Loading...
1// SPDX-License-Identifier: GPL-2.0
2
3#include "misc.h"
4#include "ctree.h"
5#include "space-info.h"
6#include "sysfs.h"
7#include "volumes.h"
8#include "free-space-cache.h"
9#include "ordered-data.h"
10#include "transaction.h"
11#include "block-group.h"
12#include "zoned.h"
13#include "fs.h"
14#include "accessors.h"
15#include "extent-tree.h"
16
17/*
18 * HOW DOES SPACE RESERVATION WORK
19 *
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
22 * works generally.
23 *
24 * BASIC CONCEPTS
25 *
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.
32 *
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.
37 *
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.
43 *
44 * MAKING RESERVATIONS, THE NORMAL CASE
45 *
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.
49 *
50 * ->reserve
51 * space_info->bytes_may_reserve += num_bytes
52 *
53 * ->extent allocation
54 * Call btrfs_add_reserved_bytes() which does
55 * space_info->bytes_may_reserve -= num_bytes
56 * space_info->bytes_reserved += extent_bytes
57 *
58 * ->insert reference
59 * Call btrfs_update_block_group() which does
60 * space_info->bytes_reserved -= extent_bytes
61 * space_info->bytes_used += extent_bytes
62 *
63 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority)
64 *
65 * Assume we are unable to simply make the reservation because we do not have
66 * enough space
67 *
68 * -> __reserve_bytes
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
71 *
72 * ->handle_reserve_ticket
73 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set
74 * on the ticket.
75 *
76 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space
77 * Flushes various things attempting to free up space.
78 *
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.
86 *
87 * -> ticket wakeup
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
90 * were interrupted.)
91 *
92 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY
93 *
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.
98 *
99 * THE FLUSHING STATES
100 *
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.
107 *
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
114 * metadata space.
115 *
116 * FLUSH_DELALLOC
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.
121 *
122 * FLUSH_DELAYED_REFS
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.
128 *
129 * ALLOC_CHUNK
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.
133 *
134 * RUN_DELAYED_IPUTS
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.
138 *
139 * COMMIT_TRANS
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.
147 *
148 * OVERCOMMIT
149 *
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
153 * overcommitting.
154 *
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.
159 *
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.
163 */
164
165u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
166 bool may_use_included)
167{
168 ASSERT(s_info);
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);
173}
174
175/*
176 * after adding space to the filesystem, we need to clear the full flags
177 * on all the space infos.
178 */
179void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
180{
181 struct list_head *head = &info->space_info;
182 struct btrfs_space_info *found;
183
184 list_for_each_entry(found, head, list)
185 found->full = 0;
186}
187
188/*
189 * Block groups with more than this value (percents) of unusable space will be
190 * scheduled for background reclaim.
191 */
192#define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH (75)
193
194/*
195 * Calculate chunk size depending on volume type (regular or zoned).
196 */
197static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags)
198{
199 if (btrfs_is_zoned(fs_info))
200 return fs_info->zone_size;
201
202 ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
203
204 if (flags & BTRFS_BLOCK_GROUP_DATA)
205 return BTRFS_MAX_DATA_CHUNK_SIZE;
206 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
207 return SZ_32M;
208
209 /* Handle BTRFS_BLOCK_GROUP_METADATA */
210 if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G)
211 return SZ_1G;
212
213 return SZ_256M;
214}
215
216/*
217 * Update default chunk size.
218 */
219void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info,
220 u64 chunk_size)
221{
222 WRITE_ONCE(space_info->chunk_size, chunk_size);
223}
224
225static int create_space_info(struct btrfs_fs_info *info, u64 flags)
226{
227
228 struct btrfs_space_info *space_info;
229 int i;
230 int ret;
231
232 space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
233 if (!space_info)
234 return -ENOMEM;
235
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));
247
248 if (btrfs_is_zoned(info))
249 space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH;
250
251 ret = btrfs_sysfs_add_space_info_type(info, space_info);
252 if (ret)
253 return ret;
254
255 list_add(&space_info->list, &info->space_info);
256 if (flags & BTRFS_BLOCK_GROUP_DATA)
257 info->data_sinfo = space_info;
258
259 return ret;
260}
261
262int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
263{
264 struct btrfs_super_block *disk_super;
265 u64 features;
266 u64 flags;
267 int mixed = 0;
268 int ret;
269
270 disk_super = fs_info->super_copy;
271 if (!btrfs_super_root(disk_super))
272 return -EINVAL;
273
274 features = btrfs_super_incompat_flags(disk_super);
275 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
276 mixed = 1;
277
278 flags = BTRFS_BLOCK_GROUP_SYSTEM;
279 ret = create_space_info(fs_info, flags);
280 if (ret)
281 goto out;
282
283 if (mixed) {
284 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
285 ret = create_space_info(fs_info, flags);
286 } else {
287 flags = BTRFS_BLOCK_GROUP_METADATA;
288 ret = create_space_info(fs_info, flags);
289 if (ret)
290 goto out;
291
292 flags = BTRFS_BLOCK_GROUP_DATA;
293 ret = create_space_info(fs_info, flags);
294 }
295out:
296 return ret;
297}
298
299void btrfs_add_bg_to_space_info(struct btrfs_fs_info *info,
300 struct btrfs_block_group *block_group)
301{
302 struct btrfs_space_info *found;
303 int factor, index;
304
305 factor = btrfs_bg_type_to_factor(block_group->flags);
306
307 found = btrfs_find_space_info(info, block_group->flags);
308 ASSERT(found);
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)
317 found->full = 0;
318 btrfs_try_granting_tickets(info, found);
319 spin_unlock(&found->lock);
320
321 block_group->space_info = found;
322
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);
327}
328
329struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
330 u64 flags)
331{
332 struct list_head *head = &info->space_info;
333 struct btrfs_space_info *found;
334
335 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
336
337 list_for_each_entry(found, head, list) {
338 if (found->flags & flags)
339 return found;
340 }
341 return NULL;
342}
343
344static 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)
347{
348 struct btrfs_space_info *data_sinfo;
349 u64 profile;
350 u64 avail;
351 u64 data_chunk_size;
352 int factor;
353
354 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
355 profile = btrfs_system_alloc_profile(fs_info);
356 else
357 profile = btrfs_metadata_alloc_profile(fs_info);
358
359 avail = atomic64_read(&fs_info->free_chunk_space);
360
361 /*
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
365 * change the math
366 */
367 factor = btrfs_bg_type_to_factor(profile);
368 avail = div_u64(avail, factor);
369 if (avail == 0)
370 return 0;
371
372 /*
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.
377 */
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);
382
383 /*
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.
389 *
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.
395 */
396 if (avail <= data_chunk_size)
397 return 0;
398 avail -= data_chunk_size;
399
400 /*
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.
404 */
405 if (flush == BTRFS_RESERVE_FLUSH_ALL)
406 avail >>= 3;
407 else
408 avail >>= 1;
409 return avail;
410}
411
412int 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)
415{
416 u64 avail;
417 u64 used;
418
419 /* Don't overcommit when in mixed mode */
420 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
421 return 0;
422
423 used = btrfs_space_info_used(space_info, true);
424 avail = calc_available_free_space(fs_info, space_info, flush);
425
426 if (used + bytes < space_info->total_bytes + avail)
427 return 1;
428 return 0;
429}
430
431static void remove_ticket(struct btrfs_space_info *space_info,
432 struct reserve_ticket *ticket)
433{
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;
438 }
439}
440
441/*
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.
444 */
445void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
446 struct btrfs_space_info *space_info)
447{
448 struct list_head *head;
449 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
450
451 lockdep_assert_held(&space_info->lock);
452
453 head = &space_info->priority_tickets;
454again:
455 while (!list_empty(head)) {
456 struct reserve_ticket *ticket;
457 u64 used = btrfs_space_info_used(space_info, true);
458
459 ticket = list_first_entry(head, struct reserve_ticket, list);
460
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,
464 flush)) {
465 btrfs_space_info_update_bytes_may_use(fs_info,
466 space_info,
467 ticket->bytes);
468 remove_ticket(space_info, ticket);
469 ticket->bytes = 0;
470 space_info->tickets_id++;
471 wake_up(&ticket->wait);
472 } else {
473 break;
474 }
475 }
476
477 if (head == &space_info->priority_tickets) {
478 head = &space_info->tickets;
479 flush = BTRFS_RESERVE_FLUSH_ALL;
480 goto again;
481 }
482}
483
484#define DUMP_BLOCK_RSV(fs_info, rsv_name) \
485do { \
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); \
491} while (0)
492
493static const char *space_info_flag_to_str(const struct btrfs_space_info *space_info)
494{
495 switch (space_info->flags) {
496 case BTRFS_BLOCK_GROUP_SYSTEM:
497 return "SYSTEM";
498 case BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA:
499 return "DATA+METADATA";
500 case BTRFS_BLOCK_GROUP_DATA:
501 return "DATA";
502 case BTRFS_BLOCK_GROUP_METADATA:
503 return "METADATA";
504 default:
505 return "UNKNOWN";
506 }
507}
508
509static void dump_global_block_rsv(struct btrfs_fs_info *fs_info)
510{
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);
516}
517
518static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
519 struct btrfs_space_info *info)
520{
521 const char *flag_str = space_info_flag_to_str(info);
522 lockdep_assert_held(&info->lock);
523
524 /* The free space could be negative in case of overcommit */
525 btrfs_info(fs_info, "space_info %s has %lld free, is %sfull",
526 flag_str,
527 (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
528 info->full ? "" : "not ");
529 btrfs_info(fs_info,
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);
534}
535
536void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
537 struct btrfs_space_info *info, u64 bytes,
538 int dump_block_groups)
539{
540 struct btrfs_block_group *cache;
541 u64 total_avail = 0;
542 int index = 0;
543
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);
548
549 if (!dump_block_groups)
550 return;
551
552 down_read(&info->groups_sem);
553again:
554 list_for_each_entry(cache, &info->block_groups[index], list) {
555 u64 avail;
556
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;
561 btrfs_info(fs_info,
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;
570 }
571 if (++index < BTRFS_NR_RAID_TYPES)
572 goto again;
573 up_read(&info->groups_sem);
574
575 btrfs_info(fs_info, "%llu bytes available across all block groups", total_avail);
576}
577
578static inline u64 calc_reclaim_items_nr(const struct btrfs_fs_info *fs_info,
579 u64 to_reclaim)
580{
581 u64 bytes;
582 u64 nr;
583
584 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
585 nr = div64_u64(to_reclaim, bytes);
586 if (!nr)
587 nr = 1;
588 return nr;
589}
590
591#define EXTENT_SIZE_PER_ITEM SZ_256K
592
593/*
594 * shrink metadata reservation for delalloc
595 */
596static void shrink_delalloc(struct btrfs_fs_info *fs_info,
597 struct btrfs_space_info *space_info,
598 u64 to_reclaim, bool wait_ordered,
599 bool for_preempt)
600{
601 struct btrfs_trans_handle *trans;
602 u64 delalloc_bytes;
603 u64 ordered_bytes;
604 u64 items;
605 long time_left;
606 int loops;
607
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)
611 return;
612
613 /* Calc the number of the pages we need flush for space reservation */
614 if (to_reclaim == U64_MAX) {
615 items = U64_MAX;
616 } else {
617 /*
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
627 * aggressive.
628 */
629 to_reclaim = max(to_reclaim, delalloc_bytes >> 3);
630 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
631 }
632
633 trans = current->journal_info;
634
635 /*
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.
639 */
640 if (ordered_bytes > delalloc_bytes && !for_preempt)
641 wait_ordered = true;
642
643 loops = 0;
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);
647 int async_pages;
648
649 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
650
651 /*
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.
660 *
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.
667 *
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.
671 */
672 async_pages = atomic_read(&fs_info->async_delalloc_pages);
673 if (!async_pages)
674 goto skip_async;
675
676 /*
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
680 * before continuing.
681 */
682 if (async_pages > nr_pages)
683 async_pages -= nr_pages;
684 else
685 async_pages = 0;
686 wait_event(fs_info->async_submit_wait,
687 atomic_read(&fs_info->async_delalloc_pages) <=
688 async_pages);
689skip_async:
690 loops++;
691 if (wait_ordered && !trans) {
692 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
693 } else {
694 time_left = schedule_timeout_killable(1);
695 if (time_left)
696 break;
697 }
698
699 /*
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
702 * to anymore.
703 */
704 if (for_preempt)
705 break;
706
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);
711 break;
712 }
713 spin_unlock(&space_info->lock);
714
715 delalloc_bytes = percpu_counter_sum_positive(
716 &fs_info->delalloc_bytes);
717 ordered_bytes = percpu_counter_sum_positive(
718 &fs_info->ordered_bytes);
719 }
720}
721
722/*
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.
726 */
727static 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)
730{
731 struct btrfs_root *root = fs_info->tree_root;
732 struct btrfs_trans_handle *trans;
733 int nr;
734 int ret = 0;
735
736 switch (state) {
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;
741 else
742 nr = -1;
743
744 trans = btrfs_join_transaction_nostart(root);
745 if (IS_ERR(trans)) {
746 ret = PTR_ERR(trans);
747 if (ret == -ENOENT)
748 ret = 0;
749 break;
750 }
751 ret = btrfs_run_delayed_items_nr(trans, nr);
752 btrfs_end_transaction(trans);
753 break;
754 case FLUSH_DELALLOC:
755 case FLUSH_DELALLOC_WAIT:
756 case FLUSH_DELALLOC_FULL:
757 if (state == FLUSH_DELALLOC_FULL)
758 num_bytes = U64_MAX;
759 shrink_delalloc(fs_info, space_info, num_bytes,
760 state != FLUSH_DELALLOC, for_preempt);
761 break;
762 case FLUSH_DELAYED_REFS_NR:
763 case FLUSH_DELAYED_REFS:
764 trans = btrfs_join_transaction_nostart(root);
765 if (IS_ERR(trans)) {
766 ret = PTR_ERR(trans);
767 if (ret == -ENOENT)
768 ret = 0;
769 break;
770 }
771 if (state == FLUSH_DELAYED_REFS_NR)
772 btrfs_run_delayed_refs(trans, num_bytes);
773 else
774 btrfs_run_delayed_refs(trans, 0);
775 btrfs_end_transaction(trans);
776 break;
777 case ALLOC_CHUNK:
778 case ALLOC_CHUNK_FORCE:
779 trans = btrfs_join_transaction(root);
780 if (IS_ERR(trans)) {
781 ret = PTR_ERR(trans);
782 break;
783 }
784 ret = btrfs_chunk_alloc(trans,
785 btrfs_get_alloc_profile(fs_info, space_info->flags),
786 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
787 CHUNK_ALLOC_FORCE);
788 btrfs_end_transaction(trans);
789
790 if (ret > 0 || ret == -ENOSPC)
791 ret = 0;
792 break;
793 case RUN_DELAYED_IPUTS:
794 /*
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.
798 */
799 btrfs_run_delayed_iputs(fs_info);
800 btrfs_wait_on_delayed_iputs(fs_info);
801 break;
802 case COMMIT_TRANS:
803 ASSERT(current->journal_info == NULL);
804 /*
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).
810 */
811 trans = btrfs_attach_transaction_barrier(root);
812 if (IS_ERR(trans)) {
813 ret = PTR_ERR(trans);
814 if (ret == -ENOENT)
815 ret = 0;
816 break;
817 }
818 ret = btrfs_commit_transaction(trans);
819 break;
820 default:
821 ret = -ENOSPC;
822 break;
823 }
824
825 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
826 ret, for_preempt);
827 return;
828}
829
830static inline u64
831btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
832 struct btrfs_space_info *space_info)
833{
834 u64 used;
835 u64 avail;
836 u64 to_reclaim = space_info->reclaim_size;
837
838 lockdep_assert_held(&space_info->lock);
839
840 avail = calc_available_free_space(fs_info, space_info,
841 BTRFS_RESERVE_FLUSH_ALL);
842 used = btrfs_space_info_used(space_info, true);
843
844 /*
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.
849 */
850 if (space_info->total_bytes + avail < used)
851 to_reclaim += used - (space_info->total_bytes + avail);
852
853 return to_reclaim;
854}
855
856static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
857 struct btrfs_space_info *space_info)
858{
859 const u64 global_rsv_size = btrfs_block_rsv_reserved(&fs_info->global_block_rsv);
860 u64 ordered, delalloc;
861 u64 thresh;
862 u64 used;
863
864 thresh = mult_perc(space_info->total_bytes, 90);
865
866 lockdep_assert_held(&space_info->lock);
867
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)
871 return false;
872
873 used = space_info->bytes_may_use + space_info->bytes_pinned;
874
875 /* The total flushable belongs to the global rsv, don't flush. */
876 if (global_rsv_size >= used)
877 return false;
878
879 /*
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.
883 */
884 if (used - global_rsv_size <= SZ_128M)
885 return false;
886
887 /*
888 * We have tickets queued, bail so we don't compete with the async
889 * flushers.
890 */
891 if (space_info->reclaim_size)
892 return false;
893
894 /*
895 * If we have over half of the free space occupied by reservations or
896 * pinned then we want to start flushing.
897 *
898 * We do not do the traditional thing here, which is to say
899 *
900 * if (used >= ((total_bytes + avail) / 2))
901 * return 1;
902 *
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.
907 *
908 * Our clamping range is 2^1 -> 2^8. Practically speaking that means
909 * the following:
910 *
911 * Amount of RAM Minimum threshold Maximum threshold
912 *
913 * 256GiB 1GiB 128GiB
914 * 128GiB 512MiB 64GiB
915 * 64GiB 256MiB 32GiB
916 * 32GiB 128MiB 16GiB
917 * 16GiB 64MiB 8GiB
918 *
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.
921 */
922
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;
930
931 used = space_info->bytes_pinned;
932
933 /*
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
941 * extents.
942 *
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.
949 *
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.
955 */
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);
961 else
962 used += space_info->bytes_may_use - global_rsv_size;
963
964 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
965 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
966}
967
968static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
969 struct btrfs_space_info *space_info,
970 struct reserve_ticket *ticket)
971{
972 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
973 u64 min_bytes;
974
975 if (!ticket->steal)
976 return false;
977
978 if (global_rsv->space_info != space_info)
979 return false;
980
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);
985 return false;
986 }
987 global_rsv->reserved -= ticket->bytes;
988 remove_ticket(space_info, ticket);
989 ticket->bytes = 0;
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);
995
996 return true;
997}
998
999/*
1000 * We've exhausted our flushing, start failing tickets.
1001 *
1002 * @fs_info - fs_info for this fs
1003 * @space_info - the space info we were flushing
1004 *
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.
1010 *
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
1013 * first ticket.
1014 */
1015static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
1016 struct btrfs_space_info *space_info)
1017{
1018 struct reserve_ticket *ticket;
1019 u64 tickets_id = space_info->tickets_id;
1020 const bool aborted = BTRFS_FS_ERROR(fs_info);
1021
1022 trace_btrfs_fail_all_tickets(fs_info, space_info);
1023
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);
1027 }
1028
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);
1033
1034 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket))
1035 return true;
1036
1037 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1038 btrfs_info(fs_info, "failing ticket with %llu bytes",
1039 ticket->bytes);
1040
1041 remove_ticket(space_info, ticket);
1042 if (aborted)
1043 ticket->error = -EIO;
1044 else
1045 ticket->error = -ENOSPC;
1046 wake_up(&ticket->wait);
1047
1048 /*
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
1052 * the list.
1053 */
1054 if (!aborted)
1055 btrfs_try_granting_tickets(fs_info, space_info);
1056 }
1057 return (tickets_id != space_info->tickets_id);
1058}
1059
1060/*
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.
1064 */
1065static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
1066{
1067 struct btrfs_fs_info *fs_info;
1068 struct btrfs_space_info *space_info;
1069 u64 to_reclaim;
1070 enum btrfs_flush_state flush_state;
1071 int commit_cycles = 0;
1072 u64 last_tickets_id;
1073
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);
1076
1077 spin_lock(&space_info->lock);
1078 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1079 if (!to_reclaim) {
1080 space_info->flush = 0;
1081 spin_unlock(&space_info->lock);
1082 return;
1083 }
1084 last_tickets_id = space_info->tickets_id;
1085 spin_unlock(&space_info->lock);
1086
1087 flush_state = FLUSH_DELAYED_ITEMS_NR;
1088 do {
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);
1094 return;
1095 }
1096 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
1097 space_info);
1098 if (last_tickets_id == space_info->tickets_id) {
1099 flush_state++;
1100 } else {
1101 last_tickets_id = space_info->tickets_id;
1102 flush_state = FLUSH_DELAYED_ITEMS_NR;
1103 if (commit_cycles)
1104 commit_cycles--;
1105 }
1106
1107 /*
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.
1111 */
1112 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles)
1113 flush_state++;
1114
1115 /*
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.
1124 */
1125 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
1126 flush_state++;
1127
1128 if (flush_state > COMMIT_TRANS) {
1129 commit_cycles++;
1130 if (commit_cycles > 2) {
1131 if (maybe_fail_all_tickets(fs_info, space_info)) {
1132 flush_state = FLUSH_DELAYED_ITEMS_NR;
1133 commit_cycles--;
1134 } else {
1135 space_info->flush = 0;
1136 }
1137 } else {
1138 flush_state = FLUSH_DELAYED_ITEMS_NR;
1139 }
1140 }
1141 spin_unlock(&space_info->lock);
1142 } while (flush_state <= COMMIT_TRANS);
1143}
1144
1145/*
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
1151 * largest.
1152 */
1153static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
1154{
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;
1161 int loops = 0;
1162
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;
1170
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);
1177
1178 loops++;
1179
1180 /*
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.
1186 */
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;
1193
1194 /*
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.
1198 */
1199 block_rsv_size -= global_rsv_size;
1200
1201 /*
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.
1205 */
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;
1218 } else {
1219 to_reclaim = btrfs_block_rsv_reserved(delayed_refs_rsv);
1220 flush = FLUSH_DELAYED_REFS_NR;
1221 }
1222
1223 spin_unlock(&space_info->lock);
1224
1225 /*
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.
1229 */
1230 to_reclaim >>= 2;
1231 if (!to_reclaim)
1232 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1233 flush_space(fs_info, space_info, to_reclaim, flush, true);
1234 cond_resched();
1235 spin_lock(&space_info->lock);
1236 }
1237
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);
1243}
1244
1245/*
1246 * FLUSH_DELALLOC_WAIT:
1247 * Space is freed from flushing delalloc in one of two ways.
1248 *
1249 * 1) compression is on and we allocate less space than we reserved
1250 * 2) we are overwriting existing space
1251 *
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
1255 * ->bytes_may_use.
1256 *
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.
1261 *
1262 * RUN_DELAYED_IPUTS
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.
1268 *
1269 * COMMIT_TRANS
1270 * This is where we reclaim all of the pinned space generated by running the
1271 * iputs
1272 *
1273 * ALLOC_CHUNK_FORCE
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.
1277 */
1278static const enum btrfs_flush_state data_flush_states[] = {
1279 FLUSH_DELALLOC_FULL,
1280 RUN_DELAYED_IPUTS,
1281 COMMIT_TRANS,
1282 ALLOC_CHUNK_FORCE,
1283};
1284
1285static void btrfs_async_reclaim_data_space(struct work_struct *work)
1286{
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;
1291
1292 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1293 space_info = fs_info->data_sinfo;
1294
1295 spin_lock(&space_info->lock);
1296 if (list_empty(&space_info->tickets)) {
1297 space_info->flush = 0;
1298 spin_unlock(&space_info->lock);
1299 return;
1300 }
1301 last_tickets_id = space_info->tickets_id;
1302 spin_unlock(&space_info->lock);
1303
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);
1310 return;
1311 }
1312
1313 /* Something happened, fail everything and bail. */
1314 if (BTRFS_FS_ERROR(fs_info))
1315 goto aborted_fs;
1316 last_tickets_id = space_info->tickets_id;
1317 spin_unlock(&space_info->lock);
1318 }
1319
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);
1327 return;
1328 }
1329
1330 if (last_tickets_id == space_info->tickets_id) {
1331 flush_state++;
1332 } else {
1333 last_tickets_id = space_info->tickets_id;
1334 flush_state = 0;
1335 }
1336
1337 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1338 if (space_info->full) {
1339 if (maybe_fail_all_tickets(fs_info, space_info))
1340 flush_state = 0;
1341 else
1342 space_info->flush = 0;
1343 } else {
1344 flush_state = 0;
1345 }
1346
1347 /* Something happened, fail everything and bail. */
1348 if (BTRFS_FS_ERROR(fs_info))
1349 goto aborted_fs;
1350
1351 }
1352 spin_unlock(&space_info->lock);
1353 }
1354 return;
1355
1356aborted_fs:
1357 maybe_fail_all_tickets(fs_info, space_info);
1358 space_info->flush = 0;
1359 spin_unlock(&space_info->lock);
1360}
1361
1362void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1363{
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);
1368}
1369
1370static const enum btrfs_flush_state priority_flush_states[] = {
1371 FLUSH_DELAYED_ITEMS_NR,
1372 FLUSH_DELAYED_ITEMS,
1373 ALLOC_CHUNK,
1374};
1375
1376static const enum btrfs_flush_state evict_flush_states[] = {
1377 FLUSH_DELAYED_ITEMS_NR,
1378 FLUSH_DELAYED_ITEMS,
1379 FLUSH_DELAYED_REFS_NR,
1380 FLUSH_DELAYED_REFS,
1381 FLUSH_DELALLOC,
1382 FLUSH_DELALLOC_WAIT,
1383 FLUSH_DELALLOC_FULL,
1384 ALLOC_CHUNK,
1385 COMMIT_TRANS,
1386};
1387
1388static 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,
1392 int states_nr)
1393{
1394 u64 to_reclaim;
1395 int flush_state = 0;
1396
1397 spin_lock(&space_info->lock);
1398 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1399 /*
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.
1404 */
1405 if (ticket->bytes == 0) {
1406 spin_unlock(&space_info->lock);
1407 return;
1408 }
1409
1410 while (flush_state < states_nr) {
1411 spin_unlock(&space_info->lock);
1412 flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1413 false);
1414 flush_state++;
1415 spin_lock(&space_info->lock);
1416 if (ticket->bytes == 0) {
1417 spin_unlock(&space_info->lock);
1418 return;
1419 }
1420 }
1421
1422 /*
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.
1429 */
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);
1436 }
1437
1438 /*
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.
1442 */
1443 btrfs_try_granting_tickets(fs_info, space_info);
1444 spin_unlock(&space_info->lock);
1445}
1446
1447static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1448 struct btrfs_space_info *space_info,
1449 struct reserve_ticket *ticket)
1450{
1451 spin_lock(&space_info->lock);
1452
1453 /* We could have been granted before we got here. */
1454 if (ticket->bytes == 0) {
1455 spin_unlock(&space_info->lock);
1456 return;
1457 }
1458
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);
1465 return;
1466 }
1467 }
1468
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);
1473}
1474
1475static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1476 struct btrfs_space_info *space_info,
1477 struct reserve_ticket *ticket)
1478
1479{
1480 DEFINE_WAIT(wait);
1481 int ret = 0;
1482
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);
1486 if (ret) {
1487 /*
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).
1494 */
1495 remove_ticket(space_info, ticket);
1496 ticket->error = -EINTR;
1497 break;
1498 }
1499 spin_unlock(&space_info->lock);
1500
1501 schedule();
1502
1503 finish_wait(&ticket->wait, &wait);
1504 spin_lock(&space_info->lock);
1505 }
1506 spin_unlock(&space_info->lock);
1507}
1508
1509/*
1510 * Do the appropriate flushing and waiting for a ticket.
1511 *
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
1518 *
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.
1521 */
1522static 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)
1527{
1528 int ret;
1529
1530 switch (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);
1535 break;
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));
1540 break;
1541 case BTRFS_RESERVE_FLUSH_EVICT:
1542 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1543 evict_flush_states,
1544 ARRAY_SIZE(evict_flush_states));
1545 break;
1546 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1547 priority_reclaim_data_space(fs_info, space_info, ticket);
1548 break;
1549 default:
1550 ASSERT(0);
1551 break;
1552 }
1553
1554 ret = ticket->error;
1555 ASSERT(list_empty(&ticket->list));
1556 /*
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).
1561 */
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);
1565 return ret;
1566}
1567
1568/*
1569 * This returns true if this flush state will go through the ordinary flushing
1570 * code.
1571 */
1572static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1573{
1574 return (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1575 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1576}
1577
1578static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1579 struct btrfs_space_info *space_info)
1580{
1581 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1582 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1583
1584 /*
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.
1591 */
1592 if (ordered < delalloc)
1593 space_info->clamp = min(space_info->clamp + 1, 8);
1594}
1595
1596static inline bool can_steal(enum btrfs_reserve_flush_enum flush)
1597{
1598 return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1599 flush == BTRFS_RESERVE_FLUSH_EVICT);
1600}
1601
1602/*
1603 * NO_FLUSH and FLUSH_EMERGENCY don't want to create a ticket, they just want to
1604 * fail as quickly as possible.
1605 */
1606static inline bool can_ticket(enum btrfs_reserve_flush_enum flush)
1607{
1608 return (flush != BTRFS_RESERVE_NO_FLUSH &&
1609 flush != BTRFS_RESERVE_FLUSH_EMERGENCY);
1610}
1611
1612/*
1613 * Try to reserve bytes from the block_rsv's space.
1614 *
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
1619 *
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
1625 * space already.
1626 */
1627static 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)
1630{
1631 struct work_struct *async_work;
1632 struct reserve_ticket ticket;
1633 u64 start_ns = 0;
1634 u64 used;
1635 int ret = -ENOSPC;
1636 bool pending_tickets;
1637
1638 ASSERT(orig_bytes);
1639 /*
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.
1644 */
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);
1650 }
1651
1652 if (flush == BTRFS_RESERVE_FLUSH_DATA)
1653 async_work = &fs_info->async_data_reclaim_work;
1654 else
1655 async_work = &fs_info->async_reclaim_work;
1656
1657 spin_lock(&space_info->lock);
1658 used = btrfs_space_info_used(space_info, true);
1659
1660 /*
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.
1664 */
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);
1668 else
1669 pending_tickets = !list_empty(&space_info->priority_tickets);
1670
1671 /*
1672 * Carry on if we have enough space (short-circuit) OR call
1673 * can_overcommit() to ensure we can overcommit to continue.
1674 */
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,
1679 orig_bytes);
1680 ret = 0;
1681 }
1682
1683 /*
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.
1687 */
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,
1692 orig_bytes);
1693 ret = 0;
1694 }
1695 }
1696
1697 /*
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.
1700 *
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.
1703 */
1704 if (ret && can_ticket(flush)) {
1705 ticket.bytes = orig_bytes;
1706 ticket.error = 0;
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();
1712
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) {
1718 /*
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
1723 * the workload.
1724 */
1725 maybe_clamp_preempt(fs_info, space_info);
1726
1727 space_info->flush = 1;
1728 trace_btrfs_trigger_flush(fs_info,
1729 space_info->flags,
1730 orig_bytes, flush,
1731 "enospc");
1732 queue_work(system_unbound_wq, async_work);
1733 }
1734 } else {
1735 list_add_tail(&ticket.list,
1736 &space_info->priority_tickets);
1737 }
1738 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1739 /*
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.
1743 */
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);
1751 }
1752 }
1753 spin_unlock(&space_info->lock);
1754 if (!ret || !can_ticket(flush))
1755 return ret;
1756
1757 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1758 orig_bytes, flush);
1759}
1760
1761/*
1762 * Try to reserve metadata bytes from the block_rsv's space.
1763 *
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
1768 *
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
1774 * space already.
1775 */
1776int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1777 struct btrfs_space_info *space_info,
1778 u64 orig_bytes,
1779 enum btrfs_reserve_flush_enum flush)
1780{
1781 int ret;
1782
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);
1787
1788 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1789 btrfs_dump_space_info(fs_info, space_info, orig_bytes, 0);
1790 }
1791 return ret;
1792}
1793
1794/*
1795 * Try to reserve data bytes for an allocation.
1796 *
1797 * @fs_info: the filesystem
1798 * @bytes: number of bytes we need
1799 * @flush: how we are allowed to flush
1800 *
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.
1803 */
1804int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1805 enum btrfs_reserve_flush_enum flush)
1806{
1807 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1808 int ret;
1809
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);
1814
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);
1821 }
1822 return ret;
1823}
1824
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)
1827{
1828 struct btrfs_space_info *space_info;
1829
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);
1835 }
1836 dump_global_block_rsv(fs_info);
1837}
1838
1839/*
1840 * Account the unused space of all the readonly block group in the space_info.
1841 * takes mirrors into account.
1842 */
1843u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
1844{
1845 struct btrfs_block_group *block_group;
1846 u64 free_bytes = 0;
1847 int factor;
1848
1849 /* It's df, we don't care if it's racy */
1850 if (list_empty(&sinfo->ro_bgs))
1851 return 0;
1852
1853 spin_lock(&sinfo->lock);
1854 list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
1855 spin_lock(&block_group->lock);
1856
1857 if (!block_group->ro) {
1858 spin_unlock(&block_group->lock);
1859 continue;
1860 }
1861
1862 factor = btrfs_bg_type_to_factor(block_group->flags);
1863 free_bytes += (block_group->length -
1864 block_group->used) * factor;
1865
1866 spin_unlock(&block_group->lock);
1867 }
1868 spin_unlock(&sinfo->lock);
1869
1870 return free_bytes;
1871}
1// SPDX-License-Identifier: GPL-2.0
2
3#include "misc.h"
4#include "ctree.h"
5#include "space-info.h"
6#include "sysfs.h"
7#include "volumes.h"
8#include "free-space-cache.h"
9#include "ordered-data.h"
10#include "transaction.h"
11#include "block-group.h"
12#include "fs.h"
13#include "accessors.h"
14#include "extent-tree.h"
15
16/*
17 * HOW DOES SPACE RESERVATION WORK
18 *
19 * If you want to know about delalloc specifically, there is a separate comment
20 * for that with the delalloc code. This comment is about how the whole system
21 * works generally.
22 *
23 * BASIC CONCEPTS
24 *
25 * 1) space_info. This is the ultimate arbiter of how much space we can use.
26 * There's a description of the bytes_ fields with the struct declaration,
27 * refer to that for specifics on each field. Suffice it to say that for
28 * reservations we care about total_bytes - SUM(space_info->bytes_) when
29 * determining if there is space to make an allocation. There is a space_info
30 * for METADATA, SYSTEM, and DATA areas.
31 *
32 * 2) block_rsv's. These are basically buckets for every different type of
33 * metadata reservation we have. You can see the comment in the block_rsv
34 * code on the rules for each type, but generally block_rsv->reserved is how
35 * much space is accounted for in space_info->bytes_may_use.
36 *
37 * 3) btrfs_calc*_size. These are the worst case calculations we used based
38 * on the number of items we will want to modify. We have one for changing
39 * items, and one for inserting new items. Generally we use these helpers to
40 * determine the size of the block reserves, and then use the actual bytes
41 * values to adjust the space_info counters.
42 *
43 * MAKING RESERVATIONS, THE NORMAL CASE
44 *
45 * We call into either btrfs_reserve_data_bytes() or
46 * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with
47 * num_bytes we want to reserve.
48 *
49 * ->reserve
50 * space_info->bytes_may_reserve += num_bytes
51 *
52 * ->extent allocation
53 * Call btrfs_add_reserved_bytes() which does
54 * space_info->bytes_may_reserve -= num_bytes
55 * space_info->bytes_reserved += extent_bytes
56 *
57 * ->insert reference
58 * Call btrfs_update_block_group() which does
59 * space_info->bytes_reserved -= extent_bytes
60 * space_info->bytes_used += extent_bytes
61 *
62 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority)
63 *
64 * Assume we are unable to simply make the reservation because we do not have
65 * enough space
66 *
67 * -> __reserve_bytes
68 * create a reserve_ticket with ->bytes set to our reservation, add it to
69 * the tail of space_info->tickets, kick async flush thread
70 *
71 * ->handle_reserve_ticket
72 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set
73 * on the ticket.
74 *
75 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space
76 * Flushes various things attempting to free up space.
77 *
78 * -> btrfs_try_granting_tickets()
79 * This is called by anything that either subtracts space from
80 * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the
81 * space_info->total_bytes. This loops through the ->priority_tickets and
82 * then the ->tickets list checking to see if the reservation can be
83 * completed. If it can the space is added to space_info->bytes_may_use and
84 * the ticket is woken up.
85 *
86 * -> ticket wakeup
87 * Check if ->bytes == 0, if it does we got our reservation and we can carry
88 * on, if not return the appropriate error (ENOSPC, but can be EINTR if we
89 * were interrupted.)
90 *
91 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY
92 *
93 * Same as the above, except we add ourselves to the
94 * space_info->priority_tickets, and we do not use ticket->wait, we simply
95 * call flush_space() ourselves for the states that are safe for us to call
96 * without deadlocking and hope for the best.
97 *
98 * THE FLUSHING STATES
99 *
100 * Generally speaking we will have two cases for each state, a "nice" state
101 * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to
102 * reduce the locking over head on the various trees, and even to keep from
103 * doing any work at all in the case of delayed refs. Each of these delayed
104 * things however hold reservations, and so letting them run allows us to
105 * reclaim space so we can make new reservations.
106 *
107 * FLUSH_DELAYED_ITEMS
108 * Every inode has a delayed item to update the inode. Take a simple write
109 * for example, we would update the inode item at write time to update the
110 * mtime, and then again at finish_ordered_io() time in order to update the
111 * isize or bytes. We keep these delayed items to coalesce these operations
112 * into a single operation done on demand. These are an easy way to reclaim
113 * metadata space.
114 *
115 * FLUSH_DELALLOC
116 * Look at the delalloc comment to get an idea of how much space is reserved
117 * for delayed allocation. We can reclaim some of this space simply by
118 * running delalloc, but usually we need to wait for ordered extents to
119 * reclaim the bulk of this space.
120 *
121 * FLUSH_DELAYED_REFS
122 * We have a block reserve for the outstanding delayed refs space, and every
123 * delayed ref operation holds a reservation. Running these is a quick way
124 * to reclaim space, but we want to hold this until the end because COW can
125 * churn a lot and we can avoid making some extent tree modifications if we
126 * are able to delay for as long as possible.
127 *
128 * ALLOC_CHUNK
129 * We will skip this the first time through space reservation, because of
130 * overcommit and we don't want to have a lot of useless metadata space when
131 * our worst case reservations will likely never come true.
132 *
133 * RUN_DELAYED_IPUTS
134 * If we're freeing inodes we're likely freeing checksums, file extent
135 * items, and extent tree items. Loads of space could be freed up by these
136 * operations, however they won't be usable until the transaction commits.
137 *
138 * COMMIT_TRANS
139 * This will commit the transaction. Historically we had a lot of logic
140 * surrounding whether or not we'd commit the transaction, but this waits born
141 * out of a pre-tickets era where we could end up committing the transaction
142 * thousands of times in a row without making progress. Now thanks to our
143 * ticketing system we know if we're not making progress and can error
144 * everybody out after a few commits rather than burning the disk hoping for
145 * a different answer.
146 *
147 * OVERCOMMIT
148 *
149 * Because we hold so many reservations for metadata we will allow you to
150 * reserve more space than is currently free in the currently allocate
151 * metadata space. This only happens with metadata, data does not allow
152 * overcommitting.
153 *
154 * You can see the current logic for when we allow overcommit in
155 * btrfs_can_overcommit(), but it only applies to unallocated space. If there
156 * is no unallocated space to be had, all reservations are kept within the
157 * free space in the allocated metadata chunks.
158 *
159 * Because of overcommitting, you generally want to use the
160 * btrfs_can_overcommit() logic for metadata allocations, as it does the right
161 * thing with or without extra unallocated space.
162 */
163
164u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
165 bool may_use_included)
166{
167 ASSERT(s_info);
168 return s_info->bytes_used + s_info->bytes_reserved +
169 s_info->bytes_pinned + s_info->bytes_readonly +
170 s_info->bytes_zone_unusable +
171 (may_use_included ? s_info->bytes_may_use : 0);
172}
173
174/*
175 * after adding space to the filesystem, we need to clear the full flags
176 * on all the space infos.
177 */
178void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
179{
180 struct list_head *head = &info->space_info;
181 struct btrfs_space_info *found;
182
183 list_for_each_entry(found, head, list)
184 found->full = 0;
185}
186
187/*
188 * Block groups with more than this value (percents) of unusable space will be
189 * scheduled for background reclaim.
190 */
191#define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH (75)
192
193/*
194 * Calculate chunk size depending on volume type (regular or zoned).
195 */
196static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags)
197{
198 if (btrfs_is_zoned(fs_info))
199 return fs_info->zone_size;
200
201 ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
202
203 if (flags & BTRFS_BLOCK_GROUP_DATA)
204 return BTRFS_MAX_DATA_CHUNK_SIZE;
205 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
206 return SZ_32M;
207
208 /* Handle BTRFS_BLOCK_GROUP_METADATA */
209 if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G)
210 return SZ_1G;
211
212 return SZ_256M;
213}
214
215/*
216 * Update default chunk size.
217 */
218void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info,
219 u64 chunk_size)
220{
221 WRITE_ONCE(space_info->chunk_size, chunk_size);
222}
223
224static int create_space_info(struct btrfs_fs_info *info, u64 flags)
225{
226
227 struct btrfs_space_info *space_info;
228 int i;
229 int ret;
230
231 space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
232 if (!space_info)
233 return -ENOMEM;
234
235 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
236 INIT_LIST_HEAD(&space_info->block_groups[i]);
237 init_rwsem(&space_info->groups_sem);
238 spin_lock_init(&space_info->lock);
239 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
240 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
241 INIT_LIST_HEAD(&space_info->ro_bgs);
242 INIT_LIST_HEAD(&space_info->tickets);
243 INIT_LIST_HEAD(&space_info->priority_tickets);
244 space_info->clamp = 1;
245 btrfs_update_space_info_chunk_size(space_info, calc_chunk_size(info, flags));
246
247 if (btrfs_is_zoned(info))
248 space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH;
249
250 ret = btrfs_sysfs_add_space_info_type(info, space_info);
251 if (ret)
252 return ret;
253
254 list_add(&space_info->list, &info->space_info);
255 if (flags & BTRFS_BLOCK_GROUP_DATA)
256 info->data_sinfo = space_info;
257
258 return ret;
259}
260
261int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
262{
263 struct btrfs_super_block *disk_super;
264 u64 features;
265 u64 flags;
266 int mixed = 0;
267 int ret;
268
269 disk_super = fs_info->super_copy;
270 if (!btrfs_super_root(disk_super))
271 return -EINVAL;
272
273 features = btrfs_super_incompat_flags(disk_super);
274 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
275 mixed = 1;
276
277 flags = BTRFS_BLOCK_GROUP_SYSTEM;
278 ret = create_space_info(fs_info, flags);
279 if (ret)
280 goto out;
281
282 if (mixed) {
283 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
284 ret = create_space_info(fs_info, flags);
285 } else {
286 flags = BTRFS_BLOCK_GROUP_METADATA;
287 ret = create_space_info(fs_info, flags);
288 if (ret)
289 goto out;
290
291 flags = BTRFS_BLOCK_GROUP_DATA;
292 ret = create_space_info(fs_info, flags);
293 }
294out:
295 return ret;
296}
297
298void btrfs_add_bg_to_space_info(struct btrfs_fs_info *info,
299 struct btrfs_block_group *block_group)
300{
301 struct btrfs_space_info *found;
302 int factor, index;
303
304 factor = btrfs_bg_type_to_factor(block_group->flags);
305
306 found = btrfs_find_space_info(info, block_group->flags);
307 ASSERT(found);
308 spin_lock(&found->lock);
309 found->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)
316 found->full = 0;
317 btrfs_try_granting_tickets(info, found);
318 spin_unlock(&found->lock);
319
320 block_group->space_info = found;
321
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);
326}
327
328struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
329 u64 flags)
330{
331 struct list_head *head = &info->space_info;
332 struct btrfs_space_info *found;
333
334 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
335
336 list_for_each_entry(found, head, list) {
337 if (found->flags & flags)
338 return found;
339 }
340 return NULL;
341}
342
343static 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)
346{
347 struct btrfs_space_info *data_sinfo;
348 u64 profile;
349 u64 avail;
350 u64 data_chunk_size;
351 int factor;
352
353 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
354 profile = btrfs_system_alloc_profile(fs_info);
355 else
356 profile = btrfs_metadata_alloc_profile(fs_info);
357
358 avail = atomic64_read(&fs_info->free_chunk_space);
359
360 /*
361 * If we have dup, raid1 or raid10 then only half of the free
362 * space is actually usable. For raid56, the space info used
363 * doesn't include the parity drive, so we don't have to
364 * change the math
365 */
366 factor = btrfs_bg_type_to_factor(profile);
367 avail = div_u64(avail, factor);
368 if (avail == 0)
369 return 0;
370
371 /*
372 * Calculate the data_chunk_size, space_info->chunk_size is the
373 * "optimal" chunk size based on the fs size. However when we actually
374 * allocate the chunk we will strip this down further, making it no more
375 * than 10% of the disk or 1G, whichever is smaller.
376 */
377 data_sinfo = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
378 data_chunk_size = min(data_sinfo->chunk_size,
379 mult_perc(fs_info->fs_devices->total_rw_bytes, 10));
380 data_chunk_size = min_t(u64, data_chunk_size, SZ_1G);
381
382 /*
383 * Since data allocations immediately use block groups as part of the
384 * reservation, because we assume that data reservations will == actual
385 * usage, we could potentially overcommit and then immediately have that
386 * available space used by a data allocation, which could put us in a
387 * bind when we get close to filling the file system.
388 *
389 * To handle this simply remove the data_chunk_size from the available
390 * space. If we are relatively empty this won't affect our ability to
391 * overcommit much, and if we're very close to full it'll keep us from
392 * getting into a position where we've given ourselves very little
393 * metadata wiggle room.
394 */
395 if (avail <= data_chunk_size)
396 return 0;
397 avail -= data_chunk_size;
398
399 /*
400 * If we aren't flushing all things, let us overcommit up to
401 * 1/2th of the space. If we can flush, don't let us overcommit
402 * too much, let it overcommit up to 1/8 of the space.
403 */
404 if (flush == BTRFS_RESERVE_FLUSH_ALL)
405 avail >>= 3;
406 else
407 avail >>= 1;
408 return avail;
409}
410
411int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
412 struct btrfs_space_info *space_info, u64 bytes,
413 enum btrfs_reserve_flush_enum flush)
414{
415 u64 avail;
416 u64 used;
417
418 /* Don't overcommit when in mixed mode */
419 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
420 return 0;
421
422 used = btrfs_space_info_used(space_info, true);
423 avail = calc_available_free_space(fs_info, space_info, flush);
424
425 if (used + bytes < space_info->total_bytes + avail)
426 return 1;
427 return 0;
428}
429
430static void remove_ticket(struct btrfs_space_info *space_info,
431 struct reserve_ticket *ticket)
432{
433 if (!list_empty(&ticket->list)) {
434 list_del_init(&ticket->list);
435 ASSERT(space_info->reclaim_size >= ticket->bytes);
436 space_info->reclaim_size -= ticket->bytes;
437 }
438}
439
440/*
441 * This is for space we already have accounted in space_info->bytes_may_use, so
442 * basically when we're returning space from block_rsv's.
443 */
444void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
445 struct btrfs_space_info *space_info)
446{
447 struct list_head *head;
448 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
449
450 lockdep_assert_held(&space_info->lock);
451
452 head = &space_info->priority_tickets;
453again:
454 while (!list_empty(head)) {
455 struct reserve_ticket *ticket;
456 u64 used = btrfs_space_info_used(space_info, true);
457
458 ticket = list_first_entry(head, struct reserve_ticket, list);
459
460 /* Check and see if our ticket can be satisfied now. */
461 if ((used + ticket->bytes <= space_info->total_bytes) ||
462 btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
463 flush)) {
464 btrfs_space_info_update_bytes_may_use(fs_info,
465 space_info,
466 ticket->bytes);
467 remove_ticket(space_info, ticket);
468 ticket->bytes = 0;
469 space_info->tickets_id++;
470 wake_up(&ticket->wait);
471 } else {
472 break;
473 }
474 }
475
476 if (head == &space_info->priority_tickets) {
477 head = &space_info->tickets;
478 flush = BTRFS_RESERVE_FLUSH_ALL;
479 goto again;
480 }
481}
482
483#define DUMP_BLOCK_RSV(fs_info, rsv_name) \
484do { \
485 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \
486 spin_lock(&__rsv->lock); \
487 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \
488 __rsv->size, __rsv->reserved); \
489 spin_unlock(&__rsv->lock); \
490} while (0)
491
492static const char *space_info_flag_to_str(const struct btrfs_space_info *space_info)
493{
494 switch (space_info->flags) {
495 case BTRFS_BLOCK_GROUP_SYSTEM:
496 return "SYSTEM";
497 case BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA:
498 return "DATA+METADATA";
499 case BTRFS_BLOCK_GROUP_DATA:
500 return "DATA";
501 case BTRFS_BLOCK_GROUP_METADATA:
502 return "METADATA";
503 default:
504 return "UNKNOWN";
505 }
506}
507
508static void dump_global_block_rsv(struct btrfs_fs_info *fs_info)
509{
510 DUMP_BLOCK_RSV(fs_info, global_block_rsv);
511 DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
512 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
513 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
514 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
515}
516
517static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
518 struct btrfs_space_info *info)
519{
520 const char *flag_str = space_info_flag_to_str(info);
521 lockdep_assert_held(&info->lock);
522
523 /* The free space could be negative in case of overcommit */
524 btrfs_info(fs_info, "space_info %s has %lld free, is %sfull",
525 flag_str,
526 (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
527 info->full ? "" : "not ");
528 btrfs_info(fs_info,
529"space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
530 info->total_bytes, info->bytes_used, info->bytes_pinned,
531 info->bytes_reserved, info->bytes_may_use,
532 info->bytes_readonly, info->bytes_zone_unusable);
533}
534
535void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
536 struct btrfs_space_info *info, u64 bytes,
537 int dump_block_groups)
538{
539 struct btrfs_block_group *cache;
540 u64 total_avail = 0;
541 int index = 0;
542
543 spin_lock(&info->lock);
544 __btrfs_dump_space_info(fs_info, info);
545 dump_global_block_rsv(fs_info);
546 spin_unlock(&info->lock);
547
548 if (!dump_block_groups)
549 return;
550
551 down_read(&info->groups_sem);
552again:
553 list_for_each_entry(cache, &info->block_groups[index], list) {
554 u64 avail;
555
556 spin_lock(&cache->lock);
557 avail = cache->length - cache->used - cache->pinned -
558 cache->reserved - cache->delalloc_bytes -
559 cache->bytes_super - cache->zone_unusable;
560 btrfs_info(fs_info,
561"block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu delalloc %llu super %llu zone_unusable (%llu bytes available) %s",
562 cache->start, cache->length, cache->used, cache->pinned,
563 cache->reserved, cache->delalloc_bytes,
564 cache->bytes_super, cache->zone_unusable,
565 avail, cache->ro ? "[readonly]" : "");
566 spin_unlock(&cache->lock);
567 btrfs_dump_free_space(cache, bytes);
568 total_avail += avail;
569 }
570 if (++index < BTRFS_NR_RAID_TYPES)
571 goto again;
572 up_read(&info->groups_sem);
573
574 btrfs_info(fs_info, "%llu bytes available across all block groups", total_avail);
575}
576
577static inline u64 calc_reclaim_items_nr(const struct btrfs_fs_info *fs_info,
578 u64 to_reclaim)
579{
580 u64 bytes;
581 u64 nr;
582
583 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
584 nr = div64_u64(to_reclaim, bytes);
585 if (!nr)
586 nr = 1;
587 return nr;
588}
589
590#define EXTENT_SIZE_PER_ITEM SZ_256K
591
592/*
593 * shrink metadata reservation for delalloc
594 */
595static void shrink_delalloc(struct btrfs_fs_info *fs_info,
596 struct btrfs_space_info *space_info,
597 u64 to_reclaim, bool wait_ordered,
598 bool for_preempt)
599{
600 struct btrfs_trans_handle *trans;
601 u64 delalloc_bytes;
602 u64 ordered_bytes;
603 u64 items;
604 long time_left;
605 int loops;
606
607 delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
608 ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes);
609 if (delalloc_bytes == 0 && ordered_bytes == 0)
610 return;
611
612 /* Calc the number of the pages we need flush for space reservation */
613 if (to_reclaim == U64_MAX) {
614 items = U64_MAX;
615 } else {
616 /*
617 * to_reclaim is set to however much metadata we need to
618 * reclaim, but reclaiming that much data doesn't really track
619 * exactly. What we really want to do is reclaim full inode's
620 * worth of reservations, however that's not available to us
621 * here. We will take a fraction of the delalloc bytes for our
622 * flushing loops and hope for the best. Delalloc will expand
623 * the amount we write to cover an entire dirty extent, which
624 * will reclaim the metadata reservation for that range. If
625 * it's not enough subsequent flush stages will be more
626 * aggressive.
627 */
628 to_reclaim = max(to_reclaim, delalloc_bytes >> 3);
629 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
630 }
631
632 trans = current->journal_info;
633
634 /*
635 * If we are doing more ordered than delalloc we need to just wait on
636 * ordered extents, otherwise we'll waste time trying to flush delalloc
637 * that likely won't give us the space back we need.
638 */
639 if (ordered_bytes > delalloc_bytes && !for_preempt)
640 wait_ordered = true;
641
642 loops = 0;
643 while ((delalloc_bytes || ordered_bytes) && loops < 3) {
644 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
645 long nr_pages = min_t(u64, temp, LONG_MAX);
646 int async_pages;
647
648 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
649
650 /*
651 * We need to make sure any outstanding async pages are now
652 * processed before we continue. This is because things like
653 * sync_inode() try to be smart and skip writing if the inode is
654 * marked clean. We don't use filemap_fwrite for flushing
655 * because we want to control how many pages we write out at a
656 * time, thus this is the only safe way to make sure we've
657 * waited for outstanding compressed workers to have started
658 * their jobs and thus have ordered extents set up properly.
659 *
660 * This exists because we do not want to wait for each
661 * individual inode to finish its async work, we simply want to
662 * start the IO on everybody, and then come back here and wait
663 * for all of the async work to catch up. Once we're done with
664 * that we know we'll have ordered extents for everything and we
665 * can decide if we wait for that or not.
666 *
667 * If we choose to replace this in the future, make absolutely
668 * sure that the proper waiting is being done in the async case,
669 * as there have been bugs in that area before.
670 */
671 async_pages = atomic_read(&fs_info->async_delalloc_pages);
672 if (!async_pages)
673 goto skip_async;
674
675 /*
676 * We don't want to wait forever, if we wrote less pages in this
677 * loop than we have outstanding, only wait for that number of
678 * pages, otherwise we can wait for all async pages to finish
679 * before continuing.
680 */
681 if (async_pages > nr_pages)
682 async_pages -= nr_pages;
683 else
684 async_pages = 0;
685 wait_event(fs_info->async_submit_wait,
686 atomic_read(&fs_info->async_delalloc_pages) <=
687 async_pages);
688skip_async:
689 loops++;
690 if (wait_ordered && !trans) {
691 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
692 } else {
693 time_left = schedule_timeout_killable(1);
694 if (time_left)
695 break;
696 }
697
698 /*
699 * If we are for preemption we just want a one-shot of delalloc
700 * flushing so we can stop flushing if we decide we don't need
701 * to anymore.
702 */
703 if (for_preempt)
704 break;
705
706 spin_lock(&space_info->lock);
707 if (list_empty(&space_info->tickets) &&
708 list_empty(&space_info->priority_tickets)) {
709 spin_unlock(&space_info->lock);
710 break;
711 }
712 spin_unlock(&space_info->lock);
713
714 delalloc_bytes = percpu_counter_sum_positive(
715 &fs_info->delalloc_bytes);
716 ordered_bytes = percpu_counter_sum_positive(
717 &fs_info->ordered_bytes);
718 }
719}
720
721/*
722 * Try to flush some data based on policy set by @state. This is only advisory
723 * and may fail for various reasons. The caller is supposed to examine the
724 * state of @space_info to detect the outcome.
725 */
726static void flush_space(struct btrfs_fs_info *fs_info,
727 struct btrfs_space_info *space_info, u64 num_bytes,
728 enum btrfs_flush_state state, bool for_preempt)
729{
730 struct btrfs_root *root = fs_info->tree_root;
731 struct btrfs_trans_handle *trans;
732 int nr;
733 int ret = 0;
734
735 switch (state) {
736 case FLUSH_DELAYED_ITEMS_NR:
737 case FLUSH_DELAYED_ITEMS:
738 if (state == FLUSH_DELAYED_ITEMS_NR)
739 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
740 else
741 nr = -1;
742
743 trans = btrfs_join_transaction_nostart(root);
744 if (IS_ERR(trans)) {
745 ret = PTR_ERR(trans);
746 if (ret == -ENOENT)
747 ret = 0;
748 break;
749 }
750 ret = btrfs_run_delayed_items_nr(trans, nr);
751 btrfs_end_transaction(trans);
752 break;
753 case FLUSH_DELALLOC:
754 case FLUSH_DELALLOC_WAIT:
755 case FLUSH_DELALLOC_FULL:
756 if (state == FLUSH_DELALLOC_FULL)
757 num_bytes = U64_MAX;
758 shrink_delalloc(fs_info, space_info, num_bytes,
759 state != FLUSH_DELALLOC, for_preempt);
760 break;
761 case FLUSH_DELAYED_REFS_NR:
762 case FLUSH_DELAYED_REFS:
763 trans = btrfs_join_transaction_nostart(root);
764 if (IS_ERR(trans)) {
765 ret = PTR_ERR(trans);
766 if (ret == -ENOENT)
767 ret = 0;
768 break;
769 }
770 if (state == FLUSH_DELAYED_REFS_NR)
771 btrfs_run_delayed_refs(trans, num_bytes);
772 else
773 btrfs_run_delayed_refs(trans, 0);
774 btrfs_end_transaction(trans);
775 break;
776 case ALLOC_CHUNK:
777 case ALLOC_CHUNK_FORCE:
778 trans = btrfs_join_transaction(root);
779 if (IS_ERR(trans)) {
780 ret = PTR_ERR(trans);
781 break;
782 }
783 ret = btrfs_chunk_alloc(trans,
784 btrfs_get_alloc_profile(fs_info, space_info->flags),
785 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
786 CHUNK_ALLOC_FORCE);
787 btrfs_end_transaction(trans);
788
789 if (ret > 0 || ret == -ENOSPC)
790 ret = 0;
791 break;
792 case RUN_DELAYED_IPUTS:
793 /*
794 * If we have pending delayed iputs then we could free up a
795 * bunch of pinned space, so make sure we run the iputs before
796 * we do our pinned bytes check below.
797 */
798 btrfs_run_delayed_iputs(fs_info);
799 btrfs_wait_on_delayed_iputs(fs_info);
800 break;
801 case COMMIT_TRANS:
802 ASSERT(current->journal_info == NULL);
803 /*
804 * We don't want to start a new transaction, just attach to the
805 * current one or wait it fully commits in case its commit is
806 * happening at the moment. Note: we don't use a nostart join
807 * because that does not wait for a transaction to fully commit
808 * (only for it to be unblocked, state TRANS_STATE_UNBLOCKED).
809 */
810 trans = btrfs_attach_transaction_barrier(root);
811 if (IS_ERR(trans)) {
812 ret = PTR_ERR(trans);
813 if (ret == -ENOENT)
814 ret = 0;
815 break;
816 }
817 ret = btrfs_commit_transaction(trans);
818 break;
819 default:
820 ret = -ENOSPC;
821 break;
822 }
823
824 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
825 ret, for_preempt);
826 return;
827}
828
829static inline u64
830btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
831 struct btrfs_space_info *space_info)
832{
833 u64 used;
834 u64 avail;
835 u64 to_reclaim = space_info->reclaim_size;
836
837 lockdep_assert_held(&space_info->lock);
838
839 avail = calc_available_free_space(fs_info, space_info,
840 BTRFS_RESERVE_FLUSH_ALL);
841 used = btrfs_space_info_used(space_info, true);
842
843 /*
844 * We may be flushing because suddenly we have less space than we had
845 * before, and now we're well over-committed based on our current free
846 * space. If that's the case add in our overage so we make sure to put
847 * appropriate pressure on the flushing state machine.
848 */
849 if (space_info->total_bytes + avail < used)
850 to_reclaim += used - (space_info->total_bytes + avail);
851
852 return to_reclaim;
853}
854
855static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
856 struct btrfs_space_info *space_info)
857{
858 const u64 global_rsv_size = btrfs_block_rsv_reserved(&fs_info->global_block_rsv);
859 u64 ordered, delalloc;
860 u64 thresh;
861 u64 used;
862
863 thresh = mult_perc(space_info->total_bytes, 90);
864
865 lockdep_assert_held(&space_info->lock);
866
867 /* If we're just plain full then async reclaim just slows us down. */
868 if ((space_info->bytes_used + space_info->bytes_reserved +
869 global_rsv_size) >= thresh)
870 return false;
871
872 used = space_info->bytes_may_use + space_info->bytes_pinned;
873
874 /* The total flushable belongs to the global rsv, don't flush. */
875 if (global_rsv_size >= used)
876 return false;
877
878 /*
879 * 128MiB is 1/4 of the maximum global rsv size. If we have less than
880 * that devoted to other reservations then there's no sense in flushing,
881 * we don't have a lot of things that need flushing.
882 */
883 if (used - global_rsv_size <= SZ_128M)
884 return false;
885
886 /*
887 * We have tickets queued, bail so we don't compete with the async
888 * flushers.
889 */
890 if (space_info->reclaim_size)
891 return false;
892
893 /*
894 * If we have over half of the free space occupied by reservations or
895 * pinned then we want to start flushing.
896 *
897 * We do not do the traditional thing here, which is to say
898 *
899 * if (used >= ((total_bytes + avail) / 2))
900 * return 1;
901 *
902 * because this doesn't quite work how we want. If we had more than 50%
903 * of the space_info used by bytes_used and we had 0 available we'd just
904 * constantly run the background flusher. Instead we want it to kick in
905 * if our reclaimable space exceeds our clamped free space.
906 *
907 * Our clamping range is 2^1 -> 2^8. Practically speaking that means
908 * the following:
909 *
910 * Amount of RAM Minimum threshold Maximum threshold
911 *
912 * 256GiB 1GiB 128GiB
913 * 128GiB 512MiB 64GiB
914 * 64GiB 256MiB 32GiB
915 * 32GiB 128MiB 16GiB
916 * 16GiB 64MiB 8GiB
917 *
918 * These are the range our thresholds will fall in, corresponding to how
919 * much delalloc we need for the background flusher to kick in.
920 */
921
922 thresh = calc_available_free_space(fs_info, space_info,
923 BTRFS_RESERVE_FLUSH_ALL);
924 used = space_info->bytes_used + space_info->bytes_reserved +
925 space_info->bytes_readonly + global_rsv_size;
926 if (used < space_info->total_bytes)
927 thresh += space_info->total_bytes - used;
928 thresh >>= space_info->clamp;
929
930 used = space_info->bytes_pinned;
931
932 /*
933 * If we have more ordered bytes than delalloc bytes then we're either
934 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting
935 * around. Preemptive flushing is only useful in that it can free up
936 * space before tickets need to wait for things to finish. In the case
937 * of ordered extents, preemptively waiting on ordered extents gets us
938 * nothing, if our reservations are tied up in ordered extents we'll
939 * simply have to slow down writers by forcing them to wait on ordered
940 * extents.
941 *
942 * In the case that ordered is larger than delalloc, only include the
943 * block reserves that we would actually be able to directly reclaim
944 * from. In this case if we're heavy on metadata operations this will
945 * clearly be heavy enough to warrant preemptive flushing. In the case
946 * of heavy DIO or ordered reservations, preemptive flushing will just
947 * waste time and cause us to slow down.
948 *
949 * We want to make sure we truly are maxed out on ordered however, so
950 * cut ordered in half, and if it's still higher than delalloc then we
951 * can keep flushing. This is to avoid the case where we start
952 * flushing, and now delalloc == ordered and we stop preemptively
953 * flushing when we could still have several gigs of delalloc to flush.
954 */
955 ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1;
956 delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes);
957 if (ordered >= delalloc)
958 used += btrfs_block_rsv_reserved(&fs_info->delayed_refs_rsv) +
959 btrfs_block_rsv_reserved(&fs_info->delayed_block_rsv);
960 else
961 used += space_info->bytes_may_use - global_rsv_size;
962
963 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
964 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
965}
966
967static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
968 struct btrfs_space_info *space_info,
969 struct reserve_ticket *ticket)
970{
971 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
972 u64 min_bytes;
973
974 if (!ticket->steal)
975 return false;
976
977 if (global_rsv->space_info != space_info)
978 return false;
979
980 spin_lock(&global_rsv->lock);
981 min_bytes = mult_perc(global_rsv->size, 10);
982 if (global_rsv->reserved < min_bytes + ticket->bytes) {
983 spin_unlock(&global_rsv->lock);
984 return false;
985 }
986 global_rsv->reserved -= ticket->bytes;
987 remove_ticket(space_info, ticket);
988 ticket->bytes = 0;
989 wake_up(&ticket->wait);
990 space_info->tickets_id++;
991 if (global_rsv->reserved < global_rsv->size)
992 global_rsv->full = 0;
993 spin_unlock(&global_rsv->lock);
994
995 return true;
996}
997
998/*
999 * We've exhausted our flushing, start failing tickets.
1000 *
1001 * @fs_info - fs_info for this fs
1002 * @space_info - the space info we were flushing
1003 *
1004 * We call this when we've exhausted our flushing ability and haven't made
1005 * progress in satisfying tickets. The reservation code handles tickets in
1006 * order, so if there is a large ticket first and then smaller ones we could
1007 * very well satisfy the smaller tickets. This will attempt to wake up any
1008 * tickets in the list to catch this case.
1009 *
1010 * This function returns true if it was able to make progress by clearing out
1011 * other tickets, or if it stumbles across a ticket that was smaller than the
1012 * first ticket.
1013 */
1014static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
1015 struct btrfs_space_info *space_info)
1016{
1017 struct reserve_ticket *ticket;
1018 u64 tickets_id = space_info->tickets_id;
1019 const bool aborted = BTRFS_FS_ERROR(fs_info);
1020
1021 trace_btrfs_fail_all_tickets(fs_info, space_info);
1022
1023 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1024 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
1025 __btrfs_dump_space_info(fs_info, space_info);
1026 }
1027
1028 while (!list_empty(&space_info->tickets) &&
1029 tickets_id == space_info->tickets_id) {
1030 ticket = list_first_entry(&space_info->tickets,
1031 struct reserve_ticket, list);
1032
1033 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket))
1034 return true;
1035
1036 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1037 btrfs_info(fs_info, "failing ticket with %llu bytes",
1038 ticket->bytes);
1039
1040 remove_ticket(space_info, ticket);
1041 if (aborted)
1042 ticket->error = -EIO;
1043 else
1044 ticket->error = -ENOSPC;
1045 wake_up(&ticket->wait);
1046
1047 /*
1048 * We're just throwing tickets away, so more flushing may not
1049 * trip over btrfs_try_granting_tickets, so we need to call it
1050 * here to see if we can make progress with the next ticket in
1051 * the list.
1052 */
1053 if (!aborted)
1054 btrfs_try_granting_tickets(fs_info, space_info);
1055 }
1056 return (tickets_id != space_info->tickets_id);
1057}
1058
1059/*
1060 * This is for normal flushers, we can wait all goddamned day if we want to. We
1061 * will loop and continuously try to flush as long as we are making progress.
1062 * We count progress as clearing off tickets each time we have to loop.
1063 */
1064static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
1065{
1066 struct btrfs_fs_info *fs_info;
1067 struct btrfs_space_info *space_info;
1068 u64 to_reclaim;
1069 enum btrfs_flush_state flush_state;
1070 int commit_cycles = 0;
1071 u64 last_tickets_id;
1072
1073 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
1074 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1075
1076 spin_lock(&space_info->lock);
1077 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1078 if (!to_reclaim) {
1079 space_info->flush = 0;
1080 spin_unlock(&space_info->lock);
1081 return;
1082 }
1083 last_tickets_id = space_info->tickets_id;
1084 spin_unlock(&space_info->lock);
1085
1086 flush_state = FLUSH_DELAYED_ITEMS_NR;
1087 do {
1088 flush_space(fs_info, space_info, to_reclaim, flush_state, false);
1089 spin_lock(&space_info->lock);
1090 if (list_empty(&space_info->tickets)) {
1091 space_info->flush = 0;
1092 spin_unlock(&space_info->lock);
1093 return;
1094 }
1095 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
1096 space_info);
1097 if (last_tickets_id == space_info->tickets_id) {
1098 flush_state++;
1099 } else {
1100 last_tickets_id = space_info->tickets_id;
1101 flush_state = FLUSH_DELAYED_ITEMS_NR;
1102 if (commit_cycles)
1103 commit_cycles--;
1104 }
1105
1106 /*
1107 * We do not want to empty the system of delalloc unless we're
1108 * under heavy pressure, so allow one trip through the flushing
1109 * logic before we start doing a FLUSH_DELALLOC_FULL.
1110 */
1111 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles)
1112 flush_state++;
1113
1114 /*
1115 * We don't want to force a chunk allocation until we've tried
1116 * pretty hard to reclaim space. Think of the case where we
1117 * freed up a bunch of space and so have a lot of pinned space
1118 * to reclaim. We would rather use that than possibly create a
1119 * underutilized metadata chunk. So if this is our first run
1120 * through the flushing state machine skip ALLOC_CHUNK_FORCE and
1121 * commit the transaction. If nothing has changed the next go
1122 * around then we can force a chunk allocation.
1123 */
1124 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
1125 flush_state++;
1126
1127 if (flush_state > COMMIT_TRANS) {
1128 commit_cycles++;
1129 if (commit_cycles > 2) {
1130 if (maybe_fail_all_tickets(fs_info, space_info)) {
1131 flush_state = FLUSH_DELAYED_ITEMS_NR;
1132 commit_cycles--;
1133 } else {
1134 space_info->flush = 0;
1135 }
1136 } else {
1137 flush_state = FLUSH_DELAYED_ITEMS_NR;
1138 }
1139 }
1140 spin_unlock(&space_info->lock);
1141 } while (flush_state <= COMMIT_TRANS);
1142}
1143
1144/*
1145 * This handles pre-flushing of metadata space before we get to the point that
1146 * we need to start blocking threads on tickets. The logic here is different
1147 * from the other flush paths because it doesn't rely on tickets to tell us how
1148 * much we need to flush, instead it attempts to keep us below the 80% full
1149 * watermark of space by flushing whichever reservation pool is currently the
1150 * largest.
1151 */
1152static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
1153{
1154 struct btrfs_fs_info *fs_info;
1155 struct btrfs_space_info *space_info;
1156 struct btrfs_block_rsv *delayed_block_rsv;
1157 struct btrfs_block_rsv *delayed_refs_rsv;
1158 struct btrfs_block_rsv *global_rsv;
1159 struct btrfs_block_rsv *trans_rsv;
1160 int loops = 0;
1161
1162 fs_info = container_of(work, struct btrfs_fs_info,
1163 preempt_reclaim_work);
1164 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1165 delayed_block_rsv = &fs_info->delayed_block_rsv;
1166 delayed_refs_rsv = &fs_info->delayed_refs_rsv;
1167 global_rsv = &fs_info->global_block_rsv;
1168 trans_rsv = &fs_info->trans_block_rsv;
1169
1170 spin_lock(&space_info->lock);
1171 while (need_preemptive_reclaim(fs_info, space_info)) {
1172 enum btrfs_flush_state flush;
1173 u64 delalloc_size = 0;
1174 u64 to_reclaim, block_rsv_size;
1175 const u64 global_rsv_size = btrfs_block_rsv_reserved(global_rsv);
1176
1177 loops++;
1178
1179 /*
1180 * We don't have a precise counter for the metadata being
1181 * reserved for delalloc, so we'll approximate it by subtracting
1182 * out the block rsv's space from the bytes_may_use. If that
1183 * amount is higher than the individual reserves, then we can
1184 * assume it's tied up in delalloc reservations.
1185 */
1186 block_rsv_size = global_rsv_size +
1187 btrfs_block_rsv_reserved(delayed_block_rsv) +
1188 btrfs_block_rsv_reserved(delayed_refs_rsv) +
1189 btrfs_block_rsv_reserved(trans_rsv);
1190 if (block_rsv_size < space_info->bytes_may_use)
1191 delalloc_size = space_info->bytes_may_use - block_rsv_size;
1192
1193 /*
1194 * We don't want to include the global_rsv in our calculation,
1195 * because that's space we can't touch. Subtract it from the
1196 * block_rsv_size for the next checks.
1197 */
1198 block_rsv_size -= global_rsv_size;
1199
1200 /*
1201 * We really want to avoid flushing delalloc too much, as it
1202 * could result in poor allocation patterns, so only flush it if
1203 * it's larger than the rest of the pools combined.
1204 */
1205 if (delalloc_size > block_rsv_size) {
1206 to_reclaim = delalloc_size;
1207 flush = FLUSH_DELALLOC;
1208 } else if (space_info->bytes_pinned >
1209 (btrfs_block_rsv_reserved(delayed_block_rsv) +
1210 btrfs_block_rsv_reserved(delayed_refs_rsv))) {
1211 to_reclaim = space_info->bytes_pinned;
1212 flush = COMMIT_TRANS;
1213 } else if (btrfs_block_rsv_reserved(delayed_block_rsv) >
1214 btrfs_block_rsv_reserved(delayed_refs_rsv)) {
1215 to_reclaim = btrfs_block_rsv_reserved(delayed_block_rsv);
1216 flush = FLUSH_DELAYED_ITEMS_NR;
1217 } else {
1218 to_reclaim = btrfs_block_rsv_reserved(delayed_refs_rsv);
1219 flush = FLUSH_DELAYED_REFS_NR;
1220 }
1221
1222 spin_unlock(&space_info->lock);
1223
1224 /*
1225 * We don't want to reclaim everything, just a portion, so scale
1226 * down the to_reclaim by 1/4. If it takes us down to 0,
1227 * reclaim 1 items worth.
1228 */
1229 to_reclaim >>= 2;
1230 if (!to_reclaim)
1231 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1232 flush_space(fs_info, space_info, to_reclaim, flush, true);
1233 cond_resched();
1234 spin_lock(&space_info->lock);
1235 }
1236
1237 /* We only went through once, back off our clamping. */
1238 if (loops == 1 && !space_info->reclaim_size)
1239 space_info->clamp = max(1, space_info->clamp - 1);
1240 trace_btrfs_done_preemptive_reclaim(fs_info, space_info);
1241 spin_unlock(&space_info->lock);
1242}
1243
1244/*
1245 * FLUSH_DELALLOC_WAIT:
1246 * Space is freed from flushing delalloc in one of two ways.
1247 *
1248 * 1) compression is on and we allocate less space than we reserved
1249 * 2) we are overwriting existing space
1250 *
1251 * For #1 that extra space is reclaimed as soon as the delalloc pages are
1252 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1253 * length to ->bytes_reserved, and subtracts the reserved space from
1254 * ->bytes_may_use.
1255 *
1256 * For #2 this is trickier. Once the ordered extent runs we will drop the
1257 * extent in the range we are overwriting, which creates a delayed ref for
1258 * that freed extent. This however is not reclaimed until the transaction
1259 * commits, thus the next stages.
1260 *
1261 * RUN_DELAYED_IPUTS
1262 * If we are freeing inodes, we want to make sure all delayed iputs have
1263 * completed, because they could have been on an inode with i_nlink == 0, and
1264 * thus have been truncated and freed up space. But again this space is not
1265 * immediately re-usable, it comes in the form of a delayed ref, which must be
1266 * run and then the transaction must be committed.
1267 *
1268 * COMMIT_TRANS
1269 * This is where we reclaim all of the pinned space generated by running the
1270 * iputs
1271 *
1272 * ALLOC_CHUNK_FORCE
1273 * For data we start with alloc chunk force, however we could have been full
1274 * before, and then the transaction commit could have freed new block groups,
1275 * so if we now have space to allocate do the force chunk allocation.
1276 */
1277static const enum btrfs_flush_state data_flush_states[] = {
1278 FLUSH_DELALLOC_FULL,
1279 RUN_DELAYED_IPUTS,
1280 COMMIT_TRANS,
1281 ALLOC_CHUNK_FORCE,
1282};
1283
1284static void btrfs_async_reclaim_data_space(struct work_struct *work)
1285{
1286 struct btrfs_fs_info *fs_info;
1287 struct btrfs_space_info *space_info;
1288 u64 last_tickets_id;
1289 enum btrfs_flush_state flush_state = 0;
1290
1291 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1292 space_info = fs_info->data_sinfo;
1293
1294 spin_lock(&space_info->lock);
1295 if (list_empty(&space_info->tickets)) {
1296 space_info->flush = 0;
1297 spin_unlock(&space_info->lock);
1298 return;
1299 }
1300 last_tickets_id = space_info->tickets_id;
1301 spin_unlock(&space_info->lock);
1302
1303 while (!space_info->full) {
1304 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1305 spin_lock(&space_info->lock);
1306 if (list_empty(&space_info->tickets)) {
1307 space_info->flush = 0;
1308 spin_unlock(&space_info->lock);
1309 return;
1310 }
1311
1312 /* Something happened, fail everything and bail. */
1313 if (BTRFS_FS_ERROR(fs_info))
1314 goto aborted_fs;
1315 last_tickets_id = space_info->tickets_id;
1316 spin_unlock(&space_info->lock);
1317 }
1318
1319 while (flush_state < ARRAY_SIZE(data_flush_states)) {
1320 flush_space(fs_info, space_info, U64_MAX,
1321 data_flush_states[flush_state], false);
1322 spin_lock(&space_info->lock);
1323 if (list_empty(&space_info->tickets)) {
1324 space_info->flush = 0;
1325 spin_unlock(&space_info->lock);
1326 return;
1327 }
1328
1329 if (last_tickets_id == space_info->tickets_id) {
1330 flush_state++;
1331 } else {
1332 last_tickets_id = space_info->tickets_id;
1333 flush_state = 0;
1334 }
1335
1336 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1337 if (space_info->full) {
1338 if (maybe_fail_all_tickets(fs_info, space_info))
1339 flush_state = 0;
1340 else
1341 space_info->flush = 0;
1342 } else {
1343 flush_state = 0;
1344 }
1345
1346 /* Something happened, fail everything and bail. */
1347 if (BTRFS_FS_ERROR(fs_info))
1348 goto aborted_fs;
1349
1350 }
1351 spin_unlock(&space_info->lock);
1352 }
1353 return;
1354
1355aborted_fs:
1356 maybe_fail_all_tickets(fs_info, space_info);
1357 space_info->flush = 0;
1358 spin_unlock(&space_info->lock);
1359}
1360
1361void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1362{
1363 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1364 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1365 INIT_WORK(&fs_info->preempt_reclaim_work,
1366 btrfs_preempt_reclaim_metadata_space);
1367}
1368
1369static const enum btrfs_flush_state priority_flush_states[] = {
1370 FLUSH_DELAYED_ITEMS_NR,
1371 FLUSH_DELAYED_ITEMS,
1372 ALLOC_CHUNK,
1373};
1374
1375static const enum btrfs_flush_state evict_flush_states[] = {
1376 FLUSH_DELAYED_ITEMS_NR,
1377 FLUSH_DELAYED_ITEMS,
1378 FLUSH_DELAYED_REFS_NR,
1379 FLUSH_DELAYED_REFS,
1380 FLUSH_DELALLOC,
1381 FLUSH_DELALLOC_WAIT,
1382 FLUSH_DELALLOC_FULL,
1383 ALLOC_CHUNK,
1384 COMMIT_TRANS,
1385};
1386
1387static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1388 struct btrfs_space_info *space_info,
1389 struct reserve_ticket *ticket,
1390 const enum btrfs_flush_state *states,
1391 int states_nr)
1392{
1393 u64 to_reclaim;
1394 int flush_state = 0;
1395
1396 spin_lock(&space_info->lock);
1397 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1398 /*
1399 * This is the priority reclaim path, so to_reclaim could be >0 still
1400 * because we may have only satisfied the priority tickets and still
1401 * left non priority tickets on the list. We would then have
1402 * to_reclaim but ->bytes == 0.
1403 */
1404 if (ticket->bytes == 0) {
1405 spin_unlock(&space_info->lock);
1406 return;
1407 }
1408
1409 while (flush_state < states_nr) {
1410 spin_unlock(&space_info->lock);
1411 flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1412 false);
1413 flush_state++;
1414 spin_lock(&space_info->lock);
1415 if (ticket->bytes == 0) {
1416 spin_unlock(&space_info->lock);
1417 return;
1418 }
1419 }
1420
1421 /*
1422 * Attempt to steal from the global rsv if we can, except if the fs was
1423 * turned into error mode due to a transaction abort when flushing space
1424 * above, in that case fail with the abort error instead of returning
1425 * success to the caller if we can steal from the global rsv - this is
1426 * just to have caller fail immeditelly instead of later when trying to
1427 * modify the fs, making it easier to debug -ENOSPC problems.
1428 */
1429 if (BTRFS_FS_ERROR(fs_info)) {
1430 ticket->error = BTRFS_FS_ERROR(fs_info);
1431 remove_ticket(space_info, ticket);
1432 } else if (!steal_from_global_rsv(fs_info, space_info, ticket)) {
1433 ticket->error = -ENOSPC;
1434 remove_ticket(space_info, ticket);
1435 }
1436
1437 /*
1438 * We must run try_granting_tickets here because we could be a large
1439 * ticket in front of a smaller ticket that can now be satisfied with
1440 * the available space.
1441 */
1442 btrfs_try_granting_tickets(fs_info, space_info);
1443 spin_unlock(&space_info->lock);
1444}
1445
1446static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1447 struct btrfs_space_info *space_info,
1448 struct reserve_ticket *ticket)
1449{
1450 spin_lock(&space_info->lock);
1451
1452 /* We could have been granted before we got here. */
1453 if (ticket->bytes == 0) {
1454 spin_unlock(&space_info->lock);
1455 return;
1456 }
1457
1458 while (!space_info->full) {
1459 spin_unlock(&space_info->lock);
1460 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1461 spin_lock(&space_info->lock);
1462 if (ticket->bytes == 0) {
1463 spin_unlock(&space_info->lock);
1464 return;
1465 }
1466 }
1467
1468 ticket->error = -ENOSPC;
1469 remove_ticket(space_info, ticket);
1470 btrfs_try_granting_tickets(fs_info, space_info);
1471 spin_unlock(&space_info->lock);
1472}
1473
1474static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1475 struct btrfs_space_info *space_info,
1476 struct reserve_ticket *ticket)
1477
1478{
1479 DEFINE_WAIT(wait);
1480 int ret = 0;
1481
1482 spin_lock(&space_info->lock);
1483 while (ticket->bytes > 0 && ticket->error == 0) {
1484 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1485 if (ret) {
1486 /*
1487 * Delete us from the list. After we unlock the space
1488 * info, we don't want the async reclaim job to reserve
1489 * space for this ticket. If that would happen, then the
1490 * ticket's task would not known that space was reserved
1491 * despite getting an error, resulting in a space leak
1492 * (bytes_may_use counter of our space_info).
1493 */
1494 remove_ticket(space_info, ticket);
1495 ticket->error = -EINTR;
1496 break;
1497 }
1498 spin_unlock(&space_info->lock);
1499
1500 schedule();
1501
1502 finish_wait(&ticket->wait, &wait);
1503 spin_lock(&space_info->lock);
1504 }
1505 spin_unlock(&space_info->lock);
1506}
1507
1508/*
1509 * Do the appropriate flushing and waiting for a ticket.
1510 *
1511 * @fs_info: the filesystem
1512 * @space_info: space info for the reservation
1513 * @ticket: ticket for the reservation
1514 * @start_ns: timestamp when the reservation started
1515 * @orig_bytes: amount of bytes originally reserved
1516 * @flush: how much we can flush
1517 *
1518 * This does the work of figuring out how to flush for the ticket, waiting for
1519 * the reservation, and returning the appropriate error if there is one.
1520 */
1521static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1522 struct btrfs_space_info *space_info,
1523 struct reserve_ticket *ticket,
1524 u64 start_ns, u64 orig_bytes,
1525 enum btrfs_reserve_flush_enum flush)
1526{
1527 int ret;
1528
1529 switch (flush) {
1530 case BTRFS_RESERVE_FLUSH_DATA:
1531 case BTRFS_RESERVE_FLUSH_ALL:
1532 case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1533 wait_reserve_ticket(fs_info, space_info, ticket);
1534 break;
1535 case BTRFS_RESERVE_FLUSH_LIMIT:
1536 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1537 priority_flush_states,
1538 ARRAY_SIZE(priority_flush_states));
1539 break;
1540 case BTRFS_RESERVE_FLUSH_EVICT:
1541 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1542 evict_flush_states,
1543 ARRAY_SIZE(evict_flush_states));
1544 break;
1545 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1546 priority_reclaim_data_space(fs_info, space_info, ticket);
1547 break;
1548 default:
1549 ASSERT(0);
1550 break;
1551 }
1552
1553 ret = ticket->error;
1554 ASSERT(list_empty(&ticket->list));
1555 /*
1556 * Check that we can't have an error set if the reservation succeeded,
1557 * as that would confuse tasks and lead them to error out without
1558 * releasing reserved space (if an error happens the expectation is that
1559 * space wasn't reserved at all).
1560 */
1561 ASSERT(!(ticket->bytes == 0 && ticket->error));
1562 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes,
1563 start_ns, flush, ticket->error);
1564 return ret;
1565}
1566
1567/*
1568 * This returns true if this flush state will go through the ordinary flushing
1569 * code.
1570 */
1571static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1572{
1573 return (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1574 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1575}
1576
1577static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1578 struct btrfs_space_info *space_info)
1579{
1580 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1581 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1582
1583 /*
1584 * If we're heavy on ordered operations then clamping won't help us. We
1585 * need to clamp specifically to keep up with dirty'ing buffered
1586 * writers, because there's not a 1:1 correlation of writing delalloc
1587 * and freeing space, like there is with flushing delayed refs or
1588 * delayed nodes. If we're already more ordered than delalloc then
1589 * we're keeping up, otherwise we aren't and should probably clamp.
1590 */
1591 if (ordered < delalloc)
1592 space_info->clamp = min(space_info->clamp + 1, 8);
1593}
1594
1595static inline bool can_steal(enum btrfs_reserve_flush_enum flush)
1596{
1597 return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1598 flush == BTRFS_RESERVE_FLUSH_EVICT);
1599}
1600
1601/*
1602 * NO_FLUSH and FLUSH_EMERGENCY don't want to create a ticket, they just want to
1603 * fail as quickly as possible.
1604 */
1605static inline bool can_ticket(enum btrfs_reserve_flush_enum flush)
1606{
1607 return (flush != BTRFS_RESERVE_NO_FLUSH &&
1608 flush != BTRFS_RESERVE_FLUSH_EMERGENCY);
1609}
1610
1611/*
1612 * Try to reserve bytes from the block_rsv's space.
1613 *
1614 * @fs_info: the filesystem
1615 * @space_info: space info we want to allocate from
1616 * @orig_bytes: number of bytes we want
1617 * @flush: whether or not we can flush to make our reservation
1618 *
1619 * This will reserve orig_bytes number of bytes from the space info associated
1620 * with the block_rsv. If there is not enough space it will make an attempt to
1621 * flush out space to make room. It will do this by flushing delalloc if
1622 * possible or committing the transaction. If flush is 0 then no attempts to
1623 * regain reservations will be made and this will fail if there is not enough
1624 * space already.
1625 */
1626static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1627 struct btrfs_space_info *space_info, u64 orig_bytes,
1628 enum btrfs_reserve_flush_enum flush)
1629{
1630 struct work_struct *async_work;
1631 struct reserve_ticket ticket;
1632 u64 start_ns = 0;
1633 u64 used;
1634 int ret = -ENOSPC;
1635 bool pending_tickets;
1636
1637 ASSERT(orig_bytes);
1638 /*
1639 * If have a transaction handle (current->journal_info != NULL), then
1640 * the flush method can not be neither BTRFS_RESERVE_FLUSH_ALL* nor
1641 * BTRFS_RESERVE_FLUSH_EVICT, as we could deadlock because those
1642 * flushing methods can trigger transaction commits.
1643 */
1644 if (current->journal_info) {
1645 /* One assert per line for easier debugging. */
1646 ASSERT(flush != BTRFS_RESERVE_FLUSH_ALL);
1647 ASSERT(flush != BTRFS_RESERVE_FLUSH_ALL_STEAL);
1648 ASSERT(flush != BTRFS_RESERVE_FLUSH_EVICT);
1649 }
1650
1651 if (flush == BTRFS_RESERVE_FLUSH_DATA)
1652 async_work = &fs_info->async_data_reclaim_work;
1653 else
1654 async_work = &fs_info->async_reclaim_work;
1655
1656 spin_lock(&space_info->lock);
1657 used = btrfs_space_info_used(space_info, true);
1658
1659 /*
1660 * We don't want NO_FLUSH allocations to jump everybody, they can
1661 * generally handle ENOSPC in a different way, so treat them the same as
1662 * normal flushers when it comes to skipping pending tickets.
1663 */
1664 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1665 pending_tickets = !list_empty(&space_info->tickets) ||
1666 !list_empty(&space_info->priority_tickets);
1667 else
1668 pending_tickets = !list_empty(&space_info->priority_tickets);
1669
1670 /*
1671 * Carry on if we have enough space (short-circuit) OR call
1672 * can_overcommit() to ensure we can overcommit to continue.
1673 */
1674 if (!pending_tickets &&
1675 ((used + orig_bytes <= space_info->total_bytes) ||
1676 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1677 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1678 orig_bytes);
1679 ret = 0;
1680 }
1681
1682 /*
1683 * Things are dire, we need to make a reservation so we don't abort. We
1684 * will let this reservation go through as long as we have actual space
1685 * left to allocate for the block.
1686 */
1687 if (ret && unlikely(flush == BTRFS_RESERVE_FLUSH_EMERGENCY)) {
1688 used = btrfs_space_info_used(space_info, false);
1689 if (used + orig_bytes <= space_info->total_bytes) {
1690 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1691 orig_bytes);
1692 ret = 0;
1693 }
1694 }
1695
1696 /*
1697 * If we couldn't make a reservation then setup our reservation ticket
1698 * and kick the async worker if it's not already running.
1699 *
1700 * If we are a priority flusher then we just need to add our ticket to
1701 * the list and we will do our own flushing further down.
1702 */
1703 if (ret && can_ticket(flush)) {
1704 ticket.bytes = orig_bytes;
1705 ticket.error = 0;
1706 space_info->reclaim_size += ticket.bytes;
1707 init_waitqueue_head(&ticket.wait);
1708 ticket.steal = can_steal(flush);
1709 if (trace_btrfs_reserve_ticket_enabled())
1710 start_ns = ktime_get_ns();
1711
1712 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1713 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1714 flush == BTRFS_RESERVE_FLUSH_DATA) {
1715 list_add_tail(&ticket.list, &space_info->tickets);
1716 if (!space_info->flush) {
1717 /*
1718 * We were forced to add a reserve ticket, so
1719 * our preemptive flushing is unable to keep
1720 * up. Clamp down on the threshold for the
1721 * preemptive flushing in order to keep up with
1722 * the workload.
1723 */
1724 maybe_clamp_preempt(fs_info, space_info);
1725
1726 space_info->flush = 1;
1727 trace_btrfs_trigger_flush(fs_info,
1728 space_info->flags,
1729 orig_bytes, flush,
1730 "enospc");
1731 queue_work(system_unbound_wq, async_work);
1732 }
1733 } else {
1734 list_add_tail(&ticket.list,
1735 &space_info->priority_tickets);
1736 }
1737 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1738 /*
1739 * We will do the space reservation dance during log replay,
1740 * which means we won't have fs_info->fs_root set, so don't do
1741 * the async reclaim as we will panic.
1742 */
1743 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1744 !work_busy(&fs_info->preempt_reclaim_work) &&
1745 need_preemptive_reclaim(fs_info, space_info)) {
1746 trace_btrfs_trigger_flush(fs_info, space_info->flags,
1747 orig_bytes, flush, "preempt");
1748 queue_work(system_unbound_wq,
1749 &fs_info->preempt_reclaim_work);
1750 }
1751 }
1752 spin_unlock(&space_info->lock);
1753 if (!ret || !can_ticket(flush))
1754 return ret;
1755
1756 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1757 orig_bytes, flush);
1758}
1759
1760/*
1761 * Try to reserve metadata bytes from the block_rsv's space.
1762 *
1763 * @fs_info: the filesystem
1764 * @space_info: the space_info we're allocating for
1765 * @orig_bytes: number of bytes we want
1766 * @flush: whether or not we can flush to make our reservation
1767 *
1768 * This will reserve orig_bytes number of bytes from the space info associated
1769 * with the block_rsv. If there is not enough space it will make an attempt to
1770 * flush out space to make room. It will do this by flushing delalloc if
1771 * possible or committing the transaction. If flush is 0 then no attempts to
1772 * regain reservations will be made and this will fail if there is not enough
1773 * space already.
1774 */
1775int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1776 struct btrfs_space_info *space_info,
1777 u64 orig_bytes,
1778 enum btrfs_reserve_flush_enum flush)
1779{
1780 int ret;
1781
1782 ret = __reserve_bytes(fs_info, space_info, orig_bytes, flush);
1783 if (ret == -ENOSPC) {
1784 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1785 space_info->flags, orig_bytes, 1);
1786
1787 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1788 btrfs_dump_space_info(fs_info, space_info, orig_bytes, 0);
1789 }
1790 return ret;
1791}
1792
1793/*
1794 * Try to reserve data bytes for an allocation.
1795 *
1796 * @fs_info: the filesystem
1797 * @bytes: number of bytes we need
1798 * @flush: how we are allowed to flush
1799 *
1800 * This will reserve bytes from the data space info. If there is not enough
1801 * space then we will attempt to flush space as specified by flush.
1802 */
1803int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1804 enum btrfs_reserve_flush_enum flush)
1805{
1806 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1807 int ret;
1808
1809 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1810 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE ||
1811 flush == BTRFS_RESERVE_NO_FLUSH);
1812 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1813
1814 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1815 if (ret == -ENOSPC) {
1816 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1817 data_sinfo->flags, bytes, 1);
1818 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1819 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1820 }
1821 return ret;
1822}
1823
1824/* Dump all the space infos when we abort a transaction due to ENOSPC. */
1825__cold void btrfs_dump_space_info_for_trans_abort(struct btrfs_fs_info *fs_info)
1826{
1827 struct btrfs_space_info *space_info;
1828
1829 btrfs_info(fs_info, "dumping space info:");
1830 list_for_each_entry(space_info, &fs_info->space_info, list) {
1831 spin_lock(&space_info->lock);
1832 __btrfs_dump_space_info(fs_info, space_info);
1833 spin_unlock(&space_info->lock);
1834 }
1835 dump_global_block_rsv(fs_info);
1836}
1837
1838/*
1839 * Account the unused space of all the readonly block group in the space_info.
1840 * takes mirrors into account.
1841 */
1842u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
1843{
1844 struct btrfs_block_group *block_group;
1845 u64 free_bytes = 0;
1846 int factor;
1847
1848 /* It's df, we don't care if it's racy */
1849 if (list_empty(&sinfo->ro_bgs))
1850 return 0;
1851
1852 spin_lock(&sinfo->lock);
1853 list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
1854 spin_lock(&block_group->lock);
1855
1856 if (!block_group->ro) {
1857 spin_unlock(&block_group->lock);
1858 continue;
1859 }
1860
1861 factor = btrfs_bg_type_to_factor(block_group->flags);
1862 free_bytes += (block_group->length -
1863 block_group->used) * factor;
1864
1865 spin_unlock(&block_group->lock);
1866 }
1867 spin_unlock(&sinfo->lock);
1868
1869 return free_bytes;
1870}