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)
348 struct btrfs_space_info *data_sinfo;
354 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
355 profile = btrfs_system_alloc_profile(fs_info);
357 profile = btrfs_metadata_alloc_profile(fs_info);
359 avail = atomic64_read(&fs_info->free_chunk_space);
362 * If we have dup, raid1 or raid10 then only half of the free
363 * space is actually usable. For raid56, the space info used
364 * doesn't include the parity drive, so we don't have to
367 factor = btrfs_bg_type_to_factor(profile);
368 avail = div_u64(avail, factor);
373 * Calculate the data_chunk_size, space_info->chunk_size is the
374 * "optimal" chunk size based on the fs size. However when we actually
375 * allocate the chunk we will strip this down further, making it no more
376 * than 10% of the disk or 1G, whichever is smaller.
378 data_sinfo = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
379 data_chunk_size = min(data_sinfo->chunk_size,
380 mult_perc(fs_info->fs_devices->total_rw_bytes, 10));
381 data_chunk_size = min_t(u64, data_chunk_size, SZ_1G);
384 * Since data allocations immediately use block groups as part of the
385 * reservation, because we assume that data reservations will == actual
386 * usage, we could potentially overcommit and then immediately have that
387 * available space used by a data allocation, which could put us in a
388 * bind when we get close to filling the file system.
390 * To handle this simply remove the data_chunk_size from the available
391 * space. If we are relatively empty this won't affect our ability to
392 * overcommit much, and if we're very close to full it'll keep us from
393 * getting into a position where we've given ourselves very little
394 * metadata wiggle room.
396 if (avail <= data_chunk_size)
398 avail -= data_chunk_size;
401 * If we aren't flushing all things, let us overcommit up to
402 * 1/2th of the space. If we can flush, don't let us overcommit
403 * too much, let it overcommit up to 1/8 of the space.
405 if (flush == BTRFS_RESERVE_FLUSH_ALL)
412 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
413 struct btrfs_space_info *space_info, u64 bytes,
414 enum btrfs_reserve_flush_enum flush)
419 /* Don't overcommit when in mixed mode */
420 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
423 used = btrfs_space_info_used(space_info, true);
424 avail = calc_available_free_space(fs_info, space_info, flush);
426 if (used + bytes < space_info->total_bytes + avail)
431 static void remove_ticket(struct btrfs_space_info *space_info,
432 struct reserve_ticket *ticket)
434 if (!list_empty(&ticket->list)) {
435 list_del_init(&ticket->list);
436 ASSERT(space_info->reclaim_size >= ticket->bytes);
437 space_info->reclaim_size -= ticket->bytes;
442 * This is for space we already have accounted in space_info->bytes_may_use, so
443 * basically when we're returning space from block_rsv's.
445 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
446 struct btrfs_space_info *space_info)
448 struct list_head *head;
449 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
451 lockdep_assert_held(&space_info->lock);
453 head = &space_info->priority_tickets;
455 while (!list_empty(head)) {
456 struct reserve_ticket *ticket;
457 u64 used = btrfs_space_info_used(space_info, true);
459 ticket = list_first_entry(head, struct reserve_ticket, list);
461 /* Check and see if our ticket can be satisfied now. */
462 if ((used + ticket->bytes <= space_info->total_bytes) ||
463 btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
465 btrfs_space_info_update_bytes_may_use(fs_info,
468 remove_ticket(space_info, ticket);
470 space_info->tickets_id++;
471 wake_up(&ticket->wait);
477 if (head == &space_info->priority_tickets) {
478 head = &space_info->tickets;
479 flush = BTRFS_RESERVE_FLUSH_ALL;
484 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \
486 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \
487 spin_lock(&__rsv->lock); \
488 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \
489 __rsv->size, __rsv->reserved); \
490 spin_unlock(&__rsv->lock); \
493 static const char *space_info_flag_to_str(const struct btrfs_space_info *space_info)
495 switch (space_info->flags) {
496 case BTRFS_BLOCK_GROUP_SYSTEM:
498 case BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA:
499 return "DATA+METADATA";
500 case BTRFS_BLOCK_GROUP_DATA:
502 case BTRFS_BLOCK_GROUP_METADATA:
509 static void dump_global_block_rsv(struct btrfs_fs_info *fs_info)
511 DUMP_BLOCK_RSV(fs_info, global_block_rsv);
512 DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
513 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
514 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
515 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
518 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
519 struct btrfs_space_info *info)
521 const char *flag_str = space_info_flag_to_str(info);
522 lockdep_assert_held(&info->lock);
524 /* The free space could be negative in case of overcommit */
525 btrfs_info(fs_info, "space_info %s has %lld free, is %sfull",
527 (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
528 info->full ? "" : "not ");
530 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
531 info->total_bytes, info->bytes_used, info->bytes_pinned,
532 info->bytes_reserved, info->bytes_may_use,
533 info->bytes_readonly, info->bytes_zone_unusable);
536 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
537 struct btrfs_space_info *info, u64 bytes,
538 int dump_block_groups)
540 struct btrfs_block_group *cache;
544 spin_lock(&info->lock);
545 __btrfs_dump_space_info(fs_info, info);
546 dump_global_block_rsv(fs_info);
547 spin_unlock(&info->lock);
549 if (!dump_block_groups)
552 down_read(&info->groups_sem);
554 list_for_each_entry(cache, &info->block_groups[index], list) {
557 spin_lock(&cache->lock);
558 avail = cache->length - cache->used - cache->pinned -
559 cache->reserved - cache->delalloc_bytes -
560 cache->bytes_super - cache->zone_unusable;
562 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu delalloc %llu super %llu zone_unusable (%llu bytes available) %s",
563 cache->start, cache->length, cache->used, cache->pinned,
564 cache->reserved, cache->delalloc_bytes,
565 cache->bytes_super, cache->zone_unusable,
566 avail, cache->ro ? "[readonly]" : "");
567 spin_unlock(&cache->lock);
568 btrfs_dump_free_space(cache, bytes);
569 total_avail += avail;
571 if (++index < BTRFS_NR_RAID_TYPES)
573 up_read(&info->groups_sem);
575 btrfs_info(fs_info, "%llu bytes available across all block groups", total_avail);
578 static inline u64 calc_reclaim_items_nr(const struct btrfs_fs_info *fs_info,
584 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
585 nr = div64_u64(to_reclaim, bytes);
591 #define EXTENT_SIZE_PER_ITEM SZ_256K
594 * shrink metadata reservation for delalloc
596 static void shrink_delalloc(struct btrfs_fs_info *fs_info,
597 struct btrfs_space_info *space_info,
598 u64 to_reclaim, bool wait_ordered,
601 struct btrfs_trans_handle *trans;
608 delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
609 ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes);
610 if (delalloc_bytes == 0 && ordered_bytes == 0)
613 /* Calc the number of the pages we need flush for space reservation */
614 if (to_reclaim == U64_MAX) {
618 * to_reclaim is set to however much metadata we need to
619 * reclaim, but reclaiming that much data doesn't really track
620 * exactly. What we really want to do is reclaim full inode's
621 * worth of reservations, however that's not available to us
622 * here. We will take a fraction of the delalloc bytes for our
623 * flushing loops and hope for the best. Delalloc will expand
624 * the amount we write to cover an entire dirty extent, which
625 * will reclaim the metadata reservation for that range. If
626 * it's not enough subsequent flush stages will be more
629 to_reclaim = max(to_reclaim, delalloc_bytes >> 3);
630 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
633 trans = current->journal_info;
636 * If we are doing more ordered than delalloc we need to just wait on
637 * ordered extents, otherwise we'll waste time trying to flush delalloc
638 * that likely won't give us the space back we need.
640 if (ordered_bytes > delalloc_bytes && !for_preempt)
644 while ((delalloc_bytes || ordered_bytes) && loops < 3) {
645 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
646 long nr_pages = min_t(u64, temp, LONG_MAX);
649 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
652 * We need to make sure any outstanding async pages are now
653 * processed before we continue. This is because things like
654 * sync_inode() try to be smart and skip writing if the inode is
655 * marked clean. We don't use filemap_fwrite for flushing
656 * because we want to control how many pages we write out at a
657 * time, thus this is the only safe way to make sure we've
658 * waited for outstanding compressed workers to have started
659 * their jobs and thus have ordered extents set up properly.
661 * This exists because we do not want to wait for each
662 * individual inode to finish its async work, we simply want to
663 * start the IO on everybody, and then come back here and wait
664 * for all of the async work to catch up. Once we're done with
665 * that we know we'll have ordered extents for everything and we
666 * can decide if we wait for that or not.
668 * If we choose to replace this in the future, make absolutely
669 * sure that the proper waiting is being done in the async case,
670 * as there have been bugs in that area before.
672 async_pages = atomic_read(&fs_info->async_delalloc_pages);
677 * We don't want to wait forever, if we wrote less pages in this
678 * loop than we have outstanding, only wait for that number of
679 * pages, otherwise we can wait for all async pages to finish
682 if (async_pages > nr_pages)
683 async_pages -= nr_pages;
686 wait_event(fs_info->async_submit_wait,
687 atomic_read(&fs_info->async_delalloc_pages) <=
691 if (wait_ordered && !trans) {
692 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
694 time_left = schedule_timeout_killable(1);
700 * If we are for preemption we just want a one-shot of delalloc
701 * flushing so we can stop flushing if we decide we don't need
707 spin_lock(&space_info->lock);
708 if (list_empty(&space_info->tickets) &&
709 list_empty(&space_info->priority_tickets)) {
710 spin_unlock(&space_info->lock);
713 spin_unlock(&space_info->lock);
715 delalloc_bytes = percpu_counter_sum_positive(
716 &fs_info->delalloc_bytes);
717 ordered_bytes = percpu_counter_sum_positive(
718 &fs_info->ordered_bytes);
723 * Try to flush some data based on policy set by @state. This is only advisory
724 * and may fail for various reasons. The caller is supposed to examine the
725 * state of @space_info to detect the outcome.
727 static void flush_space(struct btrfs_fs_info *fs_info,
728 struct btrfs_space_info *space_info, u64 num_bytes,
729 enum btrfs_flush_state state, bool for_preempt)
731 struct btrfs_root *root = fs_info->tree_root;
732 struct btrfs_trans_handle *trans;
737 case FLUSH_DELAYED_ITEMS_NR:
738 case FLUSH_DELAYED_ITEMS:
739 if (state == FLUSH_DELAYED_ITEMS_NR)
740 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
744 trans = btrfs_join_transaction_nostart(root);
746 ret = PTR_ERR(trans);
751 ret = btrfs_run_delayed_items_nr(trans, nr);
752 btrfs_end_transaction(trans);
755 case FLUSH_DELALLOC_WAIT:
756 case FLUSH_DELALLOC_FULL:
757 if (state == FLUSH_DELALLOC_FULL)
759 shrink_delalloc(fs_info, space_info, num_bytes,
760 state != FLUSH_DELALLOC, for_preempt);
762 case FLUSH_DELAYED_REFS_NR:
763 case FLUSH_DELAYED_REFS:
764 trans = btrfs_join_transaction_nostart(root);
766 ret = PTR_ERR(trans);
771 if (state == FLUSH_DELAYED_REFS_NR)
772 btrfs_run_delayed_refs(trans, num_bytes);
774 btrfs_run_delayed_refs(trans, 0);
775 btrfs_end_transaction(trans);
778 case ALLOC_CHUNK_FORCE:
779 trans = btrfs_join_transaction(root);
781 ret = PTR_ERR(trans);
784 ret = btrfs_chunk_alloc(trans,
785 btrfs_get_alloc_profile(fs_info, space_info->flags),
786 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
788 btrfs_end_transaction(trans);
790 if (ret > 0 || ret == -ENOSPC)
793 case RUN_DELAYED_IPUTS:
795 * If we have pending delayed iputs then we could free up a
796 * bunch of pinned space, so make sure we run the iputs before
797 * we do our pinned bytes check below.
799 btrfs_run_delayed_iputs(fs_info);
800 btrfs_wait_on_delayed_iputs(fs_info);
803 ASSERT(current->journal_info == NULL);
805 * We don't want to start a new transaction, just attach to the
806 * current one or wait it fully commits in case its commit is
807 * happening at the moment. Note: we don't use a nostart join
808 * because that does not wait for a transaction to fully commit
809 * (only for it to be unblocked, state TRANS_STATE_UNBLOCKED).
811 trans = btrfs_attach_transaction_barrier(root);
813 ret = PTR_ERR(trans);
818 ret = btrfs_commit_transaction(trans);
825 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
831 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
832 struct btrfs_space_info *space_info)
836 u64 to_reclaim = space_info->reclaim_size;
838 lockdep_assert_held(&space_info->lock);
840 avail = calc_available_free_space(fs_info, space_info,
841 BTRFS_RESERVE_FLUSH_ALL);
842 used = btrfs_space_info_used(space_info, true);
845 * We may be flushing because suddenly we have less space than we had
846 * before, and now we're well over-committed based on our current free
847 * space. If that's the case add in our overage so we make sure to put
848 * appropriate pressure on the flushing state machine.
850 if (space_info->total_bytes + avail < used)
851 to_reclaim += used - (space_info->total_bytes + avail);
856 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
857 struct btrfs_space_info *space_info)
859 const u64 global_rsv_size = btrfs_block_rsv_reserved(&fs_info->global_block_rsv);
860 u64 ordered, delalloc;
864 thresh = mult_perc(space_info->total_bytes, 90);
866 lockdep_assert_held(&space_info->lock);
868 /* If we're just plain full then async reclaim just slows us down. */
869 if ((space_info->bytes_used + space_info->bytes_reserved +
870 global_rsv_size) >= thresh)
873 used = space_info->bytes_may_use + space_info->bytes_pinned;
875 /* The total flushable belongs to the global rsv, don't flush. */
876 if (global_rsv_size >= used)
880 * 128MiB is 1/4 of the maximum global rsv size. If we have less than
881 * that devoted to other reservations then there's no sense in flushing,
882 * we don't have a lot of things that need flushing.
884 if (used - global_rsv_size <= SZ_128M)
888 * We have tickets queued, bail so we don't compete with the async
891 if (space_info->reclaim_size)
895 * If we have over half of the free space occupied by reservations or
896 * pinned then we want to start flushing.
898 * We do not do the traditional thing here, which is to say
900 * if (used >= ((total_bytes + avail) / 2))
903 * because this doesn't quite work how we want. If we had more than 50%
904 * of the space_info used by bytes_used and we had 0 available we'd just
905 * constantly run the background flusher. Instead we want it to kick in
906 * if our reclaimable space exceeds our clamped free space.
908 * Our clamping range is 2^1 -> 2^8. Practically speaking that means
911 * Amount of RAM Minimum threshold Maximum threshold
914 * 128GiB 512MiB 64GiB
919 * These are the range our thresholds will fall in, corresponding to how
920 * much delalloc we need for the background flusher to kick in.
923 thresh = calc_available_free_space(fs_info, space_info,
924 BTRFS_RESERVE_FLUSH_ALL);
925 used = space_info->bytes_used + space_info->bytes_reserved +
926 space_info->bytes_readonly + global_rsv_size;
927 if (used < space_info->total_bytes)
928 thresh += space_info->total_bytes - used;
929 thresh >>= space_info->clamp;
931 used = space_info->bytes_pinned;
934 * If we have more ordered bytes than delalloc bytes then we're either
935 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting
936 * around. Preemptive flushing is only useful in that it can free up
937 * space before tickets need to wait for things to finish. In the case
938 * of ordered extents, preemptively waiting on ordered extents gets us
939 * nothing, if our reservations are tied up in ordered extents we'll
940 * simply have to slow down writers by forcing them to wait on ordered
943 * In the case that ordered is larger than delalloc, only include the
944 * block reserves that we would actually be able to directly reclaim
945 * from. In this case if we're heavy on metadata operations this will
946 * clearly be heavy enough to warrant preemptive flushing. In the case
947 * of heavy DIO or ordered reservations, preemptive flushing will just
948 * waste time and cause us to slow down.
950 * We want to make sure we truly are maxed out on ordered however, so
951 * cut ordered in half, and if it's still higher than delalloc then we
952 * can keep flushing. This is to avoid the case where we start
953 * flushing, and now delalloc == ordered and we stop preemptively
954 * flushing when we could still have several gigs of delalloc to flush.
956 ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1;
957 delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes);
958 if (ordered >= delalloc)
959 used += btrfs_block_rsv_reserved(&fs_info->delayed_refs_rsv) +
960 btrfs_block_rsv_reserved(&fs_info->delayed_block_rsv);
962 used += space_info->bytes_may_use - global_rsv_size;
964 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
965 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
968 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
969 struct btrfs_space_info *space_info,
970 struct reserve_ticket *ticket)
972 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
978 if (global_rsv->space_info != space_info)
981 spin_lock(&global_rsv->lock);
982 min_bytes = mult_perc(global_rsv->size, 10);
983 if (global_rsv->reserved < min_bytes + ticket->bytes) {
984 spin_unlock(&global_rsv->lock);
987 global_rsv->reserved -= ticket->bytes;
988 remove_ticket(space_info, ticket);
990 wake_up(&ticket->wait);
991 space_info->tickets_id++;
992 if (global_rsv->reserved < global_rsv->size)
993 global_rsv->full = 0;
994 spin_unlock(&global_rsv->lock);
1000 * We've exhausted our flushing, start failing tickets.
1002 * @fs_info - fs_info for this fs
1003 * @space_info - the space info we were flushing
1005 * We call this when we've exhausted our flushing ability and haven't made
1006 * progress in satisfying tickets. The reservation code handles tickets in
1007 * order, so if there is a large ticket first and then smaller ones we could
1008 * very well satisfy the smaller tickets. This will attempt to wake up any
1009 * tickets in the list to catch this case.
1011 * This function returns true if it was able to make progress by clearing out
1012 * other tickets, or if it stumbles across a ticket that was smaller than the
1015 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
1016 struct btrfs_space_info *space_info)
1018 struct reserve_ticket *ticket;
1019 u64 tickets_id = space_info->tickets_id;
1020 const bool aborted = BTRFS_FS_ERROR(fs_info);
1022 trace_btrfs_fail_all_tickets(fs_info, space_info);
1024 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1025 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
1026 __btrfs_dump_space_info(fs_info, space_info);
1029 while (!list_empty(&space_info->tickets) &&
1030 tickets_id == space_info->tickets_id) {
1031 ticket = list_first_entry(&space_info->tickets,
1032 struct reserve_ticket, list);
1034 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket))
1037 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1038 btrfs_info(fs_info, "failing ticket with %llu bytes",
1041 remove_ticket(space_info, ticket);
1043 ticket->error = -EIO;
1045 ticket->error = -ENOSPC;
1046 wake_up(&ticket->wait);
1049 * We're just throwing tickets away, so more flushing may not
1050 * trip over btrfs_try_granting_tickets, so we need to call it
1051 * here to see if we can make progress with the next ticket in
1055 btrfs_try_granting_tickets(fs_info, space_info);
1057 return (tickets_id != space_info->tickets_id);
1061 * This is for normal flushers, we can wait all goddamned day if we want to. We
1062 * will loop and continuously try to flush as long as we are making progress.
1063 * We count progress as clearing off tickets each time we have to loop.
1065 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
1067 struct btrfs_fs_info *fs_info;
1068 struct btrfs_space_info *space_info;
1070 enum btrfs_flush_state flush_state;
1071 int commit_cycles = 0;
1072 u64 last_tickets_id;
1074 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
1075 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1077 spin_lock(&space_info->lock);
1078 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1080 space_info->flush = 0;
1081 spin_unlock(&space_info->lock);
1084 last_tickets_id = space_info->tickets_id;
1085 spin_unlock(&space_info->lock);
1087 flush_state = FLUSH_DELAYED_ITEMS_NR;
1089 flush_space(fs_info, space_info, to_reclaim, flush_state, false);
1090 spin_lock(&space_info->lock);
1091 if (list_empty(&space_info->tickets)) {
1092 space_info->flush = 0;
1093 spin_unlock(&space_info->lock);
1096 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
1098 if (last_tickets_id == space_info->tickets_id) {
1101 last_tickets_id = space_info->tickets_id;
1102 flush_state = FLUSH_DELAYED_ITEMS_NR;
1108 * We do not want to empty the system of delalloc unless we're
1109 * under heavy pressure, so allow one trip through the flushing
1110 * logic before we start doing a FLUSH_DELALLOC_FULL.
1112 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles)
1116 * We don't want to force a chunk allocation until we've tried
1117 * pretty hard to reclaim space. Think of the case where we
1118 * freed up a bunch of space and so have a lot of pinned space
1119 * to reclaim. We would rather use that than possibly create a
1120 * underutilized metadata chunk. So if this is our first run
1121 * through the flushing state machine skip ALLOC_CHUNK_FORCE and
1122 * commit the transaction. If nothing has changed the next go
1123 * around then we can force a chunk allocation.
1125 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
1128 if (flush_state > COMMIT_TRANS) {
1130 if (commit_cycles > 2) {
1131 if (maybe_fail_all_tickets(fs_info, space_info)) {
1132 flush_state = FLUSH_DELAYED_ITEMS_NR;
1135 space_info->flush = 0;
1138 flush_state = FLUSH_DELAYED_ITEMS_NR;
1141 spin_unlock(&space_info->lock);
1142 } while (flush_state <= COMMIT_TRANS);
1146 * This handles pre-flushing of metadata space before we get to the point that
1147 * we need to start blocking threads on tickets. The logic here is different
1148 * from the other flush paths because it doesn't rely on tickets to tell us how
1149 * much we need to flush, instead it attempts to keep us below the 80% full
1150 * watermark of space by flushing whichever reservation pool is currently the
1153 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
1155 struct btrfs_fs_info *fs_info;
1156 struct btrfs_space_info *space_info;
1157 struct btrfs_block_rsv *delayed_block_rsv;
1158 struct btrfs_block_rsv *delayed_refs_rsv;
1159 struct btrfs_block_rsv *global_rsv;
1160 struct btrfs_block_rsv *trans_rsv;
1163 fs_info = container_of(work, struct btrfs_fs_info,
1164 preempt_reclaim_work);
1165 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1166 delayed_block_rsv = &fs_info->delayed_block_rsv;
1167 delayed_refs_rsv = &fs_info->delayed_refs_rsv;
1168 global_rsv = &fs_info->global_block_rsv;
1169 trans_rsv = &fs_info->trans_block_rsv;
1171 spin_lock(&space_info->lock);
1172 while (need_preemptive_reclaim(fs_info, space_info)) {
1173 enum btrfs_flush_state flush;
1174 u64 delalloc_size = 0;
1175 u64 to_reclaim, block_rsv_size;
1176 const u64 global_rsv_size = btrfs_block_rsv_reserved(global_rsv);
1181 * We don't have a precise counter for the metadata being
1182 * reserved for delalloc, so we'll approximate it by subtracting
1183 * out the block rsv's space from the bytes_may_use. If that
1184 * amount is higher than the individual reserves, then we can
1185 * assume it's tied up in delalloc reservations.
1187 block_rsv_size = global_rsv_size +
1188 btrfs_block_rsv_reserved(delayed_block_rsv) +
1189 btrfs_block_rsv_reserved(delayed_refs_rsv) +
1190 btrfs_block_rsv_reserved(trans_rsv);
1191 if (block_rsv_size < space_info->bytes_may_use)
1192 delalloc_size = space_info->bytes_may_use - block_rsv_size;
1195 * We don't want to include the global_rsv in our calculation,
1196 * because that's space we can't touch. Subtract it from the
1197 * block_rsv_size for the next checks.
1199 block_rsv_size -= global_rsv_size;
1202 * We really want to avoid flushing delalloc too much, as it
1203 * could result in poor allocation patterns, so only flush it if
1204 * it's larger than the rest of the pools combined.
1206 if (delalloc_size > block_rsv_size) {
1207 to_reclaim = delalloc_size;
1208 flush = FLUSH_DELALLOC;
1209 } else if (space_info->bytes_pinned >
1210 (btrfs_block_rsv_reserved(delayed_block_rsv) +
1211 btrfs_block_rsv_reserved(delayed_refs_rsv))) {
1212 to_reclaim = space_info->bytes_pinned;
1213 flush = COMMIT_TRANS;
1214 } else if (btrfs_block_rsv_reserved(delayed_block_rsv) >
1215 btrfs_block_rsv_reserved(delayed_refs_rsv)) {
1216 to_reclaim = btrfs_block_rsv_reserved(delayed_block_rsv);
1217 flush = FLUSH_DELAYED_ITEMS_NR;
1219 to_reclaim = btrfs_block_rsv_reserved(delayed_refs_rsv);
1220 flush = FLUSH_DELAYED_REFS_NR;
1223 spin_unlock(&space_info->lock);
1226 * We don't want to reclaim everything, just a portion, so scale
1227 * down the to_reclaim by 1/4. If it takes us down to 0,
1228 * reclaim 1 items worth.
1232 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1233 flush_space(fs_info, space_info, to_reclaim, flush, true);
1235 spin_lock(&space_info->lock);
1238 /* We only went through once, back off our clamping. */
1239 if (loops == 1 && !space_info->reclaim_size)
1240 space_info->clamp = max(1, space_info->clamp - 1);
1241 trace_btrfs_done_preemptive_reclaim(fs_info, space_info);
1242 spin_unlock(&space_info->lock);
1246 * FLUSH_DELALLOC_WAIT:
1247 * Space is freed from flushing delalloc in one of two ways.
1249 * 1) compression is on and we allocate less space than we reserved
1250 * 2) we are overwriting existing space
1252 * For #1 that extra space is reclaimed as soon as the delalloc pages are
1253 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1254 * length to ->bytes_reserved, and subtracts the reserved space from
1257 * For #2 this is trickier. Once the ordered extent runs we will drop the
1258 * extent in the range we are overwriting, which creates a delayed ref for
1259 * that freed extent. This however is not reclaimed until the transaction
1260 * commits, thus the next stages.
1263 * If we are freeing inodes, we want to make sure all delayed iputs have
1264 * completed, because they could have been on an inode with i_nlink == 0, and
1265 * thus have been truncated and freed up space. But again this space is not
1266 * immediately re-usable, it comes in the form of a delayed ref, which must be
1267 * run and then the transaction must be committed.
1270 * This is where we reclaim all of the pinned space generated by running the
1274 * For data we start with alloc chunk force, however we could have been full
1275 * before, and then the transaction commit could have freed new block groups,
1276 * so if we now have space to allocate do the force chunk allocation.
1278 static const enum btrfs_flush_state data_flush_states[] = {
1279 FLUSH_DELALLOC_FULL,
1285 static void btrfs_async_reclaim_data_space(struct work_struct *work)
1287 struct btrfs_fs_info *fs_info;
1288 struct btrfs_space_info *space_info;
1289 u64 last_tickets_id;
1290 enum btrfs_flush_state flush_state = 0;
1292 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1293 space_info = fs_info->data_sinfo;
1295 spin_lock(&space_info->lock);
1296 if (list_empty(&space_info->tickets)) {
1297 space_info->flush = 0;
1298 spin_unlock(&space_info->lock);
1301 last_tickets_id = space_info->tickets_id;
1302 spin_unlock(&space_info->lock);
1304 while (!space_info->full) {
1305 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1306 spin_lock(&space_info->lock);
1307 if (list_empty(&space_info->tickets)) {
1308 space_info->flush = 0;
1309 spin_unlock(&space_info->lock);
1313 /* Something happened, fail everything and bail. */
1314 if (BTRFS_FS_ERROR(fs_info))
1316 last_tickets_id = space_info->tickets_id;
1317 spin_unlock(&space_info->lock);
1320 while (flush_state < ARRAY_SIZE(data_flush_states)) {
1321 flush_space(fs_info, space_info, U64_MAX,
1322 data_flush_states[flush_state], false);
1323 spin_lock(&space_info->lock);
1324 if (list_empty(&space_info->tickets)) {
1325 space_info->flush = 0;
1326 spin_unlock(&space_info->lock);
1330 if (last_tickets_id == space_info->tickets_id) {
1333 last_tickets_id = space_info->tickets_id;
1337 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1338 if (space_info->full) {
1339 if (maybe_fail_all_tickets(fs_info, space_info))
1342 space_info->flush = 0;
1347 /* Something happened, fail everything and bail. */
1348 if (BTRFS_FS_ERROR(fs_info))
1352 spin_unlock(&space_info->lock);
1357 maybe_fail_all_tickets(fs_info, space_info);
1358 space_info->flush = 0;
1359 spin_unlock(&space_info->lock);
1362 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1364 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1365 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1366 INIT_WORK(&fs_info->preempt_reclaim_work,
1367 btrfs_preempt_reclaim_metadata_space);
1370 static const enum btrfs_flush_state priority_flush_states[] = {
1371 FLUSH_DELAYED_ITEMS_NR,
1372 FLUSH_DELAYED_ITEMS,
1376 static const enum btrfs_flush_state evict_flush_states[] = {
1377 FLUSH_DELAYED_ITEMS_NR,
1378 FLUSH_DELAYED_ITEMS,
1379 FLUSH_DELAYED_REFS_NR,
1382 FLUSH_DELALLOC_WAIT,
1383 FLUSH_DELALLOC_FULL,
1388 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1389 struct btrfs_space_info *space_info,
1390 struct reserve_ticket *ticket,
1391 const enum btrfs_flush_state *states,
1395 int flush_state = 0;
1397 spin_lock(&space_info->lock);
1398 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1400 * This is the priority reclaim path, so to_reclaim could be >0 still
1401 * because we may have only satisfied the priority tickets and still
1402 * left non priority tickets on the list. We would then have
1403 * to_reclaim but ->bytes == 0.
1405 if (ticket->bytes == 0) {
1406 spin_unlock(&space_info->lock);
1410 while (flush_state < states_nr) {
1411 spin_unlock(&space_info->lock);
1412 flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1415 spin_lock(&space_info->lock);
1416 if (ticket->bytes == 0) {
1417 spin_unlock(&space_info->lock);
1423 * Attempt to steal from the global rsv if we can, except if the fs was
1424 * turned into error mode due to a transaction abort when flushing space
1425 * above, in that case fail with the abort error instead of returning
1426 * success to the caller if we can steal from the global rsv - this is
1427 * just to have caller fail immeditelly instead of later when trying to
1428 * modify the fs, making it easier to debug -ENOSPC problems.
1430 if (BTRFS_FS_ERROR(fs_info)) {
1431 ticket->error = BTRFS_FS_ERROR(fs_info);
1432 remove_ticket(space_info, ticket);
1433 } else if (!steal_from_global_rsv(fs_info, space_info, ticket)) {
1434 ticket->error = -ENOSPC;
1435 remove_ticket(space_info, ticket);
1439 * We must run try_granting_tickets here because we could be a large
1440 * ticket in front of a smaller ticket that can now be satisfied with
1441 * the available space.
1443 btrfs_try_granting_tickets(fs_info, space_info);
1444 spin_unlock(&space_info->lock);
1447 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1448 struct btrfs_space_info *space_info,
1449 struct reserve_ticket *ticket)
1451 spin_lock(&space_info->lock);
1453 /* We could have been granted before we got here. */
1454 if (ticket->bytes == 0) {
1455 spin_unlock(&space_info->lock);
1459 while (!space_info->full) {
1460 spin_unlock(&space_info->lock);
1461 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1462 spin_lock(&space_info->lock);
1463 if (ticket->bytes == 0) {
1464 spin_unlock(&space_info->lock);
1469 ticket->error = -ENOSPC;
1470 remove_ticket(space_info, ticket);
1471 btrfs_try_granting_tickets(fs_info, space_info);
1472 spin_unlock(&space_info->lock);
1475 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1476 struct btrfs_space_info *space_info,
1477 struct reserve_ticket *ticket)
1483 spin_lock(&space_info->lock);
1484 while (ticket->bytes > 0 && ticket->error == 0) {
1485 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1488 * Delete us from the list. After we unlock the space
1489 * info, we don't want the async reclaim job to reserve
1490 * space for this ticket. If that would happen, then the
1491 * ticket's task would not known that space was reserved
1492 * despite getting an error, resulting in a space leak
1493 * (bytes_may_use counter of our space_info).
1495 remove_ticket(space_info, ticket);
1496 ticket->error = -EINTR;
1499 spin_unlock(&space_info->lock);
1503 finish_wait(&ticket->wait, &wait);
1504 spin_lock(&space_info->lock);
1506 spin_unlock(&space_info->lock);
1510 * Do the appropriate flushing and waiting for a ticket.
1512 * @fs_info: the filesystem
1513 * @space_info: space info for the reservation
1514 * @ticket: ticket for the reservation
1515 * @start_ns: timestamp when the reservation started
1516 * @orig_bytes: amount of bytes originally reserved
1517 * @flush: how much we can flush
1519 * This does the work of figuring out how to flush for the ticket, waiting for
1520 * the reservation, and returning the appropriate error if there is one.
1522 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1523 struct btrfs_space_info *space_info,
1524 struct reserve_ticket *ticket,
1525 u64 start_ns, u64 orig_bytes,
1526 enum btrfs_reserve_flush_enum flush)
1531 case BTRFS_RESERVE_FLUSH_DATA:
1532 case BTRFS_RESERVE_FLUSH_ALL:
1533 case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1534 wait_reserve_ticket(fs_info, space_info, ticket);
1536 case BTRFS_RESERVE_FLUSH_LIMIT:
1537 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1538 priority_flush_states,
1539 ARRAY_SIZE(priority_flush_states));
1541 case BTRFS_RESERVE_FLUSH_EVICT:
1542 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1544 ARRAY_SIZE(evict_flush_states));
1546 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1547 priority_reclaim_data_space(fs_info, space_info, ticket);
1554 ret = ticket->error;
1555 ASSERT(list_empty(&ticket->list));
1557 * Check that we can't have an error set if the reservation succeeded,
1558 * as that would confuse tasks and lead them to error out without
1559 * releasing reserved space (if an error happens the expectation is that
1560 * space wasn't reserved at all).
1562 ASSERT(!(ticket->bytes == 0 && ticket->error));
1563 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes,
1564 start_ns, flush, ticket->error);
1569 * This returns true if this flush state will go through the ordinary flushing
1572 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1574 return (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1575 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1578 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1579 struct btrfs_space_info *space_info)
1581 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1582 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1585 * If we're heavy on ordered operations then clamping won't help us. We
1586 * need to clamp specifically to keep up with dirty'ing buffered
1587 * writers, because there's not a 1:1 correlation of writing delalloc
1588 * and freeing space, like there is with flushing delayed refs or
1589 * delayed nodes. If we're already more ordered than delalloc then
1590 * we're keeping up, otherwise we aren't and should probably clamp.
1592 if (ordered < delalloc)
1593 space_info->clamp = min(space_info->clamp + 1, 8);
1596 static inline bool can_steal(enum btrfs_reserve_flush_enum flush)
1598 return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1599 flush == BTRFS_RESERVE_FLUSH_EVICT);
1603 * NO_FLUSH and FLUSH_EMERGENCY don't want to create a ticket, they just want to
1604 * fail as quickly as possible.
1606 static inline bool can_ticket(enum btrfs_reserve_flush_enum flush)
1608 return (flush != BTRFS_RESERVE_NO_FLUSH &&
1609 flush != BTRFS_RESERVE_FLUSH_EMERGENCY);
1613 * Try to reserve bytes from the block_rsv's space.
1615 * @fs_info: the filesystem
1616 * @space_info: space info we want to allocate from
1617 * @orig_bytes: number of bytes we want
1618 * @flush: whether or not we can flush to make our reservation
1620 * This will reserve orig_bytes number of bytes from the space info associated
1621 * with the block_rsv. If there is not enough space it will make an attempt to
1622 * flush out space to make room. It will do this by flushing delalloc if
1623 * possible or committing the transaction. If flush is 0 then no attempts to
1624 * regain reservations will be made and this will fail if there is not enough
1627 static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1628 struct btrfs_space_info *space_info, u64 orig_bytes,
1629 enum btrfs_reserve_flush_enum flush)
1631 struct work_struct *async_work;
1632 struct reserve_ticket ticket;
1636 bool pending_tickets;
1640 * If have a transaction handle (current->journal_info != NULL), then
1641 * the flush method can not be neither BTRFS_RESERVE_FLUSH_ALL* nor
1642 * BTRFS_RESERVE_FLUSH_EVICT, as we could deadlock because those
1643 * flushing methods can trigger transaction commits.
1645 if (current->journal_info) {
1646 /* One assert per line for easier debugging. */
1647 ASSERT(flush != BTRFS_RESERVE_FLUSH_ALL);
1648 ASSERT(flush != BTRFS_RESERVE_FLUSH_ALL_STEAL);
1649 ASSERT(flush != BTRFS_RESERVE_FLUSH_EVICT);
1652 if (flush == BTRFS_RESERVE_FLUSH_DATA)
1653 async_work = &fs_info->async_data_reclaim_work;
1655 async_work = &fs_info->async_reclaim_work;
1657 spin_lock(&space_info->lock);
1658 used = btrfs_space_info_used(space_info, true);
1661 * We don't want NO_FLUSH allocations to jump everybody, they can
1662 * generally handle ENOSPC in a different way, so treat them the same as
1663 * normal flushers when it comes to skipping pending tickets.
1665 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1666 pending_tickets = !list_empty(&space_info->tickets) ||
1667 !list_empty(&space_info->priority_tickets);
1669 pending_tickets = !list_empty(&space_info->priority_tickets);
1672 * Carry on if we have enough space (short-circuit) OR call
1673 * can_overcommit() to ensure we can overcommit to continue.
1675 if (!pending_tickets &&
1676 ((used + orig_bytes <= space_info->total_bytes) ||
1677 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1678 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1684 * Things are dire, we need to make a reservation so we don't abort. We
1685 * will let this reservation go through as long as we have actual space
1686 * left to allocate for the block.
1688 if (ret && unlikely(flush == BTRFS_RESERVE_FLUSH_EMERGENCY)) {
1689 used = btrfs_space_info_used(space_info, false);
1690 if (used + orig_bytes <= space_info->total_bytes) {
1691 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1698 * If we couldn't make a reservation then setup our reservation ticket
1699 * and kick the async worker if it's not already running.
1701 * If we are a priority flusher then we just need to add our ticket to
1702 * the list and we will do our own flushing further down.
1704 if (ret && can_ticket(flush)) {
1705 ticket.bytes = orig_bytes;
1707 space_info->reclaim_size += ticket.bytes;
1708 init_waitqueue_head(&ticket.wait);
1709 ticket.steal = can_steal(flush);
1710 if (trace_btrfs_reserve_ticket_enabled())
1711 start_ns = ktime_get_ns();
1713 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1714 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1715 flush == BTRFS_RESERVE_FLUSH_DATA) {
1716 list_add_tail(&ticket.list, &space_info->tickets);
1717 if (!space_info->flush) {
1719 * We were forced to add a reserve ticket, so
1720 * our preemptive flushing is unable to keep
1721 * up. Clamp down on the threshold for the
1722 * preemptive flushing in order to keep up with
1725 maybe_clamp_preempt(fs_info, space_info);
1727 space_info->flush = 1;
1728 trace_btrfs_trigger_flush(fs_info,
1732 queue_work(system_unbound_wq, async_work);
1735 list_add_tail(&ticket.list,
1736 &space_info->priority_tickets);
1738 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1740 * We will do the space reservation dance during log replay,
1741 * which means we won't have fs_info->fs_root set, so don't do
1742 * the async reclaim as we will panic.
1744 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1745 !work_busy(&fs_info->preempt_reclaim_work) &&
1746 need_preemptive_reclaim(fs_info, space_info)) {
1747 trace_btrfs_trigger_flush(fs_info, space_info->flags,
1748 orig_bytes, flush, "preempt");
1749 queue_work(system_unbound_wq,
1750 &fs_info->preempt_reclaim_work);
1753 spin_unlock(&space_info->lock);
1754 if (!ret || !can_ticket(flush))
1757 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1762 * Try to reserve metadata bytes from the block_rsv's space.
1764 * @fs_info: the filesystem
1765 * @space_info: the space_info we're allocating for
1766 * @orig_bytes: number of bytes we want
1767 * @flush: whether or not we can flush to make our reservation
1769 * This will reserve orig_bytes number of bytes from the space info associated
1770 * with the block_rsv. If there is not enough space it will make an attempt to
1771 * flush out space to make room. It will do this by flushing delalloc if
1772 * possible or committing the transaction. If flush is 0 then no attempts to
1773 * regain reservations will be made and this will fail if there is not enough
1776 int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1777 struct btrfs_space_info *space_info,
1779 enum btrfs_reserve_flush_enum flush)
1783 ret = __reserve_bytes(fs_info, space_info, orig_bytes, flush);
1784 if (ret == -ENOSPC) {
1785 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1786 space_info->flags, orig_bytes, 1);
1788 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1789 btrfs_dump_space_info(fs_info, space_info, orig_bytes, 0);
1795 * Try to reserve data bytes for an allocation.
1797 * @fs_info: the filesystem
1798 * @bytes: number of bytes we need
1799 * @flush: how we are allowed to flush
1801 * This will reserve bytes from the data space info. If there is not enough
1802 * space then we will attempt to flush space as specified by flush.
1804 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1805 enum btrfs_reserve_flush_enum flush)
1807 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1810 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1811 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE ||
1812 flush == BTRFS_RESERVE_NO_FLUSH);
1813 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1815 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1816 if (ret == -ENOSPC) {
1817 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1818 data_sinfo->flags, bytes, 1);
1819 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1820 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1825 /* Dump all the space infos when we abort a transaction due to ENOSPC. */
1826 __cold void btrfs_dump_space_info_for_trans_abort(struct btrfs_fs_info *fs_info)
1828 struct btrfs_space_info *space_info;
1830 btrfs_info(fs_info, "dumping space info:");
1831 list_for_each_entry(space_info, &fs_info->space_info, list) {
1832 spin_lock(&space_info->lock);
1833 __btrfs_dump_space_info(fs_info, space_info);
1834 spin_unlock(&space_info->lock);
1836 dump_global_block_rsv(fs_info);
1840 * Account the unused space of all the readonly block group in the space_info.
1841 * takes mirrors into account.
1843 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
1845 struct btrfs_block_group *block_group;
1849 /* It's df, we don't care if it's racy */
1850 if (list_empty(&sinfo->ro_bgs))
1853 spin_lock(&sinfo->lock);
1854 list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
1855 spin_lock(&block_group->lock);
1857 if (!block_group->ro) {
1858 spin_unlock(&block_group->lock);
1862 factor = btrfs_bg_type_to_factor(block_group->flags);
1863 free_bytes += (block_group->length -
1864 block_group->used) * factor;
1866 spin_unlock(&block_group->lock);
1868 spin_unlock(&sinfo->lock);