1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright 2012 Google, Inc.
5 * Foreground allocator code: allocate buckets from freelist, and allocate in
6 * sector granularity from writepoints.
8 * bch2_bucket_alloc() allocates a single bucket from a specific device.
10 * bch2_bucket_alloc_set() allocates one or more buckets from different devices
11 * in a given filesystem.
15 #include "alloc_background.h"
16 #include "alloc_foreground.h"
17 #include "btree_iter.h"
18 #include "btree_update.h"
21 #include "buckets_waiting_for_journal.h"
24 #include "disk_groups.h"
31 #include <linux/math64.h>
32 #include <linux/rculist.h>
33 #include <linux/rcupdate.h>
34 #include <trace/events/bcachefs.h>
36 const char * const bch2_alloc_reserves[] = {
44 * Open buckets represent a bucket that's currently being allocated from. They
47 * - They track buckets that have been partially allocated, allowing for
48 * sub-bucket sized allocations - they're used by the sector allocator below
50 * - They provide a reference to the buckets they own that mark and sweep GC
51 * can find, until the new allocation has a pointer to it inserted into the
54 * When allocating some space with the sector allocator, the allocation comes
55 * with a reference to an open bucket - the caller is required to put that
56 * reference _after_ doing the index update that makes its allocation reachable.
59 static void bch2_open_bucket_hash_add(struct bch_fs *c, struct open_bucket *ob)
61 open_bucket_idx_t idx = ob - c->open_buckets;
62 open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
68 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
70 open_bucket_idx_t idx = ob - c->open_buckets;
71 open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
73 while (*slot != idx) {
75 slot = &c->open_buckets[*slot].hash;
82 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
84 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
87 bch2_ec_bucket_written(c, ob);
91 percpu_down_read(&c->mark_lock);
97 spin_unlock(&ob->lock);
98 percpu_up_read(&c->mark_lock);
100 spin_lock(&c->freelist_lock);
101 bch2_open_bucket_hash_remove(c, ob);
103 ob->freelist = c->open_buckets_freelist;
104 c->open_buckets_freelist = ob - c->open_buckets;
106 c->open_buckets_nr_free++;
107 ca->nr_open_buckets--;
108 spin_unlock(&c->freelist_lock);
110 closure_wake_up(&c->open_buckets_wait);
113 void bch2_open_bucket_write_error(struct bch_fs *c,
114 struct open_buckets *obs,
117 struct open_bucket *ob;
120 open_bucket_for_each(c, obs, ob, i)
121 if (ob->dev == dev && ob->ec)
122 bch2_ec_bucket_cancel(c, ob);
125 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
127 struct open_bucket *ob;
129 BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
131 ob = c->open_buckets + c->open_buckets_freelist;
132 c->open_buckets_freelist = ob->freelist;
133 atomic_set(&ob->pin, 1);
136 c->open_buckets_nr_free--;
140 static void open_bucket_free_unused(struct bch_fs *c,
141 struct write_point *wp,
142 struct open_bucket *ob)
144 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
145 bool may_realloc = wp->data_type == BCH_DATA_user;
147 BUG_ON(ca->open_buckets_partial_nr >
148 ARRAY_SIZE(ca->open_buckets_partial));
150 if (ca->open_buckets_partial_nr <
151 ARRAY_SIZE(ca->open_buckets_partial) &&
153 spin_lock(&c->freelist_lock);
154 ob->on_partial_list = true;
155 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
156 ob - c->open_buckets;
157 spin_unlock(&c->freelist_lock);
159 closure_wake_up(&c->open_buckets_wait);
160 closure_wake_up(&c->freelist_wait);
162 bch2_open_bucket_put(c, ob);
166 /* _only_ for allocating the journal on a new device: */
167 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
169 while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
170 u64 b = ca->new_fs_bucket_idx++;
172 if (!is_superblock_bucket(ca, b) &&
173 (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
180 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
184 case RESERVE_btree_movinggc:
186 case RESERVE_movinggc:
187 return OPEN_BUCKETS_COUNT / 4;
189 return OPEN_BUCKETS_COUNT / 2;
193 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
195 enum alloc_reserve reserve,
196 struct bch_alloc_v4 *a,
198 u64 *skipped_need_journal_commit,
202 struct open_bucket *ob;
204 if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
209 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
214 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
215 c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
216 (*skipped_need_journal_commit)++;
220 spin_lock(&c->freelist_lock);
222 if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
224 closure_wait(&c->open_buckets_wait, cl);
226 if (!c->blocked_allocate_open_bucket)
227 c->blocked_allocate_open_bucket = local_clock();
229 spin_unlock(&c->freelist_lock);
230 return ERR_PTR(-BCH_ERR_open_buckets_empty);
233 /* Recheck under lock: */
234 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
235 spin_unlock(&c->freelist_lock);
240 ob = bch2_open_bucket_alloc(c);
242 spin_lock(&ob->lock);
245 ob->sectors_free = ca->mi.bucket_size;
246 ob->alloc_reserve = reserve;
247 ob->dev = ca->dev_idx;
250 spin_unlock(&ob->lock);
252 ca->nr_open_buckets++;
253 bch2_open_bucket_hash_add(c, ob);
255 if (c->blocked_allocate_open_bucket) {
256 bch2_time_stats_update(
257 &c->times[BCH_TIME_blocked_allocate_open_bucket],
258 c->blocked_allocate_open_bucket);
259 c->blocked_allocate_open_bucket = 0;
262 if (c->blocked_allocate) {
263 bch2_time_stats_update(
264 &c->times[BCH_TIME_blocked_allocate],
265 c->blocked_allocate);
266 c->blocked_allocate = 0;
269 spin_unlock(&c->freelist_lock);
271 trace_and_count(c, bucket_alloc, ca, bch2_alloc_reserves[reserve]);
275 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
276 enum alloc_reserve reserve, u64 free_entry,
278 u64 *skipped_need_journal_commit,
280 struct bkey_s_c freespace_k,
283 struct bch_fs *c = trans->c;
284 struct btree_iter iter = { NULL };
286 struct open_bucket *ob;
287 struct bch_alloc_v4 a;
288 u64 b = free_entry & ~(~0ULL << 56);
289 unsigned genbits = free_entry >> 56;
290 struct printbuf buf = PRINTBUF;
293 if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
294 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
296 ca->mi.first_bucket, ca->mi.nbuckets);
297 bch2_bkey_val_to_text(&buf, c, freespace_k);
298 bch2_trans_inconsistent(trans, "%s", buf.buf);
303 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
304 k = bch2_btree_iter_peek_slot(&iter);
311 bch2_alloc_to_v4(k, &a);
313 if (genbits != (alloc_freespace_genbits(a) >> 56)) {
314 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
316 genbits, alloc_freespace_genbits(a) >> 56);
317 bch2_bkey_val_to_text(&buf, c, freespace_k);
318 prt_printf(&buf, "\n ");
319 bch2_bkey_val_to_text(&buf, c, k);
320 bch2_trans_inconsistent(trans, "%s", buf.buf);
326 if (a.data_type != BCH_DATA_free) {
327 prt_printf(&buf, "non free bucket in freespace btree\n"
329 bch2_bkey_val_to_text(&buf, c, freespace_k);
330 prt_printf(&buf, "\n ");
331 bch2_bkey_val_to_text(&buf, c, k);
332 bch2_trans_inconsistent(trans, "%s", buf.buf);
337 ob = __try_alloc_bucket(c, ca, b, reserve, &a,
339 skipped_need_journal_commit,
343 iter.path->preserve = false;
345 set_btree_iter_dontneed(&iter);
346 bch2_trans_iter_exit(trans, &iter);
351 static struct open_bucket *try_alloc_partial_bucket(struct bch_fs *c, struct bch_dev *ca,
352 enum alloc_reserve reserve)
354 struct open_bucket *ob;
357 spin_lock(&c->freelist_lock);
359 for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
360 ob = c->open_buckets + ca->open_buckets_partial[i];
362 if (reserve <= ob->alloc_reserve) {
363 array_remove_item(ca->open_buckets_partial,
364 ca->open_buckets_partial_nr,
366 ob->on_partial_list = false;
367 ob->alloc_reserve = reserve;
368 spin_unlock(&c->freelist_lock);
373 spin_unlock(&c->freelist_lock);
378 * This path is for before the freespace btree is initialized:
380 * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
381 * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
383 static noinline struct open_bucket *
384 bch2_bucket_alloc_early(struct btree_trans *trans,
386 enum alloc_reserve reserve,
390 u64 *skipped_need_journal_commit,
394 struct btree_iter iter;
396 struct open_bucket *ob = NULL;
399 *cur_bucket = max_t(u64, *cur_bucket, ca->mi.first_bucket);
400 *cur_bucket = max_t(u64, *cur_bucket, ca->new_fs_bucket_idx);
402 for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, *cur_bucket),
403 BTREE_ITER_SLOTS, k, ret) {
404 struct bch_alloc_v4 a;
406 if (bkey_cmp(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
409 if (ca->new_fs_bucket_idx &&
410 is_superblock_bucket(ca, k.k->p.offset))
413 bch2_alloc_to_v4(k, &a);
415 if (a.data_type != BCH_DATA_free)
420 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, reserve, &a,
422 skipped_need_journal_commit,
428 bch2_trans_iter_exit(trans, &iter);
430 *cur_bucket = iter.pos.offset;
432 return ob ?: ERR_PTR(ret ?: -BCH_ERR_no_buckets_found);
435 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
437 enum alloc_reserve reserve,
441 u64 *skipped_need_journal_commit,
445 struct btree_iter iter;
447 struct open_bucket *ob = NULL;
450 BUG_ON(ca->new_fs_bucket_idx);
454 * On transaction restart, we'd like to restart from the bucket we were
457 for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
458 POS(ca->dev_idx, *cur_bucket), 0, k, ret) {
459 if (k.k->p.inode != ca->dev_idx)
462 for (*cur_bucket = max(*cur_bucket, bkey_start_offset(k.k));
463 *cur_bucket < k.k->p.offset;
465 ret = btree_trans_too_many_iters(trans);
471 ob = try_alloc_bucket(trans, ca, reserve,
474 skipped_need_journal_commit,
484 bch2_trans_iter_exit(trans, &iter);
486 return ob ?: ERR_PTR(ret);
490 * bch_bucket_alloc - allocate a single bucket from a specific device
492 * Returns index of bucket on success, 0 on failure
494 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
496 enum alloc_reserve reserve,
497 bool may_alloc_partial,
500 struct bch_fs *c = trans->c;
501 struct open_bucket *ob = NULL;
502 struct bch_dev_usage usage;
503 bool freespace_initialized = READ_ONCE(ca->mi.freespace_initialized);
504 u64 start = freespace_initialized ? 0 : ca->bucket_alloc_trans_early_cursor;
506 u64 cur_bucket = start;
507 u64 buckets_seen = 0;
508 u64 skipped_open = 0;
509 u64 skipped_need_journal_commit = 0;
510 u64 skipped_nouse = 0;
511 bool waiting = false;
513 usage = bch2_dev_usage_read(ca);
514 avail = dev_buckets_free(ca, usage, reserve);
516 if (usage.d[BCH_DATA_need_discard].buckets > avail)
519 if (usage.d[BCH_DATA_need_gc_gens].buckets > avail)
522 if (should_invalidate_buckets(ca, usage))
523 bch2_do_invalidates(c);
526 if (cl && !waiting) {
527 closure_wait(&c->freelist_wait, cl);
532 if (!c->blocked_allocate)
533 c->blocked_allocate = local_clock();
535 ob = ERR_PTR(-BCH_ERR_freelist_empty);
540 closure_wake_up(&c->freelist_wait);
542 if (may_alloc_partial) {
543 ob = try_alloc_partial_bucket(c, ca, reserve);
548 ob = likely(ca->mi.freespace_initialized)
549 ? bch2_bucket_alloc_freelist(trans, ca, reserve,
553 &skipped_need_journal_commit,
556 : bch2_bucket_alloc_early(trans, ca, reserve,
560 &skipped_need_journal_commit,
564 if (skipped_need_journal_commit * 2 > avail)
565 bch2_journal_flush_async(&c->journal, NULL);
567 if (!ob && !freespace_initialized && start) {
568 start = cur_bucket = 0;
572 if (!freespace_initialized)
573 ca->bucket_alloc_trans_early_cursor = cur_bucket;
576 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
579 trace_and_count(c, bucket_alloc_fail,
580 ca, bch2_alloc_reserves[reserve],
581 usage.d[BCH_DATA_free].buckets,
583 bch2_copygc_wait_amount(c),
584 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
587 skipped_need_journal_commit,
590 bch2_err_str(PTR_ERR(ob)));
595 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
596 enum alloc_reserve reserve,
597 bool may_alloc_partial,
600 struct open_bucket *ob;
602 bch2_trans_do(c, NULL, NULL, 0,
603 PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
604 may_alloc_partial, cl)));
608 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
609 unsigned l, unsigned r)
611 return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
612 (stripe->next_alloc[l] < stripe->next_alloc[r]));
615 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
617 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
618 struct dev_stripe_state *stripe,
619 struct bch_devs_mask *devs)
621 struct dev_alloc_list ret = { .nr = 0 };
624 for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
625 ret.devs[ret.nr++] = i;
627 bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
631 void bch2_dev_stripe_increment(struct bch_dev *ca,
632 struct dev_stripe_state *stripe)
634 u64 *v = stripe->next_alloc + ca->dev_idx;
635 u64 free_space = dev_buckets_available(ca, RESERVE_none);
636 u64 free_space_inv = free_space
637 ? div64_u64(1ULL << 48, free_space)
641 if (*v + free_space_inv >= *v)
642 *v += free_space_inv;
646 for (v = stripe->next_alloc;
647 v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
648 *v = *v < scale ? 0 : *v - scale;
651 #define BUCKET_MAY_ALLOC_PARTIAL (1 << 0)
652 #define BUCKET_ALLOC_USE_DURABILITY (1 << 1)
654 static void add_new_bucket(struct bch_fs *c,
655 struct open_buckets *ptrs,
656 struct bch_devs_mask *devs_may_alloc,
657 unsigned *nr_effective,
660 struct open_bucket *ob)
662 unsigned durability =
663 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
665 __clear_bit(ob->dev, devs_may_alloc->d);
666 *nr_effective += (flags & BUCKET_ALLOC_USE_DURABILITY)
668 *have_cache |= !durability;
670 ob_push(c, ptrs, ob);
673 static int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
674 struct open_buckets *ptrs,
675 struct dev_stripe_state *stripe,
676 struct bch_devs_mask *devs_may_alloc,
677 unsigned nr_replicas,
678 unsigned *nr_effective,
680 enum alloc_reserve reserve,
684 struct bch_fs *c = trans->c;
685 struct dev_alloc_list devs_sorted =
686 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
692 BUG_ON(*nr_effective >= nr_replicas);
694 for (i = 0; i < devs_sorted.nr; i++) {
695 struct open_bucket *ob;
697 dev = devs_sorted.devs[i];
700 ca = rcu_dereference(c->devs[dev]);
702 percpu_ref_get(&ca->ref);
708 if (!ca->mi.durability && *have_cache) {
709 percpu_ref_put(&ca->ref);
713 ob = bch2_bucket_alloc_trans(trans, ca, reserve,
714 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
716 bch2_dev_stripe_increment(ca, stripe);
717 percpu_ref_put(&ca->ref);
719 ret = PTR_ERR_OR_ZERO(ob);
721 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
726 add_new_bucket(c, ptrs, devs_may_alloc,
727 nr_effective, have_cache, flags, ob);
729 if (*nr_effective >= nr_replicas)
733 if (*nr_effective >= nr_replicas)
736 ret = -BCH_ERR_insufficient_devices;
741 int bch2_bucket_alloc_set(struct bch_fs *c,
742 struct open_buckets *ptrs,
743 struct dev_stripe_state *stripe,
744 struct bch_devs_mask *devs_may_alloc,
745 unsigned nr_replicas,
746 unsigned *nr_effective,
748 enum alloc_reserve reserve,
752 return bch2_trans_do(c, NULL, NULL, 0,
753 bch2_bucket_alloc_set_trans(&trans, ptrs, stripe,
754 devs_may_alloc, nr_replicas,
755 nr_effective, have_cache, reserve,
759 /* Allocate from stripes: */
762 * if we can't allocate a new stripe because there are already too many
763 * partially filled stripes, force allocating from an existing stripe even when
764 * it's to a device we don't want:
767 static int bucket_alloc_from_stripe(struct bch_fs *c,
768 struct open_buckets *ptrs,
769 struct write_point *wp,
770 struct bch_devs_mask *devs_may_alloc,
772 unsigned erasure_code,
773 unsigned nr_replicas,
774 unsigned *nr_effective,
779 struct dev_alloc_list devs_sorted;
780 struct ec_stripe_head *h;
781 struct open_bucket *ob;
791 if (ec_open_bucket(c, ptrs))
794 h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
795 wp == &c->copygc_write_point,
802 devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
804 for (i = 0; i < devs_sorted.nr; i++)
805 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
806 if (!h->s->blocks[ec_idx])
809 ob = c->open_buckets + h->s->blocks[ec_idx];
810 if (ob->dev == devs_sorted.devs[i] &&
811 !test_and_set_bit(ec_idx, h->s->blocks_allocated))
816 ca = bch_dev_bkey_exists(c, ob->dev);
821 add_new_bucket(c, ptrs, devs_may_alloc,
822 nr_effective, have_cache, flags, ob);
823 atomic_inc(&h->s->pin);
825 bch2_ec_stripe_head_put(c, h);
829 /* Sector allocator */
831 static void get_buckets_from_writepoint(struct bch_fs *c,
832 struct open_buckets *ptrs,
833 struct write_point *wp,
834 struct bch_devs_mask *devs_may_alloc,
835 unsigned nr_replicas,
836 unsigned *nr_effective,
841 struct open_buckets ptrs_skip = { .nr = 0 };
842 struct open_bucket *ob;
845 open_bucket_for_each(c, &wp->ptrs, ob, i) {
846 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
848 if (*nr_effective < nr_replicas &&
849 test_bit(ob->dev, devs_may_alloc->d) &&
850 (ca->mi.durability ||
851 (wp->data_type == BCH_DATA_user && !*have_cache)) &&
852 (ob->ec || !need_ec)) {
853 add_new_bucket(c, ptrs, devs_may_alloc,
854 nr_effective, have_cache,
857 ob_push(c, &ptrs_skip, ob);
860 wp->ptrs = ptrs_skip;
863 static int open_bucket_add_buckets(struct btree_trans *trans,
864 struct open_buckets *ptrs,
865 struct write_point *wp,
866 struct bch_devs_list *devs_have,
868 unsigned erasure_code,
869 unsigned nr_replicas,
870 unsigned *nr_effective,
872 enum alloc_reserve reserve,
876 struct bch_fs *c = trans->c;
877 struct bch_devs_mask devs;
878 struct open_bucket *ob;
879 struct closure *cl = NULL;
884 devs = target_rw_devs(c, wp->data_type, target);
887 /* Don't allocate from devices we already have pointers to: */
888 for (i = 0; i < devs_have->nr; i++)
889 __clear_bit(devs_have->devs[i], devs.d);
891 open_bucket_for_each(c, ptrs, ob, i)
892 __clear_bit(ob->dev, devs.d);
895 if (!ec_open_bucket(c, ptrs)) {
896 get_buckets_from_writepoint(c, ptrs, wp, &devs,
897 nr_replicas, nr_effective,
898 have_cache, flags, true);
899 if (*nr_effective >= nr_replicas)
903 if (!ec_open_bucket(c, ptrs)) {
904 ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
905 target, erasure_code,
906 nr_replicas, nr_effective,
907 have_cache, flags, _cl);
908 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
909 bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
910 bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
912 if (*nr_effective >= nr_replicas)
917 get_buckets_from_writepoint(c, ptrs, wp, &devs,
918 nr_replicas, nr_effective,
919 have_cache, flags, false);
920 if (*nr_effective >= nr_replicas)
925 * Try nonblocking first, so that if one device is full we'll try from
928 ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
929 nr_replicas, nr_effective, have_cache,
932 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
933 !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
942 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
943 struct open_buckets *obs)
945 struct open_buckets ptrs = { .nr = 0 };
946 struct open_bucket *ob, *ob2;
949 open_bucket_for_each(c, obs, ob, i) {
950 bool drop = !ca || ob->dev == ca->dev_idx;
952 if (!drop && ob->ec) {
953 mutex_lock(&ob->ec->lock);
954 for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
955 if (!ob->ec->blocks[j])
958 ob2 = c->open_buckets + ob->ec->blocks[j];
959 drop |= ob2->dev == ca->dev_idx;
961 mutex_unlock(&ob->ec->lock);
965 bch2_open_bucket_put(c, ob);
967 ob_push(c, &ptrs, ob);
973 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
974 struct write_point *wp)
976 mutex_lock(&wp->lock);
977 bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
978 mutex_unlock(&wp->lock);
981 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
982 unsigned long write_point)
985 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
987 return &c->write_points_hash[hash];
990 static struct write_point *__writepoint_find(struct hlist_head *head,
991 unsigned long write_point)
993 struct write_point *wp;
996 hlist_for_each_entry_rcu(wp, head, node)
997 if (wp->write_point == write_point)
1005 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1007 u64 stranded = c->write_points_nr * c->bucket_size_max;
1008 u64 free = bch2_fs_usage_read_short(c).free;
1010 return stranded * factor > free;
1013 static bool try_increase_writepoints(struct bch_fs *c)
1015 struct write_point *wp;
1017 if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1018 too_many_writepoints(c, 32))
1021 wp = c->write_points + c->write_points_nr++;
1022 hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1026 static bool try_decrease_writepoints(struct bch_fs *c,
1029 struct write_point *wp;
1031 mutex_lock(&c->write_points_hash_lock);
1032 if (c->write_points_nr < old_nr) {
1033 mutex_unlock(&c->write_points_hash_lock);
1037 if (c->write_points_nr == 1 ||
1038 !too_many_writepoints(c, 8)) {
1039 mutex_unlock(&c->write_points_hash_lock);
1043 wp = c->write_points + --c->write_points_nr;
1045 hlist_del_rcu(&wp->node);
1046 mutex_unlock(&c->write_points_hash_lock);
1048 bch2_writepoint_stop(c, NULL, wp);
1052 static void bch2_trans_mutex_lock(struct btree_trans *trans,
1055 if (!mutex_trylock(lock)) {
1056 bch2_trans_unlock(trans);
1061 static struct write_point *writepoint_find(struct btree_trans *trans,
1062 unsigned long write_point)
1064 struct bch_fs *c = trans->c;
1065 struct write_point *wp, *oldest;
1066 struct hlist_head *head;
1068 if (!(write_point & 1UL)) {
1069 wp = (struct write_point *) write_point;
1070 bch2_trans_mutex_lock(trans, &wp->lock);
1074 head = writepoint_hash(c, write_point);
1076 wp = __writepoint_find(head, write_point);
1079 bch2_trans_mutex_lock(trans, &wp->lock);
1080 if (wp->write_point == write_point)
1082 mutex_unlock(&wp->lock);
1085 restart_find_oldest:
1087 for (wp = c->write_points;
1088 wp < c->write_points + c->write_points_nr; wp++)
1089 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1092 bch2_trans_mutex_lock(trans, &oldest->lock);
1093 bch2_trans_mutex_lock(trans, &c->write_points_hash_lock);
1094 if (oldest >= c->write_points + c->write_points_nr ||
1095 try_increase_writepoints(c)) {
1096 mutex_unlock(&c->write_points_hash_lock);
1097 mutex_unlock(&oldest->lock);
1098 goto restart_find_oldest;
1101 wp = __writepoint_find(head, write_point);
1102 if (wp && wp != oldest) {
1103 mutex_unlock(&c->write_points_hash_lock);
1104 mutex_unlock(&oldest->lock);
1109 hlist_del_rcu(&wp->node);
1110 wp->write_point = write_point;
1111 hlist_add_head_rcu(&wp->node, head);
1112 mutex_unlock(&c->write_points_hash_lock);
1114 wp->last_used = sched_clock();
1119 * Get us an open_bucket we can allocate from, return with it locked:
1121 struct write_point *bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1123 unsigned erasure_code,
1124 struct write_point_specifier write_point,
1125 struct bch_devs_list *devs_have,
1126 unsigned nr_replicas,
1127 unsigned nr_replicas_required,
1128 enum alloc_reserve reserve,
1132 struct bch_fs *c = trans->c;
1133 struct write_point *wp;
1134 struct open_bucket *ob;
1135 struct open_buckets ptrs;
1136 unsigned nr_effective, write_points_nr;
1137 unsigned ob_flags = 0;
1142 if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
1143 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
1145 BUG_ON(!nr_replicas || !nr_replicas_required);
1149 write_points_nr = c->write_points_nr;
1152 wp = writepoint_find(trans, write_point.v);
1154 if (wp->data_type == BCH_DATA_user)
1155 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
1157 /* metadata may not allocate on cache devices: */
1158 if (wp->data_type != BCH_DATA_user)
1161 if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1162 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1163 target, erasure_code,
1164 nr_replicas, &nr_effective,
1165 &have_cache, reserve,
1168 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1169 target, erasure_code,
1170 nr_replicas, &nr_effective,
1171 &have_cache, reserve,
1174 bch2_err_matches(ret, BCH_ERR_transaction_restart))
1177 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1179 nr_replicas, &nr_effective,
1180 &have_cache, reserve,
1184 BUG_ON(!ret && nr_effective < nr_replicas);
1186 if (erasure_code && !ec_open_bucket(c, &ptrs))
1187 pr_debug("failed to get ec bucket: ret %u", ret);
1189 if (ret == -BCH_ERR_insufficient_devices &&
1190 nr_effective >= nr_replicas_required)
1196 /* Free buckets we didn't use: */
1197 open_bucket_for_each(c, &wp->ptrs, ob, i)
1198 open_bucket_free_unused(c, wp, ob);
1202 wp->sectors_free = UINT_MAX;
1204 open_bucket_for_each(c, &wp->ptrs, ob, i)
1205 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1207 BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1211 open_bucket_for_each(c, &wp->ptrs, ob, i)
1212 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1213 ob_push(c, &ptrs, ob);
1215 open_bucket_free_unused(c, wp, ob);
1218 mutex_unlock(&wp->lock);
1220 if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1221 try_decrease_writepoints(c, write_points_nr))
1224 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1225 bch2_err_matches(ret, BCH_ERR_freelist_empty))
1226 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
1228 if (bch2_err_matches(ret, BCH_ERR_insufficient_devices))
1229 return ERR_PTR(-EROFS);
1231 return ERR_PTR(ret);
1234 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
1236 unsigned erasure_code,
1237 struct write_point_specifier write_point,
1238 struct bch_devs_list *devs_have,
1239 unsigned nr_replicas,
1240 unsigned nr_replicas_required,
1241 enum alloc_reserve reserve,
1245 struct write_point *wp;
1247 bch2_trans_do(c, NULL, NULL, 0,
1248 PTR_ERR_OR_ZERO(wp = bch2_alloc_sectors_start_trans(&trans, target,
1253 nr_replicas_required,
1260 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1262 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1264 return (struct bch_extent_ptr) {
1265 .type = 1 << BCH_EXTENT_ENTRY_ptr,
1268 .offset = bucket_to_sector(ca, ob->bucket) +
1269 ca->mi.bucket_size -
1275 * Append pointers to the space we just allocated to @k, and mark @sectors space
1276 * as allocated out of @ob
1278 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1279 struct bkey_i *k, unsigned sectors,
1283 struct open_bucket *ob;
1286 BUG_ON(sectors > wp->sectors_free);
1287 wp->sectors_free -= sectors;
1289 open_bucket_for_each(c, &wp->ptrs, ob, i) {
1290 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1291 struct bch_extent_ptr ptr = bch2_ob_ptr(c, ob);
1293 ptr.cached = cached ||
1294 (!ca->mi.durability &&
1295 wp->data_type == BCH_DATA_user);
1297 bch2_bkey_append_ptr(k, ptr);
1299 BUG_ON(sectors > ob->sectors_free);
1300 ob->sectors_free -= sectors;
1305 * Append pointers to the space we just allocated to @k, and mark @sectors space
1306 * as allocated out of @ob
1308 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1310 struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
1311 struct open_bucket *ob;
1314 open_bucket_for_each(c, &wp->ptrs, ob, i)
1315 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
1318 mutex_unlock(&wp->lock);
1320 bch2_open_buckets_put(c, &ptrs);
1323 static inline void writepoint_init(struct write_point *wp,
1324 enum bch_data_type type)
1326 mutex_init(&wp->lock);
1327 wp->data_type = type;
1330 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1332 struct open_bucket *ob;
1333 struct write_point *wp;
1335 mutex_init(&c->write_points_hash_lock);
1336 c->write_points_nr = ARRAY_SIZE(c->write_points);
1338 /* open bucket 0 is a sentinal NULL: */
1339 spin_lock_init(&c->open_buckets[0].lock);
1341 for (ob = c->open_buckets + 1;
1342 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1343 spin_lock_init(&ob->lock);
1344 c->open_buckets_nr_free++;
1346 ob->freelist = c->open_buckets_freelist;
1347 c->open_buckets_freelist = ob - c->open_buckets;
1350 writepoint_init(&c->btree_write_point, BCH_DATA_btree);
1351 writepoint_init(&c->rebalance_write_point, BCH_DATA_user);
1352 writepoint_init(&c->copygc_write_point, BCH_DATA_user);
1354 for (wp = c->write_points;
1355 wp < c->write_points + c->write_points_nr; wp++) {
1356 writepoint_init(wp, BCH_DATA_user);
1358 wp->last_used = sched_clock();
1359 wp->write_point = (unsigned long) wp;
1360 hlist_add_head_rcu(&wp->node,
1361 writepoint_hash(c, wp->write_point));
1365 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1367 struct open_bucket *ob;
1369 for (ob = c->open_buckets;
1370 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1372 spin_lock(&ob->lock);
1373 if (ob->valid && !ob->on_partial_list) {
1374 prt_printf(out, "%zu ref %u type %s %u:%llu:%u\n",
1375 ob - c->open_buckets,
1376 atomic_read(&ob->pin),
1377 bch2_data_types[ob->data_type],
1378 ob->dev, ob->bucket, ob->gen);
1380 spin_unlock(&ob->lock);