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 "backpointers.h"
18 #include "btree_iter.h"
19 #include "btree_update.h"
22 #include "buckets_waiting_for_journal.h"
25 #include "disk_groups.h"
32 #include <linux/math64.h>
33 #include <linux/rculist.h>
34 #include <linux/rcupdate.h>
35 #include <trace/events/bcachefs.h>
37 const char * const bch2_alloc_reserves[] = {
45 * Open buckets represent a bucket that's currently being allocated from. They
48 * - They track buckets that have been partially allocated, allowing for
49 * sub-bucket sized allocations - they're used by the sector allocator below
51 * - They provide a reference to the buckets they own that mark and sweep GC
52 * can find, until the new allocation has a pointer to it inserted into the
55 * When allocating some space with the sector allocator, the allocation comes
56 * with a reference to an open bucket - the caller is required to put that
57 * reference _after_ doing the index update that makes its allocation reachable.
60 static void bch2_open_bucket_hash_add(struct bch_fs *c, struct open_bucket *ob)
62 open_bucket_idx_t idx = ob - c->open_buckets;
63 open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
69 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
71 open_bucket_idx_t idx = ob - c->open_buckets;
72 open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
74 while (*slot != idx) {
76 slot = &c->open_buckets[*slot].hash;
83 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
85 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
88 bch2_ec_bucket_written(c, ob);
92 percpu_down_read(&c->mark_lock);
98 spin_unlock(&ob->lock);
99 percpu_up_read(&c->mark_lock);
101 spin_lock(&c->freelist_lock);
102 bch2_open_bucket_hash_remove(c, ob);
104 ob->freelist = c->open_buckets_freelist;
105 c->open_buckets_freelist = ob - c->open_buckets;
107 c->open_buckets_nr_free++;
108 ca->nr_open_buckets--;
109 spin_unlock(&c->freelist_lock);
111 closure_wake_up(&c->open_buckets_wait);
114 void bch2_open_bucket_write_error(struct bch_fs *c,
115 struct open_buckets *obs,
118 struct open_bucket *ob;
121 open_bucket_for_each(c, obs, ob, i)
122 if (ob->dev == dev && ob->ec)
123 bch2_ec_bucket_cancel(c, ob);
126 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
128 struct open_bucket *ob;
130 BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
132 ob = c->open_buckets + c->open_buckets_freelist;
133 c->open_buckets_freelist = ob->freelist;
134 atomic_set(&ob->pin, 1);
137 c->open_buckets_nr_free--;
141 static void open_bucket_free_unused(struct bch_fs *c,
142 struct write_point *wp,
143 struct open_bucket *ob)
145 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
146 bool may_realloc = wp->data_type == BCH_DATA_user;
148 BUG_ON(ca->open_buckets_partial_nr >
149 ARRAY_SIZE(ca->open_buckets_partial));
151 if (ca->open_buckets_partial_nr <
152 ARRAY_SIZE(ca->open_buckets_partial) &&
154 spin_lock(&c->freelist_lock);
155 ob->on_partial_list = true;
156 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
157 ob - c->open_buckets;
158 spin_unlock(&c->freelist_lock);
160 closure_wake_up(&c->open_buckets_wait);
161 closure_wake_up(&c->freelist_wait);
163 bch2_open_bucket_put(c, ob);
167 /* _only_ for allocating the journal on a new device: */
168 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
170 while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
171 u64 b = ca->new_fs_bucket_idx++;
173 if (!is_superblock_bucket(ca, b) &&
174 (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
181 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
185 case RESERVE_btree_movinggc:
187 case RESERVE_movinggc:
188 return OPEN_BUCKETS_COUNT / 4;
190 return OPEN_BUCKETS_COUNT / 2;
194 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
196 enum alloc_reserve reserve,
197 const struct bch_alloc_v4 *a,
198 struct bucket_alloc_state *s,
201 struct open_bucket *ob;
203 if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
208 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
213 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
214 c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
215 s->skipped_need_journal_commit++;
219 if (bch2_bucket_nocow_is_locked(&c->nocow_locks, POS(ca->dev_idx, bucket))) {
224 spin_lock(&c->freelist_lock);
226 if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
228 closure_wait(&c->open_buckets_wait, cl);
230 if (!c->blocked_allocate_open_bucket)
231 c->blocked_allocate_open_bucket = local_clock();
233 spin_unlock(&c->freelist_lock);
234 return ERR_PTR(-BCH_ERR_open_buckets_empty);
237 /* Recheck under lock: */
238 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
239 spin_unlock(&c->freelist_lock);
244 ob = bch2_open_bucket_alloc(c);
246 spin_lock(&ob->lock);
249 ob->sectors_free = ca->mi.bucket_size;
250 ob->alloc_reserve = reserve;
251 ob->dev = ca->dev_idx;
254 spin_unlock(&ob->lock);
256 ca->nr_open_buckets++;
257 bch2_open_bucket_hash_add(c, ob);
259 if (c->blocked_allocate_open_bucket) {
260 bch2_time_stats_update(
261 &c->times[BCH_TIME_blocked_allocate_open_bucket],
262 c->blocked_allocate_open_bucket);
263 c->blocked_allocate_open_bucket = 0;
266 if (c->blocked_allocate) {
267 bch2_time_stats_update(
268 &c->times[BCH_TIME_blocked_allocate],
269 c->blocked_allocate);
270 c->blocked_allocate = 0;
273 spin_unlock(&c->freelist_lock);
278 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
279 enum alloc_reserve reserve, u64 free_entry,
280 struct bucket_alloc_state *s,
281 struct bkey_s_c freespace_k,
284 struct bch_fs *c = trans->c;
285 struct btree_iter iter = { NULL };
287 struct open_bucket *ob;
288 struct bch_alloc_v4 a_convert;
289 const struct bch_alloc_v4 *a;
290 u64 b = free_entry & ~(~0ULL << 56);
291 unsigned genbits = free_entry >> 56;
292 struct printbuf buf = PRINTBUF;
295 if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
296 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
298 ca->mi.first_bucket, ca->mi.nbuckets);
299 bch2_bkey_val_to_text(&buf, c, freespace_k);
300 bch2_trans_inconsistent(trans, "%s", buf.buf);
305 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
306 k = bch2_btree_iter_peek_slot(&iter);
313 a = bch2_alloc_to_v4(k, &a_convert);
315 if (a->data_type != BCH_DATA_free) {
316 if (!test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
321 prt_printf(&buf, "non free bucket in freespace btree\n"
323 bch2_bkey_val_to_text(&buf, c, freespace_k);
324 prt_printf(&buf, "\n ");
325 bch2_bkey_val_to_text(&buf, c, k);
326 bch2_trans_inconsistent(trans, "%s", buf.buf);
331 if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
332 test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
333 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
335 genbits, alloc_freespace_genbits(*a) >> 56);
336 bch2_bkey_val_to_text(&buf, c, freespace_k);
337 prt_printf(&buf, "\n ");
338 bch2_bkey_val_to_text(&buf, c, k);
339 bch2_trans_inconsistent(trans, "%s", buf.buf);
345 if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
346 struct bch_backpointer bp;
349 ret = bch2_get_next_backpointer(trans, POS(ca->dev_idx, b), -1,
351 BTREE_ITER_NOPRESERVE);
357 if (bp_offset != U64_MAX) {
359 * Bucket may have data in it - we don't call
360 * bc2h_trans_inconnsistent() because fsck hasn't
368 ob = __try_alloc_bucket(c, ca, b, reserve, a, s, cl);
370 iter.path->preserve = false;
372 set_btree_iter_dontneed(&iter);
373 bch2_trans_iter_exit(trans, &iter);
378 static struct open_bucket *try_alloc_partial_bucket(struct bch_fs *c, struct bch_dev *ca,
379 enum alloc_reserve reserve)
381 struct open_bucket *ob;
384 spin_lock(&c->freelist_lock);
386 for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
387 ob = c->open_buckets + ca->open_buckets_partial[i];
389 if (reserve <= ob->alloc_reserve) {
390 array_remove_item(ca->open_buckets_partial,
391 ca->open_buckets_partial_nr,
393 ob->on_partial_list = false;
394 ob->alloc_reserve = reserve;
395 spin_unlock(&c->freelist_lock);
400 spin_unlock(&c->freelist_lock);
405 * This path is for before the freespace btree is initialized:
407 * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
408 * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
410 static noinline struct open_bucket *
411 bch2_bucket_alloc_early(struct btree_trans *trans,
413 enum alloc_reserve reserve,
414 struct bucket_alloc_state *s,
417 struct btree_iter iter;
419 struct open_bucket *ob = NULL;
422 s->cur_bucket = max_t(u64, s->cur_bucket, ca->mi.first_bucket);
423 s->cur_bucket = max_t(u64, s->cur_bucket, ca->new_fs_bucket_idx);
425 for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, s->cur_bucket),
426 BTREE_ITER_SLOTS, k, ret) {
427 struct bch_alloc_v4 a_convert;
428 const struct bch_alloc_v4 *a;
430 if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
433 if (ca->new_fs_bucket_idx &&
434 is_superblock_bucket(ca, k.k->p.offset))
437 a = bch2_alloc_to_v4(k, &a_convert);
439 if (a->data_type != BCH_DATA_free)
444 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, reserve, a, s, cl);
448 bch2_trans_iter_exit(trans, &iter);
450 s->cur_bucket = iter.pos.offset;
452 return ob ?: ERR_PTR(ret ?: -BCH_ERR_no_buckets_found);
455 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
457 enum alloc_reserve reserve,
458 struct bucket_alloc_state *s,
461 struct btree_iter iter;
463 struct open_bucket *ob = NULL;
466 BUG_ON(ca->new_fs_bucket_idx);
470 * On transaction restart, we'd like to restart from the bucket we were
473 for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
474 POS(ca->dev_idx, s->cur_bucket), 0, k, ret) {
475 if (k.k->p.inode != ca->dev_idx)
478 for (s->cur_bucket = max(s->cur_bucket, bkey_start_offset(k.k));
479 s->cur_bucket < k.k->p.offset;
481 ret = btree_trans_too_many_iters(trans);
487 ob = try_alloc_bucket(trans, ca, reserve,
488 s->cur_bucket, s, k, cl);
496 bch2_trans_iter_exit(trans, &iter);
498 return ob ?: ERR_PTR(ret);
502 * bch_bucket_alloc - allocate a single bucket from a specific device
504 * Returns index of bucket on success, 0 on failure
506 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
508 enum alloc_reserve reserve,
509 bool may_alloc_partial,
511 struct bch_dev_usage *usage)
513 struct bch_fs *c = trans->c;
514 struct open_bucket *ob = NULL;
515 bool freespace = READ_ONCE(ca->mi.freespace_initialized);
516 u64 start = freespace ? 0 : ca->bucket_alloc_trans_early_cursor;
518 struct bucket_alloc_state s = { .cur_bucket = start };
519 bool waiting = false;
521 bch2_dev_usage_read_fast(ca, usage);
522 avail = dev_buckets_free(ca, *usage, reserve);
524 if (usage->d[BCH_DATA_need_discard].buckets > avail)
527 if (usage->d[BCH_DATA_need_gc_gens].buckets > avail)
530 if (should_invalidate_buckets(ca, *usage))
531 bch2_do_invalidates(c);
534 if (cl && !waiting) {
535 closure_wait(&c->freelist_wait, cl);
540 if (!c->blocked_allocate)
541 c->blocked_allocate = local_clock();
543 ob = ERR_PTR(-BCH_ERR_freelist_empty);
548 closure_wake_up(&c->freelist_wait);
550 if (may_alloc_partial) {
551 ob = try_alloc_partial_bucket(c, ca, reserve);
556 ob = likely(freespace)
557 ? bch2_bucket_alloc_freelist(trans, ca, reserve, &s, cl)
558 : bch2_bucket_alloc_early(trans, ca, reserve, &s, cl);
560 if (s.skipped_need_journal_commit * 2 > avail)
561 bch2_journal_flush_async(&c->journal, NULL);
563 if (!ob && !freespace && start) {
564 start = s.cur_bucket = 0;
568 if (!ob && freespace && !test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
574 ca->bucket_alloc_trans_early_cursor = s.cur_bucket;
577 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
580 trace_and_count(c, bucket_alloc, ca, bch2_alloc_reserves[reserve],
581 may_alloc_partial, ob->bucket);
582 else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
583 trace_and_count(c, bucket_alloc_fail,
584 ca, bch2_alloc_reserves[reserve],
585 usage->d[BCH_DATA_free].buckets,
587 bch2_copygc_wait_amount(c),
588 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
591 bch2_err_str(PTR_ERR(ob)));
596 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
597 enum alloc_reserve reserve,
598 bool may_alloc_partial,
601 struct bch_dev_usage usage;
602 struct open_bucket *ob;
604 bch2_trans_do(c, NULL, NULL, 0,
605 PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
606 may_alloc_partial, cl, &usage)));
610 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
611 unsigned l, unsigned r)
613 return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
614 (stripe->next_alloc[l] < stripe->next_alloc[r]));
617 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
619 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
620 struct dev_stripe_state *stripe,
621 struct bch_devs_mask *devs)
623 struct dev_alloc_list ret = { .nr = 0 };
626 for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
627 ret.devs[ret.nr++] = i;
629 bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
633 static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
634 struct dev_stripe_state *stripe,
635 struct bch_dev_usage *usage)
637 u64 *v = stripe->next_alloc + ca->dev_idx;
638 u64 free_space = dev_buckets_available(ca, RESERVE_none);
639 u64 free_space_inv = free_space
640 ? div64_u64(1ULL << 48, free_space)
644 if (*v + free_space_inv >= *v)
645 *v += free_space_inv;
649 for (v = stripe->next_alloc;
650 v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
651 *v = *v < scale ? 0 : *v - scale;
654 void bch2_dev_stripe_increment(struct bch_dev *ca,
655 struct dev_stripe_state *stripe)
657 struct bch_dev_usage usage;
659 bch2_dev_usage_read_fast(ca, &usage);
660 bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
663 #define BUCKET_MAY_ALLOC_PARTIAL (1 << 0)
664 #define BUCKET_ALLOC_USE_DURABILITY (1 << 1)
666 static void add_new_bucket(struct bch_fs *c,
667 struct open_buckets *ptrs,
668 struct bch_devs_mask *devs_may_alloc,
669 unsigned *nr_effective,
672 struct open_bucket *ob)
674 unsigned durability =
675 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
677 __clear_bit(ob->dev, devs_may_alloc->d);
678 *nr_effective += (flags & BUCKET_ALLOC_USE_DURABILITY)
680 *have_cache |= !durability;
682 ob_push(c, ptrs, ob);
685 static int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
686 struct open_buckets *ptrs,
687 struct dev_stripe_state *stripe,
688 struct bch_devs_mask *devs_may_alloc,
689 unsigned nr_replicas,
690 unsigned *nr_effective,
692 enum alloc_reserve reserve,
696 struct bch_fs *c = trans->c;
697 struct dev_alloc_list devs_sorted =
698 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
701 int ret = -BCH_ERR_insufficient_devices;
704 BUG_ON(*nr_effective >= nr_replicas);
706 for (i = 0; i < devs_sorted.nr; i++) {
707 struct bch_dev_usage usage;
708 struct open_bucket *ob;
710 dev = devs_sorted.devs[i];
713 ca = rcu_dereference(c->devs[dev]);
715 percpu_ref_get(&ca->ref);
721 if (!ca->mi.durability && *have_cache) {
722 percpu_ref_put(&ca->ref);
726 ob = bch2_bucket_alloc_trans(trans, ca, reserve,
727 flags & BUCKET_MAY_ALLOC_PARTIAL, cl, &usage);
729 bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
730 percpu_ref_put(&ca->ref);
734 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
739 add_new_bucket(c, ptrs, devs_may_alloc,
740 nr_effective, have_cache, flags, ob);
742 if (*nr_effective >= nr_replicas) {
751 int bch2_bucket_alloc_set(struct bch_fs *c,
752 struct open_buckets *ptrs,
753 struct dev_stripe_state *stripe,
754 struct bch_devs_mask *devs_may_alloc,
755 unsigned nr_replicas,
756 unsigned *nr_effective,
758 enum alloc_reserve reserve,
762 return bch2_trans_do(c, NULL, NULL, 0,
763 bch2_bucket_alloc_set_trans(&trans, ptrs, stripe,
764 devs_may_alloc, nr_replicas,
765 nr_effective, have_cache, reserve,
769 /* Allocate from stripes: */
772 * if we can't allocate a new stripe because there are already too many
773 * partially filled stripes, force allocating from an existing stripe even when
774 * it's to a device we don't want:
777 static int bucket_alloc_from_stripe(struct bch_fs *c,
778 struct open_buckets *ptrs,
779 struct write_point *wp,
780 struct bch_devs_mask *devs_may_alloc,
782 unsigned erasure_code,
783 unsigned nr_replicas,
784 unsigned *nr_effective,
789 struct dev_alloc_list devs_sorted;
790 struct ec_stripe_head *h;
791 struct open_bucket *ob;
801 if (ec_open_bucket(c, ptrs))
804 h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
805 wp == &c->copygc_write_point,
812 devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
814 for (i = 0; i < devs_sorted.nr; i++)
815 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
816 if (!h->s->blocks[ec_idx])
819 ob = c->open_buckets + h->s->blocks[ec_idx];
820 if (ob->dev == devs_sorted.devs[i] &&
821 !test_and_set_bit(ec_idx, h->s->blocks_allocated))
826 ca = bch_dev_bkey_exists(c, ob->dev);
831 add_new_bucket(c, ptrs, devs_may_alloc,
832 nr_effective, have_cache, flags, ob);
833 atomic_inc(&h->s->pin);
835 bch2_ec_stripe_head_put(c, h);
839 /* Sector allocator */
841 static void get_buckets_from_writepoint(struct bch_fs *c,
842 struct open_buckets *ptrs,
843 struct write_point *wp,
844 struct bch_devs_mask *devs_may_alloc,
845 unsigned nr_replicas,
846 unsigned *nr_effective,
851 struct open_buckets ptrs_skip = { .nr = 0 };
852 struct open_bucket *ob;
855 open_bucket_for_each(c, &wp->ptrs, ob, i) {
856 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
858 if (*nr_effective < nr_replicas &&
859 test_bit(ob->dev, devs_may_alloc->d) &&
860 (ca->mi.durability ||
861 (wp->data_type == BCH_DATA_user && !*have_cache)) &&
862 (ob->ec || !need_ec)) {
863 add_new_bucket(c, ptrs, devs_may_alloc,
864 nr_effective, have_cache,
867 ob_push(c, &ptrs_skip, ob);
870 wp->ptrs = ptrs_skip;
873 static int open_bucket_add_buckets(struct btree_trans *trans,
874 struct open_buckets *ptrs,
875 struct write_point *wp,
876 struct bch_devs_list *devs_have,
878 unsigned erasure_code,
879 unsigned nr_replicas,
880 unsigned *nr_effective,
882 enum alloc_reserve reserve,
886 struct bch_fs *c = trans->c;
887 struct bch_devs_mask devs;
888 struct open_bucket *ob;
889 struct closure *cl = NULL;
894 devs = target_rw_devs(c, wp->data_type, target);
897 /* Don't allocate from devices we already have pointers to: */
898 for (i = 0; i < devs_have->nr; i++)
899 __clear_bit(devs_have->devs[i], devs.d);
901 open_bucket_for_each(c, ptrs, ob, i)
902 __clear_bit(ob->dev, devs.d);
905 if (!ec_open_bucket(c, ptrs)) {
906 get_buckets_from_writepoint(c, ptrs, wp, &devs,
907 nr_replicas, nr_effective,
908 have_cache, flags, true);
909 if (*nr_effective >= nr_replicas)
913 if (!ec_open_bucket(c, ptrs)) {
914 ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
915 target, erasure_code,
916 nr_replicas, nr_effective,
917 have_cache, flags, _cl);
918 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
919 bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
920 bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
922 if (*nr_effective >= nr_replicas)
927 get_buckets_from_writepoint(c, ptrs, wp, &devs,
928 nr_replicas, nr_effective,
929 have_cache, flags, false);
930 if (*nr_effective >= nr_replicas)
935 * Try nonblocking first, so that if one device is full we'll try from
938 ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
939 nr_replicas, nr_effective, have_cache,
942 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
943 !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
952 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
953 struct open_buckets *obs)
955 struct open_buckets ptrs = { .nr = 0 };
956 struct open_bucket *ob, *ob2;
959 open_bucket_for_each(c, obs, ob, i) {
960 bool drop = !ca || ob->dev == ca->dev_idx;
962 if (!drop && ob->ec) {
963 mutex_lock(&ob->ec->lock);
964 for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
965 if (!ob->ec->blocks[j])
968 ob2 = c->open_buckets + ob->ec->blocks[j];
969 drop |= ob2->dev == ca->dev_idx;
971 mutex_unlock(&ob->ec->lock);
975 bch2_open_bucket_put(c, ob);
977 ob_push(c, &ptrs, ob);
983 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
984 struct write_point *wp)
986 mutex_lock(&wp->lock);
987 bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
988 mutex_unlock(&wp->lock);
991 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
992 unsigned long write_point)
995 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
997 return &c->write_points_hash[hash];
1000 static struct write_point *__writepoint_find(struct hlist_head *head,
1001 unsigned long write_point)
1003 struct write_point *wp;
1006 hlist_for_each_entry_rcu(wp, head, node)
1007 if (wp->write_point == write_point)
1015 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1017 u64 stranded = c->write_points_nr * c->bucket_size_max;
1018 u64 free = bch2_fs_usage_read_short(c).free;
1020 return stranded * factor > free;
1023 static bool try_increase_writepoints(struct bch_fs *c)
1025 struct write_point *wp;
1027 if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1028 too_many_writepoints(c, 32))
1031 wp = c->write_points + c->write_points_nr++;
1032 hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1036 static bool try_decrease_writepoints(struct bch_fs *c,
1039 struct write_point *wp;
1041 mutex_lock(&c->write_points_hash_lock);
1042 if (c->write_points_nr < old_nr) {
1043 mutex_unlock(&c->write_points_hash_lock);
1047 if (c->write_points_nr == 1 ||
1048 !too_many_writepoints(c, 8)) {
1049 mutex_unlock(&c->write_points_hash_lock);
1053 wp = c->write_points + --c->write_points_nr;
1055 hlist_del_rcu(&wp->node);
1056 mutex_unlock(&c->write_points_hash_lock);
1058 bch2_writepoint_stop(c, NULL, wp);
1062 static void bch2_trans_mutex_lock(struct btree_trans *trans,
1065 if (!mutex_trylock(lock)) {
1066 bch2_trans_unlock(trans);
1071 static struct write_point *writepoint_find(struct btree_trans *trans,
1072 unsigned long write_point)
1074 struct bch_fs *c = trans->c;
1075 struct write_point *wp, *oldest;
1076 struct hlist_head *head;
1078 if (!(write_point & 1UL)) {
1079 wp = (struct write_point *) write_point;
1080 bch2_trans_mutex_lock(trans, &wp->lock);
1084 head = writepoint_hash(c, write_point);
1086 wp = __writepoint_find(head, write_point);
1089 bch2_trans_mutex_lock(trans, &wp->lock);
1090 if (wp->write_point == write_point)
1092 mutex_unlock(&wp->lock);
1095 restart_find_oldest:
1097 for (wp = c->write_points;
1098 wp < c->write_points + c->write_points_nr; wp++)
1099 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1102 bch2_trans_mutex_lock(trans, &oldest->lock);
1103 bch2_trans_mutex_lock(trans, &c->write_points_hash_lock);
1104 if (oldest >= c->write_points + c->write_points_nr ||
1105 try_increase_writepoints(c)) {
1106 mutex_unlock(&c->write_points_hash_lock);
1107 mutex_unlock(&oldest->lock);
1108 goto restart_find_oldest;
1111 wp = __writepoint_find(head, write_point);
1112 if (wp && wp != oldest) {
1113 mutex_unlock(&c->write_points_hash_lock);
1114 mutex_unlock(&oldest->lock);
1119 hlist_del_rcu(&wp->node);
1120 wp->write_point = write_point;
1121 hlist_add_head_rcu(&wp->node, head);
1122 mutex_unlock(&c->write_points_hash_lock);
1124 wp->last_used = local_clock();
1129 * Get us an open_bucket we can allocate from, return with it locked:
1131 int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1133 unsigned erasure_code,
1134 struct write_point_specifier write_point,
1135 struct bch_devs_list *devs_have,
1136 unsigned nr_replicas,
1137 unsigned nr_replicas_required,
1138 enum alloc_reserve reserve,
1141 struct write_point **wp_ret)
1143 struct bch_fs *c = trans->c;
1144 struct write_point *wp;
1145 struct open_bucket *ob;
1146 struct open_buckets ptrs;
1147 unsigned nr_effective, write_points_nr;
1148 unsigned ob_flags = 0;
1153 if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
1154 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
1156 BUG_ON(!nr_replicas || !nr_replicas_required);
1160 write_points_nr = c->write_points_nr;
1163 *wp_ret = wp = writepoint_find(trans, write_point.v);
1165 if (wp->data_type == BCH_DATA_user)
1166 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
1168 /* metadata may not allocate on cache devices: */
1169 if (wp->data_type != BCH_DATA_user)
1172 if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1173 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1174 target, erasure_code,
1175 nr_replicas, &nr_effective,
1176 &have_cache, reserve,
1179 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1180 target, erasure_code,
1181 nr_replicas, &nr_effective,
1182 &have_cache, reserve,
1185 bch2_err_matches(ret, BCH_ERR_transaction_restart))
1188 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1190 nr_replicas, &nr_effective,
1191 &have_cache, reserve,
1195 BUG_ON(!ret && nr_effective < nr_replicas);
1197 if (erasure_code && !ec_open_bucket(c, &ptrs))
1198 pr_debug("failed to get ec bucket: ret %u", ret);
1200 if (ret == -BCH_ERR_insufficient_devices &&
1201 nr_effective >= nr_replicas_required)
1207 /* Free buckets we didn't use: */
1208 open_bucket_for_each(c, &wp->ptrs, ob, i)
1209 open_bucket_free_unused(c, wp, ob);
1213 wp->sectors_free = UINT_MAX;
1215 open_bucket_for_each(c, &wp->ptrs, ob, i)
1216 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1218 BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1222 open_bucket_for_each(c, &wp->ptrs, ob, i)
1223 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1224 ob_push(c, &ptrs, ob);
1226 open_bucket_free_unused(c, wp, ob);
1229 mutex_unlock(&wp->lock);
1231 if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1232 try_decrease_writepoints(c, write_points_nr))
1235 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1236 bch2_err_matches(ret, BCH_ERR_freelist_empty))
1238 ? -BCH_ERR_bucket_alloc_blocked
1239 : -BCH_ERR_ENOSPC_bucket_alloc;
1244 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1246 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1248 return (struct bch_extent_ptr) {
1249 .type = 1 << BCH_EXTENT_ENTRY_ptr,
1252 .offset = bucket_to_sector(ca, ob->bucket) +
1253 ca->mi.bucket_size -
1258 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1259 struct bkey_i *k, unsigned sectors,
1262 bch2_alloc_sectors_append_ptrs_inlined(c, wp, k, sectors, cached);
1266 * Append pointers to the space we just allocated to @k, and mark @sectors space
1267 * as allocated out of @ob
1269 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1271 bch2_alloc_sectors_done_inlined(c, wp);
1274 static inline void writepoint_init(struct write_point *wp,
1275 enum bch_data_type type)
1277 mutex_init(&wp->lock);
1278 wp->data_type = type;
1280 INIT_WORK(&wp->index_update_work, bch2_write_point_do_index_updates);
1281 INIT_LIST_HEAD(&wp->writes);
1282 spin_lock_init(&wp->writes_lock);
1285 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1287 struct open_bucket *ob;
1288 struct write_point *wp;
1290 mutex_init(&c->write_points_hash_lock);
1291 c->write_points_nr = ARRAY_SIZE(c->write_points);
1293 /* open bucket 0 is a sentinal NULL: */
1294 spin_lock_init(&c->open_buckets[0].lock);
1296 for (ob = c->open_buckets + 1;
1297 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1298 spin_lock_init(&ob->lock);
1299 c->open_buckets_nr_free++;
1301 ob->freelist = c->open_buckets_freelist;
1302 c->open_buckets_freelist = ob - c->open_buckets;
1305 writepoint_init(&c->btree_write_point, BCH_DATA_btree);
1306 writepoint_init(&c->rebalance_write_point, BCH_DATA_user);
1307 writepoint_init(&c->copygc_write_point, BCH_DATA_user);
1309 for (wp = c->write_points;
1310 wp < c->write_points + c->write_points_nr; wp++) {
1311 writepoint_init(wp, BCH_DATA_user);
1313 wp->last_used = local_clock();
1314 wp->write_point = (unsigned long) wp;
1315 hlist_add_head_rcu(&wp->node,
1316 writepoint_hash(c, wp->write_point));
1320 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1322 struct open_bucket *ob;
1324 for (ob = c->open_buckets;
1325 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1327 spin_lock(&ob->lock);
1328 if (ob->valid && !ob->on_partial_list) {
1329 prt_printf(out, "%zu ref %u type %s %u:%llu:%u\n",
1330 ob - c->open_buckets,
1331 atomic_read(&ob->pin),
1332 bch2_data_types[ob->data_type],
1333 ob->dev, ob->bucket, ob->gen);
1335 spin_unlock(&ob->lock);