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)
364 static void remove_ticket(struct btrfs_space_info *space_info,
365 struct reserve_ticket *ticket)
367 if (!list_empty(&ticket->list)) {
368 list_del_init(&ticket->list);
369 ASSERT(space_info->reclaim_size >= ticket->bytes);
370 space_info->reclaim_size -= ticket->bytes;
375 * This is for space we already have accounted in space_info->bytes_may_use, so
376 * basically when we're returning space from block_rsv's.
378 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
379 struct btrfs_space_info *space_info)
381 struct list_head *head;
382 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
384 lockdep_assert_held(&space_info->lock);
386 head = &space_info->priority_tickets;
388 while (!list_empty(head)) {
389 struct reserve_ticket *ticket;
390 u64 used = btrfs_space_info_used(space_info, true);
392 ticket = list_first_entry(head, struct reserve_ticket, list);
394 /* Check and see if our ticket can be satisified now. */
395 if ((used + ticket->bytes <= space_info->total_bytes) ||
396 btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
398 btrfs_space_info_update_bytes_may_use(fs_info,
401 remove_ticket(space_info, ticket);
403 space_info->tickets_id++;
404 wake_up(&ticket->wait);
410 if (head == &space_info->priority_tickets) {
411 head = &space_info->tickets;
412 flush = BTRFS_RESERVE_FLUSH_ALL;
417 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \
419 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \
420 spin_lock(&__rsv->lock); \
421 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \
422 __rsv->size, __rsv->reserved); \
423 spin_unlock(&__rsv->lock); \
426 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
427 struct btrfs_space_info *info)
429 lockdep_assert_held(&info->lock);
431 btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
433 info->total_bytes - btrfs_space_info_used(info, true),
434 info->full ? "" : "not ");
436 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
437 info->total_bytes, info->bytes_used, info->bytes_pinned,
438 info->bytes_reserved, info->bytes_may_use,
439 info->bytes_readonly);
441 DUMP_BLOCK_RSV(fs_info, global_block_rsv);
442 DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
443 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
444 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
445 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
449 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
450 struct btrfs_space_info *info, u64 bytes,
451 int dump_block_groups)
453 struct btrfs_block_group *cache;
456 spin_lock(&info->lock);
457 __btrfs_dump_space_info(fs_info, info);
458 spin_unlock(&info->lock);
460 if (!dump_block_groups)
463 down_read(&info->groups_sem);
465 list_for_each_entry(cache, &info->block_groups[index], list) {
466 spin_lock(&cache->lock);
468 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
469 cache->start, cache->length, cache->used, cache->pinned,
470 cache->reserved, cache->ro ? "[readonly]" : "");
471 spin_unlock(&cache->lock);
472 btrfs_dump_free_space(cache, bytes);
474 if (++index < BTRFS_NR_RAID_TYPES)
476 up_read(&info->groups_sem);
479 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
485 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
486 nr = div64_u64(to_reclaim, bytes);
492 #define EXTENT_SIZE_PER_ITEM SZ_256K
495 * shrink metadata reservation for delalloc
497 static void shrink_delalloc(struct btrfs_fs_info *fs_info,
498 struct btrfs_space_info *space_info,
499 u64 to_reclaim, bool wait_ordered)
501 struct btrfs_trans_handle *trans;
508 /* Calc the number of the pages we need flush for space reservation */
509 if (to_reclaim == U64_MAX) {
513 * to_reclaim is set to however much metadata we need to
514 * reclaim, but reclaiming that much data doesn't really track
515 * exactly, so increase the amount to reclaim by 2x in order to
516 * make sure we're flushing enough delalloc to hopefully reclaim
517 * some metadata reservations.
519 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
520 to_reclaim = items * EXTENT_SIZE_PER_ITEM;
523 trans = (struct btrfs_trans_handle *)current->journal_info;
525 delalloc_bytes = percpu_counter_sum_positive(
526 &fs_info->delalloc_bytes);
527 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
528 if (delalloc_bytes == 0 && dio_bytes == 0) {
532 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
537 * If we are doing more ordered than delalloc we need to just wait on
538 * ordered extents, otherwise we'll waste time trying to flush delalloc
539 * that likely won't give us the space back we need.
541 if (dio_bytes > delalloc_bytes)
545 while ((delalloc_bytes || dio_bytes) && loops < 3) {
546 btrfs_start_delalloc_roots(fs_info, items);
549 if (wait_ordered && !trans) {
550 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
552 time_left = schedule_timeout_killable(1);
557 spin_lock(&space_info->lock);
558 if (list_empty(&space_info->tickets) &&
559 list_empty(&space_info->priority_tickets)) {
560 spin_unlock(&space_info->lock);
563 spin_unlock(&space_info->lock);
565 delalloc_bytes = percpu_counter_sum_positive(
566 &fs_info->delalloc_bytes);
567 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
572 * maybe_commit_transaction - possibly commit the transaction if its ok to
573 * @root - the root we're allocating for
574 * @bytes - the number of bytes we want to reserve
575 * @force - force the commit
577 * This will check to make sure that committing the transaction will actually
578 * get us somewhere and then commit the transaction if it does. Otherwise it
579 * will return -ENOSPC.
581 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
582 struct btrfs_space_info *space_info,
585 struct reserve_ticket *ticket = NULL;
586 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
587 struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
588 struct btrfs_block_rsv *trans_rsv = &fs_info->trans_block_rsv;
589 struct btrfs_trans_handle *trans;
590 u64 reclaim_bytes = 0;
591 u64 cur_free_bytes = 0;
592 bool do_commit = false;
594 trans = (struct btrfs_trans_handle *)current->journal_info;
599 * If we are data and have passed in U64_MAX we just want to
600 * unconditionally commit the transaction to match the previous data
603 if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) &&
604 bytes_needed == U64_MAX) {
609 spin_lock(&space_info->lock);
610 cur_free_bytes = btrfs_space_info_used(space_info, true);
611 if (cur_free_bytes < space_info->total_bytes)
612 cur_free_bytes = space_info->total_bytes - cur_free_bytes;
616 if (!list_empty(&space_info->priority_tickets))
617 ticket = list_first_entry(&space_info->priority_tickets,
618 struct reserve_ticket, list);
619 else if (!list_empty(&space_info->tickets))
620 ticket = list_first_entry(&space_info->tickets,
621 struct reserve_ticket, list);
623 bytes_needed = ticket->bytes;
625 if (bytes_needed > cur_free_bytes)
626 bytes_needed -= cur_free_bytes;
629 spin_unlock(&space_info->lock);
635 trans = btrfs_join_transaction(fs_info->extent_root);
637 return PTR_ERR(trans);
640 * See if there is enough pinned space to make this reservation, or if
641 * we have block groups that are going to be freed, allowing us to
642 * possibly do a chunk allocation the next loop through.
645 test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
646 __percpu_counter_compare(&space_info->total_bytes_pinned,
648 BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
652 * See if there is some space in the delayed insertion reserve for this
653 * reservation. If the space_info's don't match (like for DATA or
654 * SYSTEM) then just go enospc, reclaiming this space won't recover any
655 * space to satisfy those reservations.
657 if (space_info != delayed_rsv->space_info)
660 spin_lock(&delayed_rsv->lock);
661 reclaim_bytes += delayed_rsv->reserved;
662 spin_unlock(&delayed_rsv->lock);
664 spin_lock(&delayed_refs_rsv->lock);
665 reclaim_bytes += delayed_refs_rsv->reserved;
666 spin_unlock(&delayed_refs_rsv->lock);
668 spin_lock(&trans_rsv->lock);
669 reclaim_bytes += trans_rsv->reserved;
670 spin_unlock(&trans_rsv->lock);
672 if (reclaim_bytes >= bytes_needed)
674 bytes_needed -= reclaim_bytes;
676 if (__percpu_counter_compare(&space_info->total_bytes_pinned,
678 BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
682 return btrfs_commit_transaction(trans);
684 btrfs_end_transaction(trans);
689 * Try to flush some data based on policy set by @state. This is only advisory
690 * and may fail for various reasons. The caller is supposed to examine the
691 * state of @space_info to detect the outcome.
693 static void flush_space(struct btrfs_fs_info *fs_info,
694 struct btrfs_space_info *space_info, u64 num_bytes,
697 struct btrfs_root *root = fs_info->extent_root;
698 struct btrfs_trans_handle *trans;
703 case FLUSH_DELAYED_ITEMS_NR:
704 case FLUSH_DELAYED_ITEMS:
705 if (state == FLUSH_DELAYED_ITEMS_NR)
706 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
710 trans = btrfs_join_transaction(root);
712 ret = PTR_ERR(trans);
715 ret = btrfs_run_delayed_items_nr(trans, nr);
716 btrfs_end_transaction(trans);
719 case FLUSH_DELALLOC_WAIT:
720 shrink_delalloc(fs_info, space_info, num_bytes,
721 state == FLUSH_DELALLOC_WAIT);
723 case FLUSH_DELAYED_REFS_NR:
724 case FLUSH_DELAYED_REFS:
725 trans = btrfs_join_transaction(root);
727 ret = PTR_ERR(trans);
730 if (state == FLUSH_DELAYED_REFS_NR)
731 nr = calc_reclaim_items_nr(fs_info, num_bytes);
734 btrfs_run_delayed_refs(trans, nr);
735 btrfs_end_transaction(trans);
738 case ALLOC_CHUNK_FORCE:
739 trans = btrfs_join_transaction(root);
741 ret = PTR_ERR(trans);
744 ret = btrfs_chunk_alloc(trans,
745 btrfs_get_alloc_profile(fs_info, space_info->flags),
746 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
748 btrfs_end_transaction(trans);
749 if (ret > 0 || ret == -ENOSPC)
752 case RUN_DELAYED_IPUTS:
754 * If we have pending delayed iputs then we could free up a
755 * bunch of pinned space, so make sure we run the iputs before
756 * we do our pinned bytes check below.
758 btrfs_run_delayed_iputs(fs_info);
759 btrfs_wait_on_delayed_iputs(fs_info);
762 ret = may_commit_transaction(fs_info, space_info, num_bytes);
769 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
775 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
776 struct btrfs_space_info *space_info)
781 u64 to_reclaim = space_info->reclaim_size;
783 lockdep_assert_held(&space_info->lock);
785 avail = calc_available_free_space(fs_info, space_info,
786 BTRFS_RESERVE_FLUSH_ALL);
787 used = btrfs_space_info_used(space_info, true);
790 * We may be flushing because suddenly we have less space than we had
791 * before, and now we're well over-committed based on our current free
792 * space. If that's the case add in our overage so we make sure to put
793 * appropriate pressure on the flushing state machine.
795 if (space_info->total_bytes + avail < used)
796 to_reclaim += used - (space_info->total_bytes + avail);
801 to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
802 if (btrfs_can_overcommit(fs_info, space_info, to_reclaim,
803 BTRFS_RESERVE_FLUSH_ALL))
806 used = btrfs_space_info_used(space_info, true);
808 if (btrfs_can_overcommit(fs_info, space_info, SZ_1M,
809 BTRFS_RESERVE_FLUSH_ALL))
810 expected = div_factor_fine(space_info->total_bytes, 95);
812 expected = div_factor_fine(space_info->total_bytes, 90);
815 to_reclaim = used - expected;
818 to_reclaim = min(to_reclaim, space_info->bytes_may_use +
819 space_info->bytes_reserved);
823 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
824 struct btrfs_space_info *space_info,
827 u64 thresh = div_factor_fine(space_info->total_bytes, 98);
829 /* If we're just plain full then async reclaim just slows us down. */
830 if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
833 if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info))
836 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
837 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
840 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
841 struct btrfs_space_info *space_info,
842 struct reserve_ticket *ticket)
844 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
847 if (global_rsv->space_info != space_info)
850 spin_lock(&global_rsv->lock);
851 min_bytes = div_factor(global_rsv->size, 1);
852 if (global_rsv->reserved < min_bytes + ticket->bytes) {
853 spin_unlock(&global_rsv->lock);
856 global_rsv->reserved -= ticket->bytes;
857 remove_ticket(space_info, ticket);
859 wake_up(&ticket->wait);
860 space_info->tickets_id++;
861 if (global_rsv->reserved < global_rsv->size)
862 global_rsv->full = 0;
863 spin_unlock(&global_rsv->lock);
869 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
870 * @fs_info - fs_info for this fs
871 * @space_info - the space info we were flushing
873 * We call this when we've exhausted our flushing ability and haven't made
874 * progress in satisfying tickets. The reservation code handles tickets in
875 * order, so if there is a large ticket first and then smaller ones we could
876 * very well satisfy the smaller tickets. This will attempt to wake up any
877 * tickets in the list to catch this case.
879 * This function returns true if it was able to make progress by clearing out
880 * other tickets, or if it stumbles across a ticket that was smaller than the
883 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
884 struct btrfs_space_info *space_info)
886 struct reserve_ticket *ticket;
887 u64 tickets_id = space_info->tickets_id;
888 u64 first_ticket_bytes = 0;
890 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
891 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
892 __btrfs_dump_space_info(fs_info, space_info);
895 while (!list_empty(&space_info->tickets) &&
896 tickets_id == space_info->tickets_id) {
897 ticket = list_first_entry(&space_info->tickets,
898 struct reserve_ticket, list);
901 steal_from_global_rsv(fs_info, space_info, ticket))
905 * may_commit_transaction will avoid committing the transaction
906 * if it doesn't feel like the space reclaimed by the commit
907 * would result in the ticket succeeding. However if we have a
908 * smaller ticket in the queue it may be small enough to be
909 * satisified by committing the transaction, so if any
910 * subsequent ticket is smaller than the first ticket go ahead
911 * and send us back for another loop through the enospc flushing
914 if (first_ticket_bytes == 0)
915 first_ticket_bytes = ticket->bytes;
916 else if (first_ticket_bytes > ticket->bytes)
919 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
920 btrfs_info(fs_info, "failing ticket with %llu bytes",
923 remove_ticket(space_info, ticket);
924 ticket->error = -ENOSPC;
925 wake_up(&ticket->wait);
928 * We're just throwing tickets away, so more flushing may not
929 * trip over btrfs_try_granting_tickets, so we need to call it
930 * here to see if we can make progress with the next ticket in
933 btrfs_try_granting_tickets(fs_info, space_info);
935 return (tickets_id != space_info->tickets_id);
939 * This is for normal flushers, we can wait all goddamned day if we want to. We
940 * will loop and continuously try to flush as long as we are making progress.
941 * We count progress as clearing off tickets each time we have to loop.
943 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
945 struct btrfs_fs_info *fs_info;
946 struct btrfs_space_info *space_info;
949 int commit_cycles = 0;
952 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
953 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
955 spin_lock(&space_info->lock);
956 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
958 space_info->flush = 0;
959 spin_unlock(&space_info->lock);
962 last_tickets_id = space_info->tickets_id;
963 spin_unlock(&space_info->lock);
965 flush_state = FLUSH_DELAYED_ITEMS_NR;
967 flush_space(fs_info, space_info, to_reclaim, flush_state);
968 spin_lock(&space_info->lock);
969 if (list_empty(&space_info->tickets)) {
970 space_info->flush = 0;
971 spin_unlock(&space_info->lock);
974 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
976 if (last_tickets_id == space_info->tickets_id) {
979 last_tickets_id = space_info->tickets_id;
980 flush_state = FLUSH_DELAYED_ITEMS_NR;
986 * We don't want to force a chunk allocation until we've tried
987 * pretty hard to reclaim space. Think of the case where we
988 * freed up a bunch of space and so have a lot of pinned space
989 * to reclaim. We would rather use that than possibly create a
990 * underutilized metadata chunk. So if this is our first run
991 * through the flushing state machine skip ALLOC_CHUNK_FORCE and
992 * commit the transaction. If nothing has changed the next go
993 * around then we can force a chunk allocation.
995 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
998 if (flush_state > COMMIT_TRANS) {
1000 if (commit_cycles > 2) {
1001 if (maybe_fail_all_tickets(fs_info, space_info)) {
1002 flush_state = FLUSH_DELAYED_ITEMS_NR;
1005 space_info->flush = 0;
1008 flush_state = FLUSH_DELAYED_ITEMS_NR;
1011 spin_unlock(&space_info->lock);
1012 } while (flush_state <= COMMIT_TRANS);
1015 void btrfs_init_async_reclaim_work(struct work_struct *work)
1017 INIT_WORK(work, btrfs_async_reclaim_metadata_space);
1020 static const enum btrfs_flush_state priority_flush_states[] = {
1021 FLUSH_DELAYED_ITEMS_NR,
1022 FLUSH_DELAYED_ITEMS,
1026 static const enum btrfs_flush_state evict_flush_states[] = {
1027 FLUSH_DELAYED_ITEMS_NR,
1028 FLUSH_DELAYED_ITEMS,
1029 FLUSH_DELAYED_REFS_NR,
1032 FLUSH_DELALLOC_WAIT,
1037 static const enum btrfs_flush_state data_flush_states[] = {
1038 FLUSH_DELALLOC_WAIT,
1043 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1044 struct btrfs_space_info *space_info,
1045 struct reserve_ticket *ticket,
1046 const enum btrfs_flush_state *states,
1052 spin_lock(&space_info->lock);
1053 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1055 spin_unlock(&space_info->lock);
1058 spin_unlock(&space_info->lock);
1062 flush_space(fs_info, space_info, to_reclaim, states[flush_state]);
1064 spin_lock(&space_info->lock);
1065 if (ticket->bytes == 0) {
1066 spin_unlock(&space_info->lock);
1069 spin_unlock(&space_info->lock);
1070 } while (flush_state < states_nr);
1073 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1074 struct btrfs_space_info *space_info,
1075 struct reserve_ticket *ticket)
1081 spin_lock(&space_info->lock);
1082 while (ticket->bytes > 0 && ticket->error == 0) {
1083 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1086 * Delete us from the list. After we unlock the space
1087 * info, we don't want the async reclaim job to reserve
1088 * space for this ticket. If that would happen, then the
1089 * ticket's task would not known that space was reserved
1090 * despite getting an error, resulting in a space leak
1091 * (bytes_may_use counter of our space_info).
1093 remove_ticket(space_info, ticket);
1094 ticket->error = -EINTR;
1097 spin_unlock(&space_info->lock);
1101 finish_wait(&ticket->wait, &wait);
1102 spin_lock(&space_info->lock);
1104 spin_unlock(&space_info->lock);
1108 * handle_reserve_ticket - do the appropriate flushing and waiting for a ticket
1110 * @space_info - the space_info for the reservation
1111 * @ticket - the ticket for the reservation
1112 * @flush - how much we can flush
1114 * This does the work of figuring out how to flush for the ticket, waiting for
1115 * the reservation, and returning the appropriate error if there is one.
1117 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1118 struct btrfs_space_info *space_info,
1119 struct reserve_ticket *ticket,
1120 enum btrfs_reserve_flush_enum flush)
1125 case BTRFS_RESERVE_FLUSH_ALL:
1126 case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1127 wait_reserve_ticket(fs_info, space_info, ticket);
1129 case BTRFS_RESERVE_FLUSH_LIMIT:
1130 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1131 priority_flush_states,
1132 ARRAY_SIZE(priority_flush_states));
1134 case BTRFS_RESERVE_FLUSH_EVICT:
1135 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1137 ARRAY_SIZE(evict_flush_states));
1144 spin_lock(&space_info->lock);
1145 ret = ticket->error;
1146 if (ticket->bytes || ticket->error) {
1148 * We were a priority ticket, so we need to delete ourselves
1149 * from the list. Because we could have other priority tickets
1150 * behind us that require less space, run
1151 * btrfs_try_granting_tickets() to see if their reservations can
1154 if (!list_empty(&ticket->list)) {
1155 remove_ticket(space_info, ticket);
1156 btrfs_try_granting_tickets(fs_info, space_info);
1162 spin_unlock(&space_info->lock);
1163 ASSERT(list_empty(&ticket->list));
1165 * Check that we can't have an error set if the reservation succeeded,
1166 * as that would confuse tasks and lead them to error out without
1167 * releasing reserved space (if an error happens the expectation is that
1168 * space wasn't reserved at all).
1170 ASSERT(!(ticket->bytes == 0 && ticket->error));
1175 * This returns true if this flush state will go through the ordinary flushing
1178 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1180 return (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1181 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1185 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
1186 * @root - the root we're allocating for
1187 * @space_info - the space info we want to allocate from
1188 * @orig_bytes - the number of bytes we want
1189 * @flush - whether or not we can flush to make our reservation
1191 * This will reserve orig_bytes number of bytes from the space info associated
1192 * with the block_rsv. If there is not enough space it will make an attempt to
1193 * flush out space to make room. It will do this by flushing delalloc if
1194 * possible or committing the transaction. If flush is 0 then no attempts to
1195 * regain reservations will be made and this will fail if there is not enough
1198 static int __reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1199 struct btrfs_space_info *space_info,
1201 enum btrfs_reserve_flush_enum flush)
1203 struct reserve_ticket ticket;
1206 bool pending_tickets;
1209 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1211 spin_lock(&space_info->lock);
1213 used = btrfs_space_info_used(space_info, true);
1216 * We don't want NO_FLUSH allocations to jump everybody, they can
1217 * generally handle ENOSPC in a different way, so treat them the same as
1218 * normal flushers when it comes to skipping pending tickets.
1220 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1221 pending_tickets = !list_empty(&space_info->tickets) ||
1222 !list_empty(&space_info->priority_tickets);
1224 pending_tickets = !list_empty(&space_info->priority_tickets);
1227 * Carry on if we have enough space (short-circuit) OR call
1228 * can_overcommit() to ensure we can overcommit to continue.
1230 if (!pending_tickets &&
1231 ((used + orig_bytes <= space_info->total_bytes) ||
1232 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1233 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1239 * If we couldn't make a reservation then setup our reservation ticket
1240 * and kick the async worker if it's not already running.
1242 * If we are a priority flusher then we just need to add our ticket to
1243 * the list and we will do our own flushing further down.
1245 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
1246 ticket.bytes = orig_bytes;
1248 space_info->reclaim_size += ticket.bytes;
1249 init_waitqueue_head(&ticket.wait);
1250 ticket.steal = (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1251 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1252 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL) {
1253 list_add_tail(&ticket.list, &space_info->tickets);
1254 if (!space_info->flush) {
1255 space_info->flush = 1;
1256 trace_btrfs_trigger_flush(fs_info,
1260 queue_work(system_unbound_wq,
1261 &fs_info->async_reclaim_work);
1264 list_add_tail(&ticket.list,
1265 &space_info->priority_tickets);
1267 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1270 * We will do the space reservation dance during log replay,
1271 * which means we won't have fs_info->fs_root set, so don't do
1272 * the async reclaim as we will panic.
1274 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1275 need_do_async_reclaim(fs_info, space_info, used) &&
1276 !work_busy(&fs_info->async_reclaim_work)) {
1277 trace_btrfs_trigger_flush(fs_info, space_info->flags,
1278 orig_bytes, flush, "preempt");
1279 queue_work(system_unbound_wq,
1280 &fs_info->async_reclaim_work);
1283 spin_unlock(&space_info->lock);
1284 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
1287 return handle_reserve_ticket(fs_info, space_info, &ticket, flush);
1291 * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
1292 * @root - the root we're allocating for
1293 * @block_rsv - the block_rsv we're allocating for
1294 * @orig_bytes - the number of bytes we want
1295 * @flush - whether or not we can flush to make our reservation
1297 * This will reserve orig_bytes number of bytes from the space info associated
1298 * with the block_rsv. If there is not enough space it will make an attempt to
1299 * flush out space to make room. It will do this by flushing delalloc if
1300 * possible or committing the transaction. If flush is 0 then no attempts to
1301 * regain reservations will be made and this will fail if there is not enough
1304 int btrfs_reserve_metadata_bytes(struct btrfs_root *root,
1305 struct btrfs_block_rsv *block_rsv,
1307 enum btrfs_reserve_flush_enum flush)
1309 struct btrfs_fs_info *fs_info = root->fs_info;
1310 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
1313 ret = __reserve_metadata_bytes(fs_info, block_rsv->space_info,
1315 if (ret == -ENOSPC &&
1316 unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
1317 if (block_rsv != global_rsv &&
1318 !btrfs_block_rsv_use_bytes(global_rsv, orig_bytes))
1321 if (ret == -ENOSPC) {
1322 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1323 block_rsv->space_info->flags,
1326 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1327 btrfs_dump_space_info(fs_info, block_rsv->space_info,