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"
14 * HOW DOES SPACE RESERVATION WORK
16 * If you want to know about delalloc specifically, there is a separate comment
17 * for that with the delalloc code. This comment is about how the whole system
22 * 1) space_info. This is the ultimate arbiter of how much space we can use.
23 * There's a description of the bytes_ fields with the struct declaration,
24 * refer to that for specifics on each field. Suffice it to say that for
25 * reservations we care about total_bytes - SUM(space_info->bytes_) when
26 * determining if there is space to make an allocation. There is a space_info
27 * for METADATA, SYSTEM, and DATA areas.
29 * 2) block_rsv's. These are basically buckets for every different type of
30 * metadata reservation we have. You can see the comment in the block_rsv
31 * code on the rules for each type, but generally block_rsv->reserved is how
32 * much space is accounted for in space_info->bytes_may_use.
34 * 3) btrfs_calc*_size. These are the worst case calculations we used based
35 * on the number of items we will want to modify. We have one for changing
36 * items, and one for inserting new items. Generally we use these helpers to
37 * determine the size of the block reserves, and then use the actual bytes
38 * values to adjust the space_info counters.
40 * MAKING RESERVATIONS, THE NORMAL CASE
42 * We call into either btrfs_reserve_data_bytes() or
43 * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with
44 * num_bytes we want to reserve.
47 * space_info->bytes_may_reserve += num_bytes
50 * Call btrfs_add_reserved_bytes() which does
51 * space_info->bytes_may_reserve -= num_bytes
52 * space_info->bytes_reserved += extent_bytes
55 * Call btrfs_update_block_group() which does
56 * space_info->bytes_reserved -= extent_bytes
57 * space_info->bytes_used += extent_bytes
59 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority)
61 * Assume we are unable to simply make the reservation because we do not have
65 * create a reserve_ticket with ->bytes set to our reservation, add it to
66 * the tail of space_info->tickets, kick async flush thread
68 * ->handle_reserve_ticket
69 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set
72 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space
73 * Flushes various things attempting to free up space.
75 * -> btrfs_try_granting_tickets()
76 * This is called by anything that either subtracts space from
77 * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the
78 * space_info->total_bytes. This loops through the ->priority_tickets and
79 * then the ->tickets list checking to see if the reservation can be
80 * completed. If it can the space is added to space_info->bytes_may_use and
81 * the ticket is woken up.
84 * Check if ->bytes == 0, if it does we got our reservation and we can carry
85 * on, if not return the appropriate error (ENOSPC, but can be EINTR if we
88 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY
90 * Same as the above, except we add ourselves to the
91 * space_info->priority_tickets, and we do not use ticket->wait, we simply
92 * call flush_space() ourselves for the states that are safe for us to call
93 * without deadlocking and hope for the best.
97 * Generally speaking we will have two cases for each state, a "nice" state
98 * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to
99 * reduce the locking over head on the various trees, and even to keep from
100 * doing any work at all in the case of delayed refs. Each of these delayed
101 * things however hold reservations, and so letting them run allows us to
102 * reclaim space so we can make new reservations.
104 * FLUSH_DELAYED_ITEMS
105 * Every inode has a delayed item to update the inode. Take a simple write
106 * for example, we would update the inode item at write time to update the
107 * mtime, and then again at finish_ordered_io() time in order to update the
108 * isize or bytes. We keep these delayed items to coalesce these operations
109 * into a single operation done on demand. These are an easy way to reclaim
113 * Look at the delalloc comment to get an idea of how much space is reserved
114 * for delayed allocation. We can reclaim some of this space simply by
115 * running delalloc, but usually we need to wait for ordered extents to
116 * reclaim the bulk of this space.
119 * We have a block reserve for the outstanding delayed refs space, and every
120 * delayed ref operation holds a reservation. Running these is a quick way
121 * to reclaim space, but we want to hold this until the end because COW can
122 * churn a lot and we can avoid making some extent tree modifications if we
123 * are able to delay for as long as possible.
126 * We will skip this the first time through space reservation, because of
127 * overcommit and we don't want to have a lot of useless metadata space when
128 * our worst case reservations will likely never come true.
131 * If we're freeing inodes we're likely freeing checksums, file extent
132 * items, and extent tree items. Loads of space could be freed up by these
133 * operations, however they won't be usable until the transaction commits.
136 * may_commit_transaction() is the ultimate arbiter on whether we commit the
137 * transaction or not. In order to avoid constantly churning we do all the
138 * above flushing first and then commit the transaction as the last resort.
139 * However we need to take into account things like pinned space that would
140 * be freed, plus any delayed work we may not have gotten rid of in the case
145 * Because we hold so many reservations for metadata we will allow you to
146 * reserve more space than is currently free in the currently allocate
147 * metadata space. This only happens with metadata, data does not allow
150 * You can see the current logic for when we allow overcommit in
151 * btrfs_can_overcommit(), but it only applies to unallocated space. If there
152 * is no unallocated space to be had, all reservations are kept within the
153 * free space in the allocated metadata chunks.
155 * Because of overcommitting, you generally want to use the
156 * btrfs_can_overcommit() logic for metadata allocations, as it does the right
157 * thing with or without extra unallocated space.
160 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
161 bool may_use_included)
164 return s_info->bytes_used + s_info->bytes_reserved +
165 s_info->bytes_pinned + s_info->bytes_readonly +
166 (may_use_included ? s_info->bytes_may_use : 0);
170 * after adding space to the filesystem, we need to clear the full flags
171 * on all the space infos.
173 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
175 struct list_head *head = &info->space_info;
176 struct btrfs_space_info *found;
179 list_for_each_entry_rcu(found, head, list)
184 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
187 struct btrfs_space_info *space_info;
191 space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
195 ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
202 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
203 INIT_LIST_HEAD(&space_info->block_groups[i]);
204 init_rwsem(&space_info->groups_sem);
205 spin_lock_init(&space_info->lock);
206 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
207 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
208 INIT_LIST_HEAD(&space_info->ro_bgs);
209 INIT_LIST_HEAD(&space_info->tickets);
210 INIT_LIST_HEAD(&space_info->priority_tickets);
212 ret = btrfs_sysfs_add_space_info_type(info, space_info);
216 list_add_rcu(&space_info->list, &info->space_info);
217 if (flags & BTRFS_BLOCK_GROUP_DATA)
218 info->data_sinfo = space_info;
223 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
225 struct btrfs_super_block *disk_super;
231 disk_super = fs_info->super_copy;
232 if (!btrfs_super_root(disk_super))
235 features = btrfs_super_incompat_flags(disk_super);
236 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
239 flags = BTRFS_BLOCK_GROUP_SYSTEM;
240 ret = create_space_info(fs_info, flags);
245 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
246 ret = create_space_info(fs_info, flags);
248 flags = BTRFS_BLOCK_GROUP_METADATA;
249 ret = create_space_info(fs_info, flags);
253 flags = BTRFS_BLOCK_GROUP_DATA;
254 ret = create_space_info(fs_info, flags);
260 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags,
261 u64 total_bytes, u64 bytes_used,
263 struct btrfs_space_info **space_info)
265 struct btrfs_space_info *found;
268 factor = btrfs_bg_type_to_factor(flags);
270 found = btrfs_find_space_info(info, flags);
272 spin_lock(&found->lock);
273 found->total_bytes += total_bytes;
274 found->disk_total += total_bytes * factor;
275 found->bytes_used += bytes_used;
276 found->disk_used += bytes_used * factor;
277 found->bytes_readonly += bytes_readonly;
280 btrfs_try_granting_tickets(info, found);
281 spin_unlock(&found->lock);
285 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
288 struct list_head *head = &info->space_info;
289 struct btrfs_space_info *found;
291 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
294 list_for_each_entry_rcu(found, head, list) {
295 if (found->flags & flags) {
304 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
306 return (global->size << 1);
309 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
310 struct btrfs_space_info *space_info,
311 enum btrfs_reserve_flush_enum flush)
317 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
318 profile = btrfs_system_alloc_profile(fs_info);
320 profile = btrfs_metadata_alloc_profile(fs_info);
322 avail = atomic64_read(&fs_info->free_chunk_space);
325 * If we have dup, raid1 or raid10 then only half of the free
326 * space is actually usable. For raid56, the space info used
327 * doesn't include the parity drive, so we don't have to
330 factor = btrfs_bg_type_to_factor(profile);
331 avail = div_u64(avail, factor);
334 * If we aren't flushing all things, let us overcommit up to
335 * 1/2th of the space. If we can flush, don't let us overcommit
336 * too much, let it overcommit up to 1/8 of the space.
338 if (flush == BTRFS_RESERVE_FLUSH_ALL)
345 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
346 struct btrfs_space_info *space_info, u64 bytes,
347 enum btrfs_reserve_flush_enum flush)
352 /* Don't overcommit when in mixed mode */
353 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
356 used = btrfs_space_info_used(space_info, true);
357 avail = calc_available_free_space(fs_info, space_info, flush);
359 if (used + bytes < space_info->total_bytes + avail)
365 * This is for space we already have accounted in space_info->bytes_may_use, so
366 * basically when we're returning space from block_rsv's.
368 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
369 struct btrfs_space_info *space_info)
371 struct list_head *head;
372 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
374 lockdep_assert_held(&space_info->lock);
376 head = &space_info->priority_tickets;
378 while (!list_empty(head)) {
379 struct reserve_ticket *ticket;
380 u64 used = btrfs_space_info_used(space_info, true);
382 ticket = list_first_entry(head, struct reserve_ticket, list);
384 /* Check and see if our ticket can be satisified now. */
385 if ((used + ticket->bytes <= space_info->total_bytes) ||
386 btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
388 btrfs_space_info_update_bytes_may_use(fs_info,
391 list_del_init(&ticket->list);
392 ASSERT(space_info->reclaim_size >= ticket->bytes);
393 space_info->reclaim_size -= ticket->bytes;
395 space_info->tickets_id++;
396 wake_up(&ticket->wait);
402 if (head == &space_info->priority_tickets) {
403 head = &space_info->tickets;
404 flush = BTRFS_RESERVE_FLUSH_ALL;
409 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \
411 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \
412 spin_lock(&__rsv->lock); \
413 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \
414 __rsv->size, __rsv->reserved); \
415 spin_unlock(&__rsv->lock); \
418 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
419 struct btrfs_space_info *info)
421 lockdep_assert_held(&info->lock);
423 btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
425 info->total_bytes - btrfs_space_info_used(info, true),
426 info->full ? "" : "not ");
428 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
429 info->total_bytes, info->bytes_used, info->bytes_pinned,
430 info->bytes_reserved, info->bytes_may_use,
431 info->bytes_readonly);
433 DUMP_BLOCK_RSV(fs_info, global_block_rsv);
434 DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
435 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
436 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
437 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
441 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
442 struct btrfs_space_info *info, u64 bytes,
443 int dump_block_groups)
445 struct btrfs_block_group *cache;
448 spin_lock(&info->lock);
449 __btrfs_dump_space_info(fs_info, info);
450 spin_unlock(&info->lock);
452 if (!dump_block_groups)
455 down_read(&info->groups_sem);
457 list_for_each_entry(cache, &info->block_groups[index], list) {
458 spin_lock(&cache->lock);
460 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
461 cache->start, cache->length, cache->used, cache->pinned,
462 cache->reserved, cache->ro ? "[readonly]" : "");
463 btrfs_dump_free_space(cache, bytes);
464 spin_unlock(&cache->lock);
466 if (++index < BTRFS_NR_RAID_TYPES)
468 up_read(&info->groups_sem);
471 static void btrfs_writeback_inodes_sb_nr(struct btrfs_fs_info *fs_info,
472 unsigned long nr_pages, int nr_items)
474 struct super_block *sb = fs_info->sb;
476 if (down_read_trylock(&sb->s_umount)) {
477 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
478 up_read(&sb->s_umount);
481 * We needn't worry the filesystem going from r/w to r/o though
482 * we don't acquire ->s_umount mutex, because the filesystem
483 * should guarantee the delalloc inodes list be empty after
484 * the filesystem is readonly(all dirty pages are written to
487 btrfs_start_delalloc_roots(fs_info, nr_items);
488 if (!current->journal_info)
489 btrfs_wait_ordered_roots(fs_info, nr_items, 0, (u64)-1);
493 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
499 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
500 nr = div64_u64(to_reclaim, bytes);
506 #define EXTENT_SIZE_PER_ITEM SZ_256K
509 * shrink metadata reservation for delalloc
511 static void shrink_delalloc(struct btrfs_fs_info *fs_info, u64 to_reclaim,
512 u64 orig, bool wait_ordered)
514 struct btrfs_space_info *space_info;
515 struct btrfs_trans_handle *trans;
521 unsigned long nr_pages;
524 /* Calc the number of the pages we need flush for space reservation */
525 items = calc_reclaim_items_nr(fs_info, to_reclaim);
526 to_reclaim = items * EXTENT_SIZE_PER_ITEM;
528 trans = (struct btrfs_trans_handle *)current->journal_info;
529 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
531 delalloc_bytes = percpu_counter_sum_positive(
532 &fs_info->delalloc_bytes);
533 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
534 if (delalloc_bytes == 0 && dio_bytes == 0) {
538 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
543 * If we are doing more ordered than delalloc we need to just wait on
544 * ordered extents, otherwise we'll waste time trying to flush delalloc
545 * that likely won't give us the space back we need.
547 if (dio_bytes > delalloc_bytes)
551 while ((delalloc_bytes || dio_bytes) && loops < 3) {
552 nr_pages = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
555 * Triggers inode writeback for up to nr_pages. This will invoke
556 * ->writepages callback and trigger delalloc filling
557 * (btrfs_run_delalloc_range()).
559 btrfs_writeback_inodes_sb_nr(fs_info, nr_pages, items);
562 * We need to wait for the compressed pages to start before
565 async_pages = atomic_read(&fs_info->async_delalloc_pages);
570 * Calculate how many compressed pages we want to be written
571 * before we continue. I.e if there are more async pages than we
572 * require wait_event will wait until nr_pages are written.
574 if (async_pages <= nr_pages)
577 async_pages -= nr_pages;
579 wait_event(fs_info->async_submit_wait,
580 atomic_read(&fs_info->async_delalloc_pages) <=
583 spin_lock(&space_info->lock);
584 if (list_empty(&space_info->tickets) &&
585 list_empty(&space_info->priority_tickets)) {
586 spin_unlock(&space_info->lock);
589 spin_unlock(&space_info->lock);
592 if (wait_ordered && !trans) {
593 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
595 time_left = schedule_timeout_killable(1);
599 delalloc_bytes = percpu_counter_sum_positive(
600 &fs_info->delalloc_bytes);
601 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
606 * maybe_commit_transaction - possibly commit the transaction if its ok to
607 * @root - the root we're allocating for
608 * @bytes - the number of bytes we want to reserve
609 * @force - force the commit
611 * This will check to make sure that committing the transaction will actually
612 * get us somewhere and then commit the transaction if it does. Otherwise it
613 * will return -ENOSPC.
615 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
616 struct btrfs_space_info *space_info)
618 struct reserve_ticket *ticket = NULL;
619 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
620 struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
621 struct btrfs_trans_handle *trans;
623 u64 reclaim_bytes = 0;
624 u64 cur_free_bytes = 0;
626 trans = (struct btrfs_trans_handle *)current->journal_info;
630 spin_lock(&space_info->lock);
631 cur_free_bytes = btrfs_space_info_used(space_info, true);
632 if (cur_free_bytes < space_info->total_bytes)
633 cur_free_bytes = space_info->total_bytes - cur_free_bytes;
637 if (!list_empty(&space_info->priority_tickets))
638 ticket = list_first_entry(&space_info->priority_tickets,
639 struct reserve_ticket, list);
640 else if (!list_empty(&space_info->tickets))
641 ticket = list_first_entry(&space_info->tickets,
642 struct reserve_ticket, list);
643 bytes_needed = (ticket) ? ticket->bytes : 0;
645 if (bytes_needed > cur_free_bytes)
646 bytes_needed -= cur_free_bytes;
649 spin_unlock(&space_info->lock);
654 trans = btrfs_join_transaction(fs_info->extent_root);
656 return PTR_ERR(trans);
659 * See if there is enough pinned space to make this reservation, or if
660 * we have block groups that are going to be freed, allowing us to
661 * possibly do a chunk allocation the next loop through.
663 if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
664 __percpu_counter_compare(&space_info->total_bytes_pinned,
666 BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
670 * See if there is some space in the delayed insertion reservation for
673 if (space_info != delayed_rsv->space_info)
676 spin_lock(&delayed_rsv->lock);
677 reclaim_bytes += delayed_rsv->reserved;
678 spin_unlock(&delayed_rsv->lock);
680 spin_lock(&delayed_refs_rsv->lock);
681 reclaim_bytes += delayed_refs_rsv->reserved;
682 spin_unlock(&delayed_refs_rsv->lock);
683 if (reclaim_bytes >= bytes_needed)
685 bytes_needed -= reclaim_bytes;
687 if (__percpu_counter_compare(&space_info->total_bytes_pinned,
689 BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
693 return btrfs_commit_transaction(trans);
695 btrfs_end_transaction(trans);
700 * Try to flush some data based on policy set by @state. This is only advisory
701 * and may fail for various reasons. The caller is supposed to examine the
702 * state of @space_info to detect the outcome.
704 static void flush_space(struct btrfs_fs_info *fs_info,
705 struct btrfs_space_info *space_info, u64 num_bytes,
708 struct btrfs_root *root = fs_info->extent_root;
709 struct btrfs_trans_handle *trans;
714 case FLUSH_DELAYED_ITEMS_NR:
715 case FLUSH_DELAYED_ITEMS:
716 if (state == FLUSH_DELAYED_ITEMS_NR)
717 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
721 trans = btrfs_join_transaction(root);
723 ret = PTR_ERR(trans);
726 ret = btrfs_run_delayed_items_nr(trans, nr);
727 btrfs_end_transaction(trans);
730 case FLUSH_DELALLOC_WAIT:
731 shrink_delalloc(fs_info, num_bytes * 2, num_bytes,
732 state == FLUSH_DELALLOC_WAIT);
734 case FLUSH_DELAYED_REFS_NR:
735 case FLUSH_DELAYED_REFS:
736 trans = btrfs_join_transaction(root);
738 ret = PTR_ERR(trans);
741 if (state == FLUSH_DELAYED_REFS_NR)
742 nr = calc_reclaim_items_nr(fs_info, num_bytes);
745 btrfs_run_delayed_refs(trans, nr);
746 btrfs_end_transaction(trans);
749 case ALLOC_CHUNK_FORCE:
750 trans = btrfs_join_transaction(root);
752 ret = PTR_ERR(trans);
755 ret = btrfs_chunk_alloc(trans,
756 btrfs_metadata_alloc_profile(fs_info),
757 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
759 btrfs_end_transaction(trans);
760 if (ret > 0 || ret == -ENOSPC)
763 case RUN_DELAYED_IPUTS:
765 * If we have pending delayed iputs then we could free up a
766 * bunch of pinned space, so make sure we run the iputs before
767 * we do our pinned bytes check below.
769 btrfs_run_delayed_iputs(fs_info);
770 btrfs_wait_on_delayed_iputs(fs_info);
773 ret = may_commit_transaction(fs_info, space_info);
780 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
786 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
787 struct btrfs_space_info *space_info)
792 u64 to_reclaim = space_info->reclaim_size;
794 lockdep_assert_held(&space_info->lock);
796 avail = calc_available_free_space(fs_info, space_info,
797 BTRFS_RESERVE_FLUSH_ALL);
798 used = btrfs_space_info_used(space_info, true);
801 * We may be flushing because suddenly we have less space than we had
802 * before, and now we're well over-committed based on our current free
803 * space. If that's the case add in our overage so we make sure to put
804 * appropriate pressure on the flushing state machine.
806 if (space_info->total_bytes + avail < used)
807 to_reclaim += used - (space_info->total_bytes + avail);
812 to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
813 if (btrfs_can_overcommit(fs_info, space_info, to_reclaim,
814 BTRFS_RESERVE_FLUSH_ALL))
817 used = btrfs_space_info_used(space_info, true);
819 if (btrfs_can_overcommit(fs_info, space_info, SZ_1M,
820 BTRFS_RESERVE_FLUSH_ALL))
821 expected = div_factor_fine(space_info->total_bytes, 95);
823 expected = div_factor_fine(space_info->total_bytes, 90);
826 to_reclaim = used - expected;
829 to_reclaim = min(to_reclaim, space_info->bytes_may_use +
830 space_info->bytes_reserved);
834 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
835 struct btrfs_space_info *space_info,
838 u64 thresh = div_factor_fine(space_info->total_bytes, 98);
840 /* If we're just plain full then async reclaim just slows us down. */
841 if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
844 if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info))
847 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
848 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
852 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
853 * @fs_info - fs_info for this fs
854 * @space_info - the space info we were flushing
856 * We call this when we've exhausted our flushing ability and haven't made
857 * progress in satisfying tickets. The reservation code handles tickets in
858 * order, so if there is a large ticket first and then smaller ones we could
859 * very well satisfy the smaller tickets. This will attempt to wake up any
860 * tickets in the list to catch this case.
862 * This function returns true if it was able to make progress by clearing out
863 * other tickets, or if it stumbles across a ticket that was smaller than the
866 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
867 struct btrfs_space_info *space_info)
869 struct reserve_ticket *ticket;
870 u64 tickets_id = space_info->tickets_id;
871 u64 first_ticket_bytes = 0;
873 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
874 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
875 __btrfs_dump_space_info(fs_info, space_info);
878 while (!list_empty(&space_info->tickets) &&
879 tickets_id == space_info->tickets_id) {
880 ticket = list_first_entry(&space_info->tickets,
881 struct reserve_ticket, list);
884 * may_commit_transaction will avoid committing the transaction
885 * if it doesn't feel like the space reclaimed by the commit
886 * would result in the ticket succeeding. However if we have a
887 * smaller ticket in the queue it may be small enough to be
888 * satisified by committing the transaction, so if any
889 * subsequent ticket is smaller than the first ticket go ahead
890 * and send us back for another loop through the enospc flushing
893 if (first_ticket_bytes == 0)
894 first_ticket_bytes = ticket->bytes;
895 else if (first_ticket_bytes > ticket->bytes)
898 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
899 btrfs_info(fs_info, "failing ticket with %llu bytes",
902 list_del_init(&ticket->list);
903 ticket->error = -ENOSPC;
904 wake_up(&ticket->wait);
907 * We're just throwing tickets away, so more flushing may not
908 * trip over btrfs_try_granting_tickets, so we need to call it
909 * here to see if we can make progress with the next ticket in
912 btrfs_try_granting_tickets(fs_info, space_info);
914 return (tickets_id != space_info->tickets_id);
918 * This is for normal flushers, we can wait all goddamned day if we want to. We
919 * will loop and continuously try to flush as long as we are making progress.
920 * We count progress as clearing off tickets each time we have to loop.
922 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
924 struct btrfs_fs_info *fs_info;
925 struct btrfs_space_info *space_info;
928 int commit_cycles = 0;
931 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
932 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
934 spin_lock(&space_info->lock);
935 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
937 space_info->flush = 0;
938 spin_unlock(&space_info->lock);
941 last_tickets_id = space_info->tickets_id;
942 spin_unlock(&space_info->lock);
944 flush_state = FLUSH_DELAYED_ITEMS_NR;
946 flush_space(fs_info, space_info, to_reclaim, flush_state);
947 spin_lock(&space_info->lock);
948 if (list_empty(&space_info->tickets)) {
949 space_info->flush = 0;
950 spin_unlock(&space_info->lock);
953 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
955 if (last_tickets_id == space_info->tickets_id) {
958 last_tickets_id = space_info->tickets_id;
959 flush_state = FLUSH_DELAYED_ITEMS_NR;
965 * We don't want to force a chunk allocation until we've tried
966 * pretty hard to reclaim space. Think of the case where we
967 * freed up a bunch of space and so have a lot of pinned space
968 * to reclaim. We would rather use that than possibly create a
969 * underutilized metadata chunk. So if this is our first run
970 * through the flushing state machine skip ALLOC_CHUNK_FORCE and
971 * commit the transaction. If nothing has changed the next go
972 * around then we can force a chunk allocation.
974 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
977 if (flush_state > COMMIT_TRANS) {
979 if (commit_cycles > 2) {
980 if (maybe_fail_all_tickets(fs_info, space_info)) {
981 flush_state = FLUSH_DELAYED_ITEMS_NR;
984 space_info->flush = 0;
987 flush_state = FLUSH_DELAYED_ITEMS_NR;
990 spin_unlock(&space_info->lock);
991 } while (flush_state <= COMMIT_TRANS);
994 void btrfs_init_async_reclaim_work(struct work_struct *work)
996 INIT_WORK(work, btrfs_async_reclaim_metadata_space);
999 static const enum btrfs_flush_state priority_flush_states[] = {
1000 FLUSH_DELAYED_ITEMS_NR,
1001 FLUSH_DELAYED_ITEMS,
1005 static const enum btrfs_flush_state evict_flush_states[] = {
1006 FLUSH_DELAYED_ITEMS_NR,
1007 FLUSH_DELAYED_ITEMS,
1008 FLUSH_DELAYED_REFS_NR,
1011 FLUSH_DELALLOC_WAIT,
1016 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1017 struct btrfs_space_info *space_info,
1018 struct reserve_ticket *ticket,
1019 const enum btrfs_flush_state *states,
1025 spin_lock(&space_info->lock);
1026 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1028 spin_unlock(&space_info->lock);
1031 spin_unlock(&space_info->lock);
1035 flush_space(fs_info, space_info, to_reclaim, states[flush_state]);
1037 spin_lock(&space_info->lock);
1038 if (ticket->bytes == 0) {
1039 spin_unlock(&space_info->lock);
1042 spin_unlock(&space_info->lock);
1043 } while (flush_state < states_nr);
1046 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1047 struct btrfs_space_info *space_info,
1048 struct reserve_ticket *ticket)
1054 spin_lock(&space_info->lock);
1055 while (ticket->bytes > 0 && ticket->error == 0) {
1056 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1059 * Delete us from the list. After we unlock the space
1060 * info, we don't want the async reclaim job to reserve
1061 * space for this ticket. If that would happen, then the
1062 * ticket's task would not known that space was reserved
1063 * despite getting an error, resulting in a space leak
1064 * (bytes_may_use counter of our space_info).
1066 list_del_init(&ticket->list);
1067 ticket->error = -EINTR;
1070 spin_unlock(&space_info->lock);
1074 finish_wait(&ticket->wait, &wait);
1075 spin_lock(&space_info->lock);
1077 spin_unlock(&space_info->lock);
1081 * handle_reserve_ticket - do the appropriate flushing and waiting for a ticket
1083 * @space_info - the space_info for the reservation
1084 * @ticket - the ticket for the reservation
1085 * @flush - how much we can flush
1087 * This does the work of figuring out how to flush for the ticket, waiting for
1088 * the reservation, and returning the appropriate error if there is one.
1090 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1091 struct btrfs_space_info *space_info,
1092 struct reserve_ticket *ticket,
1093 enum btrfs_reserve_flush_enum flush)
1098 case BTRFS_RESERVE_FLUSH_ALL:
1099 wait_reserve_ticket(fs_info, space_info, ticket);
1101 case BTRFS_RESERVE_FLUSH_LIMIT:
1102 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1103 priority_flush_states,
1104 ARRAY_SIZE(priority_flush_states));
1106 case BTRFS_RESERVE_FLUSH_EVICT:
1107 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1109 ARRAY_SIZE(evict_flush_states));
1116 spin_lock(&space_info->lock);
1117 ret = ticket->error;
1118 if (ticket->bytes || ticket->error) {
1120 * Need to delete here for priority tickets. For regular tickets
1121 * either the async reclaim job deletes the ticket from the list
1122 * or we delete it ourselves at wait_reserve_ticket().
1124 list_del_init(&ticket->list);
1128 spin_unlock(&space_info->lock);
1129 ASSERT(list_empty(&ticket->list));
1131 * Check that we can't have an error set if the reservation succeeded,
1132 * as that would confuse tasks and lead them to error out without
1133 * releasing reserved space (if an error happens the expectation is that
1134 * space wasn't reserved at all).
1136 ASSERT(!(ticket->bytes == 0 && ticket->error));
1141 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
1142 * @root - the root we're allocating for
1143 * @space_info - the space info we want to allocate from
1144 * @orig_bytes - the number of bytes we want
1145 * @flush - whether or not we can flush to make our reservation
1147 * This will reserve orig_bytes number of bytes from the space info associated
1148 * with the block_rsv. If there is not enough space it will make an attempt to
1149 * flush out space to make room. It will do this by flushing delalloc if
1150 * possible or committing the transaction. If flush is 0 then no attempts to
1151 * regain reservations will be made and this will fail if there is not enough
1154 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1155 struct btrfs_space_info *space_info,
1157 enum btrfs_reserve_flush_enum flush)
1159 struct reserve_ticket ticket;
1162 bool pending_tickets;
1165 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1167 spin_lock(&space_info->lock);
1169 used = btrfs_space_info_used(space_info, true);
1170 pending_tickets = !list_empty(&space_info->tickets) ||
1171 !list_empty(&space_info->priority_tickets);
1174 * Carry on if we have enough space (short-circuit) OR call
1175 * can_overcommit() to ensure we can overcommit to continue.
1177 if (!pending_tickets &&
1178 ((used + orig_bytes <= space_info->total_bytes) ||
1179 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1180 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1186 * If we couldn't make a reservation then setup our reservation ticket
1187 * and kick the async worker if it's not already running.
1189 * If we are a priority flusher then we just need to add our ticket to
1190 * the list and we will do our own flushing further down.
1192 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
1193 ASSERT(space_info->reclaim_size >= 0);
1194 ticket.bytes = orig_bytes;
1196 space_info->reclaim_size += ticket.bytes;
1197 init_waitqueue_head(&ticket.wait);
1198 if (flush == BTRFS_RESERVE_FLUSH_ALL) {
1199 list_add_tail(&ticket.list, &space_info->tickets);
1200 if (!space_info->flush) {
1201 space_info->flush = 1;
1202 trace_btrfs_trigger_flush(fs_info,
1206 queue_work(system_unbound_wq,
1207 &fs_info->async_reclaim_work);
1210 list_add_tail(&ticket.list,
1211 &space_info->priority_tickets);
1213 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1216 * We will do the space reservation dance during log replay,
1217 * which means we won't have fs_info->fs_root set, so don't do
1218 * the async reclaim as we will panic.
1220 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1221 need_do_async_reclaim(fs_info, space_info, used) &&
1222 !work_busy(&fs_info->async_reclaim_work)) {
1223 trace_btrfs_trigger_flush(fs_info, space_info->flags,
1224 orig_bytes, flush, "preempt");
1225 queue_work(system_unbound_wq,
1226 &fs_info->async_reclaim_work);
1229 spin_unlock(&space_info->lock);
1230 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
1233 return handle_reserve_ticket(fs_info, space_info, &ticket, flush);
1237 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
1238 * @root - the root we're allocating for
1239 * @block_rsv - the block_rsv we're allocating for
1240 * @orig_bytes - the number of bytes we want
1241 * @flush - whether or not we can flush to make our reservation
1243 * This will reserve orig_bytes number of bytes from the space info associated
1244 * with the block_rsv. If there is not enough space it will make an attempt to
1245 * flush out space to make room. It will do this by flushing delalloc if
1246 * possible or committing the transaction. If flush is 0 then no attempts to
1247 * regain reservations will be made and this will fail if there is not enough
1250 int btrfs_reserve_metadata_bytes(struct btrfs_root *root,
1251 struct btrfs_block_rsv *block_rsv,
1253 enum btrfs_reserve_flush_enum flush)
1255 struct btrfs_fs_info *fs_info = root->fs_info;
1256 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
1259 ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
1261 if (ret == -ENOSPC &&
1262 unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
1263 if (block_rsv != global_rsv &&
1264 !btrfs_block_rsv_use_bytes(global_rsv, orig_bytes))
1267 if (ret == -ENOSPC) {
1268 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1269 block_rsv->space_info->flags,
1272 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1273 btrfs_dump_space_info(fs_info, block_rsv->space_info,