o though
326 * we don't acquire ->s_umount mutex, because the filesystem
327 * should guarantee the delalloc inodes list be empty after
328 * the filesystem is readonly(all dirty pages are written to
329 * the disk).
330 */
331 btrfs_start_delalloc_roots(fs_info, nr_items);
332 if (!current->journal_info)
333 btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
334 }
335}
336
337static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
338 u64 to_reclaim)
339{
340 u64 bytes;
341 u64 nr;
342
343 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
344 nr = div64_u64(to_reclaim, bytes);
345 if (!nr)
346 nr = 1;
347 return nr;
348}
349
350#define EXTENT_SIZE_PER_ITEM SZ_256K
351
352/*
353 * shrink metadata reservation for delalloc
354 */
355static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
356 u64 orig, bool wait_ordered)
357{
358 struct btrfs_space_info *space_info;
359 struct btrfs_trans_handle *trans;
360 u64 delalloc_bytes;
361 u64 dio_bytes;
362 u64 async_pages;
363 u64 items;
364 long time_left;
365 unsigned long nr_pages;
366 int loops;
367
368 /* Calc the number of the pages we need flush for space reservation */
369 items = calc_reclaim_items_nr(fs_info, to_reclaim);
370 to_reclaim = items * EXTENT_SIZE_PER_ITEM;
371
372 trans = (struct btrfs_trans_handle *)current->journal_info;
373 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
374
375 delalloc_bytes = percpu_counter_sum_positive(
376 &fs_info->delalloc_bytes);
377 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
378 if (delalloc_bytes == 0 && dio_bytes == 0) {
379 if (trans)
380 return;
381 if (wait_ordered)
382 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
383 return;
384 }
385
386 /*
387 * If we are doing more ordered than delalloc we need to just wait on
388 * ordered extents, otherwise we'll waste time trying to flush delalloc
389 * that likely won't give us the space back we need.
390 */
391 if (dio_bytes > delalloc_bytes)
392 wait_ordered = true;
393
394 loops = 0;
395 while ((delalloc_bytes || dio_bytes) && loops < 3) {
396 nr_pages = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
397
398 /*
399 * Triggers inode writeback for up to nr_pages. This will invoke
400 * ->writepages callback and trigger delalloc filling
401 * (btrfs_run_delalloc_range()).
402 */
403 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
404
405 /*
406 * We need to wait for the compressed pages to start before
407 * we continue.
408 */
409 async_pages = atomic_read(&fs_info->async_delalloc_pages);
410 if (!async_pages)
411 goto skip_async;
412
413 /*
414 * Calculate how many compressed pages we want to be written
415 * before we continue. I.e if there are more async pages than we
416 * require wait_event will wait until nr_pages are written.
417 */
418 if (async_pages <= nr_pages)
419 async_pages = 0;
420 else
421 async_pages -= nr_pages;
422
423 wait_event(fs_info->async_submit_wait,
424 atomic_read(&fs_info->async_delalloc_pages) <=
425 (int)async_pages);
426skip_async:
427 spin_lock(&space_info->lock);
428 if (list_empty(&space_info->tickets) &&
429 list_empty(&space_info->priority_tickets)) {
430 spin_unlock(&space_info->lock);
431 break;
432 }
433 spin_unlock(&space_info->lock);
434
435 loops++;
436 if (wait_ordered && !trans) {
437 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
438 } else {
439 time_left = schedule_timeout_killable(1);
440 if (time_left)
441 break;
442 }
443 delalloc_bytes = percpu_counter_sum_positive(
444 &fs_info->delalloc_bytes);
445 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
446 }
447}
448
449/**
450 * maybe_commit_transaction - possibly commit the transaction if its ok to
451 * @root - the root we're allocating for
452 * @bytes - the number of bytes we want to reserve
453 * @force - force the commit
454 *
455 * This will check to make sure that committing the transaction will actually
456 * get us somewhere and then commit the transaction if it does. Otherwise it
457 * will return -ENOSPC.
458 */
459static int may_commit_transaction(struct btrfs_fs_info *fs_info,
460 struct btrfs_space_info *space_info)
461{
462 struct reserve_ticket *ticket = NULL;
463 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
464 struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
465 struct btrfs_trans_handle *trans;
466 u64 bytes_needed;
467 u64 reclaim_bytes = 0;
468 u64 cur_free_bytes = 0;
469
470 trans = (struct btrfs_trans_handle *)current->journal_info;
471 if (trans)
472 return -EAGAIN;
473
474 spin_lock(&space_info->lock);
475 cur_free_bytes = btrfs_space_info_used(space_info, true);
476 if (cur_free_bytes < space_info->total_bytes)
477 cur_free_bytes = space_info->total_bytes - cur_free_bytes;
478 else
479 cur_free_bytes = 0;
480
481 if (!list_empty(&space_info->priority_tickets))
482 ticket = list_first_entry(&space_info->priority_tickets,
483 struct reserve_ticket, list);
484 else if (!list_empty(&space_info->tickets))
485 ticket = list_first_entry(&space_info->tickets,
486 struct reserve_ticket, list);
487 bytes_needed = (ticket) ? ticket->bytes : 0;
488
489 if (bytes_needed > cur_free_bytes)
490 bytes_needed -= cur_free_bytes;
491 else
492 bytes_needed = 0;
493 spin_unlock(&space_info->lock);
494
495 if (!bytes_needed)
496 return 0;
497
498 trans = btrfs_join_transaction(fs_info->extent_root);
499 if (IS_ERR(trans))
500 return PTR_ERR(trans);
501
502 /*
503 * See if there is enough pinned space to make this reservation, or if
504 * we have block groups that are going to be freed, allowing us to
505 * possibly do a chunk allocation the next loop through.
506 */
507 if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
508 __percpu_counter_compare(&space_info->total_bytes_pinned,
509 bytes_needed,
510 BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
511 goto commit;
512
513 /*
514 * See if there is some space in the delayed insertion reservation for
515 * this reservation.
516 */
517 if (space_info != delayed_rsv->space_info)
518 goto enospc;
519
520 spin_lock(&delayed_rsv->lock);
521 reclaim_bytes += delayed_rsv->reserved;
522 spin_unlock(&delayed_rsv->lock);
523
524 spin_lock(&delayed_refs_rsv->lock);
525 reclaim_bytes += delayed_refs_rsv->reserved;
526 spin_unlock(&delayed_refs_rsv->lock);
527 if (reclaim_bytes >= bytes_needed)
528 goto commit;
529 bytes_needed -= reclaim_bytes;
530
531 if (__percpu_counter_compare(&space_info->total_bytes_pinned,
532 bytes_needed,
533 BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
534 goto enospc;
535
536commit:
537 return btrfs_commit_transaction(trans);
538enospc:
539 btrfs_end_transaction(trans);
540 return -ENOSPC;
541}
542
543/*
544 * Try to flush some data based on policy set by @state. This is only advisory
545 * and may fail for various reasons. The caller is supposed to examine the
546 * state of @space_info to detect the outcome.
547 */
548static void flush_space(struct btrfs_fs_info *fs_info,
549 struct btrfs_space_info *space_info, u64 num_bytes,
550 int state)
551{
552 struct btrfs_root *root = fs_info->extent_root;
553 struct btrfs_trans_handle *trans;
554 int nr;
555 int ret = 0;
556
557 switch (state) {
558 case FLUSH_DELAYED_ITEMS_NR:
559 case FLUSH_DELAYED_ITEMS:
560 if (state == FLUSH_DELAYED_ITEMS_NR)
561 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
562 else
563 nr = -1;
564
565 trans = btrfs_join_transaction(root);
566 if (IS_ERR(trans)) {
567 ret = PTR_ERR(trans);
568 break;
569 }
570 ret = btrfs_run_delayed_items_nr(trans, nr);
571 btrfs_end_transaction(trans);
572 break;
573 case FLUSH_DELALLOC:
574 case FLUSH_DELALLOC_WAIT:
575 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
576 state == FLUSH_DELALLOC_WAIT);
577 break;
578 case FLUSH_DELAYED_REFS_NR:
579 case FLUSH_DELAYED_REFS:
580 trans = btrfs_join_transaction(root);
581 if (IS_ERR(trans)) {
582 ret = PTR_ERR(trans);
583 break;
584 }
585 if (state == FLUSH_DELAYED_REFS_NR)
586 nr = calc_reclaim_items_nr(fs_info, num_bytes);
587 else
588 nr = 0;
589 btrfs_run_delayed_refs(trans, nr);
590 btrfs_end_transaction(trans);
591 break;
592 case ALLOC_CHUNK:
593 case ALLOC_CHUNK_FORCE:
594 trans = btrfs_join_transaction(root);
595 if (IS_ERR(trans)) {
596 ret = PTR_ERR(trans);
597 break;
598 }
599 ret = btrfs_chunk_alloc(trans,
600 btrfs_metadata_alloc_profile(fs_info),
601 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
602 CHUNK_ALLOC_FORCE);
603 btrfs_end_transaction(trans);
604 if (ret > 0 || ret == -ENOSPC)
605 ret = 0;
606 break;
607 case RUN_DELAYED_IPUTS:
608 /*
609 * If we have pending delayed iputs then we could free up a
610 * bunch of pinned space, so make sure we run the iputs before
611 * we do our pinned bytes check below.
612 */
613 btrfs_run_delayed_iputs(fs_info);
614 btrfs_wait_on_delayed_iputs(fs_info);
615 break;
616 case COMMIT_TRANS:
617 ret = may_commit_transaction(fs_info, space_info);
618 break;
619 default:
620 ret = -ENOSPC;
621 break;
622 }
623
624 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
625 ret);
626 return;
627}
628
629static inline u64
630btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
631 struct btrfs_space_info *space_info,
632 bool system_chunk)
633{
634 struct reserve_ticket *ticket;
635 u64 used;
636 u64 expected;
637 u64 to_reclaim = 0;
638
639 list_for_each_entry(ticket, &space_info->tickets, list)
640 to_reclaim += ticket->bytes;
641 list_for_each_entry(ticket, &space_info->priority_tickets, list)
642 to_reclaim += ticket->bytes;
643 if (to_reclaim)
644 return to_reclaim;
645
646 to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
647 if (can_overcommit(fs_info, space_info, to_reclaim,
648 BTRFS_RESERVE_FLUSH_ALL, system_chunk))
649 return 0;
650
651 used = btrfs_space_info_used(space_info, true);
652
653 if (can_overcommit(fs_info, space_info, SZ_1M,
654 BTRFS_RESERVE_FLUSH_ALL, system_chunk))
655 expected = div_factor_fine(space_info->total_bytes, 95);
656 else
657 expected = div_factor_fine(space_info->total_bytes, 90);
658
659 if (used > expected)
660 to_reclaim = used - expected;
661 else
662 to_reclaim = 0;
663 to_reclaim = min(to_reclaim, space_info->bytes_may_use +
664 space_info->bytes_reserved);
665 return to_reclaim;
666}
667
668static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
669 struct btrfs_space_info *space_info,
670 u64 used, bool system_chunk)
671{
672 u64 thresh = div_factor_fine(space_info->total_bytes, 98);
673
674 /* If we're just plain full then async reclaim just slows us down. */
675 if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
676 return 0;
677
678 if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info,
679 system_chunk))
680 return 0;
681
682 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
683 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
684}
685
686/*
687 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
688 * @fs_info - fs_info for this fs
689 * @space_info - the space info we were flushing
690 *
691 * We call this when we've exhausted our flushing ability and haven't made
692 * progress in satisfying tickets. The reservation code handles tickets in
693 * order, so if there is a large ticket first and then smaller ones we could
694 * very well satisfy the smaller tickets. This will attempt to wake up any
695 * tickets in the list to catch this case.
696 *
697 * This function returns true if it was able to make progress by clearing out
698 * other tickets, or if it stumbles across a ticket that was smaller than the
699 * first ticket.
700 */
701static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
702 struct btrfs_space_info *space_info)
703{
704 struct reserve_ticket *ticket;
705 u64 tickets_id = space_info->tickets_id;
706 u64 first_ticket_bytes = 0;
707
708 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
709 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
710 __btrfs_dump_space_info(fs_info, space_info);
711 }
712
713 while (!list_empty(&space_info->tickets) &&
714 tickets_id == space_info->tickets_id) {
715 ticket = list_first_entry(&space_info->tickets,
716 struct reserve_ticket, list);
717
718 /*
719 * may_commit_transaction will avoid committing the transaction
720 * if it doesn't feel like the space reclaimed by the commit
721 * would result in the ticket succeeding. However if we have a
722 * smaller ticket in the queue it may be small enough to be
723 * satisified by committing the transaction, so if any
724 * subsequent ticket is smaller than the first ticket go ahead
725 * and send us back for another loop through the enospc flushing
726 * code.
727 */
728 if (first_ticket_bytes == 0)
729 first_ticket_bytes = ticket->bytes;
730 else if (first_ticket_bytes > ticket->bytes)
731 return true;
732
733 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
734 btrfs_info(fs_info, "failing ticket with %llu bytes",
735 ticket->bytes);
736
737 list_del_init(&ticket->list);
738 ticket->error = -ENOSPC;
739 wake_up(&ticket->wait);
740
741 /*
742 * We're just throwing tickets away, so more flushing may not
743 * trip over btrfs_try_granting_tickets, so we need to call it
744 * here to see if we can make progress with the next ticket in
745 * the list.
746 */
747 btrfs_try_granting_tickets(fs_info, space_info);
748 }
749 return (tickets_id != space_info->tickets_id);
750}
751
752/*
753 * This is for normal flushers, we can wait all goddamned day if we want to. We
754 * will loop and continuously try to flush as long as we are making progress.
755 * We count progress as clearing off tickets each time we have to loop.
756 */
757static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
758{
759 struct btrfs_fs_info *fs_info;
760 struct btrfs_space_info *space_info;
761 u64 to_reclaim;
762 int flush_state;
763 int commit_cycles = 0;
764 u64 last_tickets_id;
765
766 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
767 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
768
769 spin_lock(&space_info->lock);
770 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
771 false);
772 if (!to_reclaim) {
773 space_info->flush = 0;
774 spin_unlock(&space_info->lock);
775 return;
776 }
777 last_tickets_id = space_info->tickets_id;
778 spin_unlock(&space_info->lock);
779
780 flush_state = FLUSH_DELAYED_ITEMS_NR;
781 do {
782 flush_space(fs_info, space_info, to_reclaim, flush_state);
783 spin_lock(&space_info->lock);
784 if (list_empty(&space_info->tickets)) {
785 space_info->flush = 0;
786 spin_unlock(&space_info->lock);
787 return;
788 }
789 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
790 space_info,
791 false);
792 if (last_tickets_id == space_info->tickets_id) {
793 flush_state++;
794 } else {
795 last_tickets_id = space_info->tickets_id;
796 flush_state = FLUSH_DELAYED_ITEMS_NR;
797 if (commit_cycles)
798 commit_cycles--;
799 }
800
801 /*
802 * We don't want to force a chunk allocation until we've tried
803 * pretty hard to reclaim space. Think of the case where we
804 * freed up a bunch of space and so have a lot of pinned space
805 * to reclaim. We would rather use that than possibly create a
806 * underutilized metadata chunk. So if this is our first run
807 * through the flushing state machine skip ALLOC_CHUNK_FORCE and
808 * commit the transaction. If nothing has changed the next go
809 * around then we can force a chunk allocation.
810 */
811 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
812 flush_state++;
813
814 if (flush_state > COMMIT_TRANS) {
815 commit_cycles++;
816 if (commit_cycles > 2) {
817 if (maybe_fail_all_tickets(fs_info, space_info)) {
818 flush_state = FLUSH_DELAYED_ITEMS_NR;
819 commit_cycles--;
820 } else {
821 space_info->flush = 0;
822 }
823 } else {
824 flush_state = FLUSH_DELAYED_ITEMS_NR;
825 }
826 }
827 spin_unlock(&space_info->lock);
828 } while (flush_state <= COMMIT_TRANS);
829}
830
831void btrfs_init_async_reclaim_work(struct work_struct *work)
832{
833 INIT_WORK(work, btrfs_async_reclaim_metadata_space);
834}
835
836static const enum btrfs_flush_state priority_flush_states[] = {
837 FLUSH_DELAYED_ITEMS_NR,
838 FLUSH_DELAYED_ITEMS,
839 ALLOC_CHUNK,
840};
841
842static const enum btrfs_flush_state evict_flush_states[] = {
843 FLUSH_DELAYED_ITEMS_NR,
844 FLUSH_DELAYED_ITEMS,
845 FLUSH_DELAYED_REFS_NR,
846 FLUSH_DELAYED_REFS,
847 FLUSH_DELALLOC,
848 FLUSH_DELALLOC_WAIT,
849 ALLOC_CHUNK,
850 COMMIT_TRANS,
851};
852
853static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
854 struct btrfs_space_info *space_info,
855 struct reserve_ticket *ticket,
856 const enum btrfs_flush_state *states,
857 int states_nr)
858{
859 u64 to_reclaim;
860 int flush_state;
861
862 spin_lock(&space_info->lock);
863 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info,
864 false);
865 if (!to_reclaim) {
866 spin_unlock(&space_info->lock);
867 return;
868 }
869 spin_unlock(&space_info->lock);
870
871 flush_state = 0;
872 do {
873 flush_space(fs_info, space_info, to_reclaim, states[flush_state]);
874 flush_state++;
875 spin_lock(&space_info->lock);
876 if (ticket->bytes == 0) {
877 spin_unlock(&space_info->lock);
878 return;
879 }
880 spin_unlock(&space_info->lock);
881 } while (flush_state < states_nr);
882}
883
884static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
885 struct btrfs_space_info *space_info,
886 struct reserve_ticket *ticket)
887
888{
889 DEFINE_WAIT(wait);
890 int ret = 0;
891
892 spin_lock(&space_info->lock);
893 while (ticket->bytes > 0 && ticket->error == 0) {
894 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
895 if (ret) {
896 /*
897 * Delete us from the list. After we unlock the space
898 * info, we don't want the async reclaim job to reserve
899 * space for this ticket. If that would happen, then the
900 * ticket's task would not known that space was reserved
901 * despite getting an error, resulting in a space leak
902 * (bytes_may_use counter of our space_info).
903 */
904 list_del_init(&ticket->list);
905 ticket->error = -EINTR;
906 break;
907 }
908 spin_unlock(&space_info->lock);
909
910 schedule();
911
912 finish_wait(&ticket->wait, &wait);
913 spin_lock(&space_info->lock);
914 }
915 spin_unlock(&space_info->lock);
916}
917
918/**
919 * handle_reserve_ticket - do the appropriate flushing and waiting for a ticket
920 * @fs_info - the fs
921 * @space_info - the space_info for the reservation
922 * @ticket - the ticket for the reservation
923 * @flush - how much we can flush
924 *
925 * This does the work of figuring out how to flush for the ticket, waiting for
926 * the reservation, and returning the appropriate error if there is one.
927 */
928static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
929 struct btrfs_space_info *space_info,
930 struct reserve_ticket *ticket,
931 enum btrfs_reserve_flush_enum flush)
932{
933 int ret;
934
935 switch (flush) {
936 case BTRFS_RESERVE_FLUSH_ALL:
937 wait_reserve_ticket(fs_info, space_info, ticket);
938 break;
939 case BTRFS_RESERVE_FLUSH_LIMIT:
940 priority_reclaim_metadata_space(fs_info, space_info, ticket,
941 priority_flush_states,
942 ARRAY_SIZE(priority_flush_states));
943 break;
944 case BTRFS_RESERVE_FLUSH_EVICT:
945 priority_reclaim_metadata_space(fs_info, space_info, ticket,
946 evict_flush_states,
947 ARRAY_SIZE(evict_flush_states));
948 break;
949 default:
950 ASSERT(0);
951 break;
952 }
953
954 spin_lock(&space_info->lock);
955 ret = ticket->error;
956 if (ticket->bytes || ticket->error) {
957 /*
958 * Need to delete here for priority tickets. For regular tickets
959 * either the async reclaim job deletes the ticket from the list
960 * or we delete it ourselves at wait_reserve_ticket().
961 */
962 list_del_init(&ticket->list);
963 if (!ret)
964 ret = -ENOSPC;
965 }
966 spin_unlock(&space_info->lock);
967 ASSERT(list_empty(&ticket->list));
968 /*
969 * Check that we can't have an error set if the reservation succeeded,
970 * as that would confuse tasks and lead them to error out without
971 * releasing reserved space (if an error happens the expectation is that
972 * space wasn't reserved at all).
973 */
974 ASSERT(!(ticket->bytes == 0 && ticket->error));
975 return ret;
976}
977
978/**
979 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
980 * @root - the root we're allocating for
981 * @space_info - the space info we want to allocate from
982 * @orig_bytes - the number of bytes we want
983 * @flush - whether or not we can flush to make our reservation
984 *
985 * This will reserve orig_bytes number of bytes from the space info associated
986 * with the block_rsv. If there is not enough space it will make an attempt to
987 * flush out space to make room. It will do this by flushing delalloc if
988 * possible or committing the transaction. If flush is 0 then no attempts to
989 * regain reservations will be made and this will fail if there is not enough
990 * space already.
991 */
992static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
993 struct btrfs_space_info *space_info,
994 u64 orig_bytes,
995 enum btrfs_reserve_flush_enum flush,
996 bool system_chunk)
997{
998 struct reserve_ticket ticket;
999 u64 used;
1000 int ret = 0;
1001 bool pending_tickets;
1002
1003 ASSERT(orig_bytes);
1004 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1005
1006 spin_lock(&space_info->lock);
1007 ret = -ENOSPC;
1008 used = btrfs_space_info_used(space_info, true);
1009 pending_tickets = !list_empty(&space_info->tickets) ||
1010 !list_empty(&space_info->priority_tickets);
1011
1012 /*
1013 * Carry on if we have enough space (short-circuit) OR call
1014 * can_overcommit() to ensure we can overcommit to continue.
1015 */
1016 if (!pending_tickets &&
1017 ((used + orig_bytes <= space_info->total_bytes) ||
1018 can_overcommit(fs_info, space_info, orig_bytes, flush,
1019 system_chunk))) {
1020 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1021 orig_bytes);
1022 ret = 0;
1023 }
1024
1025 /*
1026 * If we couldn't make a reservation then setup our reservation ticket
1027 * and kick the async worker if it's not already running.
1028 *
1029 * If we are a priority flusher then we just need to add our ticket to
1030 * the list and we will do our own flushing further down.
1031 */
1032 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
1033 ticket.bytes = orig_bytes;
1034 ticket.error = 0;
1035 init_waitqueue_head(&ticket.wait);
1036 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
1037 list_add_tail(&ticket.list, &space_info->tickets);
1038 if (!space_info->flush) {
1039 space_info->flush = 1;
1040 trace_btrfs_trigger_flush(fs_info,
1041 space_info->flags,
1042 orig_bytes, flush,
1043 "enospc");
1044 queue_work(system_unbound_wq,
1045 &fs_info->async_reclaim_work);
1046 }
1047 } else {
1048 list_add_tail(&ticket.list,
1049 &space_info->priority_tickets);
1050 }
1051 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1052 used += orig_bytes;
1053 /*
1054 * We will do the space reservation dance during log replay,
1055 * which means we won't have fs_info->fs_root set, so don't do
1056 * the async reclaim as we will panic.
1057 */
1058 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1059 need_do_async_reclaim(fs_info, space_info,
1060 used, system_chunk) &&
1061 !work_busy(&fs_info->async_reclaim_work)) {
1062 trace_btrfs_trigger_flush(fs_info, space_info->flags,
1063 orig_bytes, flush, "preempt");
1064 queue_work(system_unbound_wq,
1065 &fs_info->async_reclaim_work);
1066 }
1067 }
1068 spin_unlock(&space_info->lock);
1069 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
1070 return ret;
1071
1072 return handle_reserve_ticket(fs_info, space_info, &ticket, flush);
1073}
1074
1075/**
1076 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
1077 * @root - the root we're allocating for
1078 * @block_rsv - the block_rsv we're allocating for
1079 * @orig_bytes - the number of bytes we want
1080 * @flush - whether or not we can flush to make our reservation
1081 *
1082 * This will reserve orig_bytes number of bytes from the space info associated
1083 * with the block_rsv. If there is not enough space it will make an attempt to
1084 * flush out space to make room. It will do this by flushing delalloc if
1085 * possible or committing the transaction. If flush is 0 then no attempts to
1086 * regain reservations will be made and this will fail if there is not enough
1087 * space already.
1088 */
1089int btrfs_reserve_metadata_bytes(struct btrfs_root *root,
1090 struct btrfs_block_rsv *block_rsv,
1091 u64 orig_bytes,
1092 enum btrfs_reserve_flush_enum flush)
1093{
1094 struct btrfs_fs_info *fs_info = root->fs_info;
1095 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
1096 int ret;
1097 bool system_chunk = (root == fs_info->chunk_root);
1098
1099 ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
1100 orig_bytes, flush, system_chunk);
1101 if (ret == -ENOSPC &&
1102 unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
1103 if (block_rsv != global_rsv &&
1104 !btrfs_block_rsv_use_bytes(global_rsv, orig_bytes))
1105 ret = 0;
1106 }
1107 if (ret == -ENOSPC) {
1108 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1109 block_rsv->space_info->flags,
1110 orig_bytes, 1);
1111
1112 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1113 btrfs_dump_space_info(fs_info, block_rsv->space_info,
1114 orig_bytes, 0);
1115 }
1116 return ret;
1117}
1// SPDX-License-Identifier: GPL-2.0
2
3#include "misc.h"
4#include "ctree.h"
5#include "space-info.h"
6#include "sysfs.h"
7#include "volumes.h"
8#include "free-space-cache.h"
9#include "ordered-data.h"
10#include "transaction.h"
11#include "block-group.h"
12#include "zoned.h"
13#include "fs.h"
14#include "accessors.h"
15#include "extent-tree.h"
16
17/*
18 * HOW DOES SPACE RESERVATION WORK
19 *
20 * If you want to know about delalloc specifically, there is a separate comment
21 * for that with the delalloc code. This comment is about how the whole system
22 * works generally.
23 *
24 * BASIC CONCEPTS
25 *
26 * 1) space_info. This is the ultimate arbiter of how much space we can use.
27 * There's a description of the bytes_ fields with the struct declaration,
28 * refer to that for specifics on each field. Suffice it to say that for
29 * reservations we care about total_bytes - SUM(space_info->bytes_) when
30 * determining if there is space to make an allocation. There is a space_info
31 * for METADATA, SYSTEM, and DATA areas.
32 *
33 * 2) block_rsv's. These are basically buckets for every different type of
34 * metadata reservation we have. You can see the comment in the block_rsv
35 * code on the rules for each type, but generally block_rsv->reserved is how
36 * much space is accounted for in space_info->bytes_may_use.
37 *
38 * 3) btrfs_calc*_size. These are the worst case calculations we used based
39 * on the number of items we will want to modify. We have one for changing
40 * items, and one for inserting new items. Generally we use these helpers to
41 * determine the size of the block reserves, and then use the actual bytes
42 * values to adjust the space_info counters.
43 *
44 * MAKING RESERVATIONS, THE NORMAL CASE
45 *
46 * We call into either btrfs_reserve_data_bytes() or
47 * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with
48 * num_bytes we want to reserve.
49 *
50 * ->reserve
51 * space_info->bytes_may_reserve += num_bytes
52 *
53 * ->extent allocation
54 * Call btrfs_add_reserved_bytes() which does
55 * space_info->bytes_may_reserve -= num_bytes
56 * space_info->bytes_reserved += extent_bytes
57 *
58 * ->insert reference
59 * Call btrfs_update_block_group() which does
60 * space_info->bytes_reserved -= extent_bytes
61 * space_info->bytes_used += extent_bytes
62 *
63 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority)
64 *
65 * Assume we are unable to simply make the reservation because we do not have
66 * enough space
67 *
68 * -> __reserve_bytes
69 * create a reserve_ticket with ->bytes set to our reservation, add it to
70 * the tail of space_info->tickets, kick async flush thread
71 *
72 * ->handle_reserve_ticket
73 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set
74 * on the ticket.
75 *
76 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space
77 * Flushes various things attempting to free up space.
78 *
79 * -> btrfs_try_granting_tickets()
80 * This is called by anything that either subtracts space from
81 * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the
82 * space_info->total_bytes. This loops through the ->priority_tickets and
83 * then the ->tickets list checking to see if the reservation can be
84 * completed. If it can the space is added to space_info->bytes_may_use and
85 * the ticket is woken up.
86 *
87 * -> ticket wakeup
88 * Check if ->bytes == 0, if it does we got our reservation and we can carry
89 * on, if not return the appropriate error (ENOSPC, but can be EINTR if we
90 * were interrupted.)
91 *
92 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY
93 *
94 * Same as the above, except we add ourselves to the
95 * space_info->priority_tickets, and we do not use ticket->wait, we simply
96 * call flush_space() ourselves for the states that are safe for us to call
97 * without deadlocking and hope for the best.
98 *
99 * THE FLUSHING STATES
100 *
101 * Generally speaking we will have two cases for each state, a "nice" state
102 * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to
103 * reduce the locking over head on the various trees, and even to keep from
104 * doing any work at all in the case of delayed refs. Each of these delayed
105 * things however hold reservations, and so letting them run allows us to
106 * reclaim space so we can make new reservations.
107 *
108 * FLUSH_DELAYED_ITEMS
109 * Every inode has a delayed item to update the inode. Take a simple write
110 * for example, we would update the inode item at write time to update the
111 * mtime, and then again at finish_ordered_io() time in order to update the
112 * isize or bytes. We keep these delayed items to coalesce these operations
113 * into a single operation done on demand. These are an easy way to reclaim
114 * metadata space.
115 *
116 * FLUSH_DELALLOC
117 * Look at the delalloc comment to get an idea of how much space is reserved
118 * for delayed allocation. We can reclaim some of this space simply by
119 * running delalloc, but usually we need to wait for ordered extents to
120 * reclaim the bulk of this space.
121 *
122 * FLUSH_DELAYED_REFS
123 * We have a block reserve for the outstanding delayed refs space, and every
124 * delayed ref operation holds a reservation. Running these is a quick way
125 * to reclaim space, but we want to hold this until the end because COW can
126 * churn a lot and we can avoid making some extent tree modifications if we
127 * are able to delay for as long as possible.
128 *
129 * ALLOC_CHUNK
130 * We will skip this the first time through space reservation, because of
131 * overcommit and we don't want to have a lot of useless metadata space when
132 * our worst case reservations will likely never come true.
133 *
134 * RUN_DELAYED_IPUTS
135 * If we're freeing inodes we're likely freeing checksums, file extent
136 * items, and extent tree items. Loads of space could be freed up by these
137 * operations, however they won't be usable until the transaction commits.
138 *
139 * COMMIT_TRANS
140 * This will commit the transaction. Historically we had a lot of logic
141 * surrounding whether or not we'd commit the transaction, but this waits born
142 * out of a pre-tickets era where we could end up committing the transaction
143 * thousands of times in a row without making progress. Now thanks to our
144 * ticketing system we know if we're not making progress and can error
145 * everybody out after a few commits rather than burning the disk hoping for
146 * a different answer.
147 *
148 * OVERCOMMIT
149 *
150 * Because we hold so many reservations for metadata we will allow you to
151 * reserve more space than is currently free in the currently allocate
152 * metadata space. This only happens with metadata, data does not allow
153 * overcommitting.
154 *
155 * You can see the current logic for when we allow overcommit in
156 * btrfs_can_overcommit(), but it only applies to unallocated space. If there
157 * is no unallocated space to be had, all reservations are kept within the
158 * free space in the allocated metadata chunks.
159 *
160 * Because of overcommitting, you generally want to use the
161 * btrfs_can_overcommit() logic for metadata allocations, as it does the right
162 * thing with or without extra unallocated space.
163 */
164
165u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
166 bool may_use_included)
167{
168 ASSERT(s_info);
169 return s_info->bytes_used + s_info->bytes_reserved +
170 s_info->bytes_pinned + s_info->bytes_readonly +
171 s_info->bytes_zone_unusable +
172 (may_use_included ? s_info->bytes_may_use : 0);
173}
174
175/*
176 * after adding space to the filesystem, we need to clear the full flags
177 * on all the space infos.
178 */
179void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
180{
181 struct list_head *head = &info->space_info;
182 struct btrfs_space_info *found;
183
184 list_for_each_entry(found, head, list)
185 found->full = 0;
186}
187
188/*
189 * Block groups with more than this value (percents) of unusable space will be
190 * scheduled for background reclaim.
191 */
192#define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH (75)
193
194/*
195 * Calculate chunk size depending on volume type (regular or zoned).
196 */
197static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags)
198{
199 if (btrfs_is_zoned(fs_info))
200 return fs_info->zone_size;
201
202 ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
203
204 if (flags & BTRFS_BLOCK_GROUP_DATA)
205 return BTRFS_MAX_DATA_CHUNK_SIZE;
206 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
207 return SZ_32M;
208
209 /* Handle BTRFS_BLOCK_GROUP_METADATA */
210 if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G)
211 return SZ_1G;
212
213 return SZ_256M;
214}
215
216/*
217 * Update default chunk size.
218 */
219void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info,
220 u64 chunk_size)
221{
222 WRITE_ONCE(space_info->chunk_size, chunk_size);
223}
224
225static int create_space_info(struct btrfs_fs_info *info, u64 flags)
226{
227
228 struct btrfs_space_info *space_info;
229 int i;
230 int ret;
231
232 space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
233 if (!space_info)
234 return -ENOMEM;
235
236 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
237 INIT_LIST_HEAD(&space_info->block_groups[i]);
238 init_rwsem(&space_info->groups_sem);
239 spin_lock_init(&space_info->lock);
240 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
241 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
242 INIT_LIST_HEAD(&space_info->ro_bgs);
243 INIT_LIST_HEAD(&space_info->tickets);
244 INIT_LIST_HEAD(&space_info->priority_tickets);
245 space_info->clamp = 1;
246 btrfs_update_space_info_chunk_size(space_info, calc_chunk_size(info, flags));
247
248 if (btrfs_is_zoned(info))
249 space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH;
250
251 ret = btrfs_sysfs_add_space_info_type(info, space_info);
252 if (ret)
253 return ret;
254
255 list_add(&space_info->list, &info->space_info);
256 if (flags & BTRFS_BLOCK_GROUP_DATA)
257 info->data_sinfo = space_info;
258
259 return ret;
260}
261
262int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
263{
264 struct btrfs_super_block *disk_super;
265 u64 features;
266 u64 flags;
267 int mixed = 0;
268 int ret;
269
270 disk_super = fs_info->super_copy;
271 if (!btrfs_super_root(disk_super))
272 return -EINVAL;
273
274 features = btrfs_super_incompat_flags(disk_super);
275 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
276 mixed = 1;
277
278 flags = BTRFS_BLOCK_GROUP_SYSTEM;
279 ret = create_space_info(fs_info, flags);
280 if (ret)
281 goto out;
282
283 if (mixed) {
284 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
285 ret = create_space_info(fs_info, flags);
286 } else {
287 flags = BTRFS_BLOCK_GROUP_METADATA;
288 ret = create_space_info(fs_info, flags);
289 if (ret)
290 goto out;
291
292 flags = BTRFS_BLOCK_GROUP_DATA;
293 ret = create_space_info(fs_info, flags);
294 }
295out:
296 return ret;
297}
298
299void btrfs_add_bg_to_space_info(struct btrfs_fs_info *info,
300 struct btrfs_block_group *block_group)
301{
302 struct btrfs_space_info *found;
303 int factor, index;
304
305 factor = btrfs_bg_type_to_factor(block_group->flags);
306
307 found = btrfs_find_space_info(info, block_group->flags);
308 ASSERT(found);
309 spin_lock(&found->lock);
310 found->total_bytes += block_group->length;
311 if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags))
312 found->active_total_bytes += block_group->length;
313 found->disk_total += block_group->length * factor;
314 found->bytes_used += block_group->used;
315 found->disk_used += block_group->used * factor;
316 found->bytes_readonly += block_group->bytes_super;
317 found->bytes_zone_unusable += block_group->zone_unusable;
318 if (block_group->length > 0)
319 found->full = 0;
320 btrfs_try_granting_tickets(info, found);
321 spin_unlock(&found->lock);
322
323 block_group->space_info = found;
324
325 index = btrfs_bg_flags_to_raid_index(block_group->flags);
326 down_write(&found->groups_sem);
327 list_add_tail(&block_group->list, &found->block_groups[index]);
328 up_write(&found->groups_sem);
329}
330
331struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
332 u64 flags)
333{
334 struct list_head *head = &info->space_info;
335 struct btrfs_space_info *found;
336
337 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
338
339 list_for_each_entry(found, head, list) {
340 if (found->flags & flags)
341 return found;
342 }
343 return NULL;
344}
345
346static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
347 struct btrfs_space_info *space_info,
348 enum btrfs_reserve_flush_enum flush)
349{
350 u64 profile;
351 u64 avail;
352 int factor;
353
354 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
355 profile = btrfs_system_alloc_profile(fs_info);
356 else
357 profile = btrfs_metadata_alloc_profile(fs_info);
358
359 avail = atomic64_read(&fs_info->free_chunk_space);
360
361 /*
362 * If we have dup, raid1 or raid10 then only half of the free
363 * space is actually usable. For raid56, the space info used
364 * doesn't include the parity drive, so we don't have to
365 * change the math
366 */
367 factor = btrfs_bg_type_to_factor(profile);
368 avail = div_u64(avail, factor);
369
370 /*
371 * If we aren't flushing all things, let us overcommit up to
372 * 1/2th of the space. If we can flush, don't let us overcommit
373 * too much, let it overcommit up to 1/8 of the space.
374 */
375 if (flush == BTRFS_RESERVE_FLUSH_ALL)
376 avail >>= 3;
377 else
378 avail >>= 1;
379 return avail;
380}
381
382static inline u64 writable_total_bytes(struct btrfs_fs_info *fs_info,
383 struct btrfs_space_info *space_info)
384{
385 /*
386 * On regular filesystem, all total_bytes are always writable. On zoned
387 * filesystem, there may be a limitation imposed by max_active_zones.
388 * For metadata allocation, we cannot finish an existing active block
389 * group to avoid a deadlock. Thus, we need to consider only the active
390 * groups to be writable for metadata space.
391 */
392 if (!btrfs_is_zoned(fs_info) || (space_info->flags & BTRFS_BLOCK_GROUP_DATA))
393 return space_info->total_bytes;
394
395 return space_info->active_total_bytes;
396}
397
398int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
399 struct btrfs_space_info *space_info, u64 bytes,
400 enum btrfs_reserve_flush_enum flush)
401{
402 u64 avail;
403 u64 used;
404
405 /* Don't overcommit when in mixed mode */
406 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
407 return 0;
408
409 used = btrfs_space_info_used(space_info, true);
410 if (test_bit(BTRFS_FS_NO_OVERCOMMIT, &fs_info->flags) &&
411 (space_info->flags & BTRFS_BLOCK_GROUP_METADATA))
412 avail = 0;
413 else
414 avail = calc_available_free_space(fs_info, space_info, flush);
415
416 if (used + bytes < writable_total_bytes(fs_info, space_info) + avail)
417 return 1;
418 return 0;
419}
420
421static void remove_ticket(struct btrfs_space_info *space_info,
422 struct reserve_ticket *ticket)
423{
424 if (!list_empty(&ticket->list)) {
425 list_del_init(&ticket->list);
426 ASSERT(space_info->reclaim_size >= ticket->bytes);
427 space_info->reclaim_size -= ticket->bytes;
428 }
429}
430
431/*
432 * This is for space we already have accounted in space_info->bytes_may_use, so
433 * basically when we're returning space from block_rsv's.
434 */
435void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
436 struct btrfs_space_info *space_info)
437{
438 struct list_head *head;
439 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
440
441 lockdep_assert_held(&space_info->lock);
442
443 head = &space_info->priority_tickets;
444again:
445 while (!list_empty(head)) {
446 struct reserve_ticket *ticket;
447 u64 used = btrfs_space_info_used(space_info, true);
448
449 ticket = list_first_entry(head, struct reserve_ticket, list);
450
451 /* Check and see if our ticket can be satisfied now. */
452 if ((used + ticket->bytes <= writable_total_bytes(fs_info, space_info)) ||
453 btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
454 flush)) {
455 btrfs_space_info_update_bytes_may_use(fs_info,
456 space_info,
457 ticket->bytes);
458 remove_ticket(space_info, ticket);
459 ticket->bytes = 0;
460 space_info->tickets_id++;
461 wake_up(&ticket->wait);
462 } else {
463 break;
464 }
465 }
466
467 if (head == &space_info->priority_tickets) {
468 head = &space_info->tickets;
469 flush = BTRFS_RESERVE_FLUSH_ALL;
470 goto again;
471 }
472}
473
474#define DUMP_BLOCK_RSV(fs_info, rsv_name) \
475do { \
476 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \
477 spin_lock(&__rsv->lock); \
478 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \
479 __rsv->size, __rsv->reserved); \
480 spin_unlock(&__rsv->lock); \
481} while (0)
482
483static const char *space_info_flag_to_str(const struct btrfs_space_info *space_info)
484{
485 switch (space_info->flags) {
486 case BTRFS_BLOCK_GROUP_SYSTEM:
487 return "SYSTEM";
488 case BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA:
489 return "DATA+METADATA";
490 case BTRFS_BLOCK_GROUP_DATA:
491 return "DATA";
492 case BTRFS_BLOCK_GROUP_METADATA:
493 return "METADATA";
494 default:
495 return "UNKNOWN";
496 }
497}
498
499static void dump_global_block_rsv(struct btrfs_fs_info *fs_info)
500{
501 DUMP_BLOCK_RSV(fs_info, global_block_rsv);
502 DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
503 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
504 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
505 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
506}
507
508static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
509 struct btrfs_space_info *info)
510{
511 const char *flag_str = space_info_flag_to_str(info);
512 lockdep_assert_held(&info->lock);
513
514 /* The free space could be negative in case of overcommit */
515 btrfs_info(fs_info, "space_info %s has %lld free, is %sfull",
516 flag_str,
517 (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
518 info->full ? "" : "not ");
519 btrfs_info(fs_info,
520"space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
521 info->total_bytes, info->bytes_used, info->bytes_pinned,
522 info->bytes_reserved, info->bytes_may_use,
523 info->bytes_readonly, info->bytes_zone_unusable);
524}
525
526void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
527 struct btrfs_space_info *info, u64 bytes,
528 int dump_block_groups)
529{
530 struct btrfs_block_group *cache;
531 int index = 0;
532
533 spin_lock(&info->lock);
534 __btrfs_dump_space_info(fs_info, info);
535 dump_global_block_rsv(fs_info);
536 spin_unlock(&info->lock);
537
538 if (!dump_block_groups)
539 return;
540
541 down_read(&info->groups_sem);
542again:
543 list_for_each_entry(cache, &info->block_groups[index], list) {
544 spin_lock(&cache->lock);
545 btrfs_info(fs_info,
546 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu zone_unusable %s",
547 cache->start, cache->length, cache->used, cache->pinned,
548 cache->reserved, cache->zone_unusable,
549 cache->ro ? "[readonly]" : "");
550 spin_unlock(&cache->lock);
551 btrfs_dump_free_space(cache, bytes);
552 }
553 if (++index < BTRFS_NR_RAID_TYPES)
554 goto again;
555 up_read(&info->groups_sem);
556}
557
558static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
559 u64 to_reclaim)
560{
561 u64 bytes;
562 u64 nr;
563
564 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
565 nr = div64_u64(to_reclaim, bytes);
566 if (!nr)
567 nr = 1;
568 return nr;
569}
570
571#define EXTENT_SIZE_PER_ITEM SZ_256K
572
573/*
574 * shrink metadata reservation for delalloc
575 */
576static void shrink_delalloc(struct btrfs_fs_info *fs_info,
577 struct btrfs_space_info *space_info,
578 u64 to_reclaim, bool wait_ordered,
579 bool for_preempt)
580{
581 struct btrfs_trans_handle *trans;
582 u64 delalloc_bytes;
583 u64 ordered_bytes;
584 u64 items;
585 long time_left;
586 int loops;
587
588 delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
589 ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes);
590 if (delalloc_bytes == 0 && ordered_bytes == 0)
591 return;
592
593 /* Calc the number of the pages we need flush for space reservation */
594 if (to_reclaim == U64_MAX) {
595 items = U64_MAX;
596 } else {
597 /*
598 * to_reclaim is set to however much metadata we need to
599 * reclaim, but reclaiming that much data doesn't really track
600 * exactly. What we really want to do is reclaim full inode's
601 * worth of reservations, however that's not available to us
602 * here. We will take a fraction of the delalloc bytes for our
603 * flushing loops and hope for the best. Delalloc will expand
604 * the amount we write to cover an entire dirty extent, which
605 * will reclaim the metadata reservation for that range. If
606 * it's not enough subsequent flush stages will be more
607 * aggressive.
608 */
609 to_reclaim = max(to_reclaim, delalloc_bytes >> 3);
610 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
611 }
612
613 trans = current->journal_info;
614
615 /*
616 * If we are doing more ordered than delalloc we need to just wait on
617 * ordered extents, otherwise we'll waste time trying to flush delalloc
618 * that likely won't give us the space back we need.
619 */
620 if (ordered_bytes > delalloc_bytes && !for_preempt)
621 wait_ordered = true;
622
623 loops = 0;
624 while ((delalloc_bytes || ordered_bytes) && loops < 3) {
625 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
626 long nr_pages = min_t(u64, temp, LONG_MAX);
627 int async_pages;
628
629 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
630
631 /*
632 * We need to make sure any outstanding async pages are now
633 * processed before we continue. This is because things like
634 * sync_inode() try to be smart and skip writing if the inode is
635 * marked clean. We don't use filemap_fwrite for flushing
636 * because we want to control how many pages we write out at a
637 * time, thus this is the only safe way to make sure we've
638 * waited for outstanding compressed workers to have started
639 * their jobs and thus have ordered extents set up properly.
640 *
641 * This exists because we do not want to wait for each
642 * individual inode to finish its async work, we simply want to
643 * start the IO on everybody, and then come back here and wait
644 * for all of the async work to catch up. Once we're done with
645 * that we know we'll have ordered extents for everything and we
646 * can decide if we wait for that or not.
647 *
648 * If we choose to replace this in the future, make absolutely
649 * sure that the proper waiting is being done in the async case,
650 * as there have been bugs in that area before.
651 */
652 async_pages = atomic_read(&fs_info->async_delalloc_pages);
653 if (!async_pages)
654 goto skip_async;
655
656 /*
657 * We don't want to wait forever, if we wrote less pages in this
658 * loop than we have outstanding, only wait for that number of
659 * pages, otherwise we can wait for all async pages to finish
660 * before continuing.
661 */
662 if (async_pages > nr_pages)
663 async_pages -= nr_pages;
664 else
665 async_pages = 0;
666 wait_event(fs_info->async_submit_wait,
667 atomic_read(&fs_info->async_delalloc_pages) <=
668 async_pages);
669skip_async:
670 loops++;
671 if (wait_ordered && !trans) {
672 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
673 } else {
674 time_left = schedule_timeout_killable(1);
675 if (time_left)
676 break;
677 }
678
679 /*
680 * If we are for preemption we just want a one-shot of delalloc
681 * flushing so we can stop flushing if we decide we don't need
682 * to anymore.
683 */
684 if (for_preempt)
685 break;
686
687 spin_lock(&space_info->lock);
688 if (list_empty(&space_info->tickets) &&
689 list_empty(&space_info->priority_tickets)) {
690 spin_unlock(&space_info->lock);
691 break;
692 }
693 spin_unlock(&space_info->lock);
694
695 delalloc_bytes = percpu_counter_sum_positive(
696 &fs_info->delalloc_bytes);
697 ordered_bytes = percpu_counter_sum_positive(
698 &fs_info->ordered_bytes);
699 }
700}
701
702/*
703 * Try to flush some data based on policy set by @state. This is only advisory
704 * and may fail for various reasons. The caller is supposed to examine the
705 * state of @space_info to detect the outcome.
706 */
707static void flush_space(struct btrfs_fs_info *fs_info,
708 struct btrfs_space_info *space_info, u64 num_bytes,
709 enum btrfs_flush_state state, bool for_preempt)
710{
711 struct btrfs_root *root = fs_info->tree_root;
712 struct btrfs_trans_handle *trans;
713 int nr;
714 int ret = 0;
715
716 switch (state) {
717 case FLUSH_DELAYED_ITEMS_NR:
718 case FLUSH_DELAYED_ITEMS:
719 if (state == FLUSH_DELAYED_ITEMS_NR)
720 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
721 else
722 nr = -1;
723
724 trans = btrfs_join_transaction(root);
725 if (IS_ERR(trans)) {
726 ret = PTR_ERR(trans);
727 break;
728 }
729 ret = btrfs_run_delayed_items_nr(trans, nr);
730 btrfs_end_transaction(trans);
731 break;
732 case FLUSH_DELALLOC:
733 case FLUSH_DELALLOC_WAIT:
734 case FLUSH_DELALLOC_FULL:
735 if (state == FLUSH_DELALLOC_FULL)
736 num_bytes = U64_MAX;
737 shrink_delalloc(fs_info, space_info, num_bytes,
738 state != FLUSH_DELALLOC, for_preempt);
739 break;
740 case FLUSH_DELAYED_REFS_NR:
741 case FLUSH_DELAYED_REFS:
742 trans = btrfs_join_transaction(root);
743 if (IS_ERR(trans)) {
744 ret = PTR_ERR(trans);
745 break;
746 }
747 if (state == FLUSH_DELAYED_REFS_NR)
748 nr = calc_reclaim_items_nr(fs_info, num_bytes);
749 else
750 nr = 0;
751 btrfs_run_delayed_refs(trans, nr);
752 btrfs_end_transaction(trans);
753 break;
754 case ALLOC_CHUNK:
755 case ALLOC_CHUNK_FORCE:
756 /*
757 * For metadata space on zoned filesystem, reaching here means we
758 * don't have enough space left in active_total_bytes. Try to
759 * activate a block group first, because we may have inactive
760 * block group already allocated.
761 */
762 ret = btrfs_zoned_activate_one_bg(fs_info, space_info, false);
763 if (ret < 0)
764 break;
765 else if (ret == 1)
766 break;
767
768 trans = btrfs_join_transaction(root);
769 if (IS_ERR(trans)) {
770 ret = PTR_ERR(trans);
771 break;
772 }
773 ret = btrfs_chunk_alloc(trans,
774 btrfs_get_alloc_profile(fs_info, space_info->flags),
775 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
776 CHUNK_ALLOC_FORCE);
777 btrfs_end_transaction(trans);
778
779 /*
780 * For metadata space on zoned filesystem, allocating a new chunk
781 * is not enough. We still need to activate the block * group.
782 * Active the newly allocated block group by (maybe) finishing
783 * a block group.
784 */
785 if (ret == 1) {
786 ret = btrfs_zoned_activate_one_bg(fs_info, space_info, true);
787 /*
788 * Revert to the original ret regardless we could finish
789 * one block group or not.
790 */
791 if (ret >= 0)
792 ret = 1;
793 }
794
795 if (ret > 0 || ret == -ENOSPC)
796 ret = 0;
797 break;
798 case RUN_DELAYED_IPUTS:
799 /*
800 * If we have pending delayed iputs then we could free up a
801 * bunch of pinned space, so make sure we run the iputs before
802 * we do our pinned bytes check below.
803 */
804 btrfs_run_delayed_iputs(fs_info);
805 btrfs_wait_on_delayed_iputs(fs_info);
806 break;
807 case COMMIT_TRANS:
808 ASSERT(current->journal_info == NULL);
809 trans = btrfs_join_transaction(root);
810 if (IS_ERR(trans)) {
811 ret = PTR_ERR(trans);
812 break;
813 }
814 ret = btrfs_commit_transaction(trans);
815 break;
816 default:
817 ret = -ENOSPC;
818 break;
819 }
820
821 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
822 ret, for_preempt);
823 return;
824}
825
826static inline u64
827btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
828 struct btrfs_space_info *space_info)
829{
830 u64 used;
831 u64 avail;
832 u64 total;
833 u64 to_reclaim = space_info->reclaim_size;
834
835 lockdep_assert_held(&space_info->lock);
836
837 avail = calc_available_free_space(fs_info, space_info,
838 BTRFS_RESERVE_FLUSH_ALL);
839 used = btrfs_space_info_used(space_info, true);
840
841 /*
842 * We may be flushing because suddenly we have less space than we had
843 * before, and now we're well over-committed based on our current free
844 * space. If that's the case add in our overage so we make sure to put
845 * appropriate pressure on the flushing state machine.
846 */
847 total = writable_total_bytes(fs_info, space_info);
848 if (total + avail < used)
849 to_reclaim += used - (total + avail);
850
851 return to_reclaim;
852}
853
854static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
855 struct btrfs_space_info *space_info)
856{
857 u64 global_rsv_size = fs_info->global_block_rsv.reserved;
858 u64 ordered, delalloc;
859 u64 total = writable_total_bytes(fs_info, space_info);
860 u64 thresh;
861 u64 used;
862
863 thresh = mult_perc(total, 90);
864
865 lockdep_assert_held(&space_info->lock);
866
867 /* If we're just plain full then async reclaim just slows us down. */
868 if ((space_info->bytes_used + space_info->bytes_reserved +
869 global_rsv_size) >= thresh)
870 return false;
871
872 used = space_info->bytes_may_use + space_info->bytes_pinned;
873
874 /* The total flushable belongs to the global rsv, don't flush. */
875 if (global_rsv_size >= used)
876 return false;
877
878 /*
879 * 128MiB is 1/4 of the maximum global rsv size. If we have less than
880 * that devoted to other reservations then there's no sense in flushing,
881 * we don't have a lot of things that need flushing.
882 */
883 if (used - global_rsv_size <= SZ_128M)
884 return false;
885
886 /*
887 * We have tickets queued, bail so we don't compete with the async
888 * flushers.
889 */
890 if (space_info->reclaim_size)
891 return false;
892
893 /*
894 * If we have over half of the free space occupied by reservations or
895 * pinned then we want to start flushing.
896 *
897 * We do not do the traditional thing here, which is to say
898 *
899 * if (used >= ((total_bytes + avail) / 2))
900 * return 1;
901 *
902 * because this doesn't quite work how we want. If we had more than 50%
903 * of the space_info used by bytes_used and we had 0 available we'd just
904 * constantly run the background flusher. Instead we want it to kick in
905 * if our reclaimable space exceeds our clamped free space.
906 *
907 * Our clamping range is 2^1 -> 2^8. Practically speaking that means
908 * the following:
909 *
910 * Amount of RAM Minimum threshold Maximum threshold
911 *
912 * 256GiB 1GiB 128GiB
913 * 128GiB 512MiB 64GiB
914 * 64GiB 256MiB 32GiB
915 * 32GiB 128MiB 16GiB
916 * 16GiB 64MiB 8GiB
917 *
918 * These are the range our thresholds will fall in, corresponding to how
919 * much delalloc we need for the background flusher to kick in.
920 */
921
922 thresh = calc_available_free_space(fs_info, space_info,
923 BTRFS_RESERVE_FLUSH_ALL);
924 used = space_info->bytes_used + space_info->bytes_reserved +
925 space_info->bytes_readonly + global_rsv_size;
926 if (used < total)
927 thresh += total - used;
928 thresh >>= space_info->clamp;
929
930 used = space_info->bytes_pinned;
931
932 /*
933 * If we have more ordered bytes than delalloc bytes then we're either
934 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting
935 * around. Preemptive flushing is only useful in that it can free up
936 * space before tickets need to wait for things to finish. In the case
937 * of ordered extents, preemptively waiting on ordered extents gets us
938 * nothing, if our reservations are tied up in ordered extents we'll
939 * simply have to slow down writers by forcing them to wait on ordered
940 * extents.
941 *
942 * In the case that ordered is larger than delalloc, only include the
943 * block reserves that we would actually be able to directly reclaim
944 * from. In this case if we're heavy on metadata operations this will
945 * clearly be heavy enough to warrant preemptive flushing. In the case
946 * of heavy DIO or ordered reservations, preemptive flushing will just
947 * waste time and cause us to slow down.
948 *
949 * We want to make sure we truly are maxed out on ordered however, so
950 * cut ordered in half, and if it's still higher than delalloc then we
951 * can keep flushing. This is to avoid the case where we start
952 * flushing, and now delalloc == ordered and we stop preemptively
953 * flushing when we could still have several gigs of delalloc to flush.
954 */
955 ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1;
956 delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes);
957 if (ordered >= delalloc)
958 used += fs_info->delayed_refs_rsv.reserved +
959 fs_info->delayed_block_rsv.reserved;
960 else
961 used += space_info->bytes_may_use - global_rsv_size;
962
963 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
964 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
965}
966
967static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
968 struct btrfs_space_info *space_info,
969 struct reserve_ticket *ticket)
970{
971 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
972 u64 min_bytes;
973
974 if (!ticket->steal)
975 return false;
976
977 if (global_rsv->space_info != space_info)
978 return false;
979
980 spin_lock(&global_rsv->lock);
981 min_bytes = mult_perc(global_rsv->size, 10);
982 if (global_rsv->reserved < min_bytes + ticket->bytes) {
983 spin_unlock(&global_rsv->lock);
984 return false;
985 }
986 global_rsv->reserved -= ticket->bytes;
987 remove_ticket(space_info, ticket);
988 ticket->bytes = 0;
989 wake_up(&ticket->wait);
990 space_info->tickets_id++;
991 if (global_rsv->reserved < global_rsv->size)
992 global_rsv->full = 0;
993 spin_unlock(&global_rsv->lock);
994
995 return true;
996}
997
998/*
999 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
1000 * @fs_info - fs_info for this fs
1001 * @space_info - the space info we were flushing
1002 *
1003 * We call this when we've exhausted our flushing ability and haven't made
1004 * progress in satisfying tickets. The reservation code handles tickets in
1005 * order, so if there is a large ticket first and then smaller ones we could
1006 * very well satisfy the smaller tickets. This will attempt to wake up any
1007 * tickets in the list to catch this case.
1008 *
1009 * This function returns true if it was able to make progress by clearing out
1010 * other tickets, or if it stumbles across a ticket that was smaller than the
1011 * first ticket.
1012 */
1013static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
1014 struct btrfs_space_info *space_info)
1015{
1016 struct reserve_ticket *ticket;
1017 u64 tickets_id = space_info->tickets_id;
1018 const bool aborted = BTRFS_FS_ERROR(fs_info);
1019
1020 trace_btrfs_fail_all_tickets(fs_info, space_info);
1021
1022 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1023 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
1024 __btrfs_dump_space_info(fs_info, space_info);
1025 }
1026
1027 while (!list_empty(&space_info->tickets) &&
1028 tickets_id == space_info->tickets_id) {
1029 ticket = list_first_entry(&space_info->tickets,
1030 struct reserve_ticket, list);
1031
1032 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket))
1033 return true;
1034
1035 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1036 btrfs_info(fs_info, "failing ticket with %llu bytes",
1037 ticket->bytes);
1038
1039 remove_ticket(space_info, ticket);
1040 if (aborted)
1041 ticket->error = -EIO;
1042 else
1043 ticket->error = -ENOSPC;
1044 wake_up(&ticket->wait);
1045
1046 /*
1047 * We're just throwing tickets away, so more flushing may not
1048 * trip over btrfs_try_granting_tickets, so we need to call it
1049 * here to see if we can make progress with the next ticket in
1050 * the list.
1051 */
1052 if (!aborted)
1053 btrfs_try_granting_tickets(fs_info, space_info);
1054 }
1055 return (tickets_id != space_info->tickets_id);
1056}
1057
1058/*
1059 * This is for normal flushers, we can wait all goddamned day if we want to. We
1060 * will loop and continuously try to flush as long as we are making progress.
1061 * We count progress as clearing off tickets each time we have to loop.
1062 */
1063static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
1064{
1065 struct btrfs_fs_info *fs_info;
1066 struct btrfs_space_info *space_info;
1067 u64 to_reclaim;
1068 enum btrfs_flush_state flush_state;
1069 int commit_cycles = 0;
1070 u64 last_tickets_id;
1071
1072 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
1073 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1074
1075 spin_lock(&space_info->lock);
1076 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1077 if (!to_reclaim) {
1078 space_info->flush = 0;
1079 spin_unlock(&space_info->lock);
1080 return;
1081 }
1082 last_tickets_id = space_info->tickets_id;
1083 spin_unlock(&space_info->lock);
1084
1085 flush_state = FLUSH_DELAYED_ITEMS_NR;
1086 do {
1087 flush_space(fs_info, space_info, to_reclaim, flush_state, false);
1088 spin_lock(&space_info->lock);
1089 if (list_empty(&space_info->tickets)) {
1090 space_info->flush = 0;
1091 spin_unlock(&space_info->lock);
1092 return;
1093 }
1094 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
1095 space_info);
1096 if (last_tickets_id == space_info->tickets_id) {
1097 flush_state++;
1098 } else {
1099 last_tickets_id = space_info->tickets_id;
1100 flush_state = FLUSH_DELAYED_ITEMS_NR;
1101 if (commit_cycles)
1102 commit_cycles--;
1103 }
1104
1105 /*
1106 * We do not want to empty the system of delalloc unless we're
1107 * under heavy pressure, so allow one trip through the flushing
1108 * logic before we start doing a FLUSH_DELALLOC_FULL.
1109 */
1110 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles)
1111 flush_state++;
1112
1113 /*
1114 * We don't want to force a chunk allocation until we've tried
1115 * pretty hard to reclaim space. Think of the case where we
1116 * freed up a bunch of space and so have a lot of pinned space
1117 * to reclaim. We would rather use that than possibly create a
1118 * underutilized metadata chunk. So if this is our first run
1119 * through the flushing state machine skip ALLOC_CHUNK_FORCE and
1120 * commit the transaction. If nothing has changed the next go
1121 * around then we can force a chunk allocation.
1122 */
1123 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
1124 flush_state++;
1125
1126 if (flush_state > COMMIT_TRANS) {
1127 commit_cycles++;
1128 if (commit_cycles > 2) {
1129 if (maybe_fail_all_tickets(fs_info, space_info)) {
1130 flush_state = FLUSH_DELAYED_ITEMS_NR;
1131 commit_cycles--;
1132 } else {
1133 space_info->flush = 0;
1134 }
1135 } else {
1136 flush_state = FLUSH_DELAYED_ITEMS_NR;
1137 }
1138 }
1139 spin_unlock(&space_info->lock);
1140 } while (flush_state <= COMMIT_TRANS);
1141}
1142
1143/*
1144 * This handles pre-flushing of metadata space before we get to the point that
1145 * we need to start blocking threads on tickets. The logic here is different
1146 * from the other flush paths because it doesn't rely on tickets to tell us how
1147 * much we need to flush, instead it attempts to keep us below the 80% full
1148 * watermark of space by flushing whichever reservation pool is currently the
1149 * largest.
1150 */
1151static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
1152{
1153 struct btrfs_fs_info *fs_info;
1154 struct btrfs_space_info *space_info;
1155 struct btrfs_block_rsv *delayed_block_rsv;
1156 struct btrfs_block_rsv *delayed_refs_rsv;
1157 struct btrfs_block_rsv *global_rsv;
1158 struct btrfs_block_rsv *trans_rsv;
1159 int loops = 0;
1160
1161 fs_info = container_of(work, struct btrfs_fs_info,
1162 preempt_reclaim_work);
1163 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1164 delayed_block_rsv = &fs_info->delayed_block_rsv;
1165 delayed_refs_rsv = &fs_info->delayed_refs_rsv;
1166 global_rsv = &fs_info->global_block_rsv;
1167 trans_rsv = &fs_info->trans_block_rsv;
1168
1169 spin_lock(&space_info->lock);
1170 while (need_preemptive_reclaim(fs_info, space_info)) {
1171 enum btrfs_flush_state flush;
1172 u64 delalloc_size = 0;
1173 u64 to_reclaim, block_rsv_size;
1174 u64 global_rsv_size = global_rsv->reserved;
1175
1176 loops++;
1177
1178 /*
1179 * We don't have a precise counter for the metadata being
1180 * reserved for delalloc, so we'll approximate it by subtracting
1181 * out the block rsv's space from the bytes_may_use. If that
1182 * amount is higher than the individual reserves, then we can
1183 * assume it's tied up in delalloc reservations.
1184 */
1185 block_rsv_size = global_rsv_size +
1186 delayed_block_rsv->reserved +
1187 delayed_refs_rsv->reserved +
1188 trans_rsv->reserved;
1189 if (block_rsv_size < space_info->bytes_may_use)
1190 delalloc_size = space_info->bytes_may_use - block_rsv_size;
1191
1192 /*
1193 * We don't want to include the global_rsv in our calculation,
1194 * because that's space we can't touch. Subtract it from the
1195 * block_rsv_size for the next checks.
1196 */
1197 block_rsv_size -= global_rsv_size;
1198
1199 /*
1200 * We really want to avoid flushing delalloc too much, as it
1201 * could result in poor allocation patterns, so only flush it if
1202 * it's larger than the rest of the pools combined.
1203 */
1204 if (delalloc_size > block_rsv_size) {
1205 to_reclaim = delalloc_size;
1206 flush = FLUSH_DELALLOC;
1207 } else if (space_info->bytes_pinned >
1208 (delayed_block_rsv->reserved +
1209 delayed_refs_rsv->reserved)) {
1210 to_reclaim = space_info->bytes_pinned;
1211 flush = COMMIT_TRANS;
1212 } else if (delayed_block_rsv->reserved >
1213 delayed_refs_rsv->reserved) {
1214 to_reclaim = delayed_block_rsv->reserved;
1215 flush = FLUSH_DELAYED_ITEMS_NR;
1216 } else {
1217 to_reclaim = delayed_refs_rsv->reserved;
1218 flush = FLUSH_DELAYED_REFS_NR;
1219 }
1220
1221 spin_unlock(&space_info->lock);
1222
1223 /*
1224 * We don't want to reclaim everything, just a portion, so scale
1225 * down the to_reclaim by 1/4. If it takes us down to 0,
1226 * reclaim 1 items worth.
1227 */
1228 to_reclaim >>= 2;
1229 if (!to_reclaim)
1230 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1231 flush_space(fs_info, space_info, to_reclaim, flush, true);
1232 cond_resched();
1233 spin_lock(&space_info->lock);
1234 }
1235
1236 /* We only went through once, back off our clamping. */
1237 if (loops == 1 && !space_info->reclaim_size)
1238 space_info->clamp = max(1, space_info->clamp - 1);
1239 trace_btrfs_done_preemptive_reclaim(fs_info, space_info);
1240 spin_unlock(&space_info->lock);
1241}
1242
1243/*
1244 * FLUSH_DELALLOC_WAIT:
1245 * Space is freed from flushing delalloc in one of two ways.
1246 *
1247 * 1) compression is on and we allocate less space than we reserved
1248 * 2) we are overwriting existing space
1249 *
1250 * For #1 that extra space is reclaimed as soon as the delalloc pages are
1251 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1252 * length to ->bytes_reserved, and subtracts the reserved space from
1253 * ->bytes_may_use.
1254 *
1255 * For #2 this is trickier. Once the ordered extent runs we will drop the
1256 * extent in the range we are overwriting, which creates a delayed ref for
1257 * that freed extent. This however is not reclaimed until the transaction
1258 * commits, thus the next stages.
1259 *
1260 * RUN_DELAYED_IPUTS
1261 * If we are freeing inodes, we want to make sure all delayed iputs have
1262 * completed, because they could have been on an inode with i_nlink == 0, and
1263 * thus have been truncated and freed up space. But again this space is not
1264 * immediately re-usable, it comes in the form of a delayed ref, which must be
1265 * run and then the transaction must be committed.
1266 *
1267 * COMMIT_TRANS
1268 * This is where we reclaim all of the pinned space generated by running the
1269 * iputs
1270 *
1271 * ALLOC_CHUNK_FORCE
1272 * For data we start with alloc chunk force, however we could have been full
1273 * before, and then the transaction commit could have freed new block groups,
1274 * so if we now have space to allocate do the force chunk allocation.
1275 */
1276static const enum btrfs_flush_state data_flush_states[] = {
1277 FLUSH_DELALLOC_FULL,
1278 RUN_DELAYED_IPUTS,
1279 COMMIT_TRANS,
1280 ALLOC_CHUNK_FORCE,
1281};
1282
1283static void btrfs_async_reclaim_data_space(struct work_struct *work)
1284{
1285 struct btrfs_fs_info *fs_info;
1286 struct btrfs_space_info *space_info;
1287 u64 last_tickets_id;
1288 enum btrfs_flush_state flush_state = 0;
1289
1290 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1291 space_info = fs_info->data_sinfo;
1292
1293 spin_lock(&space_info->lock);
1294 if (list_empty(&space_info->tickets)) {
1295 space_info->flush = 0;
1296 spin_unlock(&space_info->lock);
1297 return;
1298 }
1299 last_tickets_id = space_info->tickets_id;
1300 spin_unlock(&space_info->lock);
1301
1302 while (!space_info->full) {
1303 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1304 spin_lock(&space_info->lock);
1305 if (list_empty(&space_info->tickets)) {
1306 space_info->flush = 0;
1307 spin_unlock(&space_info->lock);
1308 return;
1309 }
1310
1311 /* Something happened, fail everything and bail. */
1312 if (BTRFS_FS_ERROR(fs_info))
1313 goto aborted_fs;
1314 last_tickets_id = space_info->tickets_id;
1315 spin_unlock(&space_info->lock);
1316 }
1317
1318 while (flush_state < ARRAY_SIZE(data_flush_states)) {
1319 flush_space(fs_info, space_info, U64_MAX,
1320 data_flush_states[flush_state], false);
1321 spin_lock(&space_info->lock);
1322 if (list_empty(&space_info->tickets)) {
1323 space_info->flush = 0;
1324 spin_unlock(&space_info->lock);
1325 return;
1326 }
1327
1328 if (last_tickets_id == space_info->tickets_id) {
1329 flush_state++;
1330 } else {
1331 last_tickets_id = space_info->tickets_id;
1332 flush_state = 0;
1333 }
1334
1335 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1336 if (space_info->full) {
1337 if (maybe_fail_all_tickets(fs_info, space_info))
1338 flush_state = 0;
1339 else
1340 space_info->flush = 0;
1341 } else {
1342 flush_state = 0;
1343 }
1344
1345 /* Something happened, fail everything and bail. */
1346 if (BTRFS_FS_ERROR(fs_info))
1347 goto aborted_fs;
1348
1349 }
1350 spin_unlock(&space_info->lock);
1351 }
1352 return;
1353
1354aborted_fs:
1355 maybe_fail_all_tickets(fs_info, space_info);
1356 space_info->flush = 0;
1357 spin_unlock(&space_info->lock);
1358}
1359
1360void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1361{
1362 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1363 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1364 INIT_WORK(&fs_info->preempt_reclaim_work,
1365 btrfs_preempt_reclaim_metadata_space);
1366}
1367
1368static const enum btrfs_flush_state priority_flush_states[] = {
1369 FLUSH_DELAYED_ITEMS_NR,
1370 FLUSH_DELAYED_ITEMS,
1371 ALLOC_CHUNK,
1372};
1373
1374static const enum btrfs_flush_state evict_flush_states[] = {
1375 FLUSH_DELAYED_ITEMS_NR,
1376 FLUSH_DELAYED_ITEMS,
1377 FLUSH_DELAYED_REFS_NR,
1378 FLUSH_DELAYED_REFS,
1379 FLUSH_DELALLOC,
1380 FLUSH_DELALLOC_WAIT,
1381 FLUSH_DELALLOC_FULL,
1382 ALLOC_CHUNK,
1383 COMMIT_TRANS,
1384};
1385
1386static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1387 struct btrfs_space_info *space_info,
1388 struct reserve_ticket *ticket,
1389 const enum btrfs_flush_state *states,
1390 int states_nr)
1391{
1392 u64 to_reclaim;
1393 int flush_state = 0;
1394
1395 spin_lock(&space_info->lock);
1396 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1397 /*
1398 * This is the priority reclaim path, so to_reclaim could be >0 still
1399 * because we may have only satisfied the priority tickets and still
1400 * left non priority tickets on the list. We would then have
1401 * to_reclaim but ->bytes == 0.
1402 */
1403 if (ticket->bytes == 0) {
1404 spin_unlock(&space_info->lock);
1405 return;
1406 }
1407
1408 while (flush_state < states_nr) {
1409 spin_unlock(&space_info->lock);
1410 flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1411 false);
1412 flush_state++;
1413 spin_lock(&space_info->lock);
1414 if (ticket->bytes == 0) {
1415 spin_unlock(&space_info->lock);
1416 return;
1417 }
1418 }
1419
1420 /* Attempt to steal from the global rsv if we can. */
1421 if (!steal_from_global_rsv(fs_info, space_info, ticket)) {
1422 ticket->error = -ENOSPC;
1423 remove_ticket(space_info, ticket);
1424 }
1425
1426 /*
1427 * We must run try_granting_tickets here because we could be a large
1428 * ticket in front of a smaller ticket that can now be satisfied with
1429 * the available space.
1430 */
1431 btrfs_try_granting_tickets(fs_info, space_info);
1432 spin_unlock(&space_info->lock);
1433}
1434
1435static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1436 struct btrfs_space_info *space_info,
1437 struct reserve_ticket *ticket)
1438{
1439 spin_lock(&space_info->lock);
1440
1441 /* We could have been granted before we got here. */
1442 if (ticket->bytes == 0) {
1443 spin_unlock(&space_info->lock);
1444 return;
1445 }
1446
1447 while (!space_info->full) {
1448 spin_unlock(&space_info->lock);
1449 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1450 spin_lock(&space_info->lock);
1451 if (ticket->bytes == 0) {
1452 spin_unlock(&space_info->lock);
1453 return;
1454 }
1455 }
1456
1457 ticket->error = -ENOSPC;
1458 remove_ticket(space_info, ticket);
1459 btrfs_try_granting_tickets(fs_info, space_info);
1460 spin_unlock(&space_info->lock);
1461}
1462
1463static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1464 struct btrfs_space_info *space_info,
1465 struct reserve_ticket *ticket)
1466
1467{
1468 DEFINE_WAIT(wait);
1469 int ret = 0;
1470
1471 spin_lock(&space_info->lock);
1472 while (ticket->bytes > 0 && ticket->error == 0) {
1473 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1474 if (ret) {
1475 /*
1476 * Delete us from the list. After we unlock the space
1477 * info, we don't want the async reclaim job to reserve
1478 * space for this ticket. If that would happen, then the
1479 * ticket's task would not known that space was reserved
1480 * despite getting an error, resulting in a space leak
1481 * (bytes_may_use counter of our space_info).
1482 */
1483 remove_ticket(space_info, ticket);
1484 ticket->error = -EINTR;
1485 break;
1486 }
1487 spin_unlock(&space_info->lock);
1488
1489 schedule();
1490
1491 finish_wait(&ticket->wait, &wait);
1492 spin_lock(&space_info->lock);
1493 }
1494 spin_unlock(&space_info->lock);
1495}
1496
1497/*
1498 * Do the appropriate flushing and waiting for a ticket.
1499 *
1500 * @fs_info: the filesystem
1501 * @space_info: space info for the reservation
1502 * @ticket: ticket for the reservation
1503 * @start_ns: timestamp when the reservation started
1504 * @orig_bytes: amount of bytes originally reserved
1505 * @flush: how much we can flush
1506 *
1507 * This does the work of figuring out how to flush for the ticket, waiting for
1508 * the reservation, and returning the appropriate error if there is one.
1509 */
1510static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1511 struct btrfs_space_info *space_info,
1512 struct reserve_ticket *ticket,
1513 u64 start_ns, u64 orig_bytes,
1514 enum btrfs_reserve_flush_enum flush)
1515{
1516 int ret;
1517
1518 switch (flush) {
1519 case BTRFS_RESERVE_FLUSH_DATA:
1520 case BTRFS_RESERVE_FLUSH_ALL:
1521 case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1522 wait_reserve_ticket(fs_info, space_info, ticket);
1523 break;
1524 case BTRFS_RESERVE_FLUSH_LIMIT:
1525 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1526 priority_flush_states,
1527 ARRAY_SIZE(priority_flush_states));
1528 break;
1529 case BTRFS_RESERVE_FLUSH_EVICT:
1530 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1531 evict_flush_states,
1532 ARRAY_SIZE(evict_flush_states));
1533 break;
1534 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1535 priority_reclaim_data_space(fs_info, space_info, ticket);
1536 break;
1537 default:
1538 ASSERT(0);
1539 break;
1540 }
1541
1542 ret = ticket->error;
1543 ASSERT(list_empty(&ticket->list));
1544 /*
1545 * Check that we can't have an error set if the reservation succeeded,
1546 * as that would confuse tasks and lead them to error out without
1547 * releasing reserved space (if an error happens the expectation is that
1548 * space wasn't reserved at all).
1549 */
1550 ASSERT(!(ticket->bytes == 0 && ticket->error));
1551 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes,
1552 start_ns, flush, ticket->error);
1553 return ret;
1554}
1555
1556/*
1557 * This returns true if this flush state will go through the ordinary flushing
1558 * code.
1559 */
1560static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1561{
1562 return (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1563 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1564}
1565
1566static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1567 struct btrfs_space_info *space_info)
1568{
1569 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1570 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1571
1572 /*
1573 * If we're heavy on ordered operations then clamping won't help us. We
1574 * need to clamp specifically to keep up with dirty'ing buffered
1575 * writers, because there's not a 1:1 correlation of writing delalloc
1576 * and freeing space, like there is with flushing delayed refs or
1577 * delayed nodes. If we're already more ordered than delalloc then
1578 * we're keeping up, otherwise we aren't and should probably clamp.
1579 */
1580 if (ordered < delalloc)
1581 space_info->clamp = min(space_info->clamp + 1, 8);
1582}
1583
1584static inline bool can_steal(enum btrfs_reserve_flush_enum flush)
1585{
1586 return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1587 flush == BTRFS_RESERVE_FLUSH_EVICT);
1588}
1589
1590/*
1591 * NO_FLUSH and FLUSH_EMERGENCY don't want to create a ticket, they just want to
1592 * fail as quickly as possible.
1593 */
1594static inline bool can_ticket(enum btrfs_reserve_flush_enum flush)
1595{
1596 return (flush != BTRFS_RESERVE_NO_FLUSH &&
1597 flush != BTRFS_RESERVE_FLUSH_EMERGENCY);
1598}
1599
1600/*
1601 * Try to reserve bytes from the block_rsv's space.
1602 *
1603 * @fs_info: the filesystem
1604 * @space_info: space info we want to allocate from
1605 * @orig_bytes: number of bytes we want
1606 * @flush: whether or not we can flush to make our reservation
1607 *
1608 * This will reserve orig_bytes number of bytes from the space info associated
1609 * with the block_rsv. If there is not enough space it will make an attempt to
1610 * flush out space to make room. It will do this by flushing delalloc if
1611 * possible or committing the transaction. If flush is 0 then no attempts to
1612 * regain reservations will be made and this will fail if there is not enough
1613 * space already.
1614 */
1615static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1616 struct btrfs_space_info *space_info, u64 orig_bytes,
1617 enum btrfs_reserve_flush_enum flush)
1618{
1619 struct work_struct *async_work;
1620 struct reserve_ticket ticket;
1621 u64 start_ns = 0;
1622 u64 used;
1623 int ret = 0;
1624 bool pending_tickets;
1625
1626 ASSERT(orig_bytes);
1627 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1628
1629 if (flush == BTRFS_RESERVE_FLUSH_DATA)
1630 async_work = &fs_info->async_data_reclaim_work;
1631 else
1632 async_work = &fs_info->async_reclaim_work;
1633
1634 spin_lock(&space_info->lock);
1635 ret = -ENOSPC;
1636 used = btrfs_space_info_used(space_info, true);
1637
1638 /*
1639 * We don't want NO_FLUSH allocations to jump everybody, they can
1640 * generally handle ENOSPC in a different way, so treat them the same as
1641 * normal flushers when it comes to skipping pending tickets.
1642 */
1643 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1644 pending_tickets = !list_empty(&space_info->tickets) ||
1645 !list_empty(&space_info->priority_tickets);
1646 else
1647 pending_tickets = !list_empty(&space_info->priority_tickets);
1648
1649 /*
1650 * Carry on if we have enough space (short-circuit) OR call
1651 * can_overcommit() to ensure we can overcommit to continue.
1652 */
1653 if (!pending_tickets &&
1654 ((used + orig_bytes <= writable_total_bytes(fs_info, space_info)) ||
1655 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1656 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1657 orig_bytes);
1658 ret = 0;
1659 }
1660
1661 /*
1662 * Things are dire, we need to make a reservation so we don't abort. We
1663 * will let this reservation go through as long as we have actual space
1664 * left to allocate for the block.
1665 */
1666 if (ret && unlikely(flush == BTRFS_RESERVE_FLUSH_EMERGENCY)) {
1667 used = btrfs_space_info_used(space_info, false);
1668 if (used + orig_bytes <=
1669 writable_total_bytes(fs_info, space_info)) {
1670 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1671 orig_bytes);
1672 ret = 0;
1673 }
1674 }
1675
1676 /*
1677 * If we couldn't make a reservation then setup our reservation ticket
1678 * and kick the async worker if it's not already running.
1679 *
1680 * If we are a priority flusher then we just need to add our ticket to
1681 * the list and we will do our own flushing further down.
1682 */
1683 if (ret && can_ticket(flush)) {
1684 ticket.bytes = orig_bytes;
1685 ticket.error = 0;
1686 space_info->reclaim_size += ticket.bytes;
1687 init_waitqueue_head(&ticket.wait);
1688 ticket.steal = can_steal(flush);
1689 if (trace_btrfs_reserve_ticket_enabled())
1690 start_ns = ktime_get_ns();
1691
1692 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1693 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1694 flush == BTRFS_RESERVE_FLUSH_DATA) {
1695 list_add_tail(&ticket.list, &space_info->tickets);
1696 if (!space_info->flush) {
1697 /*
1698 * We were forced to add a reserve ticket, so
1699 * our preemptive flushing is unable to keep
1700 * up. Clamp down on the threshold for the
1701 * preemptive flushing in order to keep up with
1702 * the workload.
1703 */
1704 maybe_clamp_preempt(fs_info, space_info);
1705
1706 space_info->flush = 1;
1707 trace_btrfs_trigger_flush(fs_info,
1708 space_info->flags,
1709 orig_bytes, flush,
1710 "enospc");
1711 queue_work(system_unbound_wq, async_work);
1712 }
1713 } else {
1714 list_add_tail(&ticket.list,
1715 &space_info->priority_tickets);
1716 }
1717 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1718 /*
1719 * We will do the space reservation dance during log replay,
1720 * which means we won't have fs_info->fs_root set, so don't do
1721 * the async reclaim as we will panic.
1722 */
1723 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1724 !work_busy(&fs_info->preempt_reclaim_work) &&
1725 need_preemptive_reclaim(fs_info, space_info)) {
1726 trace_btrfs_trigger_flush(fs_info, space_info->flags,
1727 orig_bytes, flush, "preempt");
1728 queue_work(system_unbound_wq,
1729 &fs_info->preempt_reclaim_work);
1730 }
1731 }
1732 spin_unlock(&space_info->lock);
1733 if (!ret || !can_ticket(flush))
1734 return ret;
1735
1736 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1737 orig_bytes, flush);
1738}
1739
1740/*
1741 * Try to reserve metadata bytes from the block_rsv's space.
1742 *
1743 * @fs_info: the filesystem
1744 * @block_rsv: block_rsv we're allocating for
1745 * @orig_bytes: number of bytes we want
1746 * @flush: whether or not we can flush to make our reservation
1747 *
1748 * This will reserve orig_bytes number of bytes from the space info associated
1749 * with the block_rsv. If there is not enough space it will make an attempt to
1750 * flush out space to make room. It will do this by flushing delalloc if
1751 * possible or committing the transaction. If flush is 0 then no attempts to
1752 * regain reservations will be made and this will fail if there is not enough
1753 * space already.
1754 */
1755int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1756 struct btrfs_block_rsv *block_rsv,
1757 u64 orig_bytes,
1758 enum btrfs_reserve_flush_enum flush)
1759{
1760 int ret;
1761
1762 ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush);
1763 if (ret == -ENOSPC) {
1764 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1765 block_rsv->space_info->flags,
1766 orig_bytes, 1);
1767
1768 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1769 btrfs_dump_space_info(fs_info, block_rsv->space_info,
1770 orig_bytes, 0);
1771 }
1772 return ret;
1773}
1774
1775/*
1776 * Try to reserve data bytes for an allocation.
1777 *
1778 * @fs_info: the filesystem
1779 * @bytes: number of bytes we need
1780 * @flush: how we are allowed to flush
1781 *
1782 * This will reserve bytes from the data space info. If there is not enough
1783 * space then we will attempt to flush space as specified by flush.
1784 */
1785int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1786 enum btrfs_reserve_flush_enum flush)
1787{
1788 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1789 int ret;
1790
1791 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1792 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE ||
1793 flush == BTRFS_RESERVE_NO_FLUSH);
1794 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1795
1796 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1797 if (ret == -ENOSPC) {
1798 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1799 data_sinfo->flags, bytes, 1);
1800 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1801 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1802 }
1803 return ret;
1804}
1805
1806/* Dump all the space infos when we abort a transaction due to ENOSPC. */
1807__cold void btrfs_dump_space_info_for_trans_abort(struct btrfs_fs_info *fs_info)
1808{
1809 struct btrfs_space_info *space_info;
1810
1811 btrfs_info(fs_info, "dumping space info:");
1812 list_for_each_entry(space_info, &fs_info->space_info, list) {
1813 spin_lock(&space_info->lock);
1814 __btrfs_dump_space_info(fs_info, space_info);
1815 spin_unlock(&space_info->lock);
1816 }
1817 dump_global_block_rsv(fs_info);
1818}
1819
1820/*
1821 * Account the unused space of all the readonly block group in the space_info.
1822 * takes mirrors into account.
1823 */
1824u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
1825{
1826 struct btrfs_block_group *block_group;
1827 u64 free_bytes = 0;
1828 int factor;
1829
1830 /* It's df, we don't care if it's racy */
1831 if (list_empty(&sinfo->ro_bgs))
1832 return 0;
1833
1834 spin_lock(&sinfo->lock);
1835 list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
1836 spin_lock(&block_group->lock);
1837
1838 if (!block_group->ro) {
1839 spin_unlock(&block_group->lock);
1840 continue;
1841 }
1842
1843 factor = btrfs_bg_type_to_factor(block_group->flags);
1844 free_bytes += (block_group->length -
1845 block_group->used) * factor;
1846
1847 spin_unlock(&block_group->lock);
1848 }
1849 spin_unlock(&sinfo->lock);
1850
1851 return free_bytes;
1852}