1 // SPDX-License-Identifier: GPL-2.0
3 #include "alloc_background.h"
4 #include "alloc_foreground.h"
5 #include "btree_cache.h"
7 #include "btree_key_cache.h"
8 #include "btree_update.h"
9 #include "btree_update_interior.h"
19 #include <linux/kthread.h>
20 #include <linux/math64.h>
21 #include <linux/random.h>
22 #include <linux/rculist.h>
23 #include <linux/rcupdate.h>
24 #include <linux/sched/task.h>
25 #include <linux/sort.h>
26 #include <trace/events/bcachefs.h>
28 const char * const bch2_allocator_states[] = {
35 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
36 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
41 /* Persistent alloc info: */
43 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
44 const void **p, unsigned field)
46 unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
49 if (!(a->fields & (1 << field)))
54 v = *((const u8 *) *p);
73 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
74 unsigned field, u64 v)
76 unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
81 a->v.fields |= 1 << field;
88 *((__le16 *) *p) = cpu_to_le16(v);
91 *((__le32 *) *p) = cpu_to_le32(v);
94 *((__le64 *) *p) = cpu_to_le64(v);
103 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
106 const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
107 const void *d = in->data;
112 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
113 BCH_ALLOC_FIELDS_V1()
117 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
120 struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
121 const u8 *in = a.v->data;
122 const u8 *end = bkey_val_end(a);
123 unsigned fieldnr = 0;
128 out->oldest_gen = a.v->oldest_gen;
129 out->data_type = a.v->data_type;
131 #define x(_name, _bits) \
132 if (fieldnr < a.v->nr_fields) { \
133 ret = bch2_varint_decode(in, end, &v); \
141 if (v != out->_name) \
145 BCH_ALLOC_FIELDS_V2()
150 static void bch2_alloc_pack_v2(struct bkey_alloc_buf *dst,
151 const struct bkey_alloc_unpacked src)
153 struct bkey_i_alloc_v2 *a = bkey_alloc_v2_init(&dst->k);
154 unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
156 u8 *end = (void *) &dst[1];
157 u8 *last_nonzero_field = out;
160 a->k.p = POS(src.dev, src.bucket);
162 a->v.oldest_gen = src.oldest_gen;
163 a->v.data_type = src.data_type;
165 #define x(_name, _bits) \
169 out += bch2_varint_encode(out, src._name); \
171 last_nonzero_field = out; \
172 last_nonzero_fieldnr = nr_fields; \
177 BCH_ALLOC_FIELDS_V2()
181 out = last_nonzero_field;
182 a->v.nr_fields = last_nonzero_fieldnr;
184 bytes = (u8 *) out - (u8 *) &a->v;
185 set_bkey_val_bytes(&a->k, bytes);
186 memset_u64s_tail(&a->v, 0, bytes);
189 struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
191 struct bkey_alloc_unpacked ret = {
193 .bucket = k.k->p.offset,
197 if (k.k->type == KEY_TYPE_alloc_v2)
198 bch2_alloc_unpack_v2(&ret, k);
199 else if (k.k->type == KEY_TYPE_alloc)
200 bch2_alloc_unpack_v1(&ret, k);
205 void bch2_alloc_pack(struct bch_fs *c,
206 struct bkey_alloc_buf *dst,
207 const struct bkey_alloc_unpacked src)
209 bch2_alloc_pack_v2(dst, src);
212 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
214 unsigned i, bytes = offsetof(struct bch_alloc, data);
216 for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
217 if (a->fields & (1 << i))
218 bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
220 return DIV_ROUND_UP(bytes, sizeof(u64));
223 const char *bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k)
225 struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
227 if (k.k->p.inode >= c->sb.nr_devices ||
228 !c->devs[k.k->p.inode])
229 return "invalid device";
231 /* allow for unknown fields */
232 if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v))
233 return "incorrect value size";
238 const char *bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k)
240 struct bkey_alloc_unpacked u;
242 if (k.k->p.inode >= c->sb.nr_devices ||
243 !c->devs[k.k->p.inode])
244 return "invalid device";
246 if (bch2_alloc_unpack_v2(&u, k))
247 return "unpack error";
252 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c,
255 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
257 pr_buf(out, "gen %u oldest_gen %u data_type %s",
258 u.gen, u.oldest_gen, bch2_data_types[u.data_type]);
259 #define x(_name, ...) pr_buf(out, " " #_name " %llu", (u64) u._name);
260 BCH_ALLOC_FIELDS_V2()
264 static int bch2_alloc_read_fn(struct bch_fs *c, struct bkey_s_c k)
268 struct bkey_alloc_unpacked u;
270 if (k.k->type != KEY_TYPE_alloc &&
271 k.k->type != KEY_TYPE_alloc_v2)
274 ca = bch_dev_bkey_exists(c, k.k->p.inode);
275 g = bucket(ca, k.k->p.offset);
276 u = bch2_alloc_unpack(k);
278 g->_mark.gen = u.gen;
279 g->_mark.data_type = u.data_type;
280 g->_mark.dirty_sectors = u.dirty_sectors;
281 g->_mark.cached_sectors = u.cached_sectors;
282 g->io_time[READ] = u.read_time;
283 g->io_time[WRITE] = u.write_time;
284 g->oldest_gen = u.oldest_gen;
290 int bch2_alloc_read(struct bch_fs *c)
294 down_read(&c->gc_lock);
295 ret = bch2_btree_and_journal_walk(c, BTREE_ID_alloc, bch2_alloc_read_fn);
296 up_read(&c->gc_lock);
298 bch_err(c, "error reading alloc info: %i", ret);
305 static int bch2_alloc_write_key(struct btree_trans *trans,
306 struct btree_iter *iter,
309 struct bch_fs *c = trans->c;
313 struct bucket_mark m;
314 struct bkey_alloc_unpacked old_u, new_u;
315 struct bkey_alloc_buf a;
318 bch2_trans_begin(trans);
320 ret = bch2_btree_key_cache_flush(trans,
321 BTREE_ID_alloc, iter->pos);
325 k = bch2_btree_iter_peek_slot(iter);
330 old_u = bch2_alloc_unpack(k);
332 percpu_down_read(&c->mark_lock);
333 ca = bch_dev_bkey_exists(c, iter->pos.inode);
334 g = bucket(ca, iter->pos.offset);
335 m = READ_ONCE(g->mark);
336 new_u = alloc_mem_to_key(iter, g, m);
337 percpu_up_read(&c->mark_lock);
339 if (!bkey_alloc_unpacked_cmp(old_u, new_u))
342 bch2_alloc_pack(c, &a, new_u);
343 bch2_trans_update(trans, iter, &a.k,
344 BTREE_TRIGGER_NORUN);
345 ret = bch2_trans_commit(trans, NULL, NULL,
346 BTREE_INSERT_NOFAIL|flags);
353 int bch2_alloc_write(struct bch_fs *c, unsigned flags)
355 struct btree_trans trans;
356 struct btree_iter *iter;
361 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
362 iter = bch2_trans_get_iter(&trans, BTREE_ID_alloc, POS_MIN,
363 BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
365 for_each_member_device(ca, c, i) {
366 bch2_btree_iter_set_pos(iter,
367 POS(ca->dev_idx, ca->mi.first_bucket));
369 while (iter->pos.offset < ca->mi.nbuckets) {
370 bch2_trans_cond_resched(&trans);
372 ret = bch2_alloc_write_key(&trans, iter, flags);
374 percpu_ref_put(&ca->ref);
377 bch2_btree_iter_next_slot(iter);
381 bch2_trans_iter_put(&trans, iter);
382 bch2_trans_exit(&trans);
386 /* Bucket IO clocks: */
388 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
389 size_t bucket_nr, int rw)
391 struct bch_fs *c = trans->c;
392 struct bch_dev *ca = bch_dev_bkey_exists(c, dev);
393 struct btree_iter *iter;
395 struct bkey_alloc_buf *a;
396 struct bkey_alloc_unpacked u;
400 iter = bch2_trans_get_iter(trans, BTREE_ID_alloc, POS(dev, bucket_nr),
402 BTREE_ITER_CACHED_NOFILL|
404 ret = bch2_btree_iter_traverse(iter);
408 a = bch2_trans_kmalloc(trans, sizeof(struct bkey_alloc_buf));
409 ret = PTR_ERR_OR_ZERO(a);
413 percpu_down_read(&c->mark_lock);
414 g = bucket(ca, bucket_nr);
415 u = alloc_mem_to_key(iter, g, READ_ONCE(g->mark));
416 percpu_up_read(&c->mark_lock);
418 time = rw == READ ? &u.read_time : &u.write_time;
419 now = atomic64_read(&c->io_clock[rw].now);
425 bch2_alloc_pack(c, a, u);
426 ret = bch2_trans_update(trans, iter, &a->k, 0) ?:
427 bch2_trans_commit(trans, NULL, NULL, 0);
429 bch2_trans_iter_put(trans, iter);
433 /* Background allocator thread: */
436 * Scans for buckets to be invalidated, invalidates them, rewrites prios/gens
437 * (marking them as invalidated on disk), then optionally issues discard
438 * commands to the newly free buckets, then puts them on the various freelists.
441 static bool bch2_can_invalidate_bucket(struct bch_dev *ca, size_t b,
442 struct bucket_mark m)
446 if (!is_available_bucket(m))
449 if (m.owned_by_allocator)
452 if (ca->buckets_nouse &&
453 test_bit(b, ca->buckets_nouse))
456 gc_gen = bucket_gc_gen(bucket(ca, b));
458 ca->inc_gen_needs_gc += gc_gen >= BUCKET_GC_GEN_MAX / 2;
459 ca->inc_gen_really_needs_gc += gc_gen >= BUCKET_GC_GEN_MAX;
461 return gc_gen < BUCKET_GC_GEN_MAX;
465 * Determines what order we're going to reuse buckets, smallest bucket_key()
469 static unsigned bucket_sort_key(struct bucket *g, struct bucket_mark m,
470 u64 now, u64 last_seq_ondisk)
472 unsigned used = bucket_sectors_used(m);
476 * Prefer to keep buckets that have been read more recently, and
477 * buckets that have more data in them:
479 u64 last_read = max_t(s64, 0, now - g->io_time[READ]);
480 u32 last_read_scaled = max_t(u64, U32_MAX, div_u64(last_read, used));
482 return -last_read_scaled;
485 * Prefer to use buckets with smaller gc_gen so that we don't
486 * have to walk the btree and recalculate oldest_gen - but shift
487 * off the low bits so that buckets will still have equal sort
488 * keys when there's only a small difference, so that we can
489 * keep sequential buckets together:
491 return (bucket_needs_journal_commit(m, last_seq_ondisk) << 4)|
492 (bucket_gc_gen(g) >> 4);
496 static inline int bucket_alloc_cmp(alloc_heap *h,
497 struct alloc_heap_entry l,
498 struct alloc_heap_entry r)
500 return cmp_int(l.key, r.key) ?:
501 cmp_int(r.nr, l.nr) ?:
502 cmp_int(l.bucket, r.bucket);
505 static inline int bucket_idx_cmp(const void *_l, const void *_r)
507 const struct alloc_heap_entry *l = _l, *r = _r;
509 return cmp_int(l->bucket, r->bucket);
512 static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
514 struct bucket_array *buckets;
515 struct alloc_heap_entry e = { 0 };
516 u64 now, last_seq_ondisk;
519 down_read(&ca->bucket_lock);
521 buckets = bucket_array(ca);
522 ca->alloc_heap.used = 0;
523 now = atomic64_read(&c->io_clock[READ].now);
524 last_seq_ondisk = c->journal.last_seq_ondisk;
527 * Find buckets with lowest read priority, by building a maxheap sorted
528 * by read priority and repeatedly replacing the maximum element until
529 * all buckets have been visited.
531 for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) {
532 struct bucket *g = &buckets->b[b];
533 struct bucket_mark m = READ_ONCE(g->mark);
534 unsigned key = bucket_sort_key(g, m, now, last_seq_ondisk);
538 if (!bch2_can_invalidate_bucket(ca, b, m))
541 if (e.nr && e.bucket + e.nr == b && e.key == key) {
545 heap_add_or_replace(&ca->alloc_heap, e,
546 -bucket_alloc_cmp, NULL);
548 e = (struct alloc_heap_entry) {
557 heap_add_or_replace(&ca->alloc_heap, e,
558 -bucket_alloc_cmp, NULL);
560 for (i = 0; i < ca->alloc_heap.used; i++)
561 nr += ca->alloc_heap.data[i].nr;
563 while (nr - ca->alloc_heap.data[0].nr >= ALLOC_SCAN_BATCH(ca)) {
564 nr -= ca->alloc_heap.data[0].nr;
565 heap_pop(&ca->alloc_heap, e, -bucket_alloc_cmp, NULL);
568 up_read(&ca->bucket_lock);
571 static void find_reclaimable_buckets_fifo(struct bch_fs *c, struct bch_dev *ca)
573 struct bucket_array *buckets = bucket_array(ca);
574 struct bucket_mark m;
577 if (ca->fifo_last_bucket < ca->mi.first_bucket ||
578 ca->fifo_last_bucket >= ca->mi.nbuckets)
579 ca->fifo_last_bucket = ca->mi.first_bucket;
581 start = ca->fifo_last_bucket;
584 ca->fifo_last_bucket++;
585 if (ca->fifo_last_bucket == ca->mi.nbuckets)
586 ca->fifo_last_bucket = ca->mi.first_bucket;
588 b = ca->fifo_last_bucket;
589 m = READ_ONCE(buckets->b[b].mark);
591 if (bch2_can_invalidate_bucket(ca, b, m)) {
592 struct alloc_heap_entry e = { .bucket = b, .nr = 1, };
594 heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
595 if (heap_full(&ca->alloc_heap))
600 } while (ca->fifo_last_bucket != start);
603 static void find_reclaimable_buckets_random(struct bch_fs *c, struct bch_dev *ca)
605 struct bucket_array *buckets = bucket_array(ca);
606 struct bucket_mark m;
610 checked < ca->mi.nbuckets / 2;
612 size_t b = bch2_rand_range(ca->mi.nbuckets -
613 ca->mi.first_bucket) +
616 m = READ_ONCE(buckets->b[b].mark);
618 if (bch2_can_invalidate_bucket(ca, b, m)) {
619 struct alloc_heap_entry e = { .bucket = b, .nr = 1, };
621 heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
622 if (heap_full(&ca->alloc_heap))
629 sort(ca->alloc_heap.data,
631 sizeof(ca->alloc_heap.data[0]),
632 bucket_idx_cmp, NULL);
634 /* remove duplicates: */
635 for (i = 0; i + 1 < ca->alloc_heap.used; i++)
636 if (ca->alloc_heap.data[i].bucket ==
637 ca->alloc_heap.data[i + 1].bucket)
638 ca->alloc_heap.data[i].nr = 0;
641 static size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca)
645 ca->inc_gen_needs_gc = 0;
646 ca->inc_gen_really_needs_gc = 0;
648 switch (ca->mi.replacement) {
649 case BCH_CACHE_REPLACEMENT_lru:
650 find_reclaimable_buckets_lru(c, ca);
652 case BCH_CACHE_REPLACEMENT_fifo:
653 find_reclaimable_buckets_fifo(c, ca);
655 case BCH_CACHE_REPLACEMENT_random:
656 find_reclaimable_buckets_random(c, ca);
660 heap_resort(&ca->alloc_heap, bucket_alloc_cmp, NULL);
662 for (i = 0; i < ca->alloc_heap.used; i++)
663 nr += ca->alloc_heap.data[i].nr;
669 * returns sequence number of most recent journal entry that updated this
672 static u64 bucket_journal_seq(struct bch_fs *c, struct bucket_mark m)
674 if (m.journal_seq_valid) {
675 u64 journal_seq = atomic64_read(&c->journal.seq);
676 u64 bucket_seq = journal_seq;
678 bucket_seq &= ~((u64) U16_MAX);
679 bucket_seq |= m.journal_seq;
681 if (bucket_seq > journal_seq)
682 bucket_seq -= 1 << 16;
690 static int bucket_invalidate_btree(struct btree_trans *trans,
691 struct bch_dev *ca, u64 b)
693 struct bch_fs *c = trans->c;
694 struct bkey_alloc_buf *a;
695 struct bkey_alloc_unpacked u;
697 struct bucket_mark m;
698 struct btree_iter *iter =
699 bch2_trans_get_iter(trans, BTREE_ID_alloc,
702 BTREE_ITER_CACHED_NOFILL|
706 a = bch2_trans_kmalloc(trans, sizeof(*a));
707 ret = PTR_ERR_OR_ZERO(a);
711 ret = bch2_btree_iter_traverse(iter);
715 percpu_down_read(&c->mark_lock);
717 m = READ_ONCE(g->mark);
718 u = alloc_mem_to_key(iter, g, m);
719 percpu_up_read(&c->mark_lock);
724 u.cached_sectors = 0;
725 u.read_time = atomic64_read(&c->io_clock[READ].now);
726 u.write_time = atomic64_read(&c->io_clock[WRITE].now);
728 bch2_alloc_pack(c, a, u);
729 bch2_trans_update(trans, iter, &a->k, BTREE_TRIGGER_BUCKET_INVALIDATE);
731 bch2_trans_iter_put(trans, iter);
735 static int bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
736 u64 *journal_seq, unsigned flags)
739 struct bucket_mark m;
743 BUG_ON(!ca->alloc_heap.used ||
744 !ca->alloc_heap.data[0].nr);
745 b = ca->alloc_heap.data[0].bucket;
747 /* first, put on free_inc and mark as owned by allocator: */
748 percpu_down_read(&c->mark_lock);
750 m = READ_ONCE(g->mark);
752 BUG_ON(m.dirty_sectors);
754 bch2_mark_alloc_bucket(c, ca, b, true);
756 spin_lock(&c->freelist_lock);
757 verify_not_on_freelist(c, ca, b);
758 BUG_ON(!fifo_push(&ca->free_inc, b));
759 spin_unlock(&c->freelist_lock);
762 * If we're not invalidating cached data, we only increment the bucket
763 * gen in memory here, the incremented gen will be updated in the btree
764 * by bch2_trans_mark_pointer():
766 if (!m.cached_sectors &&
767 !bucket_needs_journal_commit(m, c->journal.last_seq_ondisk)) {
769 bucket_cmpxchg(g, m, m.gen++);
770 percpu_up_read(&c->mark_lock);
774 percpu_up_read(&c->mark_lock);
777 * If the read-only path is trying to shut down, we can't be generating
780 if (test_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags)) {
785 ret = bch2_trans_do(c, NULL, journal_seq,
786 BTREE_INSERT_NOCHECK_RW|
788 BTREE_INSERT_JOURNAL_RESERVED|
790 bucket_invalidate_btree(&trans, ca, b));
793 /* remove from alloc_heap: */
794 struct alloc_heap_entry e, *top = ca->alloc_heap.data;
800 heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
803 * Make sure we flush the last journal entry that updated this
804 * bucket (i.e. deleting the last reference) before writing to
807 *journal_seq = max(*journal_seq, bucket_journal_seq(c, m));
811 /* remove from free_inc: */
812 percpu_down_read(&c->mark_lock);
813 spin_lock(&c->freelist_lock);
815 bch2_mark_alloc_bucket(c, ca, b, false);
817 BUG_ON(!fifo_pop_back(&ca->free_inc, b2));
820 spin_unlock(&c->freelist_lock);
821 percpu_up_read(&c->mark_lock);
824 return ret < 0 ? ret : 0;
828 * Pull buckets off ca->alloc_heap, invalidate them, move them to ca->free_inc:
830 static int bch2_invalidate_buckets(struct bch_fs *c, struct bch_dev *ca)
835 /* Only use nowait if we've already invalidated at least one bucket: */
837 !fifo_full(&ca->free_inc) &&
838 ca->alloc_heap.used) {
839 ret = bch2_invalidate_one_bucket(c, ca, &journal_seq,
840 (!fifo_empty(&ca->free_inc)
841 ? BTREE_INSERT_NOWAIT : 0));
843 * We only want to batch up invalidates when they're going to
844 * require flushing the journal:
850 /* If we used NOWAIT, don't return the error: */
851 if (!fifo_empty(&ca->free_inc))
854 bch_err(ca, "error invalidating buckets: %i", ret);
859 ret = bch2_journal_flush_seq(&c->journal, journal_seq);
861 bch_err(ca, "journal error: %i", ret);
868 static void alloc_thread_set_state(struct bch_dev *ca, unsigned new_state)
870 if (ca->allocator_state != new_state) {
871 ca->allocator_state = new_state;
872 closure_wake_up(&ca->fs->freelist_wait);
876 static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, u64 b)
881 spin_lock(&c->freelist_lock);
882 for (i = 0; i < RESERVE_NR; i++) {
884 * Don't strand buckets on the copygc freelist until
885 * after recovery is finished:
887 if (i == RESERVE_MOVINGGC &&
888 !test_bit(BCH_FS_STARTED, &c->flags))
891 if (fifo_push(&ca->free[i], b)) {
892 fifo_pop(&ca->free_inc, b);
897 spin_unlock(&c->freelist_lock);
899 ca->allocator_state = ret
901 : ALLOCATOR_blocked_full;
902 closure_wake_up(&c->freelist_wait);
906 static void discard_one_bucket(struct bch_fs *c, struct bch_dev *ca, u64 b)
908 if (ca->mi.discard &&
909 blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
910 blkdev_issue_discard(ca->disk_sb.bdev, bucket_to_sector(ca, b),
911 ca->mi.bucket_size, GFP_NOFS, 0);
914 static bool allocator_thread_running(struct bch_dev *ca)
916 unsigned state = ca->mi.state == BCH_MEMBER_STATE_rw &&
917 test_bit(BCH_FS_ALLOCATOR_RUNNING, &ca->fs->flags)
920 alloc_thread_set_state(ca, state);
921 return state == ALLOCATOR_running;
924 static int buckets_available(struct bch_dev *ca, unsigned long gc_count)
926 s64 available = dev_buckets_reclaimable(ca) -
927 (gc_count == ca->fs->gc_count ? ca->inc_gen_really_needs_gc : 0);
928 bool ret = available > 0;
930 alloc_thread_set_state(ca, ret
932 : ALLOCATOR_blocked);
937 * bch_allocator_thread - move buckets from free_inc to reserves
939 * The free_inc FIFO is populated by find_reclaimable_buckets(), and
940 * the reserves are depleted by bucket allocation. When we run out
941 * of free_inc, try to invalidate some buckets and write out
944 static int bch2_allocator_thread(void *arg)
946 struct bch_dev *ca = arg;
947 struct bch_fs *c = ca->fs;
948 unsigned long gc_count = c->gc_count;
955 ret = kthread_wait_freezable(allocator_thread_running(ca));
959 while (!ca->alloc_heap.used) {
962 ret = kthread_wait_freezable(buckets_available(ca, gc_count));
966 gc_count = c->gc_count;
967 nr = find_reclaimable_buckets(c, ca);
969 trace_alloc_scan(ca, nr, ca->inc_gen_needs_gc,
970 ca->inc_gen_really_needs_gc);
972 if ((ca->inc_gen_needs_gc >= ALLOC_SCAN_BATCH(ca) ||
973 ca->inc_gen_really_needs_gc) &&
975 atomic_inc(&c->kick_gc);
976 wake_up_process(c->gc_thread);
980 ret = bch2_invalidate_buckets(c, ca);
984 while (!fifo_empty(&ca->free_inc)) {
985 u64 b = fifo_peek(&ca->free_inc);
987 discard_one_bucket(c, ca, b);
989 ret = kthread_wait_freezable(push_invalidated_bucket(c, ca, b));
995 alloc_thread_set_state(ca, ALLOCATOR_stopped);
999 /* Startup/shutdown (ro/rw): */
1001 void bch2_recalc_capacity(struct bch_fs *c)
1004 u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1005 unsigned bucket_size_max = 0;
1006 unsigned long ra_pages = 0;
1009 lockdep_assert_held(&c->state_lock);
1011 for_each_online_member(ca, c, i) {
1012 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_bdi;
1014 ra_pages += bdi->ra_pages;
1017 bch2_set_ra_pages(c, ra_pages);
1019 for_each_rw_member(ca, c, i) {
1020 u64 dev_reserve = 0;
1023 * We need to reserve buckets (from the number
1024 * of currently available buckets) against
1025 * foreground writes so that mainly copygc can
1026 * make forward progress.
1028 * We need enough to refill the various reserves
1029 * from scratch - copygc will use its entire
1030 * reserve all at once, then run against when
1031 * its reserve is refilled (from the formerly
1032 * available buckets).
1034 * This reserve is just used when considering if
1035 * allocations for foreground writes must wait -
1036 * not -ENOSPC calculations.
1038 for (j = 0; j < RESERVE_NONE; j++)
1039 dev_reserve += ca->free[j].size;
1041 dev_reserve += 1; /* btree write point */
1042 dev_reserve += 1; /* copygc write point */
1043 dev_reserve += 1; /* rebalance write point */
1045 dev_reserve *= ca->mi.bucket_size;
1047 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1048 ca->mi.first_bucket);
1050 reserved_sectors += dev_reserve * 2;
1052 bucket_size_max = max_t(unsigned, bucket_size_max,
1053 ca->mi.bucket_size);
1056 gc_reserve = c->opts.gc_reserve_bytes
1057 ? c->opts.gc_reserve_bytes >> 9
1058 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1060 reserved_sectors = max(gc_reserve, reserved_sectors);
1062 reserved_sectors = min(reserved_sectors, capacity);
1064 c->capacity = capacity - reserved_sectors;
1066 c->bucket_size_max = bucket_size_max;
1068 /* Wake up case someone was waiting for buckets */
1069 closure_wake_up(&c->freelist_wait);
1072 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1074 struct open_bucket *ob;
1077 for (ob = c->open_buckets;
1078 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1080 spin_lock(&ob->lock);
1081 if (ob->valid && !ob->on_partial_list &&
1082 ob->ptr.dev == ca->dev_idx)
1084 spin_unlock(&ob->lock);
1090 /* device goes ro: */
1091 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1095 BUG_ON(ca->alloc_thread);
1097 /* First, remove device from allocation groups: */
1099 for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1100 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1103 * Capacity is calculated based off of devices in allocation groups:
1105 bch2_recalc_capacity(c);
1107 /* Next, close write points that point to this device... */
1108 for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1109 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1111 bch2_writepoint_stop(c, ca, &c->copygc_write_point);
1112 bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1113 bch2_writepoint_stop(c, ca, &c->btree_write_point);
1115 mutex_lock(&c->btree_reserve_cache_lock);
1116 while (c->btree_reserve_cache_nr) {
1117 struct btree_alloc *a =
1118 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1120 bch2_open_buckets_put(c, &a->ob);
1122 mutex_unlock(&c->btree_reserve_cache_lock);
1125 struct open_bucket *ob;
1127 spin_lock(&c->freelist_lock);
1128 if (!ca->open_buckets_partial_nr) {
1129 spin_unlock(&c->freelist_lock);
1132 ob = c->open_buckets +
1133 ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1134 ob->on_partial_list = false;
1135 spin_unlock(&c->freelist_lock);
1137 bch2_open_bucket_put(c, ob);
1140 bch2_ec_stop_dev(c, ca);
1143 * Wake up threads that were blocked on allocation, so they can notice
1144 * the device can no longer be removed and the capacity has changed:
1146 closure_wake_up(&c->freelist_wait);
1149 * journal_res_get() can block waiting for free space in the journal -
1150 * it needs to notice there may not be devices to allocate from anymore:
1152 wake_up(&c->journal.wait);
1154 /* Now wait for any in flight writes: */
1156 closure_wait_event(&c->open_buckets_wait,
1157 !bch2_dev_has_open_write_point(c, ca));
1160 /* device goes rw: */
1161 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1165 for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1166 if (ca->mi.data_allowed & (1 << i))
1167 set_bit(ca->dev_idx, c->rw_devs[i].d);
1170 void bch2_dev_allocator_quiesce(struct bch_fs *c, struct bch_dev *ca)
1172 if (ca->alloc_thread)
1173 closure_wait_event(&c->freelist_wait,
1174 ca->allocator_state != ALLOCATOR_running);
1177 /* stop allocator thread: */
1178 void bch2_dev_allocator_stop(struct bch_dev *ca)
1180 struct task_struct *p;
1182 p = rcu_dereference_protected(ca->alloc_thread, 1);
1183 ca->alloc_thread = NULL;
1186 * We need an rcu barrier between setting ca->alloc_thread = NULL and
1187 * the thread shutting down to avoid bch2_wake_allocator() racing:
1189 * XXX: it would be better to have the rcu barrier be asynchronous
1190 * instead of blocking us here
1200 /* start allocator thread: */
1201 int bch2_dev_allocator_start(struct bch_dev *ca)
1203 struct task_struct *p;
1206 * allocator thread already started?
1208 if (ca->alloc_thread)
1211 p = kthread_create(bch2_allocator_thread, ca,
1212 "bch-alloc/%s", ca->name);
1214 bch_err(ca->fs, "error creating allocator thread: %li",
1220 rcu_assign_pointer(ca->alloc_thread, p);
1225 void bch2_fs_allocator_background_init(struct bch_fs *c)
1227 spin_lock_init(&c->freelist_lock);