1 // SPDX-License-Identifier: GPL-2.0
5 #include "space-info.h"
8 #include "free-space-cache.h"
9 #include "ordered-data.h"
10 #include "transaction.h"
11 #include "block-group.h"
13 u64 btrfs_space_info_used(struct btrfs_space_info *s_info,
14 bool may_use_included)
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);
23 * after adding space to the filesystem, we need to clear the full flags
24 * on all the space infos.
26 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
28 struct list_head *head = &info->space_info;
29 struct btrfs_space_info *found;
32 list_for_each_entry_rcu(found, head, list)
37 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
40 struct btrfs_space_info *space_info;
44 space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
48 ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
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);
66 ret = btrfs_sysfs_add_space_info_type(info, space_info);
70 list_add_rcu(&space_info->list, &info->space_info);
71 if (flags & BTRFS_BLOCK_GROUP_DATA)
72 info->data_sinfo = space_info;
77 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
79 struct btrfs_super_block *disk_super;
85 disk_super = fs_info->super_copy;
86 if (!btrfs_super_root(disk_super))
89 features = btrfs_super_incompat_flags(disk_super);
90 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
93 flags = BTRFS_BLOCK_GROUP_SYSTEM;
94 ret = create_space_info(fs_info, flags);
99 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
100 ret = create_space_info(fs_info, flags);
102 flags = BTRFS_BLOCK_GROUP_METADATA;
103 ret = create_space_info(fs_info, flags);
107 flags = BTRFS_BLOCK_GROUP_DATA;
108 ret = create_space_info(fs_info, flags);
114 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags,
115 u64 total_bytes, u64 bytes_used,
117 struct btrfs_space_info **space_info)
119 struct btrfs_space_info *found;
122 factor = btrfs_bg_type_to_factor(flags);
124 found = btrfs_find_space_info(info, flags);
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;
134 btrfs_try_granting_tickets(info, found);
135 spin_unlock(&found->lock);
139 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
142 struct list_head *head = &info->space_info;
143 struct btrfs_space_info *found;
145 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
148 list_for_each_entry_rcu(found, head, list) {
149 if (found->flags & flags) {
158 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
160 return (global->size << 1);
163 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
164 struct btrfs_space_info *space_info, u64 bytes,
165 enum btrfs_reserve_flush_enum flush)
172 /* Don't overcommit when in mixed mode. */
173 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
176 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
177 profile = btrfs_system_alloc_profile(fs_info);
179 profile = btrfs_metadata_alloc_profile(fs_info);
181 used = btrfs_space_info_used(space_info, true);
182 avail = atomic64_read(&fs_info->free_chunk_space);
185 * If we have dup, raid1 or raid10 then only half of the free
186 * space is actually usable. For raid56, the space info used
187 * doesn't include the parity drive, so we don't have to
190 factor = btrfs_bg_type_to_factor(profile);
191 avail = div_u64(avail, factor);
194 * If we aren't flushing all things, let us overcommit up to
195 * 1/2th of the space. If we can flush, don't let us overcommit
196 * too much, let it overcommit up to 1/8 of the space.
198 if (flush == BTRFS_RESERVE_FLUSH_ALL)
203 if (used + bytes < space_info->total_bytes + avail)
209 * This is for space we already have accounted in space_info->bytes_may_use, so
210 * basically when we're returning space from block_rsv's.
212 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
213 struct btrfs_space_info *space_info)
215 struct list_head *head;
216 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
218 lockdep_assert_held(&space_info->lock);
220 head = &space_info->priority_tickets;
222 while (!list_empty(head)) {
223 struct reserve_ticket *ticket;
224 u64 used = btrfs_space_info_used(space_info, true);
226 ticket = list_first_entry(head, struct reserve_ticket, list);
228 /* Check and see if our ticket can be satisified now. */
229 if ((used + ticket->bytes <= space_info->total_bytes) ||
230 btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
232 btrfs_space_info_update_bytes_may_use(fs_info,
235 list_del_init(&ticket->list);
237 space_info->tickets_id++;
238 wake_up(&ticket->wait);
244 if (head == &space_info->priority_tickets) {
245 head = &space_info->tickets;
246 flush = BTRFS_RESERVE_FLUSH_ALL;
251 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \
253 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \
254 spin_lock(&__rsv->lock); \
255 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \
256 __rsv->size, __rsv->reserved); \
257 spin_unlock(&__rsv->lock); \
260 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
261 struct btrfs_space_info *info)
263 lockdep_assert_held(&info->lock);
265 /* The free space could be negative in case of overcommit */
266 btrfs_info(fs_info, "space_info %llu has %lld free, is %sfull",
268 (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
269 info->full ? "" : "not ");
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);
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);
284 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
285 struct btrfs_space_info *info, u64 bytes,
286 int dump_block_groups)
288 struct btrfs_block_group_cache *cache;
291 spin_lock(&info->lock);
292 __btrfs_dump_space_info(fs_info, info);
293 spin_unlock(&info->lock);
295 if (!dump_block_groups)
298 down_read(&info->groups_sem);
300 list_for_each_entry(cache, &info->block_groups[index], list) {
301 spin_lock(&cache->lock);
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 spin_unlock(&cache->lock);
308 btrfs_dump_free_space(cache, bytes);
310 if (++index < BTRFS_NR_RAID_TYPES)
312 up_read(&info->groups_sem);
315 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
316 unsigned long nr_pages, int nr_items)
318 struct super_block *sb = fs_info->sb;
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);
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
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);
337 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
343 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
344 nr = div64_u64(to_reclaim, bytes);
350 #define EXTENT_SIZE_PER_ITEM SZ_256K
353 * shrink metadata reservation for delalloc
355 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
356 u64 orig, bool wait_ordered)
358 struct btrfs_space_info *space_info;
359 struct btrfs_trans_handle *trans;
365 unsigned long nr_pages;
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;
372 trans = (struct btrfs_trans_handle *)current->journal_info;
373 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
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) {
382 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
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.
391 if (dio_bytes > delalloc_bytes)
395 while ((delalloc_bytes || dio_bytes) && loops < 3) {
396 nr_pages = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
399 * Triggers inode writeback for up to nr_pages. This will invoke
400 * ->writepages callback and trigger delalloc filling
401 * (btrfs_run_delalloc_range()).
403 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
406 * We need to wait for the compressed pages to start before
409 async_pages = atomic_read(&fs_info->async_delalloc_pages);
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.
418 if (async_pages <= nr_pages)
421 async_pages -= nr_pages;
423 wait_event(fs_info->async_submit_wait,
424 atomic_read(&fs_info->async_delalloc_pages) <=
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);
433 spin_unlock(&space_info->lock);
436 if (wait_ordered && !trans) {
437 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
439 time_left = schedule_timeout_killable(1);
443 delalloc_bytes = percpu_counter_sum_positive(
444 &fs_info->delalloc_bytes);
445 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
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
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.
459 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
460 struct btrfs_space_info *space_info)
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_block_rsv *trans_rsv = &fs_info->trans_block_rsv;
466 struct btrfs_trans_handle *trans;
468 u64 reclaim_bytes = 0;
469 u64 cur_free_bytes = 0;
471 trans = (struct btrfs_trans_handle *)current->journal_info;
475 spin_lock(&space_info->lock);
476 cur_free_bytes = btrfs_space_info_used(space_info, true);
477 if (cur_free_bytes < space_info->total_bytes)
478 cur_free_bytes = space_info->total_bytes - cur_free_bytes;
482 if (!list_empty(&space_info->priority_tickets))
483 ticket = list_first_entry(&space_info->priority_tickets,
484 struct reserve_ticket, list);
485 else if (!list_empty(&space_info->tickets))
486 ticket = list_first_entry(&space_info->tickets,
487 struct reserve_ticket, list);
488 bytes_needed = (ticket) ? ticket->bytes : 0;
490 if (bytes_needed > cur_free_bytes)
491 bytes_needed -= cur_free_bytes;
494 spin_unlock(&space_info->lock);
499 trans = btrfs_join_transaction(fs_info->extent_root);
501 return PTR_ERR(trans);
504 * See if there is enough pinned space to make this reservation, or if
505 * we have block groups that are going to be freed, allowing us to
506 * possibly do a chunk allocation the next loop through.
508 if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
509 __percpu_counter_compare(&space_info->total_bytes_pinned,
511 BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
515 * See if there is some space in the delayed insertion reservation for
518 if (space_info != delayed_rsv->space_info)
521 spin_lock(&delayed_rsv->lock);
522 reclaim_bytes += delayed_rsv->reserved;
523 spin_unlock(&delayed_rsv->lock);
525 spin_lock(&delayed_refs_rsv->lock);
526 reclaim_bytes += delayed_refs_rsv->reserved;
527 spin_unlock(&delayed_refs_rsv->lock);
529 spin_lock(&trans_rsv->lock);
530 reclaim_bytes += trans_rsv->reserved;
531 spin_unlock(&trans_rsv->lock);
533 if (reclaim_bytes >= bytes_needed)
535 bytes_needed -= reclaim_bytes;
537 if (__percpu_counter_compare(&space_info->total_bytes_pinned,
539 BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
543 return btrfs_commit_transaction(trans);
545 btrfs_end_transaction(trans);
550 * Try to flush some data based on policy set by @state. This is only advisory
551 * and may fail for various reasons. The caller is supposed to examine the
552 * state of @space_info to detect the outcome.
554 static void flush_space(struct btrfs_fs_info *fs_info,
555 struct btrfs_space_info *space_info, u64 num_bytes,
558 struct btrfs_root *root = fs_info->extent_root;
559 struct btrfs_trans_handle *trans;
564 case FLUSH_DELAYED_ITEMS_NR:
565 case FLUSH_DELAYED_ITEMS:
566 if (state == FLUSH_DELAYED_ITEMS_NR)
567 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
571 trans = btrfs_join_transaction(root);
573 ret = PTR_ERR(trans);
576 ret = btrfs_run_delayed_items_nr(trans, nr);
577 btrfs_end_transaction(trans);
580 case FLUSH_DELALLOC_WAIT:
581 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
582 state == FLUSH_DELALLOC_WAIT);
584 case FLUSH_DELAYED_REFS_NR:
585 case FLUSH_DELAYED_REFS:
586 trans = btrfs_join_transaction(root);
588 ret = PTR_ERR(trans);
591 if (state == FLUSH_DELAYED_REFS_NR)
592 nr = calc_reclaim_items_nr(fs_info, num_bytes);
595 btrfs_run_delayed_refs(trans, nr);
596 btrfs_end_transaction(trans);
599 case ALLOC_CHUNK_FORCE:
600 trans = btrfs_join_transaction(root);
602 ret = PTR_ERR(trans);
605 ret = btrfs_chunk_alloc(trans,
606 btrfs_metadata_alloc_profile(fs_info),
607 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
609 btrfs_end_transaction(trans);
610 if (ret > 0 || ret == -ENOSPC)
613 case RUN_DELAYED_IPUTS:
615 * If we have pending delayed iputs then we could free up a
616 * bunch of pinned space, so make sure we run the iputs before
617 * we do our pinned bytes check below.
619 btrfs_run_delayed_iputs(fs_info);
620 btrfs_wait_on_delayed_iputs(fs_info);
623 ret = may_commit_transaction(fs_info, space_info);
630 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
636 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
637 struct btrfs_space_info *space_info)
639 struct reserve_ticket *ticket;
644 list_for_each_entry(ticket, &space_info->tickets, list)
645 to_reclaim += ticket->bytes;
646 list_for_each_entry(ticket, &space_info->priority_tickets, list)
647 to_reclaim += ticket->bytes;
651 to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
652 if (btrfs_can_overcommit(fs_info, space_info, to_reclaim,
653 BTRFS_RESERVE_FLUSH_ALL))
656 used = btrfs_space_info_used(space_info, true);
658 if (btrfs_can_overcommit(fs_info, space_info, SZ_1M,
659 BTRFS_RESERVE_FLUSH_ALL))
660 expected = div_factor_fine(space_info->total_bytes, 95);
662 expected = div_factor_fine(space_info->total_bytes, 90);
665 to_reclaim = used - expected;
668 to_reclaim = min(to_reclaim, space_info->bytes_may_use +
669 space_info->bytes_reserved);
673 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
674 struct btrfs_space_info *space_info,
677 u64 thresh = div_factor_fine(space_info->total_bytes, 98);
679 /* If we're just plain full then async reclaim just slows us down. */
680 if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
683 if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info))
686 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
687 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
690 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
691 struct btrfs_space_info *space_info,
692 struct reserve_ticket *ticket)
694 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
697 if (global_rsv->space_info != space_info)
700 spin_lock(&global_rsv->lock);
701 min_bytes = div_factor(global_rsv->size, 5);
702 if (global_rsv->reserved < min_bytes + ticket->bytes) {
703 spin_unlock(&global_rsv->lock);
706 global_rsv->reserved -= ticket->bytes;
708 list_del_init(&ticket->list);
709 wake_up(&ticket->wait);
710 space_info->tickets_id++;
711 if (global_rsv->reserved < global_rsv->size)
712 global_rsv->full = 0;
713 spin_unlock(&global_rsv->lock);
719 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
720 * @fs_info - fs_info for this fs
721 * @space_info - the space info we were flushing
723 * We call this when we've exhausted our flushing ability and haven't made
724 * progress in satisfying tickets. The reservation code handles tickets in
725 * order, so if there is a large ticket first and then smaller ones we could
726 * very well satisfy the smaller tickets. This will attempt to wake up any
727 * tickets in the list to catch this case.
729 * This function returns true if it was able to make progress by clearing out
730 * other tickets, or if it stumbles across a ticket that was smaller than the
733 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
734 struct btrfs_space_info *space_info)
736 struct reserve_ticket *ticket;
737 u64 tickets_id = space_info->tickets_id;
738 u64 first_ticket_bytes = 0;
740 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
741 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
742 __btrfs_dump_space_info(fs_info, space_info);
745 while (!list_empty(&space_info->tickets) &&
746 tickets_id == space_info->tickets_id) {
747 ticket = list_first_entry(&space_info->tickets,
748 struct reserve_ticket, list);
751 steal_from_global_rsv(fs_info, space_info, ticket))
755 * may_commit_transaction will avoid committing the transaction
756 * if it doesn't feel like the space reclaimed by the commit
757 * would result in the ticket succeeding. However if we have a
758 * smaller ticket in the queue it may be small enough to be
759 * satisified by committing the transaction, so if any
760 * subsequent ticket is smaller than the first ticket go ahead
761 * and send us back for another loop through the enospc flushing
764 if (first_ticket_bytes == 0)
765 first_ticket_bytes = ticket->bytes;
766 else if (first_ticket_bytes > ticket->bytes)
769 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
770 btrfs_info(fs_info, "failing ticket with %llu bytes",
773 list_del_init(&ticket->list);
774 ticket->error = -ENOSPC;
775 wake_up(&ticket->wait);
778 * We're just throwing tickets away, so more flushing may not
779 * trip over btrfs_try_granting_tickets, so we need to call it
780 * here to see if we can make progress with the next ticket in
783 btrfs_try_granting_tickets(fs_info, space_info);
785 return (tickets_id != space_info->tickets_id);
789 * This is for normal flushers, we can wait all goddamned day if we want to. We
790 * will loop and continuously try to flush as long as we are making progress.
791 * We count progress as clearing off tickets each time we have to loop.
793 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
795 struct btrfs_fs_info *fs_info;
796 struct btrfs_space_info *space_info;
799 int commit_cycles = 0;
802 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
803 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
805 spin_lock(&space_info->lock);
806 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
808 space_info->flush = 0;
809 spin_unlock(&space_info->lock);
812 last_tickets_id = space_info->tickets_id;
813 spin_unlock(&space_info->lock);
815 flush_state = FLUSH_DELAYED_ITEMS_NR;
817 flush_space(fs_info, space_info, to_reclaim, flush_state);
818 spin_lock(&space_info->lock);
819 if (list_empty(&space_info->tickets)) {
820 space_info->flush = 0;
821 spin_unlock(&space_info->lock);
824 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
826 if (last_tickets_id == space_info->tickets_id) {
829 last_tickets_id = space_info->tickets_id;
830 flush_state = FLUSH_DELAYED_ITEMS_NR;
836 * We don't want to force a chunk allocation until we've tried
837 * pretty hard to reclaim space. Think of the case where we
838 * freed up a bunch of space and so have a lot of pinned space
839 * to reclaim. We would rather use that than possibly create a
840 * underutilized metadata chunk. So if this is our first run
841 * through the flushing state machine skip ALLOC_CHUNK_FORCE and
842 * commit the transaction. If nothing has changed the next go
843 * around then we can force a chunk allocation.
845 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
848 if (flush_state > COMMIT_TRANS) {
850 if (commit_cycles > 2) {
851 if (maybe_fail_all_tickets(fs_info, space_info)) {
852 flush_state = FLUSH_DELAYED_ITEMS_NR;
855 space_info->flush = 0;
858 flush_state = FLUSH_DELAYED_ITEMS_NR;
861 spin_unlock(&space_info->lock);
862 } while (flush_state <= COMMIT_TRANS);
865 void btrfs_init_async_reclaim_work(struct work_struct *work)
867 INIT_WORK(work, btrfs_async_reclaim_metadata_space);
870 static const enum btrfs_flush_state priority_flush_states[] = {
871 FLUSH_DELAYED_ITEMS_NR,
876 static const enum btrfs_flush_state evict_flush_states[] = {
877 FLUSH_DELAYED_ITEMS_NR,
879 FLUSH_DELAYED_REFS_NR,
887 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
888 struct btrfs_space_info *space_info,
889 struct reserve_ticket *ticket,
890 const enum btrfs_flush_state *states,
896 spin_lock(&space_info->lock);
897 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
899 spin_unlock(&space_info->lock);
902 spin_unlock(&space_info->lock);
906 flush_space(fs_info, space_info, to_reclaim, states[flush_state]);
908 spin_lock(&space_info->lock);
909 if (ticket->bytes == 0) {
910 spin_unlock(&space_info->lock);
913 spin_unlock(&space_info->lock);
914 } while (flush_state < states_nr);
917 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
918 struct btrfs_space_info *space_info,
919 struct reserve_ticket *ticket)
925 spin_lock(&space_info->lock);
926 while (ticket->bytes > 0 && ticket->error == 0) {
927 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
930 * Delete us from the list. After we unlock the space
931 * info, we don't want the async reclaim job to reserve
932 * space for this ticket. If that would happen, then the
933 * ticket's task would not known that space was reserved
934 * despite getting an error, resulting in a space leak
935 * (bytes_may_use counter of our space_info).
937 list_del_init(&ticket->list);
938 ticket->error = -EINTR;
941 spin_unlock(&space_info->lock);
945 finish_wait(&ticket->wait, &wait);
946 spin_lock(&space_info->lock);
948 spin_unlock(&space_info->lock);
952 * handle_reserve_ticket - do the appropriate flushing and waiting for a ticket
954 * @space_info - the space_info for the reservation
955 * @ticket - the ticket for the reservation
956 * @flush - how much we can flush
958 * This does the work of figuring out how to flush for the ticket, waiting for
959 * the reservation, and returning the appropriate error if there is one.
961 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
962 struct btrfs_space_info *space_info,
963 struct reserve_ticket *ticket,
964 enum btrfs_reserve_flush_enum flush)
969 case BTRFS_RESERVE_FLUSH_ALL:
970 case BTRFS_RESERVE_FLUSH_ALL_STEAL:
971 wait_reserve_ticket(fs_info, space_info, ticket);
973 case BTRFS_RESERVE_FLUSH_LIMIT:
974 priority_reclaim_metadata_space(fs_info, space_info, ticket,
975 priority_flush_states,
976 ARRAY_SIZE(priority_flush_states));
978 case BTRFS_RESERVE_FLUSH_EVICT:
979 priority_reclaim_metadata_space(fs_info, space_info, ticket,
981 ARRAY_SIZE(evict_flush_states));
988 spin_lock(&space_info->lock);
990 if (ticket->bytes || ticket->error) {
992 * Need to delete here for priority tickets. For regular tickets
993 * either the async reclaim job deletes the ticket from the list
994 * or we delete it ourselves at wait_reserve_ticket().
996 list_del_init(&ticket->list);
1000 spin_unlock(&space_info->lock);
1001 ASSERT(list_empty(&ticket->list));
1003 * Check that we can't have an error set if the reservation succeeded,
1004 * as that would confuse tasks and lead them to error out without
1005 * releasing reserved space (if an error happens the expectation is that
1006 * space wasn't reserved at all).
1008 ASSERT(!(ticket->bytes == 0 && ticket->error));
1013 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
1014 * @root - the root we're allocating for
1015 * @space_info - the space info we want to allocate from
1016 * @orig_bytes - the number of bytes we want
1017 * @flush - whether or not we can flush to make our reservation
1019 * This will reserve orig_bytes number of bytes from the space info associated
1020 * with the block_rsv. If there is not enough space it will make an attempt to
1021 * flush out space to make room. It will do this by flushing delalloc if
1022 * possible or committing the transaction. If flush is 0 then no attempts to
1023 * regain reservations will be made and this will fail if there is not enough
1026 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1027 struct btrfs_space_info *space_info,
1029 enum btrfs_reserve_flush_enum flush)
1031 struct reserve_ticket ticket;
1034 bool pending_tickets;
1037 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1039 spin_lock(&space_info->lock);
1041 used = btrfs_space_info_used(space_info, true);
1042 pending_tickets = !list_empty(&space_info->tickets) ||
1043 !list_empty(&space_info->priority_tickets);
1046 * Carry on if we have enough space (short-circuit) OR call
1047 * can_overcommit() to ensure we can overcommit to continue.
1049 if (!pending_tickets &&
1050 ((used + orig_bytes <= space_info->total_bytes) ||
1051 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1052 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1058 * If we couldn't make a reservation then setup our reservation ticket
1059 * and kick the async worker if it's not already running.
1061 * If we are a priority flusher then we just need to add our ticket to
1062 * the list and we will do our own flushing further down.
1064 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
1065 ticket.bytes = orig_bytes;
1067 init_waitqueue_head(&ticket.wait);
1068 ticket.steal = (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1069 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1070 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL) {
1071 list_add_tail(&ticket.list, &space_info->tickets);
1072 if (!space_info->flush) {
1073 space_info->flush = 1;
1074 trace_btrfs_trigger_flush(fs_info,
1078 queue_work(system_unbound_wq,
1079 &fs_info->async_reclaim_work);
1082 list_add_tail(&ticket.list,
1083 &space_info->priority_tickets);
1085 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1088 * We will do the space reservation dance during log replay,
1089 * which means we won't have fs_info->fs_root set, so don't do
1090 * the async reclaim as we will panic.
1092 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1093 need_do_async_reclaim(fs_info, space_info, used) &&
1094 !work_busy(&fs_info->async_reclaim_work)) {
1095 trace_btrfs_trigger_flush(fs_info, space_info->flags,
1096 orig_bytes, flush, "preempt");
1097 queue_work(system_unbound_wq,
1098 &fs_info->async_reclaim_work);
1101 spin_unlock(&space_info->lock);
1102 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
1105 return handle_reserve_ticket(fs_info, space_info, &ticket, flush);
1109 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
1110 * @root - the root we're allocating for
1111 * @block_rsv - the block_rsv we're allocating for
1112 * @orig_bytes - the number of bytes we want
1113 * @flush - whether or not we can flush to make our reservation
1115 * This will reserve orig_bytes number of bytes from the space info associated
1116 * with the block_rsv. If there is not enough space it will make an attempt to
1117 * flush out space to make room. It will do this by flushing delalloc if
1118 * possible or committing the transaction. If flush is 0 then no attempts to
1119 * regain reservations will be made and this will fail if there is not enough
1122 int btrfs_reserve_metadata_bytes(struct btrfs_root *root,
1123 struct btrfs_block_rsv *block_rsv,
1125 enum btrfs_reserve_flush_enum flush)
1127 struct btrfs_fs_info *fs_info = root->fs_info;
1128 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
1131 ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
1133 if (ret == -ENOSPC &&
1134 unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
1135 if (block_rsv != global_rsv &&
1136 !btrfs_block_rsv_use_bytes(global_rsv, orig_bytes))
1139 if (ret == -ENOSPC) {
1140 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1141 block_rsv->space_info->flags,
1144 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1145 btrfs_dump_space_info(fs_info, block_rsv->space_info,