1 // SPDX-License-Identifier: GPL-2.0
2
3 #include "linux/spinlock.h"
4 #include <linux/minmax.h>
5 #include "misc.h"
6 #include "ctree.h"
7 #include "space-info.h"
8 #include "sysfs.h"
9 #include "volumes.h"
10 #include "free-space-cache.h"
11 #include "ordered-data.h"
12 #include "transaction.h"
13 #include "block-group.h"
14 #include "fs.h"
15 #include "accessors.h"
16 #include "extent-tree.h"
17
18 /*
19 * HOW DOES SPACE RESERVATION WORK
20 *
21 * If you want to know about delalloc specifically, there is a separate comment
22 * for that with the delalloc code. This comment is about how the whole system
23 * works generally.
24 *
25 * BASIC CONCEPTS
26 *
27 * 1) space_info. This is the ultimate arbiter of how much space we can use.
28 * There's a description of the bytes_ fields with the struct declaration,
29 * refer to that for specifics on each field. Suffice it to say that for
30 * reservations we care about total_bytes - SUM(space_info->bytes_) when
31 * determining if there is space to make an allocation. There is a space_info
32 * for METADATA, SYSTEM, and DATA areas.
33 *
34 * 2) block_rsv's. These are basically buckets for every different type of
35 * metadata reservation we have. You can see the comment in the block_rsv
36 * code on the rules for each type, but generally block_rsv->reserved is how
37 * much space is accounted for in space_info->bytes_may_use.
38 *
39 * 3) btrfs_calc*_size. These are the worst case calculations we used based
40 * on the number of items we will want to modify. We have one for changing
41 * items, and one for inserting new items. Generally we use these helpers to
42 * determine the size of the block reserves, and then use the actual bytes
43 * values to adjust the space_info counters.
44 *
45 * MAKING RESERVATIONS, THE NORMAL CASE
46 *
47 * We call into either btrfs_reserve_data_bytes() or
48 * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with
49 * num_bytes we want to reserve.
50 *
51 * ->reserve
52 * space_info->bytes_may_reserve += num_bytes
53 *
54 * ->extent allocation
55 * Call btrfs_add_reserved_bytes() which does
56 * space_info->bytes_may_reserve -= num_bytes
57 * space_info->bytes_reserved += extent_bytes
58 *
59 * ->insert reference
60 * Call btrfs_update_block_group() which does
61 * space_info->bytes_reserved -= extent_bytes
62 * space_info->bytes_used += extent_bytes
63 *
64 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority)
65 *
66 * Assume we are unable to simply make the reservation because we do not have
67 * enough space
68 *
69 * -> __reserve_bytes
70 * create a reserve_ticket with ->bytes set to our reservation, add it to
71 * the tail of space_info->tickets, kick async flush thread
72 *
73 * ->handle_reserve_ticket
74 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set
75 * on the ticket.
76 *
77 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space
78 * Flushes various things attempting to free up space.
79 *
80 * -> btrfs_try_granting_tickets()
81 * This is called by anything that either subtracts space from
82 * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the
83 * space_info->total_bytes. This loops through the ->priority_tickets and
84 * then the ->tickets list checking to see if the reservation can be
85 * completed. If it can the space is added to space_info->bytes_may_use and
86 * the ticket is woken up.
87 *
88 * -> ticket wakeup
89 * Check if ->bytes == 0, if it does we got our reservation and we can carry
90 * on, if not return the appropriate error (ENOSPC, but can be EINTR if we
91 * were interrupted.)
92 *
93 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY
94 *
95 * Same as the above, except we add ourselves to the
96 * space_info->priority_tickets, and we do not use ticket->wait, we simply
97 * call flush_space() ourselves for the states that are safe for us to call
98 * without deadlocking and hope for the best.
99 *
100 * THE FLUSHING STATES
101 *
102 * Generally speaking we will have two cases for each state, a "nice" state
103 * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to
104 * reduce the locking over head on the various trees, and even to keep from
105 * doing any work at all in the case of delayed refs. Each of these delayed
106 * things however hold reservations, and so letting them run allows us to
107 * reclaim space so we can make new reservations.
108 *
109 * FLUSH_DELAYED_ITEMS
110 * Every inode has a delayed item to update the inode. Take a simple write
111 * for example, we would update the inode item at write time to update the
112 * mtime, and then again at finish_ordered_io() time in order to update the
113 * isize or bytes. We keep these delayed items to coalesce these operations
114 * into a single operation done on demand. These are an easy way to reclaim
115 * metadata space.
116 *
117 * FLUSH_DELALLOC
118 * Look at the delalloc comment to get an idea of how much space is reserved
119 * for delayed allocation. We can reclaim some of this space simply by
120 * running delalloc, but usually we need to wait for ordered extents to
121 * reclaim the bulk of this space.
122 *
123 * FLUSH_DELAYED_REFS
124 * We have a block reserve for the outstanding delayed refs space, and every
125 * delayed ref operation holds a reservation. Running these is a quick way
126 * to reclaim space, but we want to hold this until the end because COW can
127 * churn a lot and we can avoid making some extent tree modifications if we
128 * are able to delay for as long as possible.
129 *
130 * ALLOC_CHUNK
131 * We will skip this the first time through space reservation, because of
132 * overcommit and we don't want to have a lot of useless metadata space when
133 * our worst case reservations will likely never come true.
134 *
135 * RUN_DELAYED_IPUTS
136 * If we're freeing inodes we're likely freeing checksums, file extent
137 * items, and extent tree items. Loads of space could be freed up by these
138 * operations, however they won't be usable until the transaction commits.
139 *
140 * COMMIT_TRANS
141 * This will commit the transaction. Historically we had a lot of logic
142 * surrounding whether or not we'd commit the transaction, but this waits born
143 * out of a pre-tickets era where we could end up committing the transaction
144 * thousands of times in a row without making progress. Now thanks to our
145 * ticketing system we know if we're not making progress and can error
146 * everybody out after a few commits rather than burning the disk hoping for
147 * a different answer.
148 *
149 * OVERCOMMIT
150 *
151 * Because we hold so many reservations for metadata we will allow you to
152 * reserve more space than is currently free in the currently allocate
153 * metadata space. This only happens with metadata, data does not allow
154 * overcommitting.
155 *
156 * You can see the current logic for when we allow overcommit in
157 * btrfs_can_overcommit(), but it only applies to unallocated space. If there
158 * is no unallocated space to be had, all reservations are kept within the
159 * free space in the allocated metadata chunks.
160 *
161 * Because of overcommitting, you generally want to use the
162 * btrfs_can_overcommit() logic for metadata allocations, as it does the right
163 * thing with or without extra unallocated space.
164 */
165
166 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
167 bool may_use_included)
168 {
169 ASSERT(s_info);
170 return s_info->bytes_used + s_info->bytes_reserved +
171 s_info->bytes_pinned + s_info->bytes_readonly +
172 s_info->bytes_zone_unusable +
173 (may_use_included ? s_info->bytes_may_use : 0);
174 }
175
176 /*
177 * after adding space to the filesystem, we need to clear the full flags
178 * on all the space infos.
179 */
180 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
181 {
182 struct list_head *head = &info->space_info;
183 struct btrfs_space_info *found;
184
185 list_for_each_entry(found, head, list)
186 found->full = 0;
187 }
188
189 /*
190 * Block groups with more than this value (percents) of unusable space will be
191 * scheduled for background reclaim.
192 */
193 #define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH (75)
194
195 #define BTRFS_UNALLOC_BLOCK_GROUP_TARGET (10ULL)
196
197 /*
198 * Calculate chunk size depending on volume type (regular or zoned).
199 */
200 static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags)
201 {
202 if (btrfs_is_zoned(fs_info))
203 return fs_info->zone_size;
204
205 ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
206
207 if (flags & BTRFS_BLOCK_GROUP_DATA)
208 return BTRFS_MAX_DATA_CHUNK_SIZE;
209 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
210 return SZ_32M;
211
212 /* Handle BTRFS_BLOCK_GROUP_METADATA */
213 if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G)
214 return SZ_1G;
215
216 return SZ_256M;
217 }
218
219 /*
220 * Update default chunk size.
221 */
222 void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info,
223 u64 chunk_size)
224 {
225 WRITE_ONCE(space_info->chunk_size, chunk_size);
226 }
227
228 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
229 {
230
231 struct btrfs_space_info *space_info;
232 int i;
233 int ret;
234
235 space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
236 if (!space_info)
237 return -ENOMEM;
238
239 space_info->fs_info = info;
240 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
241 INIT_LIST_HEAD(&space_info->block_groups[i]);
242 init_rwsem(&space_info->groups_sem);
243 spin_lock_init(&space_info->lock);
244 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
245 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
246 INIT_LIST_HEAD(&space_info->ro_bgs);
247 INIT_LIST_HEAD(&space_info->tickets);
248 INIT_LIST_HEAD(&space_info->priority_tickets);
249 space_info->clamp = 1;
250 btrfs_update_space_info_chunk_size(space_info, calc_chunk_size(info, flags));
251
252 if (btrfs_is_zoned(info))
253 space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH;
254
255 ret = btrfs_sysfs_add_space_info_type(info, space_info);
256 if (ret)
257 return ret;
258
259 list_add(&space_info->list, &info->space_info);
260 if (flags & BTRFS_BLOCK_GROUP_DATA)
261 info->data_sinfo = space_info;
262
263 return ret;
264 }
265
266 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
267 {
268 struct btrfs_super_block *disk_super;
269 u64 features;
270 u64 flags;
271 int mixed = 0;
272 int ret;
273
274 disk_super = fs_info->super_copy;
275 if (!btrfs_super_root(disk_super))
276 return -EINVAL;
277
278 features = btrfs_super_incompat_flags(disk_super);
279 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
280 mixed = 1;
281
282 flags = BTRFS_BLOCK_GROUP_SYSTEM;
283 ret = create_space_info(fs_info, flags);
284 if (ret)
285 goto out;
286
287 if (mixed) {
288 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
289 ret = create_space_info(fs_info, flags);
290 } else {
291 flags = BTRFS_BLOCK_GROUP_METADATA;
292 ret = create_space_info(fs_info, flags);
293 if (ret)
294 goto out;
295
296 flags = BTRFS_BLOCK_GROUP_DATA;
297 ret = create_space_info(fs_info, flags);
298 }
299 out:
300 return ret;
301 }
302
303 void btrfs_add_bg_to_space_info(struct btrfs_fs_info *info,
304 struct btrfs_block_group *block_group)
305 {
306 struct btrfs_space_info *found;
307 int factor, index;
308
309 factor = btrfs_bg_type_to_factor(block_group->flags);
310
311 found = btrfs_find_space_info(info, block_group->flags);
312 ASSERT(found);
313 spin_lock(&found->lock);
314 found->total_bytes += block_group->length;
315 found->disk_total += block_group->length * factor;
316 found->bytes_used += block_group->used;
317 found->disk_used += block_group->used * factor;
318 found->bytes_readonly += block_group->bytes_super;
319 btrfs_space_info_update_bytes_zone_unusable(info, found, block_group->zone_unusable);
320 if (block_group->length > 0)
321 found->full = 0;
322 btrfs_try_granting_tickets(info, found);
323 spin_unlock(&found->lock);
324
325 block_group->space_info = found;
326
327 index = btrfs_bg_flags_to_raid_index(block_group->flags);
328 down_write(&found->groups_sem);
329 list_add_tail(&block_group->list, &found->block_groups[index]);
330 up_write(&found->groups_sem);
331 }
332
333 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
334 u64 flags)
335 {
336 struct list_head *head = &info->space_info;
337 struct btrfs_space_info *found;
338
339 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
340
341 list_for_each_entry(found, head, list) {
342 if (found->flags & flags)
343 return found;
344 }
345 return NULL;
346 }
347
348 static u64 calc_effective_data_chunk_size(struct btrfs_fs_info *fs_info)
349 {
350 struct btrfs_space_info *data_sinfo;
351 u64 data_chunk_size;
352
353 /*
354 * Calculate the data_chunk_size, space_info->chunk_size is the
355 * "optimal" chunk size based on the fs size. However when we actually
356 * allocate the chunk we will strip this down further, making it no
357 * more than 10% of the disk or 1G, whichever is smaller.
358 *
359 * On the zoned mode, we need to use zone_size (= data_sinfo->chunk_size)
360 * as it is.
361 */
362 data_sinfo = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
363 if (btrfs_is_zoned(fs_info))
364 return data_sinfo->chunk_size;
365 data_chunk_size = min(data_sinfo->chunk_size,
366 mult_perc(fs_info->fs_devices->total_rw_bytes, 10));
367 return min_t(u64, data_chunk_size, SZ_1G);
368 }
369
370 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
371 struct btrfs_space_info *space_info,
372 enum btrfs_reserve_flush_enum flush)
373 {
374 u64 profile;
375 u64 avail;
376 u64 data_chunk_size;
377 int factor;
378
379 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
380 profile = btrfs_system_alloc_profile(fs_info);
381 else
382 profile = btrfs_metadata_alloc_profile(fs_info);
383
384 avail = atomic64_read(&fs_info->free_chunk_space);
385
386 /*
387 * If we have dup, raid1 or raid10 then only half of the free
388 * space is actually usable. For raid56, the space info used
389 * doesn't include the parity drive, so we don't have to
390 * change the math
391 */
392 factor = btrfs_bg_type_to_factor(profile);
393 avail = div_u64(avail, factor);
394 if (avail == 0)
395 return 0;
396
397 data_chunk_size = calc_effective_data_chunk_size(fs_info);
398
399 /*
400 * Since data allocations immediately use block groups as part of the
401 * reservation, because we assume that data reservations will == actual
402 * usage, we could potentially overcommit and then immediately have that
403 * available space used by a data allocation, which could put us in a
404 * bind when we get close to filling the file system.
405 *
406 * To handle this simply remove the data_chunk_size from the available
407 * space. If we are relatively empty this won't affect our ability to
408 * overcommit much, and if we're very close to full it'll keep us from
409 * getting into a position where we've given ourselves very little
410 * metadata wiggle room.
411 */
412 if (avail <= data_chunk_size)
413 return 0;
414 avail -= data_chunk_size;
415
416 /*
417 * If we aren't flushing all things, let us overcommit up to
418 * 1/2th of the space. If we can flush, don't let us overcommit
419 * too much, let it overcommit up to 1/8 of the space.
420 */
421 if (flush == BTRFS_RESERVE_FLUSH_ALL)
422 avail >>= 3;
423 else
424 avail >>= 1;
425
426 /*
427 * On the zoned mode, we always allocate one zone as one chunk.
428 * Returning non-zone size alingned bytes here will result in
429 * less pressure for the async metadata reclaim process, and it
430 * will over-commit too much leading to ENOSPC. Align down to the
431 * zone size to avoid that.
432 */
433 if (btrfs_is_zoned(fs_info))
434 avail = ALIGN_DOWN(avail, fs_info->zone_size);
435
436 return avail;
437 }
438
439 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
440 struct btrfs_space_info *space_info, u64 bytes,
441 enum btrfs_reserve_flush_enum flush)
442 {
443 u64 avail;
444 u64 used;
445
446 /* Don't overcommit when in mixed mode */
447 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
448 return 0;
449
450 used = btrfs_space_info_used(space_info, true);
451 avail = calc_available_free_space(fs_info, space_info, flush);
452
453 if (used + bytes < space_info->total_bytes + avail)
454 return 1;
455 return 0;
456 }
457
458 static void remove_ticket(struct btrfs_space_info *space_info,
459 struct reserve_ticket *ticket)
460 {
461 if (!list_empty(&ticket->list)) {
462 list_del_init(&ticket->list);
463 ASSERT(space_info->reclaim_size >= ticket->bytes);
464 space_info->reclaim_size -= ticket->bytes;
465 }
466 }
467
468 /*
469 * This is for space we already have accounted in space_info->bytes_may_use, so
470 * basically when we're returning space from block_rsv's.
471 */
472 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
473 struct btrfs_space_info *space_info)
474 {
475 struct list_head *head;
476 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
477
478 lockdep_assert_held(&space_info->lock);
479
480 head = &space_info->priority_tickets;
481 again:
482 while (!list_empty(head)) {
483 struct reserve_ticket *ticket;
484 u64 used = btrfs_space_info_used(space_info, true);
485
486 ticket = list_first_entry(head, struct reserve_ticket, list);
487
488 /* Check and see if our ticket can be satisfied now. */
489 if ((used + ticket->bytes <= space_info->total_bytes) ||
490 btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
491 flush)) {
492 btrfs_space_info_update_bytes_may_use(fs_info,
493 space_info,
494 ticket->bytes);
495 remove_ticket(space_info, ticket);
496 ticket->bytes = 0;
497 space_info->tickets_id++;
498 wake_up(&ticket->wait);
499 } else {
500 break;
501 }
502 }
503
504 if (head == &space_info->priority_tickets) {
505 head = &space_info->tickets;
506 flush = BTRFS_RESERVE_FLUSH_ALL;
507 goto again;
508 }
509 }
510
511 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \
512 do { \
513 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \
514 spin_lock(&__rsv->lock); \
515 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \
516 __rsv->size, __rsv->reserved); \
517 spin_unlock(&__rsv->lock); \
518 } while (0)
519
520 static const char *space_info_flag_to_str(const struct btrfs_space_info *space_info)
521 {
522 switch (space_info->flags) {
523 case BTRFS_BLOCK_GROUP_SYSTEM:
524 return "SYSTEM";
525 case BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA:
526 return "DATA+METADATA";
527 case BTRFS_BLOCK_GROUP_DATA:
528 return "DATA";
529 case BTRFS_BLOCK_GROUP_METADATA:
530 return "METADATA";
531 default:
532 return "UNKNOWN";
533 }
534 }
535
536 static void dump_global_block_rsv(struct btrfs_fs_info *fs_info)
537 {
538 DUMP_BLOCK_RSV(fs_info, global_block_rsv);
539 DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
540 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
541 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
542 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
543 }
544
545 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
546 struct btrfs_space_info *info)
547 {
548 const char *flag_str = space_info_flag_to_str(info);
549 lockdep_assert_held(&info->lock);
550
551 /* The free space could be negative in case of overcommit */
552 btrfs_info(fs_info, "space_info %s has %lld free, is %sfull",
553 flag_str,
554 (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
555 info->full ? "" : "not ");
556 btrfs_info(fs_info,
557 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
558 info->total_bytes, info->bytes_used, info->bytes_pinned,
559 info->bytes_reserved, info->bytes_may_use,
560 info->bytes_readonly, info->bytes_zone_unusable);
561 }
562
563 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
564 struct btrfs_space_info *info, u64 bytes,
565 int dump_block_groups)
566 {
567 struct btrfs_block_group *cache;
568 u64 total_avail = 0;
569 int index = 0;
570
571 spin_lock(&info->lock);
572 __btrfs_dump_space_info(fs_info, info);
573 dump_global_block_rsv(fs_info);
574 spin_unlock(&info->lock);
575
576 if (!dump_block_groups)
577 return;
578
579 down_read(&info->groups_sem);
580 again:
581 list_for_each_entry(cache, &info->block_groups[index], list) {
582 u64 avail;
583
584 spin_lock(&cache->lock);
585 avail = cache->length - cache->used - cache->pinned -
586 cache->reserved - cache->bytes_super - cache->zone_unusable;
587 btrfs_info(fs_info,
588 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu delalloc %llu super %llu zone_unusable (%llu bytes available) %s",
589 cache->start, cache->length, cache->used, cache->pinned,
590 cache->reserved, cache->delalloc_bytes,
591 cache->bytes_super, cache->zone_unusable,
592 avail, cache->ro ? "[readonly]" : "");
593 spin_unlock(&cache->lock);
594 btrfs_dump_free_space(cache, bytes);
595 total_avail += avail;
596 }
597 if (++index < BTRFS_NR_RAID_TYPES)
598 goto again;
599 up_read(&info->groups_sem);
600
601 btrfs_info(fs_info, "%llu bytes available across all block groups", total_avail);
602 }
603
604 static inline u64 calc_reclaim_items_nr(const struct btrfs_fs_info *fs_info,
605 u64 to_reclaim)
606 {
607 u64 bytes;
608 u64 nr;
609
610 bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
611 nr = div64_u64(to_reclaim, bytes);
612 if (!nr)
613 nr = 1;
614 return nr;
615 }
616
617 /*
618 * shrink metadata reservation for delalloc
619 */
620 static void shrink_delalloc(struct btrfs_fs_info *fs_info,
621 struct btrfs_space_info *space_info,
622 u64 to_reclaim, bool wait_ordered,
623 bool for_preempt)
624 {
625 struct btrfs_trans_handle *trans;
626 u64 delalloc_bytes;
627 u64 ordered_bytes;
628 u64 items;
629 long time_left;
630 int loops;
631
632 delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
633 ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes);
634 if (delalloc_bytes == 0 && ordered_bytes == 0)
635 return;
636
637 /* Calc the number of the pages we need flush for space reservation */
638 if (to_reclaim == U64_MAX) {
639 items = U64_MAX;
640 } else {
641 /*
642 * to_reclaim is set to however much metadata we need to
643 * reclaim, but reclaiming that much data doesn't really track
644 * exactly. What we really want to do is reclaim full inode's
645 * worth of reservations, however that's not available to us
646 * here. We will take a fraction of the delalloc bytes for our
647 * flushing loops and hope for the best. Delalloc will expand
648 * the amount we write to cover an entire dirty extent, which
649 * will reclaim the metadata reservation for that range. If
650 * it's not enough subsequent flush stages will be more
651 * aggressive.
652 */
653 to_reclaim = max(to_reclaim, delalloc_bytes >> 3);
654 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
655 }
656
657 trans = current->journal_info;
658
659 /*
660 * If we are doing more ordered than delalloc we need to just wait on
661 * ordered extents, otherwise we'll waste time trying to flush delalloc
662 * that likely won't give us the space back we need.
663 */
664 if (ordered_bytes > delalloc_bytes && !for_preempt)
665 wait_ordered = true;
666
667 loops = 0;
668 while ((delalloc_bytes || ordered_bytes) && loops < 3) {
669 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
670 long nr_pages = min_t(u64, temp, LONG_MAX);
671 int async_pages;
672
673 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
674
675 /*
676 * We need to make sure any outstanding async pages are now
677 * processed before we continue. This is because things like
678 * sync_inode() try to be smart and skip writing if the inode is
679 * marked clean. We don't use filemap_fwrite for flushing
680 * because we want to control how many pages we write out at a
681 * time, thus this is the only safe way to make sure we've
682 * waited for outstanding compressed workers to have started
683 * their jobs and thus have ordered extents set up properly.
684 *
685 * This exists because we do not want to wait for each
686 * individual inode to finish its async work, we simply want to
687 * start the IO on everybody, and then come back here and wait
688 * for all of the async work to catch up. Once we're done with
689 * that we know we'll have ordered extents for everything and we
690 * can decide if we wait for that or not.
691 *
692 * If we choose to replace this in the future, make absolutely
693 * sure that the proper waiting is being done in the async case,
694 * as there have been bugs in that area before.
695 */
696 async_pages = atomic_read(&fs_info->async_delalloc_pages);
697 if (!async_pages)
698 goto skip_async;
699
700 /*
701 * We don't want to wait forever, if we wrote less pages in this
702 * loop than we have outstanding, only wait for that number of
703 * pages, otherwise we can wait for all async pages to finish
704 * before continuing.
705 */
706 if (async_pages > nr_pages)
707 async_pages -= nr_pages;
708 else
709 async_pages = 0;
710 wait_event(fs_info->async_submit_wait,
711 atomic_read(&fs_info->async_delalloc_pages) <=
712 async_pages);
713 skip_async:
714 loops++;
715 if (wait_ordered && !trans) {
716 btrfs_wait_ordered_roots(fs_info, items, NULL);
717 } else {
718 time_left = schedule_timeout_killable(1);
719 if (time_left)
720 break;
721 }
722
723 /*
724 * If we are for preemption we just want a one-shot of delalloc
725 * flushing so we can stop flushing if we decide we don't need
726 * to anymore.
727 */
728 if (for_preempt)
729 break;
730
731 spin_lock(&space_info->lock);
732 if (list_empty(&space_info->tickets) &&
733 list_empty(&space_info->priority_tickets)) {
734 spin_unlock(&space_info->lock);
735 break;
736 }
737 spin_unlock(&space_info->lock);
738
739 delalloc_bytes = percpu_counter_sum_positive(
740 &fs_info->delalloc_bytes);
741 ordered_bytes = percpu_counter_sum_positive(
742 &fs_info->ordered_bytes);
743 }
744 }
745
746 /*
747 * Try to flush some data based on policy set by @state. This is only advisory
748 * and may fail for various reasons. The caller is supposed to examine the
749 * state of @space_info to detect the outcome.
750 */
751 static void flush_space(struct btrfs_fs_info *fs_info,
752 struct btrfs_space_info *space_info, u64 num_bytes,
753 enum btrfs_flush_state state, bool for_preempt)
754 {
755 struct btrfs_root *root = fs_info->tree_root;
756 struct btrfs_trans_handle *trans;
757 int nr;
758 int ret = 0;
759
760 switch (state) {
761 case FLUSH_DELAYED_ITEMS_NR:
762 case FLUSH_DELAYED_ITEMS:
763 if (state == FLUSH_DELAYED_ITEMS_NR)
764 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
765 else
766 nr = -1;
767
768 trans = btrfs_join_transaction_nostart(root);
769 if (IS_ERR(trans)) {
770 ret = PTR_ERR(trans);
771 if (ret == -ENOENT)
772 ret = 0;
773 break;
774 }
775 ret = btrfs_run_delayed_items_nr(trans, nr);
776 btrfs_end_transaction(trans);
777 break;
778 case FLUSH_DELALLOC:
779 case FLUSH_DELALLOC_WAIT:
780 case FLUSH_DELALLOC_FULL:
781 if (state == FLUSH_DELALLOC_FULL)
782 num_bytes = U64_MAX;
783 shrink_delalloc(fs_info, space_info, num_bytes,
784 state != FLUSH_DELALLOC, for_preempt);
785 break;
786 case FLUSH_DELAYED_REFS_NR:
787 case FLUSH_DELAYED_REFS:
788 trans = btrfs_join_transaction_nostart(root);
789 if (IS_ERR(trans)) {
790 ret = PTR_ERR(trans);
791 if (ret == -ENOENT)
792 ret = 0;
793 break;
794 }
795 if (state == FLUSH_DELAYED_REFS_NR)
796 btrfs_run_delayed_refs(trans, num_bytes);
797 else
798 btrfs_run_delayed_refs(trans, 0);
799 btrfs_end_transaction(trans);
800 break;
801 case ALLOC_CHUNK:
802 case ALLOC_CHUNK_FORCE:
803 trans = btrfs_join_transaction(root);
804 if (IS_ERR(trans)) {
805 ret = PTR_ERR(trans);
806 break;
807 }
808 ret = btrfs_chunk_alloc(trans,
809 btrfs_get_alloc_profile(fs_info, space_info->flags),
810 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
811 CHUNK_ALLOC_FORCE);
812 btrfs_end_transaction(trans);
813
814 if (ret > 0 || ret == -ENOSPC)
815 ret = 0;
816 break;
817 case RUN_DELAYED_IPUTS:
818 /*
819 * If we have pending delayed iputs then we could free up a
820 * bunch of pinned space, so make sure we run the iputs before
821 * we do our pinned bytes check below.
822 */
823 btrfs_run_delayed_iputs(fs_info);
824 btrfs_wait_on_delayed_iputs(fs_info);
825 break;
826 case COMMIT_TRANS:
827 ASSERT(current->journal_info == NULL);
828 /*
829 * We don't want to start a new transaction, just attach to the
830 * current one or wait it fully commits in case its commit is
831 * happening at the moment. Note: we don't use a nostart join
832 * because that does not wait for a transaction to fully commit
833 * (only for it to be unblocked, state TRANS_STATE_UNBLOCKED).
834 */
835 ret = btrfs_commit_current_transaction(root);
836 break;
837 default:
838 ret = -ENOSPC;
839 break;
840 }
841
842 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
843 ret, for_preempt);
844 return;
845 }
846
847 static inline u64
848 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
849 struct btrfs_space_info *space_info)
850 {
851 u64 used;
852 u64 avail;
853 u64 to_reclaim = space_info->reclaim_size;
854
855 lockdep_assert_held(&space_info->lock);
856
857 avail = calc_available_free_space(fs_info, space_info,
858 BTRFS_RESERVE_FLUSH_ALL);
859 used = btrfs_space_info_used(space_info, true);
860
861 /*
862 * We may be flushing because suddenly we have less space than we had
863 * before, and now we're well over-committed based on our current free
864 * space. If that's the case add in our overage so we make sure to put
865 * appropriate pressure on the flushing state machine.
866 */
867 if (space_info->total_bytes + avail < used)
868 to_reclaim += used - (space_info->total_bytes + avail);
869
870 return to_reclaim;
871 }
872
873 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
874 struct btrfs_space_info *space_info)
875 {
876 const u64 global_rsv_size = btrfs_block_rsv_reserved(&fs_info->global_block_rsv);
877 u64 ordered, delalloc;
878 u64 thresh;
879 u64 used;
880
881 thresh = mult_perc(space_info->total_bytes, 90);
882
883 lockdep_assert_held(&space_info->lock);
884
885 /* If we're just plain full then async reclaim just slows us down. */
886 if ((space_info->bytes_used + space_info->bytes_reserved +
887 global_rsv_size) >= thresh)
888 return false;
889
890 used = space_info->bytes_may_use + space_info->bytes_pinned;
891
892 /* The total flushable belongs to the global rsv, don't flush. */
893 if (global_rsv_size >= used)
894 return false;
895
896 /*
897 * 128MiB is 1/4 of the maximum global rsv size. If we have less than
898 * that devoted to other reservations then there's no sense in flushing,
899 * we don't have a lot of things that need flushing.
900 */
901 if (used - global_rsv_size <= SZ_128M)
902 return false;
903
904 /*
905 * We have tickets queued, bail so we don't compete with the async
906 * flushers.
907 */
908 if (space_info->reclaim_size)
909 return false;
910
911 /*
912 * If we have over half of the free space occupied by reservations or
913 * pinned then we want to start flushing.
914 *
915 * We do not do the traditional thing here, which is to say
916 *
917 * if (used >= ((total_bytes + avail) / 2))
918 * return 1;
919 *
920 * because this doesn't quite work how we want. If we had more than 50%
921 * of the space_info used by bytes_used and we had 0 available we'd just
922 * constantly run the background flusher. Instead we want it to kick in
923 * if our reclaimable space exceeds our clamped free space.
924 *
925 * Our clamping range is 2^1 -> 2^8. Practically speaking that means
926 * the following:
927 *
928 * Amount of RAM Minimum threshold Maximum threshold
929 *
930 * 256GiB 1GiB 128GiB
931 * 128GiB 512MiB 64GiB
932 * 64GiB 256MiB 32GiB
933 * 32GiB 128MiB 16GiB
934 * 16GiB 64MiB 8GiB
935 *
936 * These are the range our thresholds will fall in, corresponding to how
937 * much delalloc we need for the background flusher to kick in.
938 */
939
940 thresh = calc_available_free_space(fs_info, space_info,
941 BTRFS_RESERVE_FLUSH_ALL);
942 used = space_info->bytes_used + space_info->bytes_reserved +
943 space_info->bytes_readonly + global_rsv_size;
944 if (used < space_info->total_bytes)
945 thresh += space_info->total_bytes - used;
946 thresh >>= space_info->clamp;
947
948 used = space_info->bytes_pinned;
949
950 /*
951 * If we have more ordered bytes than delalloc bytes then we're either
952 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting
953 * around. Preemptive flushing is only useful in that it can free up
954 * space before tickets need to wait for things to finish. In the case
955 * of ordered extents, preemptively waiting on ordered extents gets us
956 * nothing, if our reservations are tied up in ordered extents we'll
957 * simply have to slow down writers by forcing them to wait on ordered
958 * extents.
959 *
960 * In the case that ordered is larger than delalloc, only include the
961 * block reserves that we would actually be able to directly reclaim
962 * from. In this case if we're heavy on metadata operations this will
963 * clearly be heavy enough to warrant preemptive flushing. In the case
964 * of heavy DIO or ordered reservations, preemptive flushing will just
965 * waste time and cause us to slow down.
966 *
967 * We want to make sure we truly are maxed out on ordered however, so
968 * cut ordered in half, and if it's still higher than delalloc then we
969 * can keep flushing. This is to avoid the case where we start
970 * flushing, and now delalloc == ordered and we stop preemptively
971 * flushing when we could still have several gigs of delalloc to flush.
972 */
973 ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1;
974 delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes);
975 if (ordered >= delalloc)
976 used += btrfs_block_rsv_reserved(&fs_info->delayed_refs_rsv) +
977 btrfs_block_rsv_reserved(&fs_info->delayed_block_rsv);
978 else
979 used += space_info->bytes_may_use - global_rsv_size;
980
981 return (used >= thresh && !btrfs_fs_closing(fs_info) &&
982 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
983 }
984
985 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
986 struct btrfs_space_info *space_info,
987 struct reserve_ticket *ticket)
988 {
989 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
990 u64 min_bytes;
991
992 if (!ticket->steal)
993 return false;
994
995 if (global_rsv->space_info != space_info)
996 return false;
997
998 spin_lock(&global_rsv->lock);
999 min_bytes = mult_perc(global_rsv->size, 10);
1000 if (global_rsv->reserved < min_bytes + ticket->bytes) {
1001 spin_unlock(&global_rsv->lock);
1002 return false;
1003 }
1004 global_rsv->reserved -= ticket->bytes;
1005 remove_ticket(space_info, ticket);
1006 ticket->bytes = 0;
1007 wake_up(&ticket->wait);
1008 space_info->tickets_id++;
1009 if (global_rsv->reserved < global_rsv->size)
1010 global_rsv->full = 0;
1011 spin_unlock(&global_rsv->lock);
1012
1013 return true;
1014 }
1015
1016 /*
1017 * We've exhausted our flushing, start failing tickets.
1018 *
1019 * @fs_info - fs_info for this fs
1020 * @space_info - the space info we were flushing
1021 *
1022 * We call this when we've exhausted our flushing ability and haven't made
1023 * progress in satisfying tickets. The reservation code handles tickets in
1024 * order, so if there is a large ticket first and then smaller ones we could
1025 * very well satisfy the smaller tickets. This will attempt to wake up any
1026 * tickets in the list to catch this case.
1027 *
1028 * This function returns true if it was able to make progress by clearing out
1029 * other tickets, or if it stumbles across a ticket that was smaller than the
1030 * first ticket.
1031 */
1032 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
1033 struct btrfs_space_info *space_info)
1034 {
1035 struct reserve_ticket *ticket;
1036 u64 tickets_id = space_info->tickets_id;
1037 const bool aborted = BTRFS_FS_ERROR(fs_info);
1038
1039 trace_btrfs_fail_all_tickets(fs_info, space_info);
1040
1041 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1042 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
1043 __btrfs_dump_space_info(fs_info, space_info);
1044 }
1045
1046 while (!list_empty(&space_info->tickets) &&
1047 tickets_id == space_info->tickets_id) {
1048 ticket = list_first_entry(&space_info->tickets,
1049 struct reserve_ticket, list);
1050
1051 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket))
1052 return true;
1053
1054 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1055 btrfs_info(fs_info, "failing ticket with %llu bytes",
1056 ticket->bytes);
1057
1058 remove_ticket(space_info, ticket);
1059 if (aborted)
1060 ticket->error = -EIO;
1061 else
1062 ticket->error = -ENOSPC;
1063 wake_up(&ticket->wait);
1064
1065 /*
1066 * We're just throwing tickets away, so more flushing may not
1067 * trip over btrfs_try_granting_tickets, so we need to call it
1068 * here to see if we can make progress with the next ticket in
1069 * the list.
1070 */
1071 if (!aborted)
1072 btrfs_try_granting_tickets(fs_info, space_info);
1073 }
1074 return (tickets_id != space_info->tickets_id);
1075 }
1076
1077 /*
1078 * This is for normal flushers, we can wait all goddamned day if we want to. We
1079 * will loop and continuously try to flush as long as we are making progress.
1080 * We count progress as clearing off tickets each time we have to loop.
1081 */
1082 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
1083 {
1084 struct btrfs_fs_info *fs_info;
1085 struct btrfs_space_info *space_info;
1086 u64 to_reclaim;
1087 enum btrfs_flush_state flush_state;
1088 int commit_cycles = 0;
1089 u64 last_tickets_id;
1090
1091 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
1092 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1093
1094 spin_lock(&space_info->lock);
1095 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1096 if (!to_reclaim) {
1097 space_info->flush = 0;
1098 spin_unlock(&space_info->lock);
1099 return;
1100 }
1101 last_tickets_id = space_info->tickets_id;
1102 spin_unlock(&space_info->lock);
1103
1104 flush_state = FLUSH_DELAYED_ITEMS_NR;
1105 do {
1106 flush_space(fs_info, space_info, to_reclaim, flush_state, false);
1107 spin_lock(&space_info->lock);
1108 if (list_empty(&space_info->tickets)) {
1109 space_info->flush = 0;
1110 spin_unlock(&space_info->lock);
1111 return;
1112 }
1113 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
1114 space_info);
1115 if (last_tickets_id == space_info->tickets_id) {
1116 flush_state++;
1117 } else {
1118 last_tickets_id = space_info->tickets_id;
1119 flush_state = FLUSH_DELAYED_ITEMS_NR;
1120 if (commit_cycles)
1121 commit_cycles--;
1122 }
1123
1124 /*
1125 * We do not want to empty the system of delalloc unless we're
1126 * under heavy pressure, so allow one trip through the flushing
1127 * logic before we start doing a FLUSH_DELALLOC_FULL.
1128 */
1129 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles)
1130 flush_state++;
1131
1132 /*
1133 * We don't want to force a chunk allocation until we've tried
1134 * pretty hard to reclaim space. Think of the case where we
1135 * freed up a bunch of space and so have a lot of pinned space
1136 * to reclaim. We would rather use that than possibly create a
1137 * underutilized metadata chunk. So if this is our first run
1138 * through the flushing state machine skip ALLOC_CHUNK_FORCE and
1139 * commit the transaction. If nothing has changed the next go
1140 * around then we can force a chunk allocation.
1141 */
1142 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
1143 flush_state++;
1144
1145 if (flush_state > COMMIT_TRANS) {
1146 commit_cycles++;
1147 if (commit_cycles > 2) {
1148 if (maybe_fail_all_tickets(fs_info, space_info)) {
1149 flush_state = FLUSH_DELAYED_ITEMS_NR;
1150 commit_cycles--;
1151 } else {
1152 space_info->flush = 0;
1153 }
1154 } else {
1155 flush_state = FLUSH_DELAYED_ITEMS_NR;
1156 }
1157 }
1158 spin_unlock(&space_info->lock);
1159 } while (flush_state <= COMMIT_TRANS);
1160 }
1161
1162 /*
1163 * This handles pre-flushing of metadata space before we get to the point that
1164 * we need to start blocking threads on tickets. The logic here is different
1165 * from the other flush paths because it doesn't rely on tickets to tell us how
1166 * much we need to flush, instead it attempts to keep us below the 80% full
1167 * watermark of space by flushing whichever reservation pool is currently the
1168 * largest.
1169 */
1170 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
1171 {
1172 struct btrfs_fs_info *fs_info;
1173 struct btrfs_space_info *space_info;
1174 struct btrfs_block_rsv *delayed_block_rsv;
1175 struct btrfs_block_rsv *delayed_refs_rsv;
1176 struct btrfs_block_rsv *global_rsv;
1177 struct btrfs_block_rsv *trans_rsv;
1178 int loops = 0;
1179
1180 fs_info = container_of(work, struct btrfs_fs_info,
1181 preempt_reclaim_work);
1182 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1183 delayed_block_rsv = &fs_info->delayed_block_rsv;
1184 delayed_refs_rsv = &fs_info->delayed_refs_rsv;
1185 global_rsv = &fs_info->global_block_rsv;
1186 trans_rsv = &fs_info->trans_block_rsv;
1187
1188 spin_lock(&space_info->lock);
1189 while (need_preemptive_reclaim(fs_info, space_info)) {
1190 enum btrfs_flush_state flush;
1191 u64 delalloc_size = 0;
1192 u64 to_reclaim, block_rsv_size;
1193 const u64 global_rsv_size = btrfs_block_rsv_reserved(global_rsv);
1194
1195 loops++;
1196
1197 /*
1198 * We don't have a precise counter for the metadata being
1199 * reserved for delalloc, so we'll approximate it by subtracting
1200 * out the block rsv's space from the bytes_may_use. If that
1201 * amount is higher than the individual reserves, then we can
1202 * assume it's tied up in delalloc reservations.
1203 */
1204 block_rsv_size = global_rsv_size +
1205 btrfs_block_rsv_reserved(delayed_block_rsv) +
1206 btrfs_block_rsv_reserved(delayed_refs_rsv) +
1207 btrfs_block_rsv_reserved(trans_rsv);
1208 if (block_rsv_size < space_info->bytes_may_use)
1209 delalloc_size = space_info->bytes_may_use - block_rsv_size;
1210
1211 /*
1212 * We don't want to include the global_rsv in our calculation,
1213 * because that's space we can't touch. Subtract it from the
1214 * block_rsv_size for the next checks.
1215 */
1216 block_rsv_size -= global_rsv_size;
1217
1218 /*
1219 * We really want to avoid flushing delalloc too much, as it
1220 * could result in poor allocation patterns, so only flush it if
1221 * it's larger than the rest of the pools combined.
1222 */
1223 if (delalloc_size > block_rsv_size) {
1224 to_reclaim = delalloc_size;
1225 flush = FLUSH_DELALLOC;
1226 } else if (space_info->bytes_pinned >
1227 (btrfs_block_rsv_reserved(delayed_block_rsv) +
1228 btrfs_block_rsv_reserved(delayed_refs_rsv))) {
1229 to_reclaim = space_info->bytes_pinned;
1230 flush = COMMIT_TRANS;
1231 } else if (btrfs_block_rsv_reserved(delayed_block_rsv) >
1232 btrfs_block_rsv_reserved(delayed_refs_rsv)) {
1233 to_reclaim = btrfs_block_rsv_reserved(delayed_block_rsv);
1234 flush = FLUSH_DELAYED_ITEMS_NR;
1235 } else {
1236 to_reclaim = btrfs_block_rsv_reserved(delayed_refs_rsv);
1237 flush = FLUSH_DELAYED_REFS_NR;
1238 }
1239
1240 spin_unlock(&space_info->lock);
1241
1242 /*
1243 * We don't want to reclaim everything, just a portion, so scale
1244 * down the to_reclaim by 1/4. If it takes us down to 0,
1245 * reclaim 1 items worth.
1246 */
1247 to_reclaim >>= 2;
1248 if (!to_reclaim)
1249 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1250 flush_space(fs_info, space_info, to_reclaim, flush, true);
1251 cond_resched();
1252 spin_lock(&space_info->lock);
1253 }
1254
1255 /* We only went through once, back off our clamping. */
1256 if (loops == 1 && !space_info->reclaim_size)
1257 space_info->clamp = max(1, space_info->clamp - 1);
1258 trace_btrfs_done_preemptive_reclaim(fs_info, space_info);
1259 spin_unlock(&space_info->lock);
1260 }
1261
1262 /*
1263 * FLUSH_DELALLOC_WAIT:
1264 * Space is freed from flushing delalloc in one of two ways.
1265 *
1266 * 1) compression is on and we allocate less space than we reserved
1267 * 2) we are overwriting existing space
1268 *
1269 * For #1 that extra space is reclaimed as soon as the delalloc pages are
1270 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1271 * length to ->bytes_reserved, and subtracts the reserved space from
1272 * ->bytes_may_use.
1273 *
1274 * For #2 this is trickier. Once the ordered extent runs we will drop the
1275 * extent in the range we are overwriting, which creates a delayed ref for
1276 * that freed extent. This however is not reclaimed until the transaction
1277 * commits, thus the next stages.
1278 *
1279 * RUN_DELAYED_IPUTS
1280 * If we are freeing inodes, we want to make sure all delayed iputs have
1281 * completed, because they could have been on an inode with i_nlink == 0, and
1282 * thus have been truncated and freed up space. But again this space is not
1283 * immediately re-usable, it comes in the form of a delayed ref, which must be
1284 * run and then the transaction must be committed.
1285 *
1286 * COMMIT_TRANS
1287 * This is where we reclaim all of the pinned space generated by running the
1288 * iputs
1289 *
1290 * ALLOC_CHUNK_FORCE
1291 * For data we start with alloc chunk force, however we could have been full
1292 * before, and then the transaction commit could have freed new block groups,
1293 * so if we now have space to allocate do the force chunk allocation.
1294 */
1295 static const enum btrfs_flush_state data_flush_states[] = {
1296 FLUSH_DELALLOC_FULL,
1297 RUN_DELAYED_IPUTS,
1298 COMMIT_TRANS,
1299 ALLOC_CHUNK_FORCE,
1300 };
1301
1302 static void btrfs_async_reclaim_data_space(struct work_struct *work)
1303 {
1304 struct btrfs_fs_info *fs_info;
1305 struct btrfs_space_info *space_info;
1306 u64 last_tickets_id;
1307 enum btrfs_flush_state flush_state = 0;
1308
1309 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1310 space_info = fs_info->data_sinfo;
1311
1312 spin_lock(&space_info->lock);
1313 if (list_empty(&space_info->tickets)) {
1314 space_info->flush = 0;
1315 spin_unlock(&space_info->lock);
1316 return;
1317 }
1318 last_tickets_id = space_info->tickets_id;
1319 spin_unlock(&space_info->lock);
1320
1321 while (!space_info->full) {
1322 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, 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 /* Something happened, fail everything and bail. */
1331 if (BTRFS_FS_ERROR(fs_info))
1332 goto aborted_fs;
1333 last_tickets_id = space_info->tickets_id;
1334 spin_unlock(&space_info->lock);
1335 }
1336
1337 while (flush_state < ARRAY_SIZE(data_flush_states)) {
1338 flush_space(fs_info, space_info, U64_MAX,
1339 data_flush_states[flush_state], false);
1340 spin_lock(&space_info->lock);
1341 if (list_empty(&space_info->tickets)) {
1342 space_info->flush = 0;
1343 spin_unlock(&space_info->lock);
1344 return;
1345 }
1346
1347 if (last_tickets_id == space_info->tickets_id) {
1348 flush_state++;
1349 } else {
1350 last_tickets_id = space_info->tickets_id;
1351 flush_state = 0;
1352 }
1353
1354 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1355 if (space_info->full) {
1356 if (maybe_fail_all_tickets(fs_info, space_info))
1357 flush_state = 0;
1358 else
1359 space_info->flush = 0;
1360 } else {
1361 flush_state = 0;
1362 }
1363
1364 /* Something happened, fail everything and bail. */
1365 if (BTRFS_FS_ERROR(fs_info))
1366 goto aborted_fs;
1367
1368 }
1369 spin_unlock(&space_info->lock);
1370 }
1371 return;
1372
1373 aborted_fs:
1374 maybe_fail_all_tickets(fs_info, space_info);
1375 space_info->flush = 0;
1376 spin_unlock(&space_info->lock);
1377 }
1378
1379 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1380 {
1381 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1382 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1383 INIT_WORK(&fs_info->preempt_reclaim_work,
1384 btrfs_preempt_reclaim_metadata_space);
1385 }
1386
1387 static const enum btrfs_flush_state priority_flush_states[] = {
1388 FLUSH_DELAYED_ITEMS_NR,
1389 FLUSH_DELAYED_ITEMS,
1390 ALLOC_CHUNK,
1391 };
1392
1393 static const enum btrfs_flush_state evict_flush_states[] = {
1394 FLUSH_DELAYED_ITEMS_NR,
1395 FLUSH_DELAYED_ITEMS,
1396 FLUSH_DELAYED_REFS_NR,
1397 FLUSH_DELAYED_REFS,
1398 FLUSH_DELALLOC,
1399 FLUSH_DELALLOC_WAIT,
1400 FLUSH_DELALLOC_FULL,
1401 ALLOC_CHUNK,
1402 COMMIT_TRANS,
1403 };
1404
1405 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1406 struct btrfs_space_info *space_info,
1407 struct reserve_ticket *ticket,
1408 const enum btrfs_flush_state *states,
1409 int states_nr)
1410 {
1411 u64 to_reclaim;
1412 int flush_state = 0;
1413
1414 spin_lock(&space_info->lock);
1415 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1416 /*
1417 * This is the priority reclaim path, so to_reclaim could be >0 still
1418 * because we may have only satisfied the priority tickets and still
1419 * left non priority tickets on the list. We would then have
1420 * to_reclaim but ->bytes == 0.
1421 */
1422 if (ticket->bytes == 0) {
1423 spin_unlock(&space_info->lock);
1424 return;
1425 }
1426
1427 while (flush_state < states_nr) {
1428 spin_unlock(&space_info->lock);
1429 flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1430 false);
1431 flush_state++;
1432 spin_lock(&space_info->lock);
1433 if (ticket->bytes == 0) {
1434 spin_unlock(&space_info->lock);
1435 return;
1436 }
1437 }
1438
1439 /*
1440 * Attempt to steal from the global rsv if we can, except if the fs was
1441 * turned into error mode due to a transaction abort when flushing space
1442 * above, in that case fail with the abort error instead of returning
1443 * success to the caller if we can steal from the global rsv - this is
1444 * just to have caller fail immeditelly instead of later when trying to
1445 * modify the fs, making it easier to debug -ENOSPC problems.
1446 */
1447 if (BTRFS_FS_ERROR(fs_info)) {
1448 ticket->error = BTRFS_FS_ERROR(fs_info);
1449 remove_ticket(space_info, ticket);
1450 } else if (!steal_from_global_rsv(fs_info, space_info, ticket)) {
1451 ticket->error = -ENOSPC;
1452 remove_ticket(space_info, ticket);
1453 }
1454
1455 /*
1456 * We must run try_granting_tickets here because we could be a large
1457 * ticket in front of a smaller ticket that can now be satisfied with
1458 * the available space.
1459 */
1460 btrfs_try_granting_tickets(fs_info, space_info);
1461 spin_unlock(&space_info->lock);
1462 }
1463
1464 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1465 struct btrfs_space_info *space_info,
1466 struct reserve_ticket *ticket)
1467 {
1468 spin_lock(&space_info->lock);
1469
1470 /* We could have been granted before we got here. */
1471 if (ticket->bytes == 0) {
1472 spin_unlock(&space_info->lock);
1473 return;
1474 }
1475
1476 while (!space_info->full) {
1477 spin_unlock(&space_info->lock);
1478 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1479 spin_lock(&space_info->lock);
1480 if (ticket->bytes == 0) {
1481 spin_unlock(&space_info->lock);
1482 return;
1483 }
1484 }
1485
1486 ticket->error = -ENOSPC;
1487 remove_ticket(space_info, ticket);
1488 btrfs_try_granting_tickets(fs_info, space_info);
1489 spin_unlock(&space_info->lock);
1490 }
1491
1492 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1493 struct btrfs_space_info *space_info,
1494 struct reserve_ticket *ticket)
1495
1496 {
1497 DEFINE_WAIT(wait);
1498 int ret = 0;
1499
1500 spin_lock(&space_info->lock);
1501 while (ticket->bytes > 0 && ticket->error == 0) {
1502 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1503 if (ret) {
1504 /*
1505 * Delete us from the list. After we unlock the space
1506 * info, we don't want the async reclaim job to reserve
1507 * space for this ticket. If that would happen, then the
1508 * ticket's task would not known that space was reserved
1509 * despite getting an error, resulting in a space leak
1510 * (bytes_may_use counter of our space_info).
1511 */
1512 remove_ticket(space_info, ticket);
1513 ticket->error = -EINTR;
1514 break;
1515 }
1516 spin_unlock(&space_info->lock);
1517
1518 schedule();
1519
1520 finish_wait(&ticket->wait, &wait);
1521 spin_lock(&space_info->lock);
1522 }
1523 spin_unlock(&space_info->lock);
1524 }
1525
1526 /*
1527 * Do the appropriate flushing and waiting for a ticket.
1528 *
1529 * @fs_info: the filesystem
1530 * @space_info: space info for the reservation
1531 * @ticket: ticket for the reservation
1532 * @start_ns: timestamp when the reservation started
1533 * @orig_bytes: amount of bytes originally reserved
1534 * @flush: how much we can flush
1535 *
1536 * This does the work of figuring out how to flush for the ticket, waiting for
1537 * the reservation, and returning the appropriate error if there is one.
1538 */
1539 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1540 struct btrfs_space_info *space_info,
1541 struct reserve_ticket *ticket,
1542 u64 start_ns, u64 orig_bytes,
1543 enum btrfs_reserve_flush_enum flush)
1544 {
1545 int ret;
1546
1547 switch (flush) {
1548 case BTRFS_RESERVE_FLUSH_DATA:
1549 case BTRFS_RESERVE_FLUSH_ALL:
1550 case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1551 wait_reserve_ticket(fs_info, space_info, ticket);
1552 break;
1553 case BTRFS_RESERVE_FLUSH_LIMIT:
1554 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1555 priority_flush_states,
1556 ARRAY_SIZE(priority_flush_states));
1557 break;
1558 case BTRFS_RESERVE_FLUSH_EVICT:
1559 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1560 evict_flush_states,
1561 ARRAY_SIZE(evict_flush_states));
1562 break;
1563 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1564 priority_reclaim_data_space(fs_info, space_info, ticket);
1565 break;
1566 default:
1567 ASSERT(0);
1568 break;
1569 }
1570
1571 ret = ticket->error;
1572 ASSERT(list_empty(&ticket->list));
1573 /*
1574 * Check that we can't have an error set if the reservation succeeded,
1575 * as that would confuse tasks and lead them to error out without
1576 * releasing reserved space (if an error happens the expectation is that
1577 * space wasn't reserved at all).
1578 */
1579 ASSERT(!(ticket->bytes == 0 && ticket->error));
1580 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes,
1581 start_ns, flush, ticket->error);
1582 return ret;
1583 }
1584
1585 /*
1586 * This returns true if this flush state will go through the ordinary flushing
1587 * code.
1588 */
1589 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1590 {
1591 return (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1592 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1593 }
1594
1595 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1596 struct btrfs_space_info *space_info)
1597 {
1598 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1599 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1600
1601 /*
1602 * If we're heavy on ordered operations then clamping won't help us. We
1603 * need to clamp specifically to keep up with dirty'ing buffered
1604 * writers, because there's not a 1:1 correlation of writing delalloc
1605 * and freeing space, like there is with flushing delayed refs or
1606 * delayed nodes. If we're already more ordered than delalloc then
1607 * we're keeping up, otherwise we aren't and should probably clamp.
1608 */
1609 if (ordered < delalloc)
1610 space_info->clamp = min(space_info->clamp + 1, 8);
1611 }
1612
1613 static inline bool can_steal(enum btrfs_reserve_flush_enum flush)
1614 {
1615 return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1616 flush == BTRFS_RESERVE_FLUSH_EVICT);
1617 }
1618
1619 /*
1620 * NO_FLUSH and FLUSH_EMERGENCY don't want to create a ticket, they just want to
1621 * fail as quickly as possible.
1622 */
1623 static inline bool can_ticket(enum btrfs_reserve_flush_enum flush)
1624 {
1625 return (flush != BTRFS_RESERVE_NO_FLUSH &&
1626 flush != BTRFS_RESERVE_FLUSH_EMERGENCY);
1627 }
1628
1629 /*
1630 * Try to reserve bytes from the block_rsv's space.
1631 *
1632 * @fs_info: the filesystem
1633 * @space_info: space info we want to allocate from
1634 * @orig_bytes: number of bytes we want
1635 * @flush: whether or not we can flush to make our reservation
1636 *
1637 * This will reserve orig_bytes number of bytes from the space info associated
1638 * with the block_rsv. If there is not enough space it will make an attempt to
1639 * flush out space to make room. It will do this by flushing delalloc if
1640 * possible or committing the transaction. If flush is 0 then no attempts to
1641 * regain reservations will be made and this will fail if there is not enough
1642 * space already.
1643 */
1644 static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1645 struct btrfs_space_info *space_info, u64 orig_bytes,
1646 enum btrfs_reserve_flush_enum flush)
1647 {
1648 struct work_struct *async_work;
1649 struct reserve_ticket ticket;
1650 u64 start_ns = 0;
1651 u64 used;
1652 int ret = -ENOSPC;
1653 bool pending_tickets;
1654
1655 ASSERT(orig_bytes);
1656 /*
1657 * If have a transaction handle (current->journal_info != NULL), then
1658 * the flush method can not be neither BTRFS_RESERVE_FLUSH_ALL* nor
1659 * BTRFS_RESERVE_FLUSH_EVICT, as we could deadlock because those
1660 * flushing methods can trigger transaction commits.
1661 */
1662 if (current->journal_info) {
1663 /* One assert per line for easier debugging. */
1664 ASSERT(flush != BTRFS_RESERVE_FLUSH_ALL);
1665 ASSERT(flush != BTRFS_RESERVE_FLUSH_ALL_STEAL);
1666 ASSERT(flush != BTRFS_RESERVE_FLUSH_EVICT);
1667 }
1668
1669 if (flush == BTRFS_RESERVE_FLUSH_DATA)
1670 async_work = &fs_info->async_data_reclaim_work;
1671 else
1672 async_work = &fs_info->async_reclaim_work;
1673
1674 spin_lock(&space_info->lock);
1675 used = btrfs_space_info_used(space_info, true);
1676
1677 /*
1678 * We don't want NO_FLUSH allocations to jump everybody, they can
1679 * generally handle ENOSPC in a different way, so treat them the same as
1680 * normal flushers when it comes to skipping pending tickets.
1681 */
1682 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1683 pending_tickets = !list_empty(&space_info->tickets) ||
1684 !list_empty(&space_info->priority_tickets);
1685 else
1686 pending_tickets = !list_empty(&space_info->priority_tickets);
1687
1688 /*
1689 * Carry on if we have enough space (short-circuit) OR call
1690 * can_overcommit() to ensure we can overcommit to continue.
1691 */
1692 if (!pending_tickets &&
1693 ((used + orig_bytes <= space_info->total_bytes) ||
1694 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1695 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1696 orig_bytes);
1697 ret = 0;
1698 }
1699
1700 /*
1701 * Things are dire, we need to make a reservation so we don't abort. We
1702 * will let this reservation go through as long as we have actual space
1703 * left to allocate for the block.
1704 */
1705 if (ret && unlikely(flush == BTRFS_RESERVE_FLUSH_EMERGENCY)) {
1706 used = btrfs_space_info_used(space_info, false);
1707 if (used + orig_bytes <= space_info->total_bytes) {
1708 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1709 orig_bytes);
1710 ret = 0;
1711 }
1712 }
1713
1714 /*
1715 * If we couldn't make a reservation then setup our reservation ticket
1716 * and kick the async worker if it's not already running.
1717 *
1718 * If we are a priority flusher then we just need to add our ticket to
1719 * the list and we will do our own flushing further down.
1720 */
1721 if (ret && can_ticket(flush)) {
1722 ticket.bytes = orig_bytes;
1723 ticket.error = 0;
1724 space_info->reclaim_size += ticket.bytes;
1725 init_waitqueue_head(&ticket.wait);
1726 ticket.steal = can_steal(flush);
1727 if (trace_btrfs_reserve_ticket_enabled())
1728 start_ns = ktime_get_ns();
1729
1730 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1731 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1732 flush == BTRFS_RESERVE_FLUSH_DATA) {
1733 list_add_tail(&ticket.list, &space_info->tickets);
1734 if (!space_info->flush) {
1735 /*
1736 * We were forced to add a reserve ticket, so
1737 * our preemptive flushing is unable to keep
1738 * up. Clamp down on the threshold for the
1739 * preemptive flushing in order to keep up with
1740 * the workload.
1741 */
1742 maybe_clamp_preempt(fs_info, space_info);
1743
1744 space_info->flush = 1;
1745 trace_btrfs_trigger_flush(fs_info,
1746 space_info->flags,
1747 orig_bytes, flush,
1748 "enospc");
1749 queue_work(system_unbound_wq, async_work);
1750 }
1751 } else {
1752 list_add_tail(&ticket.list,
1753 &space_info->priority_tickets);
1754 }
1755 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1756 /*
1757 * We will do the space reservation dance during log replay,
1758 * which means we won't have fs_info->fs_root set, so don't do
1759 * the async reclaim as we will panic.
1760 */
1761 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1762 !work_busy(&fs_info->preempt_reclaim_work) &&
1763 need_preemptive_reclaim(fs_info, space_info)) {
1764 trace_btrfs_trigger_flush(fs_info, space_info->flags,
1765 orig_bytes, flush, "preempt");
1766 queue_work(system_unbound_wq,
1767 &fs_info->preempt_reclaim_work);
1768 }
1769 }
1770 spin_unlock(&space_info->lock);
1771 if (!ret || !can_ticket(flush))
1772 return ret;
1773
1774 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1775 orig_bytes, flush);
1776 }
1777
1778 /*
1779 * Try to reserve metadata bytes from the block_rsv's space.
1780 *
1781 * @fs_info: the filesystem
1782 * @space_info: the space_info we're allocating for
1783 * @orig_bytes: number of bytes we want
1784 * @flush: whether or not we can flush to make our reservation
1785 *
1786 * This will reserve orig_bytes number of bytes from the space info associated
1787 * with the block_rsv. If there is not enough space it will make an attempt to
1788 * flush out space to make room. It will do this by flushing delalloc if
1789 * possible or committing the transaction. If flush is 0 then no attempts to
1790 * regain reservations will be made and this will fail if there is not enough
1791 * space already.
1792 */
1793 int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1794 struct btrfs_space_info *space_info,
1795 u64 orig_bytes,
1796 enum btrfs_reserve_flush_enum flush)
1797 {
1798 int ret;
1799
1800 ret = __reserve_bytes(fs_info, space_info, orig_bytes, flush);
1801 if (ret == -ENOSPC) {
1802 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1803 space_info->flags, orig_bytes, 1);
1804
1805 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1806 btrfs_dump_space_info(fs_info, space_info, orig_bytes, 0);
1807 }
1808 return ret;
1809 }
1810
1811 /*
1812 * Try to reserve data bytes for an allocation.
1813 *
1814 * @fs_info: the filesystem
1815 * @bytes: number of bytes we need
1816 * @flush: how we are allowed to flush
1817 *
1818 * This will reserve bytes from the data space info. If there is not enough
1819 * space then we will attempt to flush space as specified by flush.
1820 */
1821 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1822 enum btrfs_reserve_flush_enum flush)
1823 {
1824 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1825 int ret;
1826
1827 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1828 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE ||
1829 flush == BTRFS_RESERVE_NO_FLUSH);
1830 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1831
1832 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1833 if (ret == -ENOSPC) {
1834 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1835 data_sinfo->flags, bytes, 1);
1836 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1837 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1838 }
1839 return ret;
1840 }
1841
1842 /* Dump all the space infos when we abort a transaction due to ENOSPC. */
1843 __cold void btrfs_dump_space_info_for_trans_abort(struct btrfs_fs_info *fs_info)
1844 {
1845 struct btrfs_space_info *space_info;
1846
1847 btrfs_info(fs_info, "dumping space info:");
1848 list_for_each_entry(space_info, &fs_info->space_info, list) {
1849 spin_lock(&space_info->lock);
1850 __btrfs_dump_space_info(fs_info, space_info);
1851 spin_unlock(&space_info->lock);
1852 }
1853 dump_global_block_rsv(fs_info);
1854 }
1855
1856 /*
1857 * Account the unused space of all the readonly block group in the space_info.
1858 * takes mirrors into account.
1859 */
1860 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
1861 {
1862 struct btrfs_block_group *block_group;
1863 u64 free_bytes = 0;
1864 int factor;
1865
1866 /* It's df, we don't care if it's racy */
1867 if (list_empty(&sinfo->ro_bgs))
1868 return 0;
1869
1870 spin_lock(&sinfo->lock);
1871 list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
1872 spin_lock(&block_group->lock);
1873
1874 if (!block_group->ro) {
1875 spin_unlock(&block_group->lock);
1876 continue;
1877 }
1878
1879 factor = btrfs_bg_type_to_factor(block_group->flags);
1880 free_bytes += (block_group->length -
1881 block_group->used) * factor;
1882
1883 spin_unlock(&block_group->lock);
1884 }
1885 spin_unlock(&sinfo->lock);
1886
1887 return free_bytes;
1888 }
1889
1890 static u64 calc_pct_ratio(u64 x, u64 y)
1891 {
1892 int err;
1893
1894 if (!y)
1895 return 0;
1896 again:
1897 err = check_mul_overflow(100, x, &x);
1898 if (err)
1899 goto lose_precision;
1900 return div64_u64(x, y);
1901 lose_precision:
1902 x >>= 10;
1903 y >>= 10;
1904 if (!y)
1905 y = 1;
1906 goto again;
1907 }
1908
1909 /*
1910 * A reasonable buffer for unallocated space is 10 data block_groups.
1911 * If we claw this back repeatedly, we can still achieve efficient
1912 * utilization when near full, and not do too much reclaim while
1913 * always maintaining a solid buffer for workloads that quickly
1914 * allocate and pressure the unallocated space.
1915 */
1916 static u64 calc_unalloc_target(struct btrfs_fs_info *fs_info)
1917 {
1918 u64 chunk_sz = calc_effective_data_chunk_size(fs_info);
1919
1920 return BTRFS_UNALLOC_BLOCK_GROUP_TARGET * chunk_sz;
1921 }
1922
1923 /*
1924 * The fundamental goal of automatic reclaim is to protect the filesystem's
1925 * unallocated space and thus minimize the probability of the filesystem going
1926 * read only when a metadata allocation failure causes a transaction abort.
1927 *
1928 * However, relocations happen into the space_info's unused space, therefore
1929 * automatic reclaim must also back off as that space runs low. There is no
1930 * value in doing trivial "relocations" of re-writing the same block group
1931 * into a fresh one.
1932 *
1933 * Furthermore, we want to avoid doing too much reclaim even if there are good
1934 * candidates. This is because the allocator is pretty good at filling up the
1935 * holes with writes. So we want to do just enough reclaim to try and stay
1936 * safe from running out of unallocated space but not be wasteful about it.
1937 *
1938 * Therefore, the dynamic reclaim threshold is calculated as follows:
1939 * - calculate a target unallocated amount of 5 block group sized chunks
1940 * - ratchet up the intensity of reclaim depending on how far we are from
1941 * that target by using a formula of unalloc / target to set the threshold.
1942 *
1943 * Typically with 10 block groups as the target, the discrete values this comes
1944 * out to are 0, 10, 20, ... , 80, 90, and 99.
1945 */
1946 static int calc_dynamic_reclaim_threshold(struct btrfs_space_info *space_info)
1947 {
1948 struct btrfs_fs_info *fs_info = space_info->fs_info;
1949 u64 unalloc = atomic64_read(&fs_info->free_chunk_space);
1950 u64 target = calc_unalloc_target(fs_info);
1951 u64 alloc = space_info->total_bytes;
1952 u64 used = btrfs_space_info_used(space_info, false);
1953 u64 unused = alloc - used;
1954 u64 want = target > unalloc ? target - unalloc : 0;
1955 u64 data_chunk_size = calc_effective_data_chunk_size(fs_info);
1956
1957 /* If we have no unused space, don't bother, it won't work anyway. */
1958 if (unused < data_chunk_size)
1959 return 0;
1960
1961 /* Cast to int is OK because want <= target. */
1962 return calc_pct_ratio(want, target);
1963 }
1964
1965 int btrfs_calc_reclaim_threshold(struct btrfs_space_info *space_info)
1966 {
1967 lockdep_assert_held(&space_info->lock);
1968
1969 if (READ_ONCE(space_info->dynamic_reclaim))
1970 return calc_dynamic_reclaim_threshold(space_info);
1971 return READ_ONCE(space_info->bg_reclaim_threshold);
1972 }
1973
1974 /*
1975 * Under "urgent" reclaim, we will reclaim even fresh block groups that have
1976 * recently seen successful allocations, as we are desperate to reclaim
1977 * whatever we can to avoid ENOSPC in a transaction leading to a readonly fs.
1978 */
1979 static bool is_reclaim_urgent(struct btrfs_space_info *space_info)
1980 {
1981 struct btrfs_fs_info *fs_info = space_info->fs_info;
1982 u64 unalloc = atomic64_read(&fs_info->free_chunk_space);
1983 u64 data_chunk_size = calc_effective_data_chunk_size(fs_info);
1984
1985 return unalloc < data_chunk_size;
1986 }
1987
1988 static int do_reclaim_sweep(struct btrfs_fs_info *fs_info,
1989 struct btrfs_space_info *space_info, int raid)
1990 {
1991 struct btrfs_block_group *bg;
1992 int thresh_pct;
1993 bool try_again = true;
1994 bool urgent;
1995
1996 spin_lock(&space_info->lock);
1997 urgent = is_reclaim_urgent(space_info);
1998 thresh_pct = btrfs_calc_reclaim_threshold(space_info);
1999 spin_unlock(&space_info->lock);
2000
2001 down_read(&space_info->groups_sem);
2002 again:
2003 list_for_each_entry(bg, &space_info->block_groups[raid], list) {
2004 u64 thresh;
2005 bool reclaim = false;
2006
2007 btrfs_get_block_group(bg);
2008 spin_lock(&bg->lock);
2009 thresh = mult_perc(bg->length, thresh_pct);
2010 if (bg->used < thresh && bg->reclaim_mark) {
2011 try_again = false;
2012 reclaim = true;
2013 }
2014 bg->reclaim_mark++;
2015 spin_unlock(&bg->lock);
2016 if (reclaim)
2017 btrfs_mark_bg_to_reclaim(bg);
2018 btrfs_put_block_group(bg);
2019 }
2020
2021 /*
2022 * In situations where we are very motivated to reclaim (low unalloc)
2023 * use two passes to make the reclaim mark check best effort.
2024 *
2025 * If we have any staler groups, we don't touch the fresher ones, but if we
2026 * really need a block group, do take a fresh one.
2027 */
2028 if (try_again && urgent) {
2029 try_again = false;
2030 goto again;
2031 }
2032
2033 up_read(&space_info->groups_sem);
2034 return 0;
2035 }
2036
2037 void btrfs_space_info_update_reclaimable(struct btrfs_space_info *space_info, s64 bytes)
2038 {
2039 u64 chunk_sz = calc_effective_data_chunk_size(space_info->fs_info);
2040
2041 lockdep_assert_held(&space_info->lock);
2042 space_info->reclaimable_bytes += bytes;
2043
2044 if (space_info->reclaimable_bytes >= chunk_sz)
2045 btrfs_set_periodic_reclaim_ready(space_info, true);
2046 }
2047
2048 void btrfs_set_periodic_reclaim_ready(struct btrfs_space_info *space_info, bool ready)
2049 {
2050 lockdep_assert_held(&space_info->lock);
2051 if (!READ_ONCE(space_info->periodic_reclaim))
2052 return;
2053 if (ready != space_info->periodic_reclaim_ready) {
2054 space_info->periodic_reclaim_ready = ready;
2055 if (!ready)
2056 space_info->reclaimable_bytes = 0;
2057 }
2058 }
2059
2060 bool btrfs_should_periodic_reclaim(struct btrfs_space_info *space_info)
2061 {
2062 bool ret;
2063
2064 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
2065 return false;
2066 if (!READ_ONCE(space_info->periodic_reclaim))
2067 return false;
2068
2069 spin_lock(&space_info->lock);
2070 ret = space_info->periodic_reclaim_ready;
2071 btrfs_set_periodic_reclaim_ready(space_info, false);
2072 spin_unlock(&space_info->lock);
2073
2074 return ret;
2075 }
2076
2077 int btrfs_reclaim_sweep(struct btrfs_fs_info *fs_info)
2078 {
2079 int ret;
2080 int raid;
2081 struct btrfs_space_info *space_info;
2082
2083 list_for_each_entry(space_info, &fs_info->space_info, list) {
2084 if (!btrfs_should_periodic_reclaim(space_info))
2085 continue;
2086 for (raid = 0; raid < BTRFS_NR_RAID_TYPES; raid++) {
2087 ret = do_reclaim_sweep(fs_info, space_info, raid);
2088 if (ret)
2089 return ret;
2090 }
2091 }
2092
2093 return ret;
2094 }
2095
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.