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 #include "accessors.h"
15 #include "extent-tree.h"
18 * HOW DOES SPACE RESERVATION WORK
20 * If you want to know about delalloc specifically, there is a separate comment
21 * for that with the delalloc code. This comment is about how the whole system
26 * 1) space_info. This is the ultimate arbiter of how much space we can use.
27 * There's a description of the bytes_ fields with the struct declaration,
28 * refer to that for specifics on each field. Suffice it to say that for
29 * reservations we care about total_bytes - SUM(space_info->bytes_) when
30 * determining if there is space to make an allocation. There is a space_info
31 * for METADATA, SYSTEM, and DATA areas.
33 * 2) block_rsv's. These are basically buckets for every different type of
34 * metadata reservation we have. You can see the comment in the block_rsv
35 * code on the rules for each type, but generally block_rsv->reserved is how
36 * much space is accounted for in space_info->bytes_may_use.
38 * 3) btrfs_calc*_size. These are the worst case calculations we used based
39 * on the number of items we will want to modify. We have one for changing
40 * items, and one for inserting new items. Generally we use these helpers to
41 * determine the size of the block reserves, and then use the actual bytes
42 * values to adjust the space_info counters.
44 * MAKING RESERVATIONS, THE NORMAL CASE
46 * We call into either btrfs_reserve_data_bytes() or
47 * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with
48 * num_bytes we want to reserve.
51 * space_info->bytes_may_reserve += num_bytes
54 * Call btrfs_add_reserved_bytes() which does
55 * space_info->bytes_may_reserve -= num_bytes
56 * space_info->bytes_reserved += extent_bytes
59 * Call btrfs_update_block_group() which does
60 * space_info->bytes_reserved -= extent_bytes
61 * space_info->bytes_used += extent_bytes
63 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority)
65 * Assume we are unable to simply make the reservation because we do not have
69 * create a reserve_ticket with ->bytes set to our reservation, add it to
70 * the tail of space_info->tickets, kick async flush thread
72 * ->handle_reserve_ticket
73 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set
76 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space
77 * Flushes various things attempting to free up space.
79 * -> btrfs_try_granting_tickets()
80 * This is called by anything that either subtracts space from
81 * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the
82 * space_info->total_bytes. This loops through the ->priority_tickets and
83 * then the ->tickets list checking to see if the reservation can be
84 * completed. If it can the space is added to space_info->bytes_may_use and
85 * the ticket is woken up.
88 * Check if ->bytes == 0, if it does we got our reservation and we can carry
89 * on, if not return the appropriate error (ENOSPC, but can be EINTR if we
92 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY
94 * Same as the above, except we add ourselves to the
95 * space_info->priority_tickets, and we do not use ticket->wait, we simply
96 * call flush_space() ourselves for the states that are safe for us to call
97 * without deadlocking and hope for the best.
101 * Generally speaking we will have two cases for each state, a "nice" state
102 * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to
103 * reduce the locking over head on the various trees, and even to keep from
104 * doing any work at all in the case of delayed refs. Each of these delayed
105 * things however hold reservations, and so letting them run allows us to
106 * reclaim space so we can make new reservations.
108 * FLUSH_DELAYED_ITEMS
109 * Every inode has a delayed item to update the inode. Take a simple write
110 * for example, we would update the inode item at write time to update the
111 * mtime, and then again at finish_ordered_io() time in order to update the
112 * isize or bytes. We keep these delayed items to coalesce these operations
113 * into a single operation done on demand. These are an easy way to reclaim
117 * Look at the delalloc comment to get an idea of how much space is reserved
118 * for delayed allocation. We can reclaim some of this space simply by
119 * running delalloc, but usually we need to wait for ordered extents to
120 * reclaim the bulk of this space.
123 * We have a block reserve for the outstanding delayed refs space, and every
124 * delayed ref operation holds a reservation. Running these is a quick way
125 * to reclaim space, but we want to hold this until the end because COW can
126 * churn a lot and we can avoid making some extent tree modifications if we
127 * are able to delay for as long as possible.
130 * We will skip this the first time through space reservation, because of
131 * overcommit and we don't want to have a lot of useless metadata space when
132 * our worst case reservations will likely never come true.
135 * If we're freeing inodes we're likely freeing checksums, file extent
136 * items, and extent tree items. Loads of space could be freed up by these
137 * operations, however they won't be usable until the transaction commits.
140 * This will commit the transaction. Historically we had a lot of logic
141 * surrounding whether or not we'd commit the transaction, but this waits born
142 * out of a pre-tickets era where we could end up committing the transaction
143 * thousands of times in a row without making progress. Now thanks to our
144 * ticketing system we know if we're not making progress and can error
145 * everybody out after a few commits rather than burning the disk hoping for
146 * a different answer.
150 * Because we hold so many reservations for metadata we will allow you to
151 * reserve more space than is currently free in the currently allocate
152 * metadata space. This only happens with metadata, data does not allow
155 * You can see the current logic for when we allow overcommit in
156 * btrfs_can_overcommit(), but it only applies to unallocated space. If there
157 * is no unallocated space to be had, all reservations are kept within the
158 * free space in the allocated metadata chunks.
160 * Because of overcommitting, you generally want to use the
161 * btrfs_can_overcommit() logic for metadata allocations, as it does the right
162 * thing with or without extra unallocated space.
165 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
166 bool may_use_included)
169 return s_info->bytes_used + s_info->bytes_reserved +
170 s_info->bytes_pinned + s_info->bytes_readonly +
171 s_info->bytes_zone_unusable +
172 (may_use_included ? s_info->bytes_may_use : 0);
176 * after adding space to the filesystem, we need to clear the full flags
177 * on all the space infos.
179 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
181 struct list_head *head = &info->space_info;
182 struct btrfs_space_info *found;
184 list_for_each_entry(found, head, list)
189 * Block groups with more than this value (percents) of unusable space will be
190 * scheduled for background reclaim.
192 #define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH (75)
195 * Calculate chunk size depending on volume type (regular or zoned).
197 static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags)
199 if (btrfs_is_zoned(fs_info))
200 return fs_info->zone_size;
202 ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
204 if (flags & BTRFS_BLOCK_GROUP_DATA)
205 return BTRFS_MAX_DATA_CHUNK_SIZE;
206 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
209 /* Handle BTRFS_BLOCK_GROUP_METADATA */
210 if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G)
217 * Update default chunk size.
219 void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info,
222 WRITE_ONCE(space_info->chunk_size, chunk_size);
225 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
228 struct btrfs_space_info *space_info;
232 space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
236 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
237 INIT_LIST_HEAD(&space_info->block_groups[i]);
238 init_rwsem(&space_info->groups_sem);
239 spin_lock_init(&space_info->lock);
240 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
241 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
242 INIT_LIST_HEAD(&space_info->ro_bgs);
243 INIT_LIST_HEAD(&space_info->tickets);
244 INIT_LIST_HEAD(&space_info->priority_tickets);
245 space_info->clamp = 1;
246 btrfs_update_space_info_chunk_size(space_info, calc_chunk_size(info, flags));
248 if (btrfs_is_zoned(info))
249 space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH;
251 ret = btrfs_sysfs_add_space_info_type(info, space_info);
255 list_add(&space_info->list, &info->space_info);
256 if (flags & BTRFS_BLOCK_GROUP_DATA)
257 info->data_sinfo = space_info;
262 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
264 struct btrfs_super_block *disk_super;
270 disk_super = fs_info->super_copy;
271 if (!btrfs_super_root(disk_super))
274 features = btrfs_super_incompat_flags(disk_super);
275 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
278 flags = BTRFS_BLOCK_GROUP_SYSTEM;
279 ret = create_space_info(fs_info, flags);
284 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
285 ret = create_space_info(fs_info, flags);
287 flags = BTRFS_BLOCK_GROUP_METADATA;
288 ret = create_space_info(fs_info, flags);
292 flags = BTRFS_BLOCK_GROUP_DATA;
293 ret = create_space_info(fs_info, flags);
299 void btrfs_add_bg_to_space_info(struct btrfs_fs_info *info,
300 struct btrfs_block_group *block_group)
302 struct btrfs_space_info *found;
305 factor = btrfs_bg_type_to_factor(block_group->flags);
307 found = btrfs_find_space_info(info, block_group->flags);
309 spin_lock(&found->lock);
310 found->total_bytes += block_group->length;
311 found->disk_total += block_group->length * factor;
312 found->bytes_used += block_group->used;
313 found->disk_used += block_group->used * factor;
314 found->bytes_readonly += block_group->bytes_super;
315 found->bytes_zone_unusable += block_group->zone_unusable;
316 if (block_group->length > 0)
318 btrfs_try_granting_tickets(info, found);
319 spin_unlock(&found->lock);
321 block_group->space_info = found;
323 index = btrfs_bg_flags_to_raid_index(block_group->flags);
324 down_write(&found->groups_sem);
325 list_add_tail(&block_group->list, &found->block_groups[index]);
326 up_write(&found->groups_sem);
329 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
332 struct list_head *head = &info->space_info;
333 struct btrfs_space_info *found;
335 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
337 list_for_each_entry(found, head, list) {
338 if (found->flags & flags)
344 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
345 struct btrfs_space_info *space_info,
346 enum btrfs_reserve_flush_enum flush)
352 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
353 profile = btrfs_system_alloc_profile(fs_info);
355 profile = btrfs_metadata_alloc_profile(fs_info);
357 avail = atomic64_read(&fs_info->free_chunk_space);
360 * If we have dup, raid1 or raid10 then only half of the free
361 * space is actually usable. For raid56, the space info used
362 * doesn't include the parity drive, so we don't have to
365 factor = btrfs_bg_type_to_factor(profile);
366 avail = div_u64(avail, factor);
369 * If we aren't flushing all things, let us overcommit up to
370 * 1/2th of the space. If we can flush, don't let us overcommit
371 * too much, let it overcommit up to 1/8 of the space.
373 if (flush == BTRFS_RESERVE_FLUSH_ALL)
380 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
381 struct btrfs_space_info *space_info, u64 bytes,
382 enum btrfs_reserve_flush_enum flush)
387 /* Don't overcommit when in mixed mode */
388 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
391 used = btrfs_space_info_used(space_info, true);
392 avail = calc_available_free_space(fs_info, space_info, flush);
394 if (used + bytes < space_info->total_bytes + avail)
399 static void remove_ticket(struct btrfs_space_info *space_info,
400 struct reserve_ticket *ticket)
402 if (!list_empty(&ticket->list)) {
403 list_del_init(&ticket->list);
404 ASSERT(space_info->reclaim_size >= ticket->bytes);
405 space_info->reclaim_size -= ticket->bytes;
410 * This is for space we already have accounted in space_info->bytes_may_use, so
411 * basically when we're returning space from block_rsv's.
413 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
414 struct btrfs_space_info *space_info)
416 struct list_head *head;
417 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
419 lockdep_assert_held(&space_info->lock);
421 head = &space_info->priority_tickets;
423 while (!list_empty(head)) {
424 struct reserve_ticket *ticket;
425 u64 used = btrfs_space_info_used(space_info, true);
427 ticket = list_first_entry(head, struct reserve_ticket, list);
429 /* Check and see if our ticket can be satisfied now. */
430 if ((used + ticket->bytes <= space_info->total_bytes) ||
431 btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
433 btrfs_space_info_update_bytes_may_use(fs_info,
436 remove_ticket(space_info, ticket);
438 space_info->tickets_id++;
439 wake_up(&ticket->wait);
445 if (head == &space_info->priority_tickets) {
446 head = &space_info->tickets;
447 flush = BTRFS_RESERVE_FLUSH_ALL;
452 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \
454 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \
455 spin_lock(&__rsv->lock); \
456 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \
457 __rsv->size, __rsv->reserved); \
458 spin_unlock(&__rsv->lock); \
461 static const char *space_info_flag_to_str(const struct btrfs_space_info *space_info)
463 switch (space_info->flags) {
464 case BTRFS_BLOCK_GROUP_SYSTEM:
466 case BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA:
467 return "DATA+METADATA";
468 case BTRFS_BLOCK_GROUP_DATA:
470 case BTRFS_BLOCK_GROUP_METADATA:
477 static void dump_global_block_rsv(struct btrfs_fs_info *fs_info)
479 DUMP_BLOCK_RSV(fs_info, global_block_rsv);
480 DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
481 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
482 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
483 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
486 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
487 struct btrfs_space_info *info)
489 const char *flag_str = space_info_flag_to_str(info);
490 lockdep_assert_held(&info->lock);
492 /* The free space could be negative in case of overcommit */
493 btrfs_info(fs_info, "space_info %s has %lld free, is %sfull",
495 (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
496 info->full ? "" : "not ");
498 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
499 info->total_bytes, info->bytes_used, info->bytes_pinned,
500 info->bytes_reserved, info->bytes_may_use,
501 info->bytes_readonly, info->bytes_zone_unusable);
504 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
505 struct btrfs_space_info *info, u64 bytes,
506 int dump_block_groups)
508 struct btrfs_block_group *cache;
512 spin_lock(&info->lock);
513 __btrfs_dump_space_info(fs_info, info);
514 dump_global_block_rsv(fs_info);
515 spin_unlock(&info->lock);
517 if (!dump_block_groups)
520 down_read(&info->groups_sem);
522 list_for_each_entry(cache, &info->block_groups[index], list) {
525 spin_lock(&cache->lock);
526 avail = cache->length - cache->used - cache->pinned -
527 cache->reserved - cache->delalloc_bytes -
528 cache->bytes_super - cache->zone_unusable;
530 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu delalloc %llu super %llu zone_unusable (%llu bytes available) %s",
531 cache->start, cache->length, cache->used, cache->pinned,
532 cache->reserved, cache->delalloc_bytes,
533 cache->bytes_super, cache->zone_unusable,
534 avail, cache->ro ? "[readonly]" : "");
535 spin_unlock(&cache->lock);
536 btrfs_dump_free_space(cache, bytes);
537 total_avail += avail;
539 if (++index < BTRFS_NR_RAID_TYPES)
541 up_read(&info->groups_sem);
543 btrfs_info(fs_info, "%llu bytes available across all block groups", total_avail);
546 static inline u64 calc_reclaim_items_nr(const struct btrfs_fs_info *fs_info,
552 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
553 nr = div64_u64(to_reclaim, bytes);
559 static inline u64 calc_delayed_refs_nr(const struct btrfs_fs_info *fs_info,
562 const u64 bytes = btrfs_calc_delayed_ref_bytes(fs_info, 1);
565 nr = div64_u64(to_reclaim, bytes);
571 #define EXTENT_SIZE_PER_ITEM SZ_256K
574 * shrink metadata reservation for delalloc
576 static void shrink_delalloc(struct btrfs_fs_info *fs_info,
577 struct btrfs_space_info *space_info,
578 u64 to_reclaim, bool wait_ordered,
581 struct btrfs_trans_handle *trans;
588 delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
589 ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes);
590 if (delalloc_bytes == 0 && ordered_bytes == 0)
593 /* Calc the number of the pages we need flush for space reservation */
594 if (to_reclaim == U64_MAX) {
598 * to_reclaim is set to however much metadata we need to
599 * reclaim, but reclaiming that much data doesn't really track
600 * exactly. What we really want to do is reclaim full inode's
601 * worth of reservations, however that's not available to us
602 * here. We will take a fraction of the delalloc bytes for our
603 * flushing loops and hope for the best. Delalloc will expand
604 * the amount we write to cover an entire dirty extent, which
605 * will reclaim the metadata reservation for that range. If
606 * it's not enough subsequent flush stages will be more
609 to_reclaim = max(to_reclaim, delalloc_bytes >> 3);
610 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
613 trans = current->journal_info;
616 * If we are doing more ordered than delalloc we need to just wait on
617 * ordered extents, otherwise we'll waste time trying to flush delalloc
618 * that likely won't give us the space back we need.
620 if (ordered_bytes > delalloc_bytes && !for_preempt)
624 while ((delalloc_bytes || ordered_bytes) && loops < 3) {
625 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
626 long nr_pages = min_t(u64, temp, LONG_MAX);
629 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
632 * We need to make sure any outstanding async pages are now
633 * processed before we continue. This is because things like
634 * sync_inode() try to be smart and skip writing if the inode is
635 * marked clean. We don't use filemap_fwrite for flushing
636 * because we want to control how many pages we write out at a
637 * time, thus this is the only safe way to make sure we've
638 * waited for outstanding compressed workers to have started
639 * their jobs and thus have ordered extents set up properly.
641 * This exists because we do not want to wait for each
642 * individual inode to finish its async work, we simply want to
643 * start the IO on everybody, and then come back here and wait
644 * for all of the async work to catch up. Once we're done with
645 * that we know we'll have ordered extents for everything and we
646 * can decide if we wait for that or not.
648 * If we choose to replace this in the future, make absolutely
649 * sure that the proper waiting is being done in the async case,
650 * as there have been bugs in that area before.
652 async_pages = atomic_read(&fs_info->async_delalloc_pages);
657 * We don't want to wait forever, if we wrote less pages in this
658 * loop than we have outstanding, only wait for that number of
659 * pages, otherwise we can wait for all async pages to finish
662 if (async_pages > nr_pages)
663 async_pages -= nr_pages;
666 wait_event(fs_info->async_submit_wait,
667 atomic_read(&fs_info->async_delalloc_pages) <=
671 if (wait_ordered && !trans) {
672 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
674 time_left = schedule_timeout_killable(1);
680 * If we are for preemption we just want a one-shot of delalloc
681 * flushing so we can stop flushing if we decide we don't need
687 spin_lock(&space_info->lock);
688 if (list_empty(&space_info->tickets) &&
689 list_empty(&space_info->priority_tickets)) {
690 spin_unlock(&space_info->lock);
693 spin_unlock(&space_info->lock);
695 delalloc_bytes = percpu_counter_sum_positive(
696 &fs_info->delalloc_bytes);
697 ordered_bytes = percpu_counter_sum_positive(
698 &fs_info->ordered_bytes);
703 * Try to flush some data based on policy set by @state. This is only advisory
704 * and may fail for various reasons. The caller is supposed to examine the
705 * state of @space_info to detect the outcome.
707 static void flush_space(struct btrfs_fs_info *fs_info,
708 struct btrfs_space_info *space_info, u64 num_bytes,
709 enum btrfs_flush_state state, bool for_preempt)
711 struct btrfs_root *root = fs_info->tree_root;
712 struct btrfs_trans_handle *trans;
717 case FLUSH_DELAYED_ITEMS_NR:
718 case FLUSH_DELAYED_ITEMS:
719 if (state == FLUSH_DELAYED_ITEMS_NR)
720 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
724 trans = btrfs_join_transaction_nostart(root);
726 ret = PTR_ERR(trans);
731 ret = btrfs_run_delayed_items_nr(trans, nr);
732 btrfs_end_transaction(trans);
735 case FLUSH_DELALLOC_WAIT:
736 case FLUSH_DELALLOC_FULL:
737 if (state == FLUSH_DELALLOC_FULL)
739 shrink_delalloc(fs_info, space_info, num_bytes,
740 state != FLUSH_DELALLOC, for_preempt);
742 case FLUSH_DELAYED_REFS_NR:
743 case FLUSH_DELAYED_REFS:
744 trans = btrfs_join_transaction_nostart(root);
746 ret = PTR_ERR(trans);
751 if (state == FLUSH_DELAYED_REFS_NR)
752 nr = calc_delayed_refs_nr(fs_info, num_bytes);
755 btrfs_run_delayed_refs(trans, nr);
756 btrfs_end_transaction(trans);
759 case ALLOC_CHUNK_FORCE:
760 trans = btrfs_join_transaction(root);
762 ret = PTR_ERR(trans);
765 ret = btrfs_chunk_alloc(trans,
766 btrfs_get_alloc_profile(fs_info, space_info->flags),
767 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
769 btrfs_end_transaction(trans);
771 if (ret > 0 || ret == -ENOSPC)
774 case RUN_DELAYED_IPUTS:
776 * If we have pending delayed iputs then we could free up a
777 * bunch of pinned space, so make sure we run the iputs before
778 * we do our pinned bytes check below.
780 btrfs_run_delayed_iputs(fs_info);
781 btrfs_wait_on_delayed_iputs(fs_info);
784 ASSERT(current->journal_info == NULL);
786 * We don't want to start a new transaction, just attach to the
787 * current one or wait it fully commits in case its commit is
788 * happening at the moment. Note: we don't use a nostart join
789 * because that does not wait for a transaction to fully commit
790 * (only for it to be unblocked, state TRANS_STATE_UNBLOCKED).
792 trans = btrfs_attach_transaction_barrier(root);
794 ret = PTR_ERR(trans);
799 ret = btrfs_commit_transaction(trans);
806 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
812 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
813 struct btrfs_space_info *space_info)
817 u64 to_reclaim = space_info->reclaim_size;
819 lockdep_assert_held(&space_info->lock);
821 avail = calc_available_free_space(fs_info, space_info,
822 BTRFS_RESERVE_FLUSH_ALL);
823 used = btrfs_space_info_used(space_info, true);
826 * We may be flushing because suddenly we have less space than we had
827 * before, and now we're well over-committed based on our current free
828 * space. If that's the case add in our overage so we make sure to put
829 * appropriate pressure on the flushing state machine.
831 if (space_info->total_bytes + avail < used)
832 to_reclaim += used - (space_info->total_bytes + avail);
837 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
838 struct btrfs_space_info *space_info)
840 u64 global_rsv_size = fs_info->global_block_rsv.reserved;
841 u64 ordered, delalloc;
845 thresh = mult_perc(space_info->total_bytes, 90);
847 lockdep_assert_held(&space_info->lock);
849 /* If we're just plain full then async reclaim just slows us down. */
850 if ((space_info->bytes_used + space_info->bytes_reserved +
851 global_rsv_size) >= thresh)
854 used = space_info->bytes_may_use + space_info->bytes_pinned;
856 /* The total flushable belongs to the global rsv, don't flush. */
857 if (global_rsv_size >= used)
861 * 128MiB is 1/4 of the maximum global rsv size. If we have less than
862 * that devoted to other reservations then there's no sense in flushing,
863 * we don't have a lot of things that need flushing.
865 if (used - global_rsv_size <= SZ_128M)
869 * We have tickets queued, bail so we don't compete with the async
872 if (space_info->reclaim_size)
876 * If we have over half of the free space occupied by reservations or
877 * pinned then we want to start flushing.
879 * We do not do the traditional thing here, which is to say
881 * if (used >= ((total_bytes + avail) / 2))
884 * because this doesn't quite work how we want. If we had more than 50%
885 * of the space_info used by bytes_used and we had 0 available we'd just
886 * constantly run the background flusher. Instead we want it to kick in
887 * if our reclaimable space exceeds our clamped free space.
889 * Our clamping range is 2^1 -> 2^8. Practically speaking that means
892 * Amount of RAM Minimum threshold Maximum threshold
895 * 128GiB 512MiB 64GiB
900 * These are the range our thresholds will fall in, corresponding to how
901 * much delalloc we need for the background flusher to kick in.
904 thresh = calc_available_free_space(fs_info, space_info,
905 BTRFS_RESERVE_FLUSH_ALL);
906 used = space_info->bytes_used + space_info->bytes_reserved +
907 space_info->bytes_readonly + global_rsv_size;
908 if (used < space_info->total_bytes)
909 thresh += space_info->total_bytes - used;
910 thresh >>= space_info->clamp;
912 used = space_info->bytes_pinned;
915 * If we have more ordered bytes than delalloc bytes then we're either
916 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting
917 * around. Preemptive flushing is only useful in that it can free up
918 * space before tickets need to wait for things to finish. In the case
919 * of ordered extents, preemptively waiting on ordered extents gets us
920 * nothing, if our reservations are tied up in ordered extents we'll
921 * simply have to slow down writers by forcing them to wait on ordered
924 * In the case that ordered is larger than delalloc, only include the
925 * block reserves that we would actually be able to directly reclaim
926 * from. In this case if we're heavy on metadata operations this will
927 * clearly be heavy enough to warrant preemptive flushing. In the case
928 * of heavy DIO or ordered reservations, preemptive flushing will just
929 * waste time and cause us to slow down.
931 * We want to make sure we truly are maxed out on ordered however, so
932 * cut ordered in half, and if it's still higher than delalloc then we
933 * can keep flushing. This is to avoid the case where we start
934 * flushing, and now delalloc == ordered and we stop preemptively
935 * flushing when we could still have several gigs of delalloc to flush.
937 ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1;
938 delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes);
939 if (ordered >= delalloc)
940 used += fs_info->delayed_refs_rsv.reserved +
941 fs_info->delayed_block_rsv.reserved;
943 used += space_info->bytes_may_use - global_rsv_size;
945 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
946 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
949 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
950 struct btrfs_space_info *space_info,
951 struct reserve_ticket *ticket)
953 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
959 if (global_rsv->space_info != space_info)
962 spin_lock(&global_rsv->lock);
963 min_bytes = mult_perc(global_rsv->size, 10);
964 if (global_rsv->reserved < min_bytes + ticket->bytes) {
965 spin_unlock(&global_rsv->lock);
968 global_rsv->reserved -= ticket->bytes;
969 remove_ticket(space_info, ticket);
971 wake_up(&ticket->wait);
972 space_info->tickets_id++;
973 if (global_rsv->reserved < global_rsv->size)
974 global_rsv->full = 0;
975 spin_unlock(&global_rsv->lock);
981 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
982 * @fs_info - fs_info for this fs
983 * @space_info - the space info we were flushing
985 * We call this when we've exhausted our flushing ability and haven't made
986 * progress in satisfying tickets. The reservation code handles tickets in
987 * order, so if there is a large ticket first and then smaller ones we could
988 * very well satisfy the smaller tickets. This will attempt to wake up any
989 * tickets in the list to catch this case.
991 * This function returns true if it was able to make progress by clearing out
992 * other tickets, or if it stumbles across a ticket that was smaller than the
995 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
996 struct btrfs_space_info *space_info)
998 struct reserve_ticket *ticket;
999 u64 tickets_id = space_info->tickets_id;
1000 const bool aborted = BTRFS_FS_ERROR(fs_info);
1002 trace_btrfs_fail_all_tickets(fs_info, space_info);
1004 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1005 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
1006 __btrfs_dump_space_info(fs_info, space_info);
1009 while (!list_empty(&space_info->tickets) &&
1010 tickets_id == space_info->tickets_id) {
1011 ticket = list_first_entry(&space_info->tickets,
1012 struct reserve_ticket, list);
1014 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket))
1017 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1018 btrfs_info(fs_info, "failing ticket with %llu bytes",
1021 remove_ticket(space_info, ticket);
1023 ticket->error = -EIO;
1025 ticket->error = -ENOSPC;
1026 wake_up(&ticket->wait);
1029 * We're just throwing tickets away, so more flushing may not
1030 * trip over btrfs_try_granting_tickets, so we need to call it
1031 * here to see if we can make progress with the next ticket in
1035 btrfs_try_granting_tickets(fs_info, space_info);
1037 return (tickets_id != space_info->tickets_id);
1041 * This is for normal flushers, we can wait all goddamned day if we want to. We
1042 * will loop and continuously try to flush as long as we are making progress.
1043 * We count progress as clearing off tickets each time we have to loop.
1045 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
1047 struct btrfs_fs_info *fs_info;
1048 struct btrfs_space_info *space_info;
1050 enum btrfs_flush_state flush_state;
1051 int commit_cycles = 0;
1052 u64 last_tickets_id;
1054 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
1055 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1057 spin_lock(&space_info->lock);
1058 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1060 space_info->flush = 0;
1061 spin_unlock(&space_info->lock);
1064 last_tickets_id = space_info->tickets_id;
1065 spin_unlock(&space_info->lock);
1067 flush_state = FLUSH_DELAYED_ITEMS_NR;
1069 flush_space(fs_info, space_info, to_reclaim, flush_state, false);
1070 spin_lock(&space_info->lock);
1071 if (list_empty(&space_info->tickets)) {
1072 space_info->flush = 0;
1073 spin_unlock(&space_info->lock);
1076 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
1078 if (last_tickets_id == space_info->tickets_id) {
1081 last_tickets_id = space_info->tickets_id;
1082 flush_state = FLUSH_DELAYED_ITEMS_NR;
1088 * We do not want to empty the system of delalloc unless we're
1089 * under heavy pressure, so allow one trip through the flushing
1090 * logic before we start doing a FLUSH_DELALLOC_FULL.
1092 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles)
1096 * We don't want to force a chunk allocation until we've tried
1097 * pretty hard to reclaim space. Think of the case where we
1098 * freed up a bunch of space and so have a lot of pinned space
1099 * to reclaim. We would rather use that than possibly create a
1100 * underutilized metadata chunk. So if this is our first run
1101 * through the flushing state machine skip ALLOC_CHUNK_FORCE and
1102 * commit the transaction. If nothing has changed the next go
1103 * around then we can force a chunk allocation.
1105 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
1108 if (flush_state > COMMIT_TRANS) {
1110 if (commit_cycles > 2) {
1111 if (maybe_fail_all_tickets(fs_info, space_info)) {
1112 flush_state = FLUSH_DELAYED_ITEMS_NR;
1115 space_info->flush = 0;
1118 flush_state = FLUSH_DELAYED_ITEMS_NR;
1121 spin_unlock(&space_info->lock);
1122 } while (flush_state <= COMMIT_TRANS);
1126 * This handles pre-flushing of metadata space before we get to the point that
1127 * we need to start blocking threads on tickets. The logic here is different
1128 * from the other flush paths because it doesn't rely on tickets to tell us how
1129 * much we need to flush, instead it attempts to keep us below the 80% full
1130 * watermark of space by flushing whichever reservation pool is currently the
1133 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
1135 struct btrfs_fs_info *fs_info;
1136 struct btrfs_space_info *space_info;
1137 struct btrfs_block_rsv *delayed_block_rsv;
1138 struct btrfs_block_rsv *delayed_refs_rsv;
1139 struct btrfs_block_rsv *global_rsv;
1140 struct btrfs_block_rsv *trans_rsv;
1143 fs_info = container_of(work, struct btrfs_fs_info,
1144 preempt_reclaim_work);
1145 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1146 delayed_block_rsv = &fs_info->delayed_block_rsv;
1147 delayed_refs_rsv = &fs_info->delayed_refs_rsv;
1148 global_rsv = &fs_info->global_block_rsv;
1149 trans_rsv = &fs_info->trans_block_rsv;
1151 spin_lock(&space_info->lock);
1152 while (need_preemptive_reclaim(fs_info, space_info)) {
1153 enum btrfs_flush_state flush;
1154 u64 delalloc_size = 0;
1155 u64 to_reclaim, block_rsv_size;
1156 u64 global_rsv_size = global_rsv->reserved;
1161 * We don't have a precise counter for the metadata being
1162 * reserved for delalloc, so we'll approximate it by subtracting
1163 * out the block rsv's space from the bytes_may_use. If that
1164 * amount is higher than the individual reserves, then we can
1165 * assume it's tied up in delalloc reservations.
1167 block_rsv_size = global_rsv_size +
1168 delayed_block_rsv->reserved +
1169 delayed_refs_rsv->reserved +
1170 trans_rsv->reserved;
1171 if (block_rsv_size < space_info->bytes_may_use)
1172 delalloc_size = space_info->bytes_may_use - block_rsv_size;
1175 * We don't want to include the global_rsv in our calculation,
1176 * because that's space we can't touch. Subtract it from the
1177 * block_rsv_size for the next checks.
1179 block_rsv_size -= global_rsv_size;
1182 * We really want to avoid flushing delalloc too much, as it
1183 * could result in poor allocation patterns, so only flush it if
1184 * it's larger than the rest of the pools combined.
1186 if (delalloc_size > block_rsv_size) {
1187 to_reclaim = delalloc_size;
1188 flush = FLUSH_DELALLOC;
1189 } else if (space_info->bytes_pinned >
1190 (delayed_block_rsv->reserved +
1191 delayed_refs_rsv->reserved)) {
1192 to_reclaim = space_info->bytes_pinned;
1193 flush = COMMIT_TRANS;
1194 } else if (delayed_block_rsv->reserved >
1195 delayed_refs_rsv->reserved) {
1196 to_reclaim = delayed_block_rsv->reserved;
1197 flush = FLUSH_DELAYED_ITEMS_NR;
1199 to_reclaim = delayed_refs_rsv->reserved;
1200 flush = FLUSH_DELAYED_REFS_NR;
1203 spin_unlock(&space_info->lock);
1206 * We don't want to reclaim everything, just a portion, so scale
1207 * down the to_reclaim by 1/4. If it takes us down to 0,
1208 * reclaim 1 items worth.
1212 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1213 flush_space(fs_info, space_info, to_reclaim, flush, true);
1215 spin_lock(&space_info->lock);
1218 /* We only went through once, back off our clamping. */
1219 if (loops == 1 && !space_info->reclaim_size)
1220 space_info->clamp = max(1, space_info->clamp - 1);
1221 trace_btrfs_done_preemptive_reclaim(fs_info, space_info);
1222 spin_unlock(&space_info->lock);
1226 * FLUSH_DELALLOC_WAIT:
1227 * Space is freed from flushing delalloc in one of two ways.
1229 * 1) compression is on and we allocate less space than we reserved
1230 * 2) we are overwriting existing space
1232 * For #1 that extra space is reclaimed as soon as the delalloc pages are
1233 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1234 * length to ->bytes_reserved, and subtracts the reserved space from
1237 * For #2 this is trickier. Once the ordered extent runs we will drop the
1238 * extent in the range we are overwriting, which creates a delayed ref for
1239 * that freed extent. This however is not reclaimed until the transaction
1240 * commits, thus the next stages.
1243 * If we are freeing inodes, we want to make sure all delayed iputs have
1244 * completed, because they could have been on an inode with i_nlink == 0, and
1245 * thus have been truncated and freed up space. But again this space is not
1246 * immediately re-usable, it comes in the form of a delayed ref, which must be
1247 * run and then the transaction must be committed.
1250 * This is where we reclaim all of the pinned space generated by running the
1254 * For data we start with alloc chunk force, however we could have been full
1255 * before, and then the transaction commit could have freed new block groups,
1256 * so if we now have space to allocate do the force chunk allocation.
1258 static const enum btrfs_flush_state data_flush_states[] = {
1259 FLUSH_DELALLOC_FULL,
1265 static void btrfs_async_reclaim_data_space(struct work_struct *work)
1267 struct btrfs_fs_info *fs_info;
1268 struct btrfs_space_info *space_info;
1269 u64 last_tickets_id;
1270 enum btrfs_flush_state flush_state = 0;
1272 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1273 space_info = fs_info->data_sinfo;
1275 spin_lock(&space_info->lock);
1276 if (list_empty(&space_info->tickets)) {
1277 space_info->flush = 0;
1278 spin_unlock(&space_info->lock);
1281 last_tickets_id = space_info->tickets_id;
1282 spin_unlock(&space_info->lock);
1284 while (!space_info->full) {
1285 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1286 spin_lock(&space_info->lock);
1287 if (list_empty(&space_info->tickets)) {
1288 space_info->flush = 0;
1289 spin_unlock(&space_info->lock);
1293 /* Something happened, fail everything and bail. */
1294 if (BTRFS_FS_ERROR(fs_info))
1296 last_tickets_id = space_info->tickets_id;
1297 spin_unlock(&space_info->lock);
1300 while (flush_state < ARRAY_SIZE(data_flush_states)) {
1301 flush_space(fs_info, space_info, U64_MAX,
1302 data_flush_states[flush_state], false);
1303 spin_lock(&space_info->lock);
1304 if (list_empty(&space_info->tickets)) {
1305 space_info->flush = 0;
1306 spin_unlock(&space_info->lock);
1310 if (last_tickets_id == space_info->tickets_id) {
1313 last_tickets_id = space_info->tickets_id;
1317 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1318 if (space_info->full) {
1319 if (maybe_fail_all_tickets(fs_info, space_info))
1322 space_info->flush = 0;
1327 /* Something happened, fail everything and bail. */
1328 if (BTRFS_FS_ERROR(fs_info))
1332 spin_unlock(&space_info->lock);
1337 maybe_fail_all_tickets(fs_info, space_info);
1338 space_info->flush = 0;
1339 spin_unlock(&space_info->lock);
1342 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1344 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1345 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1346 INIT_WORK(&fs_info->preempt_reclaim_work,
1347 btrfs_preempt_reclaim_metadata_space);
1350 static const enum btrfs_flush_state priority_flush_states[] = {
1351 FLUSH_DELAYED_ITEMS_NR,
1352 FLUSH_DELAYED_ITEMS,
1356 static const enum btrfs_flush_state evict_flush_states[] = {
1357 FLUSH_DELAYED_ITEMS_NR,
1358 FLUSH_DELAYED_ITEMS,
1359 FLUSH_DELAYED_REFS_NR,
1362 FLUSH_DELALLOC_WAIT,
1363 FLUSH_DELALLOC_FULL,
1368 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1369 struct btrfs_space_info *space_info,
1370 struct reserve_ticket *ticket,
1371 const enum btrfs_flush_state *states,
1375 int flush_state = 0;
1377 spin_lock(&space_info->lock);
1378 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1380 * This is the priority reclaim path, so to_reclaim could be >0 still
1381 * because we may have only satisfied the priority tickets and still
1382 * left non priority tickets on the list. We would then have
1383 * to_reclaim but ->bytes == 0.
1385 if (ticket->bytes == 0) {
1386 spin_unlock(&space_info->lock);
1390 while (flush_state < states_nr) {
1391 spin_unlock(&space_info->lock);
1392 flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1395 spin_lock(&space_info->lock);
1396 if (ticket->bytes == 0) {
1397 spin_unlock(&space_info->lock);
1403 * Attempt to steal from the global rsv if we can, except if the fs was
1404 * turned into error mode due to a transaction abort when flushing space
1405 * above, in that case fail with the abort error instead of returning
1406 * success to the caller if we can steal from the global rsv - this is
1407 * just to have caller fail immeditelly instead of later when trying to
1408 * modify the fs, making it easier to debug -ENOSPC problems.
1410 if (BTRFS_FS_ERROR(fs_info)) {
1411 ticket->error = BTRFS_FS_ERROR(fs_info);
1412 remove_ticket(space_info, ticket);
1413 } else if (!steal_from_global_rsv(fs_info, space_info, ticket)) {
1414 ticket->error = -ENOSPC;
1415 remove_ticket(space_info, ticket);
1419 * We must run try_granting_tickets here because we could be a large
1420 * ticket in front of a smaller ticket that can now be satisfied with
1421 * the available space.
1423 btrfs_try_granting_tickets(fs_info, space_info);
1424 spin_unlock(&space_info->lock);
1427 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1428 struct btrfs_space_info *space_info,
1429 struct reserve_ticket *ticket)
1431 spin_lock(&space_info->lock);
1433 /* We could have been granted before we got here. */
1434 if (ticket->bytes == 0) {
1435 spin_unlock(&space_info->lock);
1439 while (!space_info->full) {
1440 spin_unlock(&space_info->lock);
1441 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1442 spin_lock(&space_info->lock);
1443 if (ticket->bytes == 0) {
1444 spin_unlock(&space_info->lock);
1449 ticket->error = -ENOSPC;
1450 remove_ticket(space_info, ticket);
1451 btrfs_try_granting_tickets(fs_info, space_info);
1452 spin_unlock(&space_info->lock);
1455 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1456 struct btrfs_space_info *space_info,
1457 struct reserve_ticket *ticket)
1463 spin_lock(&space_info->lock);
1464 while (ticket->bytes > 0 && ticket->error == 0) {
1465 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1468 * Delete us from the list. After we unlock the space
1469 * info, we don't want the async reclaim job to reserve
1470 * space for this ticket. If that would happen, then the
1471 * ticket's task would not known that space was reserved
1472 * despite getting an error, resulting in a space leak
1473 * (bytes_may_use counter of our space_info).
1475 remove_ticket(space_info, ticket);
1476 ticket->error = -EINTR;
1479 spin_unlock(&space_info->lock);
1483 finish_wait(&ticket->wait, &wait);
1484 spin_lock(&space_info->lock);
1486 spin_unlock(&space_info->lock);
1490 * Do the appropriate flushing and waiting for a ticket.
1492 * @fs_info: the filesystem
1493 * @space_info: space info for the reservation
1494 * @ticket: ticket for the reservation
1495 * @start_ns: timestamp when the reservation started
1496 * @orig_bytes: amount of bytes originally reserved
1497 * @flush: how much we can flush
1499 * This does the work of figuring out how to flush for the ticket, waiting for
1500 * the reservation, and returning the appropriate error if there is one.
1502 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1503 struct btrfs_space_info *space_info,
1504 struct reserve_ticket *ticket,
1505 u64 start_ns, u64 orig_bytes,
1506 enum btrfs_reserve_flush_enum flush)
1511 case BTRFS_RESERVE_FLUSH_DATA:
1512 case BTRFS_RESERVE_FLUSH_ALL:
1513 case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1514 wait_reserve_ticket(fs_info, space_info, ticket);
1516 case BTRFS_RESERVE_FLUSH_LIMIT:
1517 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1518 priority_flush_states,
1519 ARRAY_SIZE(priority_flush_states));
1521 case BTRFS_RESERVE_FLUSH_EVICT:
1522 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1524 ARRAY_SIZE(evict_flush_states));
1526 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1527 priority_reclaim_data_space(fs_info, space_info, ticket);
1534 ret = ticket->error;
1535 ASSERT(list_empty(&ticket->list));
1537 * Check that we can't have an error set if the reservation succeeded,
1538 * as that would confuse tasks and lead them to error out without
1539 * releasing reserved space (if an error happens the expectation is that
1540 * space wasn't reserved at all).
1542 ASSERT(!(ticket->bytes == 0 && ticket->error));
1543 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes,
1544 start_ns, flush, ticket->error);
1549 * This returns true if this flush state will go through the ordinary flushing
1552 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1554 return (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1555 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1558 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1559 struct btrfs_space_info *space_info)
1561 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1562 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1565 * If we're heavy on ordered operations then clamping won't help us. We
1566 * need to clamp specifically to keep up with dirty'ing buffered
1567 * writers, because there's not a 1:1 correlation of writing delalloc
1568 * and freeing space, like there is with flushing delayed refs or
1569 * delayed nodes. If we're already more ordered than delalloc then
1570 * we're keeping up, otherwise we aren't and should probably clamp.
1572 if (ordered < delalloc)
1573 space_info->clamp = min(space_info->clamp + 1, 8);
1576 static inline bool can_steal(enum btrfs_reserve_flush_enum flush)
1578 return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1579 flush == BTRFS_RESERVE_FLUSH_EVICT);
1583 * NO_FLUSH and FLUSH_EMERGENCY don't want to create a ticket, they just want to
1584 * fail as quickly as possible.
1586 static inline bool can_ticket(enum btrfs_reserve_flush_enum flush)
1588 return (flush != BTRFS_RESERVE_NO_FLUSH &&
1589 flush != BTRFS_RESERVE_FLUSH_EMERGENCY);
1593 * Try to reserve bytes from the block_rsv's space.
1595 * @fs_info: the filesystem
1596 * @space_info: space info we want to allocate from
1597 * @orig_bytes: number of bytes we want
1598 * @flush: whether or not we can flush to make our reservation
1600 * This will reserve orig_bytes number of bytes from the space info associated
1601 * with the block_rsv. If there is not enough space it will make an attempt to
1602 * flush out space to make room. It will do this by flushing delalloc if
1603 * possible or committing the transaction. If flush is 0 then no attempts to
1604 * regain reservations will be made and this will fail if there is not enough
1607 static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1608 struct btrfs_space_info *space_info, u64 orig_bytes,
1609 enum btrfs_reserve_flush_enum flush)
1611 struct work_struct *async_work;
1612 struct reserve_ticket ticket;
1616 bool pending_tickets;
1620 * If have a transaction handle (current->journal_info != NULL), then
1621 * the flush method can not be neither BTRFS_RESERVE_FLUSH_ALL* nor
1622 * BTRFS_RESERVE_FLUSH_EVICT, as we could deadlock because those
1623 * flushing methods can trigger transaction commits.
1625 if (current->journal_info) {
1626 /* One assert per line for easier debugging. */
1627 ASSERT(flush != BTRFS_RESERVE_FLUSH_ALL);
1628 ASSERT(flush != BTRFS_RESERVE_FLUSH_ALL_STEAL);
1629 ASSERT(flush != BTRFS_RESERVE_FLUSH_EVICT);
1632 if (flush == BTRFS_RESERVE_FLUSH_DATA)
1633 async_work = &fs_info->async_data_reclaim_work;
1635 async_work = &fs_info->async_reclaim_work;
1637 spin_lock(&space_info->lock);
1638 used = btrfs_space_info_used(space_info, true);
1641 * We don't want NO_FLUSH allocations to jump everybody, they can
1642 * generally handle ENOSPC in a different way, so treat them the same as
1643 * normal flushers when it comes to skipping pending tickets.
1645 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1646 pending_tickets = !list_empty(&space_info->tickets) ||
1647 !list_empty(&space_info->priority_tickets);
1649 pending_tickets = !list_empty(&space_info->priority_tickets);
1652 * Carry on if we have enough space (short-circuit) OR call
1653 * can_overcommit() to ensure we can overcommit to continue.
1655 if (!pending_tickets &&
1656 ((used + orig_bytes <= space_info->total_bytes) ||
1657 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1658 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1664 * Things are dire, we need to make a reservation so we don't abort. We
1665 * will let this reservation go through as long as we have actual space
1666 * left to allocate for the block.
1668 if (ret && unlikely(flush == BTRFS_RESERVE_FLUSH_EMERGENCY)) {
1669 used = btrfs_space_info_used(space_info, false);
1670 if (used + orig_bytes <= space_info->total_bytes) {
1671 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1678 * If we couldn't make a reservation then setup our reservation ticket
1679 * and kick the async worker if it's not already running.
1681 * If we are a priority flusher then we just need to add our ticket to
1682 * the list and we will do our own flushing further down.
1684 if (ret && can_ticket(flush)) {
1685 ticket.bytes = orig_bytes;
1687 space_info->reclaim_size += ticket.bytes;
1688 init_waitqueue_head(&ticket.wait);
1689 ticket.steal = can_steal(flush);
1690 if (trace_btrfs_reserve_ticket_enabled())
1691 start_ns = ktime_get_ns();
1693 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1694 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1695 flush == BTRFS_RESERVE_FLUSH_DATA) {
1696 list_add_tail(&ticket.list, &space_info->tickets);
1697 if (!space_info->flush) {
1699 * We were forced to add a reserve ticket, so
1700 * our preemptive flushing is unable to keep
1701 * up. Clamp down on the threshold for the
1702 * preemptive flushing in order to keep up with
1705 maybe_clamp_preempt(fs_info, space_info);
1707 space_info->flush = 1;
1708 trace_btrfs_trigger_flush(fs_info,
1712 queue_work(system_unbound_wq, async_work);
1715 list_add_tail(&ticket.list,
1716 &space_info->priority_tickets);
1718 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1720 * We will do the space reservation dance during log replay,
1721 * which means we won't have fs_info->fs_root set, so don't do
1722 * the async reclaim as we will panic.
1724 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1725 !work_busy(&fs_info->preempt_reclaim_work) &&
1726 need_preemptive_reclaim(fs_info, space_info)) {
1727 trace_btrfs_trigger_flush(fs_info, space_info->flags,
1728 orig_bytes, flush, "preempt");
1729 queue_work(system_unbound_wq,
1730 &fs_info->preempt_reclaim_work);
1733 spin_unlock(&space_info->lock);
1734 if (!ret || !can_ticket(flush))
1737 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1742 * Try to reserve metadata bytes from the block_rsv's space.
1744 * @fs_info: the filesystem
1745 * @block_rsv: block_rsv we're allocating for
1746 * @orig_bytes: number of bytes we want
1747 * @flush: whether or not we can flush to make our reservation
1749 * This will reserve orig_bytes number of bytes from the space info associated
1750 * with the block_rsv. If there is not enough space it will make an attempt to
1751 * flush out space to make room. It will do this by flushing delalloc if
1752 * possible or committing the transaction. If flush is 0 then no attempts to
1753 * regain reservations will be made and this will fail if there is not enough
1756 int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1757 struct btrfs_block_rsv *block_rsv,
1759 enum btrfs_reserve_flush_enum flush)
1763 ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush);
1764 if (ret == -ENOSPC) {
1765 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1766 block_rsv->space_info->flags,
1769 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1770 btrfs_dump_space_info(fs_info, block_rsv->space_info,
1777 * Try to reserve data bytes for an allocation.
1779 * @fs_info: the filesystem
1780 * @bytes: number of bytes we need
1781 * @flush: how we are allowed to flush
1783 * This will reserve bytes from the data space info. If there is not enough
1784 * space then we will attempt to flush space as specified by flush.
1786 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1787 enum btrfs_reserve_flush_enum flush)
1789 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1792 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1793 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE ||
1794 flush == BTRFS_RESERVE_NO_FLUSH);
1795 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1797 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1798 if (ret == -ENOSPC) {
1799 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1800 data_sinfo->flags, bytes, 1);
1801 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1802 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1807 /* Dump all the space infos when we abort a transaction due to ENOSPC. */
1808 __cold void btrfs_dump_space_info_for_trans_abort(struct btrfs_fs_info *fs_info)
1810 struct btrfs_space_info *space_info;
1812 btrfs_info(fs_info, "dumping space info:");
1813 list_for_each_entry(space_info, &fs_info->space_info, list) {
1814 spin_lock(&space_info->lock);
1815 __btrfs_dump_space_info(fs_info, space_info);
1816 spin_unlock(&space_info->lock);
1818 dump_global_block_rsv(fs_info);
1822 * Account the unused space of all the readonly block group in the space_info.
1823 * takes mirrors into account.
1825 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
1827 struct btrfs_block_group *block_group;
1831 /* It's df, we don't care if it's racy */
1832 if (list_empty(&sinfo->ro_bgs))
1835 spin_lock(&sinfo->lock);
1836 list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
1837 spin_lock(&block_group->lock);
1839 if (!block_group->ro) {
1840 spin_unlock(&block_group->lock);
1844 factor = btrfs_bg_type_to_factor(block_group->flags);
1845 free_bytes += (block_group->length -
1846 block_group->used) * factor;
1848 spin_unlock(&block_group->lock);
1850 spin_unlock(&sinfo->lock);