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"
15 * HOW DOES SPACE RESERVATION WORK
17 * If you want to know about delalloc specifically, there is a separate comment
18 * for that with the delalloc code. This comment is about how the whole system
23 * 1) space_info. This is the ultimate arbiter of how much space we can use.
24 * There's a description of the bytes_ fields with the struct declaration,
25 * refer to that for specifics on each field. Suffice it to say that for
26 * reservations we care about total_bytes - SUM(space_info->bytes_) when
27 * determining if there is space to make an allocation. There is a space_info
28 * for METADATA, SYSTEM, and DATA areas.
30 * 2) block_rsv's. These are basically buckets for every different type of
31 * metadata reservation we have. You can see the comment in the block_rsv
32 * code on the rules for each type, but generally block_rsv->reserved is how
33 * much space is accounted for in space_info->bytes_may_use.
35 * 3) btrfs_calc*_size. These are the worst case calculations we used based
36 * on the number of items we will want to modify. We have one for changing
37 * items, and one for inserting new items. Generally we use these helpers to
38 * determine the size of the block reserves, and then use the actual bytes
39 * values to adjust the space_info counters.
41 * MAKING RESERVATIONS, THE NORMAL CASE
43 * We call into either btrfs_reserve_data_bytes() or
44 * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with
45 * num_bytes we want to reserve.
48 * space_info->bytes_may_reserve += num_bytes
51 * Call btrfs_add_reserved_bytes() which does
52 * space_info->bytes_may_reserve -= num_bytes
53 * space_info->bytes_reserved += extent_bytes
56 * Call btrfs_update_block_group() which does
57 * space_info->bytes_reserved -= extent_bytes
58 * space_info->bytes_used += extent_bytes
60 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority)
62 * Assume we are unable to simply make the reservation because we do not have
66 * create a reserve_ticket with ->bytes set to our reservation, add it to
67 * the tail of space_info->tickets, kick async flush thread
69 * ->handle_reserve_ticket
70 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set
73 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space
74 * Flushes various things attempting to free up space.
76 * -> btrfs_try_granting_tickets()
77 * This is called by anything that either subtracts space from
78 * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the
79 * space_info->total_bytes. This loops through the ->priority_tickets and
80 * then the ->tickets list checking to see if the reservation can be
81 * completed. If it can the space is added to space_info->bytes_may_use and
82 * the ticket is woken up.
85 * Check if ->bytes == 0, if it does we got our reservation and we can carry
86 * on, if not return the appropriate error (ENOSPC, but can be EINTR if we
89 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY
91 * Same as the above, except we add ourselves to the
92 * space_info->priority_tickets, and we do not use ticket->wait, we simply
93 * call flush_space() ourselves for the states that are safe for us to call
94 * without deadlocking and hope for the best.
98 * Generally speaking we will have two cases for each state, a "nice" state
99 * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to
100 * reduce the locking over head on the various trees, and even to keep from
101 * doing any work at all in the case of delayed refs. Each of these delayed
102 * things however hold reservations, and so letting them run allows us to
103 * reclaim space so we can make new reservations.
105 * FLUSH_DELAYED_ITEMS
106 * Every inode has a delayed item to update the inode. Take a simple write
107 * for example, we would update the inode item at write time to update the
108 * mtime, and then again at finish_ordered_io() time in order to update the
109 * isize or bytes. We keep these delayed items to coalesce these operations
110 * into a single operation done on demand. These are an easy way to reclaim
114 * Look at the delalloc comment to get an idea of how much space is reserved
115 * for delayed allocation. We can reclaim some of this space simply by
116 * running delalloc, but usually we need to wait for ordered extents to
117 * reclaim the bulk of this space.
120 * We have a block reserve for the outstanding delayed refs space, and every
121 * delayed ref operation holds a reservation. Running these is a quick way
122 * to reclaim space, but we want to hold this until the end because COW can
123 * churn a lot and we can avoid making some extent tree modifications if we
124 * are able to delay for as long as possible.
127 * We will skip this the first time through space reservation, because of
128 * overcommit and we don't want to have a lot of useless metadata space when
129 * our worst case reservations will likely never come true.
132 * If we're freeing inodes we're likely freeing checksums, file extent
133 * items, and extent tree items. Loads of space could be freed up by these
134 * operations, however they won't be usable until the transaction commits.
137 * This will commit the transaction. Historically we had a lot of logic
138 * surrounding whether or not we'd commit the transaction, but this waits born
139 * out of a pre-tickets era where we could end up committing the transaction
140 * thousands of times in a row without making progress. Now thanks to our
141 * ticketing system we know if we're not making progress and can error
142 * everybody out after a few commits rather than burning the disk hoping for
143 * a different answer.
147 * Because we hold so many reservations for metadata we will allow you to
148 * reserve more space than is currently free in the currently allocate
149 * metadata space. This only happens with metadata, data does not allow
152 * You can see the current logic for when we allow overcommit in
153 * btrfs_can_overcommit(), but it only applies to unallocated space. If there
154 * is no unallocated space to be had, all reservations are kept within the
155 * free space in the allocated metadata chunks.
157 * Because of overcommitting, you generally want to use the
158 * btrfs_can_overcommit() logic for metadata allocations, as it does the right
159 * thing with or without extra unallocated space.
162 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
163 bool may_use_included)
166 return s_info->bytes_used + s_info->bytes_reserved +
167 s_info->bytes_pinned + s_info->bytes_readonly +
168 s_info->bytes_zone_unusable +
169 (may_use_included ? s_info->bytes_may_use : 0);
173 * after adding space to the filesystem, we need to clear the full flags
174 * on all the space infos.
176 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
178 struct list_head *head = &info->space_info;
179 struct btrfs_space_info *found;
181 list_for_each_entry(found, head, list)
186 * Block groups with more than this value (percents) of unusable space will be
187 * scheduled for background reclaim.
189 #define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH (75)
192 * Calculate chunk size depending on volume type (regular or zoned).
194 static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags)
196 if (btrfs_is_zoned(fs_info))
197 return fs_info->zone_size;
199 ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
201 if (flags & BTRFS_BLOCK_GROUP_DATA)
202 return BTRFS_MAX_DATA_CHUNK_SIZE;
203 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
206 /* Handle BTRFS_BLOCK_GROUP_METADATA */
207 if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G)
214 * Update default chunk size.
216 void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info,
219 WRITE_ONCE(space_info->chunk_size, chunk_size);
222 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
225 struct btrfs_space_info *space_info;
229 space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
233 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
234 INIT_LIST_HEAD(&space_info->block_groups[i]);
235 init_rwsem(&space_info->groups_sem);
236 spin_lock_init(&space_info->lock);
237 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
238 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
239 INIT_LIST_HEAD(&space_info->ro_bgs);
240 INIT_LIST_HEAD(&space_info->tickets);
241 INIT_LIST_HEAD(&space_info->priority_tickets);
242 space_info->clamp = 1;
243 btrfs_update_space_info_chunk_size(space_info, calc_chunk_size(info, flags));
245 if (btrfs_is_zoned(info))
246 space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH;
248 ret = btrfs_sysfs_add_space_info_type(info, space_info);
252 list_add(&space_info->list, &info->space_info);
253 if (flags & BTRFS_BLOCK_GROUP_DATA)
254 info->data_sinfo = space_info;
259 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
261 struct btrfs_super_block *disk_super;
267 disk_super = fs_info->super_copy;
268 if (!btrfs_super_root(disk_super))
271 features = btrfs_super_incompat_flags(disk_super);
272 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
275 flags = BTRFS_BLOCK_GROUP_SYSTEM;
276 ret = create_space_info(fs_info, flags);
281 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
282 ret = create_space_info(fs_info, flags);
284 flags = BTRFS_BLOCK_GROUP_METADATA;
285 ret = create_space_info(fs_info, flags);
289 flags = BTRFS_BLOCK_GROUP_DATA;
290 ret = create_space_info(fs_info, flags);
296 void btrfs_add_bg_to_space_info(struct btrfs_fs_info *info,
297 struct btrfs_block_group *block_group)
299 struct btrfs_space_info *found;
302 factor = btrfs_bg_type_to_factor(block_group->flags);
304 found = btrfs_find_space_info(info, block_group->flags);
306 spin_lock(&found->lock);
307 found->total_bytes += block_group->length;
308 if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags))
309 found->active_total_bytes += block_group->length;
310 found->disk_total += block_group->length * factor;
311 found->bytes_used += block_group->used;
312 found->disk_used += block_group->used * factor;
313 found->bytes_readonly += block_group->bytes_super;
314 found->bytes_zone_unusable += block_group->zone_unusable;
315 if (block_group->length > 0)
317 btrfs_try_granting_tickets(info, found);
318 spin_unlock(&found->lock);
320 block_group->space_info = found;
322 index = btrfs_bg_flags_to_raid_index(block_group->flags);
323 down_write(&found->groups_sem);
324 list_add_tail(&block_group->list, &found->block_groups[index]);
325 up_write(&found->groups_sem);
328 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
331 struct list_head *head = &info->space_info;
332 struct btrfs_space_info *found;
334 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
336 list_for_each_entry(found, head, list) {
337 if (found->flags & flags)
343 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
344 struct btrfs_space_info *space_info,
345 enum btrfs_reserve_flush_enum flush)
351 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
352 profile = btrfs_system_alloc_profile(fs_info);
354 profile = btrfs_metadata_alloc_profile(fs_info);
356 avail = atomic64_read(&fs_info->free_chunk_space);
359 * If we have dup, raid1 or raid10 then only half of the free
360 * space is actually usable. For raid56, the space info used
361 * doesn't include the parity drive, so we don't have to
364 factor = btrfs_bg_type_to_factor(profile);
365 avail = div_u64(avail, factor);
368 * If we aren't flushing all things, let us overcommit up to
369 * 1/2th of the space. If we can flush, don't let us overcommit
370 * too much, let it overcommit up to 1/8 of the space.
372 if (flush == BTRFS_RESERVE_FLUSH_ALL)
379 static inline u64 writable_total_bytes(struct btrfs_fs_info *fs_info,
380 struct btrfs_space_info *space_info)
383 * On regular filesystem, all total_bytes are always writable. On zoned
384 * filesystem, there may be a limitation imposed by max_active_zones.
385 * For metadata allocation, we cannot finish an existing active block
386 * group to avoid a deadlock. Thus, we need to consider only the active
387 * groups to be writable for metadata space.
389 if (!btrfs_is_zoned(fs_info) || (space_info->flags & BTRFS_BLOCK_GROUP_DATA))
390 return space_info->total_bytes;
392 return space_info->active_total_bytes;
395 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
396 struct btrfs_space_info *space_info, u64 bytes,
397 enum btrfs_reserve_flush_enum flush)
402 /* Don't overcommit when in mixed mode */
403 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
406 used = btrfs_space_info_used(space_info, true);
407 if (test_bit(BTRFS_FS_ACTIVE_ZONE_TRACKING, &fs_info->flags) &&
408 (space_info->flags & BTRFS_BLOCK_GROUP_METADATA))
411 avail = calc_available_free_space(fs_info, space_info, flush);
413 if (used + bytes < writable_total_bytes(fs_info, space_info) + avail)
418 static void remove_ticket(struct btrfs_space_info *space_info,
419 struct reserve_ticket *ticket)
421 if (!list_empty(&ticket->list)) {
422 list_del_init(&ticket->list);
423 ASSERT(space_info->reclaim_size >= ticket->bytes);
424 space_info->reclaim_size -= ticket->bytes;
429 * This is for space we already have accounted in space_info->bytes_may_use, so
430 * basically when we're returning space from block_rsv's.
432 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
433 struct btrfs_space_info *space_info)
435 struct list_head *head;
436 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
438 lockdep_assert_held(&space_info->lock);
440 head = &space_info->priority_tickets;
442 while (!list_empty(head)) {
443 struct reserve_ticket *ticket;
444 u64 used = btrfs_space_info_used(space_info, true);
446 ticket = list_first_entry(head, struct reserve_ticket, list);
448 /* Check and see if our ticket can be satisfied now. */
449 if ((used + ticket->bytes <= writable_total_bytes(fs_info, space_info)) ||
450 btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
452 btrfs_space_info_update_bytes_may_use(fs_info,
455 remove_ticket(space_info, ticket);
457 space_info->tickets_id++;
458 wake_up(&ticket->wait);
464 if (head == &space_info->priority_tickets) {
465 head = &space_info->tickets;
466 flush = BTRFS_RESERVE_FLUSH_ALL;
471 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \
473 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \
474 spin_lock(&__rsv->lock); \
475 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \
476 __rsv->size, __rsv->reserved); \
477 spin_unlock(&__rsv->lock); \
480 static const char *space_info_flag_to_str(const struct btrfs_space_info *space_info)
482 switch (space_info->flags) {
483 case BTRFS_BLOCK_GROUP_SYSTEM:
485 case BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA:
486 return "DATA+METADATA";
487 case BTRFS_BLOCK_GROUP_DATA:
489 case BTRFS_BLOCK_GROUP_METADATA:
496 static void dump_global_block_rsv(struct btrfs_fs_info *fs_info)
498 DUMP_BLOCK_RSV(fs_info, global_block_rsv);
499 DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
500 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
501 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
502 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
505 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
506 struct btrfs_space_info *info)
508 const char *flag_str = space_info_flag_to_str(info);
509 lockdep_assert_held(&info->lock);
511 /* The free space could be negative in case of overcommit */
512 btrfs_info(fs_info, "space_info %s has %lld free, is %sfull",
514 (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
515 info->full ? "" : "not ");
517 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
518 info->total_bytes, info->bytes_used, info->bytes_pinned,
519 info->bytes_reserved, info->bytes_may_use,
520 info->bytes_readonly, info->bytes_zone_unusable);
523 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
524 struct btrfs_space_info *info, u64 bytes,
525 int dump_block_groups)
527 struct btrfs_block_group *cache;
530 spin_lock(&info->lock);
531 __btrfs_dump_space_info(fs_info, info);
532 dump_global_block_rsv(fs_info);
533 spin_unlock(&info->lock);
535 if (!dump_block_groups)
538 down_read(&info->groups_sem);
540 list_for_each_entry(cache, &info->block_groups[index], list) {
541 spin_lock(&cache->lock);
543 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu zone_unusable %s",
544 cache->start, cache->length, cache->used, cache->pinned,
545 cache->reserved, cache->zone_unusable,
546 cache->ro ? "[readonly]" : "");
547 spin_unlock(&cache->lock);
548 btrfs_dump_free_space(cache, bytes);
550 if (++index < BTRFS_NR_RAID_TYPES)
552 up_read(&info->groups_sem);
555 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
561 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
562 nr = div64_u64(to_reclaim, bytes);
568 #define EXTENT_SIZE_PER_ITEM SZ_256K
571 * shrink metadata reservation for delalloc
573 static void shrink_delalloc(struct btrfs_fs_info *fs_info,
574 struct btrfs_space_info *space_info,
575 u64 to_reclaim, bool wait_ordered,
578 struct btrfs_trans_handle *trans;
585 delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
586 ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes);
587 if (delalloc_bytes == 0 && ordered_bytes == 0)
590 /* Calc the number of the pages we need flush for space reservation */
591 if (to_reclaim == U64_MAX) {
595 * to_reclaim is set to however much metadata we need to
596 * reclaim, but reclaiming that much data doesn't really track
597 * exactly. What we really want to do is reclaim full inode's
598 * worth of reservations, however that's not available to us
599 * here. We will take a fraction of the delalloc bytes for our
600 * flushing loops and hope for the best. Delalloc will expand
601 * the amount we write to cover an entire dirty extent, which
602 * will reclaim the metadata reservation for that range. If
603 * it's not enough subsequent flush stages will be more
606 to_reclaim = max(to_reclaim, delalloc_bytes >> 3);
607 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
610 trans = current->journal_info;
613 * If we are doing more ordered than delalloc we need to just wait on
614 * ordered extents, otherwise we'll waste time trying to flush delalloc
615 * that likely won't give us the space back we need.
617 if (ordered_bytes > delalloc_bytes && !for_preempt)
621 while ((delalloc_bytes || ordered_bytes) && loops < 3) {
622 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
623 long nr_pages = min_t(u64, temp, LONG_MAX);
626 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
629 * We need to make sure any outstanding async pages are now
630 * processed before we continue. This is because things like
631 * sync_inode() try to be smart and skip writing if the inode is
632 * marked clean. We don't use filemap_fwrite for flushing
633 * because we want to control how many pages we write out at a
634 * time, thus this is the only safe way to make sure we've
635 * waited for outstanding compressed workers to have started
636 * their jobs and thus have ordered extents set up properly.
638 * This exists because we do not want to wait for each
639 * individual inode to finish its async work, we simply want to
640 * start the IO on everybody, and then come back here and wait
641 * for all of the async work to catch up. Once we're done with
642 * that we know we'll have ordered extents for everything and we
643 * can decide if we wait for that or not.
645 * If we choose to replace this in the future, make absolutely
646 * sure that the proper waiting is being done in the async case,
647 * as there have been bugs in that area before.
649 async_pages = atomic_read(&fs_info->async_delalloc_pages);
654 * We don't want to wait forever, if we wrote less pages in this
655 * loop than we have outstanding, only wait for that number of
656 * pages, otherwise we can wait for all async pages to finish
659 if (async_pages > nr_pages)
660 async_pages -= nr_pages;
663 wait_event(fs_info->async_submit_wait,
664 atomic_read(&fs_info->async_delalloc_pages) <=
668 if (wait_ordered && !trans) {
669 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
671 time_left = schedule_timeout_killable(1);
677 * If we are for preemption we just want a one-shot of delalloc
678 * flushing so we can stop flushing if we decide we don't need
684 spin_lock(&space_info->lock);
685 if (list_empty(&space_info->tickets) &&
686 list_empty(&space_info->priority_tickets)) {
687 spin_unlock(&space_info->lock);
690 spin_unlock(&space_info->lock);
692 delalloc_bytes = percpu_counter_sum_positive(
693 &fs_info->delalloc_bytes);
694 ordered_bytes = percpu_counter_sum_positive(
695 &fs_info->ordered_bytes);
700 * Try to flush some data based on policy set by @state. This is only advisory
701 * and may fail for various reasons. The caller is supposed to examine the
702 * state of @space_info to detect the outcome.
704 static void flush_space(struct btrfs_fs_info *fs_info,
705 struct btrfs_space_info *space_info, u64 num_bytes,
706 enum btrfs_flush_state state, bool for_preempt)
708 struct btrfs_root *root = fs_info->tree_root;
709 struct btrfs_trans_handle *trans;
714 case FLUSH_DELAYED_ITEMS_NR:
715 case FLUSH_DELAYED_ITEMS:
716 if (state == FLUSH_DELAYED_ITEMS_NR)
717 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
721 trans = btrfs_join_transaction(root);
723 ret = PTR_ERR(trans);
726 ret = btrfs_run_delayed_items_nr(trans, nr);
727 btrfs_end_transaction(trans);
730 case FLUSH_DELALLOC_WAIT:
731 case FLUSH_DELALLOC_FULL:
732 if (state == FLUSH_DELALLOC_FULL)
734 shrink_delalloc(fs_info, space_info, num_bytes,
735 state != FLUSH_DELALLOC, for_preempt);
737 case FLUSH_DELAYED_REFS_NR:
738 case FLUSH_DELAYED_REFS:
739 trans = btrfs_join_transaction(root);
741 ret = PTR_ERR(trans);
744 if (state == FLUSH_DELAYED_REFS_NR)
745 nr = calc_reclaim_items_nr(fs_info, num_bytes);
748 btrfs_run_delayed_refs(trans, nr);
749 btrfs_end_transaction(trans);
752 case ALLOC_CHUNK_FORCE:
754 * For metadata space on zoned filesystem, reaching here means we
755 * don't have enough space left in active_total_bytes. Try to
756 * activate a block group first, because we may have inactive
757 * block group already allocated.
759 ret = btrfs_zoned_activate_one_bg(fs_info, space_info, false);
765 trans = btrfs_join_transaction(root);
767 ret = PTR_ERR(trans);
770 ret = btrfs_chunk_alloc(trans,
771 btrfs_get_alloc_profile(fs_info, space_info->flags),
772 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
774 btrfs_end_transaction(trans);
777 * For metadata space on zoned filesystem, allocating a new chunk
778 * is not enough. We still need to activate the block * group.
779 * Active the newly allocated block group by (maybe) finishing
783 ret = btrfs_zoned_activate_one_bg(fs_info, space_info, true);
785 * Revert to the original ret regardless we could finish
786 * one block group or not.
792 if (ret > 0 || ret == -ENOSPC)
795 case RUN_DELAYED_IPUTS:
797 * If we have pending delayed iputs then we could free up a
798 * bunch of pinned space, so make sure we run the iputs before
799 * we do our pinned bytes check below.
801 btrfs_run_delayed_iputs(fs_info);
802 btrfs_wait_on_delayed_iputs(fs_info);
805 ASSERT(current->journal_info == NULL);
806 trans = btrfs_join_transaction(root);
808 ret = PTR_ERR(trans);
811 ret = btrfs_commit_transaction(trans);
818 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
824 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
825 struct btrfs_space_info *space_info)
830 u64 to_reclaim = space_info->reclaim_size;
832 lockdep_assert_held(&space_info->lock);
834 avail = calc_available_free_space(fs_info, space_info,
835 BTRFS_RESERVE_FLUSH_ALL);
836 used = btrfs_space_info_used(space_info, true);
839 * We may be flushing because suddenly we have less space than we had
840 * before, and now we're well over-committed based on our current free
841 * space. If that's the case add in our overage so we make sure to put
842 * appropriate pressure on the flushing state machine.
844 total = writable_total_bytes(fs_info, space_info);
845 if (total + avail < used)
846 to_reclaim += used - (total + avail);
851 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
852 struct btrfs_space_info *space_info)
854 u64 global_rsv_size = fs_info->global_block_rsv.reserved;
855 u64 ordered, delalloc;
856 u64 total = writable_total_bytes(fs_info, space_info);
860 thresh = div_factor_fine(total, 90);
862 lockdep_assert_held(&space_info->lock);
864 /* If we're just plain full then async reclaim just slows us down. */
865 if ((space_info->bytes_used + space_info->bytes_reserved +
866 global_rsv_size) >= thresh)
869 used = space_info->bytes_may_use + space_info->bytes_pinned;
871 /* The total flushable belongs to the global rsv, don't flush. */
872 if (global_rsv_size >= used)
876 * 128MiB is 1/4 of the maximum global rsv size. If we have less than
877 * that devoted to other reservations then there's no sense in flushing,
878 * we don't have a lot of things that need flushing.
880 if (used - global_rsv_size <= SZ_128M)
884 * We have tickets queued, bail so we don't compete with the async
887 if (space_info->reclaim_size)
891 * If we have over half of the free space occupied by reservations or
892 * pinned then we want to start flushing.
894 * We do not do the traditional thing here, which is to say
896 * if (used >= ((total_bytes + avail) / 2))
899 * because this doesn't quite work how we want. If we had more than 50%
900 * of the space_info used by bytes_used and we had 0 available we'd just
901 * constantly run the background flusher. Instead we want it to kick in
902 * if our reclaimable space exceeds our clamped free space.
904 * Our clamping range is 2^1 -> 2^8. Practically speaking that means
907 * Amount of RAM Minimum threshold Maximum threshold
910 * 128GiB 512MiB 64GiB
915 * These are the range our thresholds will fall in, corresponding to how
916 * much delalloc we need for the background flusher to kick in.
919 thresh = calc_available_free_space(fs_info, space_info,
920 BTRFS_RESERVE_FLUSH_ALL);
921 used = space_info->bytes_used + space_info->bytes_reserved +
922 space_info->bytes_readonly + global_rsv_size;
924 thresh += total - used;
925 thresh >>= space_info->clamp;
927 used = space_info->bytes_pinned;
930 * If we have more ordered bytes than delalloc bytes then we're either
931 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting
932 * around. Preemptive flushing is only useful in that it can free up
933 * space before tickets need to wait for things to finish. In the case
934 * of ordered extents, preemptively waiting on ordered extents gets us
935 * nothing, if our reservations are tied up in ordered extents we'll
936 * simply have to slow down writers by forcing them to wait on ordered
939 * In the case that ordered is larger than delalloc, only include the
940 * block reserves that we would actually be able to directly reclaim
941 * from. In this case if we're heavy on metadata operations this will
942 * clearly be heavy enough to warrant preemptive flushing. In the case
943 * of heavy DIO or ordered reservations, preemptive flushing will just
944 * waste time and cause us to slow down.
946 * We want to make sure we truly are maxed out on ordered however, so
947 * cut ordered in half, and if it's still higher than delalloc then we
948 * can keep flushing. This is to avoid the case where we start
949 * flushing, and now delalloc == ordered and we stop preemptively
950 * flushing when we could still have several gigs of delalloc to flush.
952 ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1;
953 delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes);
954 if (ordered >= delalloc)
955 used += fs_info->delayed_refs_rsv.reserved +
956 fs_info->delayed_block_rsv.reserved;
958 used += space_info->bytes_may_use - global_rsv_size;
960 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
961 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
964 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
965 struct btrfs_space_info *space_info,
966 struct reserve_ticket *ticket)
968 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
974 if (global_rsv->space_info != space_info)
977 spin_lock(&global_rsv->lock);
978 min_bytes = div_factor(global_rsv->size, 1);
979 if (global_rsv->reserved < min_bytes + ticket->bytes) {
980 spin_unlock(&global_rsv->lock);
983 global_rsv->reserved -= ticket->bytes;
984 remove_ticket(space_info, ticket);
986 wake_up(&ticket->wait);
987 space_info->tickets_id++;
988 if (global_rsv->reserved < global_rsv->size)
989 global_rsv->full = 0;
990 spin_unlock(&global_rsv->lock);
996 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
997 * @fs_info - fs_info for this fs
998 * @space_info - the space info we were flushing
1000 * We call this when we've exhausted our flushing ability and haven't made
1001 * progress in satisfying tickets. The reservation code handles tickets in
1002 * order, so if there is a large ticket first and then smaller ones we could
1003 * very well satisfy the smaller tickets. This will attempt to wake up any
1004 * tickets in the list to catch this case.
1006 * This function returns true if it was able to make progress by clearing out
1007 * other tickets, or if it stumbles across a ticket that was smaller than the
1010 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
1011 struct btrfs_space_info *space_info)
1013 struct reserve_ticket *ticket;
1014 u64 tickets_id = space_info->tickets_id;
1015 const bool aborted = BTRFS_FS_ERROR(fs_info);
1017 trace_btrfs_fail_all_tickets(fs_info, space_info);
1019 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1020 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
1021 __btrfs_dump_space_info(fs_info, space_info);
1024 while (!list_empty(&space_info->tickets) &&
1025 tickets_id == space_info->tickets_id) {
1026 ticket = list_first_entry(&space_info->tickets,
1027 struct reserve_ticket, list);
1029 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket))
1032 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1033 btrfs_info(fs_info, "failing ticket with %llu bytes",
1036 remove_ticket(space_info, ticket);
1038 ticket->error = -EIO;
1040 ticket->error = -ENOSPC;
1041 wake_up(&ticket->wait);
1044 * We're just throwing tickets away, so more flushing may not
1045 * trip over btrfs_try_granting_tickets, so we need to call it
1046 * here to see if we can make progress with the next ticket in
1050 btrfs_try_granting_tickets(fs_info, space_info);
1052 return (tickets_id != space_info->tickets_id);
1056 * This is for normal flushers, we can wait all goddamned day if we want to. We
1057 * will loop and continuously try to flush as long as we are making progress.
1058 * We count progress as clearing off tickets each time we have to loop.
1060 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
1062 struct btrfs_fs_info *fs_info;
1063 struct btrfs_space_info *space_info;
1065 enum btrfs_flush_state flush_state;
1066 int commit_cycles = 0;
1067 u64 last_tickets_id;
1069 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
1070 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1072 spin_lock(&space_info->lock);
1073 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1075 space_info->flush = 0;
1076 spin_unlock(&space_info->lock);
1079 last_tickets_id = space_info->tickets_id;
1080 spin_unlock(&space_info->lock);
1082 flush_state = FLUSH_DELAYED_ITEMS_NR;
1084 flush_space(fs_info, space_info, to_reclaim, flush_state, false);
1085 spin_lock(&space_info->lock);
1086 if (list_empty(&space_info->tickets)) {
1087 space_info->flush = 0;
1088 spin_unlock(&space_info->lock);
1091 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
1093 if (last_tickets_id == space_info->tickets_id) {
1096 last_tickets_id = space_info->tickets_id;
1097 flush_state = FLUSH_DELAYED_ITEMS_NR;
1103 * We do not want to empty the system of delalloc unless we're
1104 * under heavy pressure, so allow one trip through the flushing
1105 * logic before we start doing a FLUSH_DELALLOC_FULL.
1107 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles)
1111 * We don't want to force a chunk allocation until we've tried
1112 * pretty hard to reclaim space. Think of the case where we
1113 * freed up a bunch of space and so have a lot of pinned space
1114 * to reclaim. We would rather use that than possibly create a
1115 * underutilized metadata chunk. So if this is our first run
1116 * through the flushing state machine skip ALLOC_CHUNK_FORCE and
1117 * commit the transaction. If nothing has changed the next go
1118 * around then we can force a chunk allocation.
1120 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
1123 if (flush_state > COMMIT_TRANS) {
1125 if (commit_cycles > 2) {
1126 if (maybe_fail_all_tickets(fs_info, space_info)) {
1127 flush_state = FLUSH_DELAYED_ITEMS_NR;
1130 space_info->flush = 0;
1133 flush_state = FLUSH_DELAYED_ITEMS_NR;
1136 spin_unlock(&space_info->lock);
1137 } while (flush_state <= COMMIT_TRANS);
1141 * This handles pre-flushing of metadata space before we get to the point that
1142 * we need to start blocking threads on tickets. The logic here is different
1143 * from the other flush paths because it doesn't rely on tickets to tell us how
1144 * much we need to flush, instead it attempts to keep us below the 80% full
1145 * watermark of space by flushing whichever reservation pool is currently the
1148 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
1150 struct btrfs_fs_info *fs_info;
1151 struct btrfs_space_info *space_info;
1152 struct btrfs_block_rsv *delayed_block_rsv;
1153 struct btrfs_block_rsv *delayed_refs_rsv;
1154 struct btrfs_block_rsv *global_rsv;
1155 struct btrfs_block_rsv *trans_rsv;
1158 fs_info = container_of(work, struct btrfs_fs_info,
1159 preempt_reclaim_work);
1160 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1161 delayed_block_rsv = &fs_info->delayed_block_rsv;
1162 delayed_refs_rsv = &fs_info->delayed_refs_rsv;
1163 global_rsv = &fs_info->global_block_rsv;
1164 trans_rsv = &fs_info->trans_block_rsv;
1166 spin_lock(&space_info->lock);
1167 while (need_preemptive_reclaim(fs_info, space_info)) {
1168 enum btrfs_flush_state flush;
1169 u64 delalloc_size = 0;
1170 u64 to_reclaim, block_rsv_size;
1171 u64 global_rsv_size = global_rsv->reserved;
1176 * We don't have a precise counter for the metadata being
1177 * reserved for delalloc, so we'll approximate it by subtracting
1178 * out the block rsv's space from the bytes_may_use. If that
1179 * amount is higher than the individual reserves, then we can
1180 * assume it's tied up in delalloc reservations.
1182 block_rsv_size = global_rsv_size +
1183 delayed_block_rsv->reserved +
1184 delayed_refs_rsv->reserved +
1185 trans_rsv->reserved;
1186 if (block_rsv_size < space_info->bytes_may_use)
1187 delalloc_size = space_info->bytes_may_use - block_rsv_size;
1190 * We don't want to include the global_rsv in our calculation,
1191 * because that's space we can't touch. Subtract it from the
1192 * block_rsv_size for the next checks.
1194 block_rsv_size -= global_rsv_size;
1197 * We really want to avoid flushing delalloc too much, as it
1198 * could result in poor allocation patterns, so only flush it if
1199 * it's larger than the rest of the pools combined.
1201 if (delalloc_size > block_rsv_size) {
1202 to_reclaim = delalloc_size;
1203 flush = FLUSH_DELALLOC;
1204 } else if (space_info->bytes_pinned >
1205 (delayed_block_rsv->reserved +
1206 delayed_refs_rsv->reserved)) {
1207 to_reclaim = space_info->bytes_pinned;
1208 flush = COMMIT_TRANS;
1209 } else if (delayed_block_rsv->reserved >
1210 delayed_refs_rsv->reserved) {
1211 to_reclaim = delayed_block_rsv->reserved;
1212 flush = FLUSH_DELAYED_ITEMS_NR;
1214 to_reclaim = delayed_refs_rsv->reserved;
1215 flush = FLUSH_DELAYED_REFS_NR;
1218 spin_unlock(&space_info->lock);
1221 * We don't want to reclaim everything, just a portion, so scale
1222 * down the to_reclaim by 1/4. If it takes us down to 0,
1223 * reclaim 1 items worth.
1227 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1228 flush_space(fs_info, space_info, to_reclaim, flush, true);
1230 spin_lock(&space_info->lock);
1233 /* We only went through once, back off our clamping. */
1234 if (loops == 1 && !space_info->reclaim_size)
1235 space_info->clamp = max(1, space_info->clamp - 1);
1236 trace_btrfs_done_preemptive_reclaim(fs_info, space_info);
1237 spin_unlock(&space_info->lock);
1241 * FLUSH_DELALLOC_WAIT:
1242 * Space is freed from flushing delalloc in one of two ways.
1244 * 1) compression is on and we allocate less space than we reserved
1245 * 2) we are overwriting existing space
1247 * For #1 that extra space is reclaimed as soon as the delalloc pages are
1248 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1249 * length to ->bytes_reserved, and subtracts the reserved space from
1252 * For #2 this is trickier. Once the ordered extent runs we will drop the
1253 * extent in the range we are overwriting, which creates a delayed ref for
1254 * that freed extent. This however is not reclaimed until the transaction
1255 * commits, thus the next stages.
1258 * If we are freeing inodes, we want to make sure all delayed iputs have
1259 * completed, because they could have been on an inode with i_nlink == 0, and
1260 * thus have been truncated and freed up space. But again this space is not
1261 * immediately re-usable, it comes in the form of a delayed ref, which must be
1262 * run and then the transaction must be committed.
1265 * This is where we reclaim all of the pinned space generated by running the
1269 * For data we start with alloc chunk force, however we could have been full
1270 * before, and then the transaction commit could have freed new block groups,
1271 * so if we now have space to allocate do the force chunk allocation.
1273 static const enum btrfs_flush_state data_flush_states[] = {
1274 FLUSH_DELALLOC_FULL,
1280 static void btrfs_async_reclaim_data_space(struct work_struct *work)
1282 struct btrfs_fs_info *fs_info;
1283 struct btrfs_space_info *space_info;
1284 u64 last_tickets_id;
1285 enum btrfs_flush_state flush_state = 0;
1287 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1288 space_info = fs_info->data_sinfo;
1290 spin_lock(&space_info->lock);
1291 if (list_empty(&space_info->tickets)) {
1292 space_info->flush = 0;
1293 spin_unlock(&space_info->lock);
1296 last_tickets_id = space_info->tickets_id;
1297 spin_unlock(&space_info->lock);
1299 while (!space_info->full) {
1300 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1301 spin_lock(&space_info->lock);
1302 if (list_empty(&space_info->tickets)) {
1303 space_info->flush = 0;
1304 spin_unlock(&space_info->lock);
1308 /* Something happened, fail everything and bail. */
1309 if (BTRFS_FS_ERROR(fs_info))
1311 last_tickets_id = space_info->tickets_id;
1312 spin_unlock(&space_info->lock);
1315 while (flush_state < ARRAY_SIZE(data_flush_states)) {
1316 flush_space(fs_info, space_info, U64_MAX,
1317 data_flush_states[flush_state], false);
1318 spin_lock(&space_info->lock);
1319 if (list_empty(&space_info->tickets)) {
1320 space_info->flush = 0;
1321 spin_unlock(&space_info->lock);
1325 if (last_tickets_id == space_info->tickets_id) {
1328 last_tickets_id = space_info->tickets_id;
1332 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1333 if (space_info->full) {
1334 if (maybe_fail_all_tickets(fs_info, space_info))
1337 space_info->flush = 0;
1342 /* Something happened, fail everything and bail. */
1343 if (BTRFS_FS_ERROR(fs_info))
1347 spin_unlock(&space_info->lock);
1352 maybe_fail_all_tickets(fs_info, space_info);
1353 space_info->flush = 0;
1354 spin_unlock(&space_info->lock);
1357 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1359 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1360 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1361 INIT_WORK(&fs_info->preempt_reclaim_work,
1362 btrfs_preempt_reclaim_metadata_space);
1365 static const enum btrfs_flush_state priority_flush_states[] = {
1366 FLUSH_DELAYED_ITEMS_NR,
1367 FLUSH_DELAYED_ITEMS,
1371 static const enum btrfs_flush_state evict_flush_states[] = {
1372 FLUSH_DELAYED_ITEMS_NR,
1373 FLUSH_DELAYED_ITEMS,
1374 FLUSH_DELAYED_REFS_NR,
1377 FLUSH_DELALLOC_WAIT,
1378 FLUSH_DELALLOC_FULL,
1383 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1384 struct btrfs_space_info *space_info,
1385 struct reserve_ticket *ticket,
1386 const enum btrfs_flush_state *states,
1390 int flush_state = 0;
1392 spin_lock(&space_info->lock);
1393 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1395 * This is the priority reclaim path, so to_reclaim could be >0 still
1396 * because we may have only satisfied the priority tickets and still
1397 * left non priority tickets on the list. We would then have
1398 * to_reclaim but ->bytes == 0.
1400 if (ticket->bytes == 0) {
1401 spin_unlock(&space_info->lock);
1405 while (flush_state < states_nr) {
1406 spin_unlock(&space_info->lock);
1407 flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1410 spin_lock(&space_info->lock);
1411 if (ticket->bytes == 0) {
1412 spin_unlock(&space_info->lock);
1417 /* Attempt to steal from the global rsv if we can. */
1418 if (!steal_from_global_rsv(fs_info, space_info, ticket)) {
1419 ticket->error = -ENOSPC;
1420 remove_ticket(space_info, ticket);
1424 * We must run try_granting_tickets here because we could be a large
1425 * ticket in front of a smaller ticket that can now be satisfied with
1426 * the available space.
1428 btrfs_try_granting_tickets(fs_info, space_info);
1429 spin_unlock(&space_info->lock);
1432 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1433 struct btrfs_space_info *space_info,
1434 struct reserve_ticket *ticket)
1436 spin_lock(&space_info->lock);
1438 /* We could have been granted before we got here. */
1439 if (ticket->bytes == 0) {
1440 spin_unlock(&space_info->lock);
1444 while (!space_info->full) {
1445 spin_unlock(&space_info->lock);
1446 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1447 spin_lock(&space_info->lock);
1448 if (ticket->bytes == 0) {
1449 spin_unlock(&space_info->lock);
1454 ticket->error = -ENOSPC;
1455 remove_ticket(space_info, ticket);
1456 btrfs_try_granting_tickets(fs_info, space_info);
1457 spin_unlock(&space_info->lock);
1460 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1461 struct btrfs_space_info *space_info,
1462 struct reserve_ticket *ticket)
1468 spin_lock(&space_info->lock);
1469 while (ticket->bytes > 0 && ticket->error == 0) {
1470 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1473 * Delete us from the list. After we unlock the space
1474 * info, we don't want the async reclaim job to reserve
1475 * space for this ticket. If that would happen, then the
1476 * ticket's task would not known that space was reserved
1477 * despite getting an error, resulting in a space leak
1478 * (bytes_may_use counter of our space_info).
1480 remove_ticket(space_info, ticket);
1481 ticket->error = -EINTR;
1484 spin_unlock(&space_info->lock);
1488 finish_wait(&ticket->wait, &wait);
1489 spin_lock(&space_info->lock);
1491 spin_unlock(&space_info->lock);
1495 * Do the appropriate flushing and waiting for a ticket
1497 * @fs_info: the filesystem
1498 * @space_info: space info for the reservation
1499 * @ticket: ticket for the reservation
1500 * @start_ns: timestamp when the reservation started
1501 * @orig_bytes: amount of bytes originally reserved
1502 * @flush: how much we can flush
1504 * This does the work of figuring out how to flush for the ticket, waiting for
1505 * the reservation, and returning the appropriate error if there is one.
1507 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1508 struct btrfs_space_info *space_info,
1509 struct reserve_ticket *ticket,
1510 u64 start_ns, u64 orig_bytes,
1511 enum btrfs_reserve_flush_enum flush)
1516 case BTRFS_RESERVE_FLUSH_DATA:
1517 case BTRFS_RESERVE_FLUSH_ALL:
1518 case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1519 wait_reserve_ticket(fs_info, space_info, ticket);
1521 case BTRFS_RESERVE_FLUSH_LIMIT:
1522 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1523 priority_flush_states,
1524 ARRAY_SIZE(priority_flush_states));
1526 case BTRFS_RESERVE_FLUSH_EVICT:
1527 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1529 ARRAY_SIZE(evict_flush_states));
1531 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1532 priority_reclaim_data_space(fs_info, space_info, ticket);
1539 ret = ticket->error;
1540 ASSERT(list_empty(&ticket->list));
1542 * Check that we can't have an error set if the reservation succeeded,
1543 * as that would confuse tasks and lead them to error out without
1544 * releasing reserved space (if an error happens the expectation is that
1545 * space wasn't reserved at all).
1547 ASSERT(!(ticket->bytes == 0 && ticket->error));
1548 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes,
1549 start_ns, flush, ticket->error);
1554 * This returns true if this flush state will go through the ordinary flushing
1557 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1559 return (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1560 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1563 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1564 struct btrfs_space_info *space_info)
1566 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1567 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1570 * If we're heavy on ordered operations then clamping won't help us. We
1571 * need to clamp specifically to keep up with dirty'ing buffered
1572 * writers, because there's not a 1:1 correlation of writing delalloc
1573 * and freeing space, like there is with flushing delayed refs or
1574 * delayed nodes. If we're already more ordered than delalloc then
1575 * we're keeping up, otherwise we aren't and should probably clamp.
1577 if (ordered < delalloc)
1578 space_info->clamp = min(space_info->clamp + 1, 8);
1581 static inline bool can_steal(enum btrfs_reserve_flush_enum flush)
1583 return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1584 flush == BTRFS_RESERVE_FLUSH_EVICT);
1588 * Try to reserve bytes from the block_rsv's space
1590 * @fs_info: the filesystem
1591 * @space_info: space info we want to allocate from
1592 * @orig_bytes: number of bytes we want
1593 * @flush: whether or not we can flush to make our reservation
1595 * This will reserve orig_bytes number of bytes from the space info associated
1596 * with the block_rsv. If there is not enough space it will make an attempt to
1597 * flush out space to make room. It will do this by flushing delalloc if
1598 * possible or committing the transaction. If flush is 0 then no attempts to
1599 * regain reservations will be made and this will fail if there is not enough
1602 static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1603 struct btrfs_space_info *space_info, u64 orig_bytes,
1604 enum btrfs_reserve_flush_enum flush)
1606 struct work_struct *async_work;
1607 struct reserve_ticket ticket;
1611 bool pending_tickets;
1614 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1616 if (flush == BTRFS_RESERVE_FLUSH_DATA)
1617 async_work = &fs_info->async_data_reclaim_work;
1619 async_work = &fs_info->async_reclaim_work;
1621 spin_lock(&space_info->lock);
1623 used = btrfs_space_info_used(space_info, true);
1626 * We don't want NO_FLUSH allocations to jump everybody, they can
1627 * generally handle ENOSPC in a different way, so treat them the same as
1628 * normal flushers when it comes to skipping pending tickets.
1630 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1631 pending_tickets = !list_empty(&space_info->tickets) ||
1632 !list_empty(&space_info->priority_tickets);
1634 pending_tickets = !list_empty(&space_info->priority_tickets);
1637 * Carry on if we have enough space (short-circuit) OR call
1638 * can_overcommit() to ensure we can overcommit to continue.
1640 if (!pending_tickets &&
1641 ((used + orig_bytes <= writable_total_bytes(fs_info, space_info)) ||
1642 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1643 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1649 * If we couldn't make a reservation then setup our reservation ticket
1650 * and kick the async worker if it's not already running.
1652 * If we are a priority flusher then we just need to add our ticket to
1653 * the list and we will do our own flushing further down.
1655 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
1656 ticket.bytes = orig_bytes;
1658 space_info->reclaim_size += ticket.bytes;
1659 init_waitqueue_head(&ticket.wait);
1660 ticket.steal = can_steal(flush);
1661 if (trace_btrfs_reserve_ticket_enabled())
1662 start_ns = ktime_get_ns();
1664 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1665 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1666 flush == BTRFS_RESERVE_FLUSH_DATA) {
1667 list_add_tail(&ticket.list, &space_info->tickets);
1668 if (!space_info->flush) {
1670 * We were forced to add a reserve ticket, so
1671 * our preemptive flushing is unable to keep
1672 * up. Clamp down on the threshold for the
1673 * preemptive flushing in order to keep up with
1676 maybe_clamp_preempt(fs_info, space_info);
1678 space_info->flush = 1;
1679 trace_btrfs_trigger_flush(fs_info,
1683 queue_work(system_unbound_wq, async_work);
1686 list_add_tail(&ticket.list,
1687 &space_info->priority_tickets);
1689 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1691 * We will do the space reservation dance during log replay,
1692 * which means we won't have fs_info->fs_root set, so don't do
1693 * the async reclaim as we will panic.
1695 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1696 !work_busy(&fs_info->preempt_reclaim_work) &&
1697 need_preemptive_reclaim(fs_info, space_info)) {
1698 trace_btrfs_trigger_flush(fs_info, space_info->flags,
1699 orig_bytes, flush, "preempt");
1700 queue_work(system_unbound_wq,
1701 &fs_info->preempt_reclaim_work);
1704 spin_unlock(&space_info->lock);
1705 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
1708 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1713 * Trye to reserve metadata bytes from the block_rsv's space
1715 * @fs_info: the filesystem
1716 * @block_rsv: block_rsv we're allocating for
1717 * @orig_bytes: number of bytes we want
1718 * @flush: whether or not we can flush to make our reservation
1720 * This will reserve orig_bytes number of bytes from the space info associated
1721 * with the block_rsv. If there is not enough space it will make an attempt to
1722 * flush out space to make room. It will do this by flushing delalloc if
1723 * possible or committing the transaction. If flush is 0 then no attempts to
1724 * regain reservations will be made and this will fail if there is not enough
1727 int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1728 struct btrfs_block_rsv *block_rsv,
1730 enum btrfs_reserve_flush_enum flush)
1734 ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush);
1735 if (ret == -ENOSPC) {
1736 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1737 block_rsv->space_info->flags,
1740 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1741 btrfs_dump_space_info(fs_info, block_rsv->space_info,
1748 * Try to reserve data bytes for an allocation
1750 * @fs_info: the filesystem
1751 * @bytes: number of bytes we need
1752 * @flush: how we are allowed to flush
1754 * This will reserve bytes from the data space info. If there is not enough
1755 * space then we will attempt to flush space as specified by flush.
1757 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1758 enum btrfs_reserve_flush_enum flush)
1760 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1763 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1764 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE ||
1765 flush == BTRFS_RESERVE_NO_FLUSH);
1766 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1768 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1769 if (ret == -ENOSPC) {
1770 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1771 data_sinfo->flags, bytes, 1);
1772 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1773 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1778 /* Dump all the space infos when we abort a transaction due to ENOSPC. */
1779 __cold void btrfs_dump_space_info_for_trans_abort(struct btrfs_fs_info *fs_info)
1781 struct btrfs_space_info *space_info;
1783 btrfs_info(fs_info, "dumping space info:");
1784 list_for_each_entry(space_info, &fs_info->space_info, list) {
1785 spin_lock(&space_info->lock);
1786 __btrfs_dump_space_info(fs_info, space_info);
1787 spin_unlock(&space_info->lock);
1789 dump_global_block_rsv(fs_info);