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"
31 #include "nocow_locking.h"
33 #include <linux/math64.h>
34 #include <linux/rculist.h>
35 #include <linux/rcupdate.h>
36 #include <trace/events/bcachefs.h>
38 const char * const bch2_alloc_reserves[] = {
46 * Open buckets represent a bucket that's currently being allocated from. They
49 * - They track buckets that have been partially allocated, allowing for
50 * sub-bucket sized allocations - they're used by the sector allocator below
52 * - They provide a reference to the buckets they own that mark and sweep GC
53 * can find, until the new allocation has a pointer to it inserted into the
56 * When allocating some space with the sector allocator, the allocation comes
57 * with a reference to an open bucket - the caller is required to put that
58 * reference _after_ doing the index update that makes its allocation reachable.
61 void bch2_reset_alloc_cursors(struct bch_fs *c)
67 for_each_member_device_rcu(ca, c, i, NULL)
72 static void bch2_open_bucket_hash_add(struct bch_fs *c, struct open_bucket *ob)
74 open_bucket_idx_t idx = ob - c->open_buckets;
75 open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
81 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
83 open_bucket_idx_t idx = ob - c->open_buckets;
84 open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
86 while (*slot != idx) {
88 slot = &c->open_buckets[*slot].hash;
95 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
97 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
100 bch2_ec_bucket_written(c, ob);
104 percpu_down_read(&c->mark_lock);
105 spin_lock(&ob->lock);
110 spin_unlock(&ob->lock);
111 percpu_up_read(&c->mark_lock);
113 spin_lock(&c->freelist_lock);
114 bch2_open_bucket_hash_remove(c, ob);
116 ob->freelist = c->open_buckets_freelist;
117 c->open_buckets_freelist = ob - c->open_buckets;
119 c->open_buckets_nr_free++;
120 ca->nr_open_buckets--;
121 spin_unlock(&c->freelist_lock);
123 closure_wake_up(&c->open_buckets_wait);
126 void bch2_open_bucket_write_error(struct bch_fs *c,
127 struct open_buckets *obs,
130 struct open_bucket *ob;
133 open_bucket_for_each(c, obs, ob, i)
134 if (ob->dev == dev && ob->ec)
135 bch2_ec_bucket_cancel(c, ob);
138 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
140 struct open_bucket *ob;
142 BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
144 ob = c->open_buckets + c->open_buckets_freelist;
145 c->open_buckets_freelist = ob->freelist;
146 atomic_set(&ob->pin, 1);
149 c->open_buckets_nr_free--;
153 static void open_bucket_free_unused(struct bch_fs *c,
154 struct write_point *wp,
155 struct open_bucket *ob)
157 BUG_ON(c->open_buckets_partial_nr >=
158 ARRAY_SIZE(c->open_buckets_partial));
160 spin_lock(&c->freelist_lock);
161 ob->on_partial_list = true;
162 c->open_buckets_partial[c->open_buckets_partial_nr++] =
163 ob - c->open_buckets;
164 spin_unlock(&c->freelist_lock);
166 closure_wake_up(&c->open_buckets_wait);
167 closure_wake_up(&c->freelist_wait);
170 /* _only_ for allocating the journal on a new device: */
171 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
173 while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
174 u64 b = ca->new_fs_bucket_idx++;
176 if (!is_superblock_bucket(ca, b) &&
177 (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
184 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
188 case RESERVE_btree_movinggc:
190 case RESERVE_movinggc:
191 return OPEN_BUCKETS_COUNT / 4;
193 return OPEN_BUCKETS_COUNT / 2;
197 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
199 enum alloc_reserve reserve,
200 const struct bch_alloc_v4 *a,
201 struct bucket_alloc_state *s,
204 struct open_bucket *ob;
206 if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
211 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
216 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
217 c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
218 s->skipped_need_journal_commit++;
222 if (bch2_bucket_nocow_is_locked(&c->nocow_locks, POS(ca->dev_idx, bucket))) {
227 spin_lock(&c->freelist_lock);
229 if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
231 closure_wait(&c->open_buckets_wait, cl);
233 if (!c->blocked_allocate_open_bucket)
234 c->blocked_allocate_open_bucket = local_clock();
236 spin_unlock(&c->freelist_lock);
237 return ERR_PTR(-BCH_ERR_open_buckets_empty);
240 /* Recheck under lock: */
241 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
242 spin_unlock(&c->freelist_lock);
247 ob = bch2_open_bucket_alloc(c);
249 spin_lock(&ob->lock);
252 ob->sectors_free = ca->mi.bucket_size;
253 ob->dev = ca->dev_idx;
256 spin_unlock(&ob->lock);
258 ca->nr_open_buckets++;
259 bch2_open_bucket_hash_add(c, ob);
261 if (c->blocked_allocate_open_bucket) {
262 bch2_time_stats_update(
263 &c->times[BCH_TIME_blocked_allocate_open_bucket],
264 c->blocked_allocate_open_bucket);
265 c->blocked_allocate_open_bucket = 0;
268 if (c->blocked_allocate) {
269 bch2_time_stats_update(
270 &c->times[BCH_TIME_blocked_allocate],
271 c->blocked_allocate);
272 c->blocked_allocate = 0;
275 spin_unlock(&c->freelist_lock);
279 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
280 enum alloc_reserve reserve, u64 free_entry,
281 struct bucket_alloc_state *s,
282 struct bkey_s_c freespace_k,
285 struct bch_fs *c = trans->c;
286 struct btree_iter iter = { NULL };
288 struct open_bucket *ob;
289 struct bch_alloc_v4 a_convert;
290 const struct bch_alloc_v4 *a;
291 u64 b = free_entry & ~(~0ULL << 56);
292 unsigned genbits = free_entry >> 56;
293 struct printbuf buf = PRINTBUF;
296 if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
297 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
299 ca->mi.first_bucket, ca->mi.nbuckets);
300 bch2_bkey_val_to_text(&buf, c, freespace_k);
301 bch2_trans_inconsistent(trans, "%s", buf.buf);
306 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
307 k = bch2_btree_iter_peek_slot(&iter);
314 a = bch2_alloc_to_v4(k, &a_convert);
316 if (a->data_type != BCH_DATA_free) {
317 if (!test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
322 prt_printf(&buf, "non free bucket in freespace btree\n"
324 bch2_bkey_val_to_text(&buf, c, freespace_k);
325 prt_printf(&buf, "\n ");
326 bch2_bkey_val_to_text(&buf, c, k);
327 bch2_trans_inconsistent(trans, "%s", buf.buf);
332 if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
333 test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
334 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
336 genbits, alloc_freespace_genbits(*a) >> 56);
337 bch2_bkey_val_to_text(&buf, c, freespace_k);
338 prt_printf(&buf, "\n ");
339 bch2_bkey_val_to_text(&buf, c, k);
340 bch2_trans_inconsistent(trans, "%s", buf.buf);
346 if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
347 struct bch_backpointer bp;
350 ret = bch2_get_next_backpointer(trans, POS(ca->dev_idx, b), -1,
352 BTREE_ITER_NOPRESERVE);
358 if (bp_offset != U64_MAX) {
360 * Bucket may have data in it - we don't call
361 * bc2h_trans_inconnsistent() because fsck hasn't
369 ob = __try_alloc_bucket(c, ca, b, reserve, a, s, cl);
371 iter.path->preserve = false;
373 set_btree_iter_dontneed(&iter);
374 bch2_trans_iter_exit(trans, &iter);
380 * This path is for before the freespace btree is initialized:
382 * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
383 * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
385 static noinline struct open_bucket *
386 bch2_bucket_alloc_early(struct btree_trans *trans,
388 enum alloc_reserve reserve,
389 struct bucket_alloc_state *s,
392 struct btree_iter iter;
394 struct open_bucket *ob = NULL;
395 u64 alloc_start = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
396 u64 alloc_cursor = max(alloc_start, READ_ONCE(ca->alloc_cursor));
399 for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, alloc_cursor),
400 BTREE_ITER_SLOTS, k, ret) {
401 struct bch_alloc_v4 a_convert;
402 const struct bch_alloc_v4 *a;
404 if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
407 if (ca->new_fs_bucket_idx &&
408 is_superblock_bucket(ca, k.k->p.offset))
411 a = bch2_alloc_to_v4(k, &a_convert);
413 if (a->data_type != BCH_DATA_free)
418 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, reserve, a, s, cl);
422 bch2_trans_iter_exit(trans, &iter);
424 ca->alloc_cursor = alloc_cursor;
429 if (!ob && alloc_cursor > alloc_start) {
430 alloc_cursor = alloc_start;
437 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
439 enum alloc_reserve reserve,
440 struct bucket_alloc_state *s,
443 struct btree_iter iter;
445 struct open_bucket *ob = NULL;
446 u64 alloc_start = max_t(u64, ca->mi.first_bucket, READ_ONCE(ca->alloc_cursor));
447 u64 alloc_cursor = alloc_start;
450 BUG_ON(ca->new_fs_bucket_idx);
452 for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
453 POS(ca->dev_idx, alloc_cursor), 0, k, ret) {
454 if (k.k->p.inode != ca->dev_idx)
457 for (alloc_cursor = max(alloc_cursor, bkey_start_offset(k.k));
458 alloc_cursor < k.k->p.offset;
460 ret = btree_trans_too_many_iters(trans);
468 ob = try_alloc_bucket(trans, ca, reserve,
469 alloc_cursor, s, k, cl);
471 iter.path->preserve = false;
479 bch2_trans_iter_exit(trans, &iter);
481 ca->alloc_cursor = alloc_cursor;
486 if (!ob && alloc_start > ca->mi.first_bucket) {
487 alloc_cursor = alloc_start = ca->mi.first_bucket;
495 * bch_bucket_alloc - allocate a single bucket from a specific device
497 * Returns index of bucket on success, 0 on failure
499 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
501 enum alloc_reserve reserve,
503 struct bch_dev_usage *usage)
505 struct bch_fs *c = trans->c;
506 struct open_bucket *ob = NULL;
507 bool freespace = READ_ONCE(ca->mi.freespace_initialized);
509 struct bucket_alloc_state s = { 0 };
510 bool waiting = false;
512 bch2_dev_usage_read_fast(ca, usage);
513 avail = dev_buckets_free(ca, *usage, reserve);
515 if (usage->d[BCH_DATA_need_discard].buckets > avail)
518 if (usage->d[BCH_DATA_need_gc_gens].buckets > avail)
521 if (should_invalidate_buckets(ca, *usage))
522 bch2_do_invalidates(c);
525 if (cl && !waiting) {
526 closure_wait(&c->freelist_wait, cl);
531 if (!c->blocked_allocate)
532 c->blocked_allocate = local_clock();
534 ob = ERR_PTR(-BCH_ERR_freelist_empty);
539 closure_wake_up(&c->freelist_wait);
541 ob = likely(freespace)
542 ? bch2_bucket_alloc_freelist(trans, ca, reserve, &s, cl)
543 : bch2_bucket_alloc_early(trans, ca, reserve, &s, cl);
545 if (s.skipped_need_journal_commit * 2 > avail)
546 bch2_journal_flush_async(&c->journal, NULL);
548 if (!ob && freespace && !test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
554 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
557 trace_and_count(c, bucket_alloc, ca,
558 bch2_alloc_reserves[reserve],
560 usage->d[BCH_DATA_free].buckets,
562 bch2_copygc_wait_amount(c),
563 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
567 else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
568 trace_and_count(c, bucket_alloc_fail, ca,
569 bch2_alloc_reserves[reserve],
571 usage->d[BCH_DATA_free].buckets,
573 bch2_copygc_wait_amount(c),
574 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
577 bch2_err_str(PTR_ERR(ob)));
582 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
583 enum alloc_reserve reserve,
586 struct bch_dev_usage usage;
587 struct open_bucket *ob;
589 bch2_trans_do(c, NULL, NULL, 0,
590 PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
595 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
596 unsigned l, unsigned r)
598 return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
599 (stripe->next_alloc[l] < stripe->next_alloc[r]));
602 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
604 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
605 struct dev_stripe_state *stripe,
606 struct bch_devs_mask *devs)
608 struct dev_alloc_list ret = { .nr = 0 };
611 for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
612 ret.devs[ret.nr++] = i;
614 bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
618 static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
619 struct dev_stripe_state *stripe,
620 struct bch_dev_usage *usage)
622 u64 *v = stripe->next_alloc + ca->dev_idx;
623 u64 free_space = dev_buckets_available(ca, RESERVE_none);
624 u64 free_space_inv = free_space
625 ? div64_u64(1ULL << 48, free_space)
629 if (*v + free_space_inv >= *v)
630 *v += free_space_inv;
634 for (v = stripe->next_alloc;
635 v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
636 *v = *v < scale ? 0 : *v - scale;
639 void bch2_dev_stripe_increment(struct bch_dev *ca,
640 struct dev_stripe_state *stripe)
642 struct bch_dev_usage usage;
644 bch2_dev_usage_read_fast(ca, &usage);
645 bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
648 static int add_new_bucket(struct bch_fs *c,
649 struct open_buckets *ptrs,
650 struct bch_devs_mask *devs_may_alloc,
651 unsigned nr_replicas,
652 unsigned *nr_effective,
655 struct open_bucket *ob)
657 unsigned durability =
658 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
660 BUG_ON(*nr_effective >= nr_replicas);
662 __clear_bit(ob->dev, devs_may_alloc->d);
663 *nr_effective += durability;
664 *have_cache |= !durability;
666 ob_push(c, ptrs, ob);
668 if (*nr_effective >= nr_replicas)
675 int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
676 struct open_buckets *ptrs,
677 struct dev_stripe_state *stripe,
678 struct bch_devs_mask *devs_may_alloc,
679 unsigned nr_replicas,
680 unsigned *nr_effective,
682 enum bch_data_type data_type,
683 enum alloc_reserve reserve,
686 struct bch_fs *c = trans->c;
687 struct dev_alloc_list devs_sorted =
688 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
691 int ret = -BCH_ERR_insufficient_devices;
694 BUG_ON(*nr_effective >= nr_replicas);
696 for (i = 0; i < devs_sorted.nr; i++) {
697 struct bch_dev_usage usage;
698 struct open_bucket *ob;
700 dev = devs_sorted.devs[i];
703 ca = rcu_dereference(c->devs[dev]);
705 percpu_ref_get(&ca->ref);
711 if (!ca->mi.durability && *have_cache) {
712 percpu_ref_put(&ca->ref);
716 ob = bch2_bucket_alloc_trans(trans, ca, reserve, cl, &usage);
718 bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
719 percpu_ref_put(&ca->ref);
723 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
728 ob->data_type = data_type;
730 if (add_new_bucket(c, ptrs, devs_may_alloc,
731 nr_replicas, nr_effective,
732 have_cache, 0, ob)) {
741 /* Allocate from stripes: */
744 * if we can't allocate a new stripe because there are already too many
745 * partially filled stripes, force allocating from an existing stripe even when
746 * it's to a device we don't want:
749 static int bucket_alloc_from_stripe(struct btree_trans *trans,
750 struct open_buckets *ptrs,
751 struct write_point *wp,
752 struct bch_devs_mask *devs_may_alloc,
754 unsigned nr_replicas,
755 unsigned *nr_effective,
760 struct bch_fs *c = trans->c;
761 struct dev_alloc_list devs_sorted;
762 struct ec_stripe_head *h;
763 struct open_bucket *ob;
771 if (ec_open_bucket(c, ptrs))
774 h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1,
775 wp == &c->copygc_write_point,
782 devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
784 for (i = 0; i < devs_sorted.nr; i++)
785 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
786 if (!h->s->blocks[ec_idx])
789 ob = c->open_buckets + h->s->blocks[ec_idx];
790 if (ob->dev == devs_sorted.devs[i] &&
791 !test_and_set_bit(ec_idx, h->s->blocks_allocated))
796 ca = bch_dev_bkey_exists(c, ob->dev);
801 ret = add_new_bucket(c, ptrs, devs_may_alloc,
802 nr_replicas, nr_effective,
803 have_cache, flags, ob);
804 atomic_inc(&h->s->pin);
806 bch2_ec_stripe_head_put(c, h);
810 /* Sector allocator */
812 static bool want_bucket(struct bch_fs *c,
813 struct write_point *wp,
814 struct bch_devs_mask *devs_may_alloc,
815 bool *have_cache, bool ec,
816 struct open_bucket *ob)
818 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
820 if (!test_bit(ob->dev, devs_may_alloc->d))
823 if (ob->data_type != wp->data_type)
826 if (!ca->mi.durability &&
827 (wp->data_type != BCH_DATA_user || !*have_cache))
830 if (ec != (ob->ec != NULL))
836 static int bucket_alloc_set_writepoint(struct bch_fs *c,
837 struct open_buckets *ptrs,
838 struct write_point *wp,
839 struct bch_devs_mask *devs_may_alloc,
840 unsigned nr_replicas,
841 unsigned *nr_effective,
843 bool ec, unsigned flags)
845 struct open_buckets ptrs_skip = { .nr = 0 };
846 struct open_bucket *ob;
850 open_bucket_for_each(c, &wp->ptrs, ob, i) {
851 if (!ret && want_bucket(c, wp, devs_may_alloc,
853 ret = add_new_bucket(c, ptrs, devs_may_alloc,
854 nr_replicas, nr_effective,
855 have_cache, flags, ob);
857 ob_push(c, &ptrs_skip, ob);
859 wp->ptrs = ptrs_skip;
864 static int bucket_alloc_set_partial(struct bch_fs *c,
865 struct open_buckets *ptrs,
866 struct write_point *wp,
867 struct bch_devs_mask *devs_may_alloc,
868 unsigned nr_replicas,
869 unsigned *nr_effective,
870 bool *have_cache, bool ec,
871 enum alloc_reserve reserve,
876 if (!c->open_buckets_partial_nr)
879 spin_lock(&c->freelist_lock);
881 for (i = c->open_buckets_partial_nr - 1; i >= 0; --i) {
882 struct open_bucket *ob = c->open_buckets + c->open_buckets_partial[i];
884 if (want_bucket(c, wp, devs_may_alloc, have_cache, ec, ob)) {
885 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
886 struct bch_dev_usage usage;
889 bch2_dev_usage_read_fast(ca, &usage);
890 avail = dev_buckets_free(ca, usage, reserve);
894 array_remove_item(c->open_buckets_partial,
895 c->open_buckets_partial_nr,
897 ob->on_partial_list = false;
899 ret = add_new_bucket(c, ptrs, devs_may_alloc,
900 nr_replicas, nr_effective,
901 have_cache, flags, ob);
907 spin_unlock(&c->freelist_lock);
911 static int __open_bucket_add_buckets(struct btree_trans *trans,
912 struct open_buckets *ptrs,
913 struct write_point *wp,
914 struct bch_devs_list *devs_have,
917 unsigned nr_replicas,
918 unsigned *nr_effective,
920 enum alloc_reserve reserve,
924 struct bch_fs *c = trans->c;
925 struct bch_devs_mask devs;
926 struct open_bucket *ob;
927 struct closure *cl = NULL;
932 devs = target_rw_devs(c, wp->data_type, target);
935 /* Don't allocate from devices we already have pointers to: */
936 for (i = 0; i < devs_have->nr; i++)
937 __clear_bit(devs_have->devs[i], devs.d);
939 open_bucket_for_each(c, ptrs, ob, i)
940 __clear_bit(ob->dev, devs.d);
942 if (erasure_code && ec_open_bucket(c, ptrs))
945 ret = bucket_alloc_set_writepoint(c, ptrs, wp, &devs,
946 nr_replicas, nr_effective,
947 have_cache, erasure_code, flags);
951 ret = bucket_alloc_set_partial(c, ptrs, wp, &devs,
952 nr_replicas, nr_effective,
953 have_cache, erasure_code, reserve, flags);
958 ret = bucket_alloc_from_stripe(trans, ptrs, wp, &devs,
960 nr_replicas, nr_effective,
961 have_cache, flags, _cl);
965 * Try nonblocking first, so that if one device is full we'll try from
968 ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
969 nr_replicas, nr_effective, have_cache,
970 wp->data_type, reserve, cl);
972 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
973 !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
984 static int open_bucket_add_buckets(struct btree_trans *trans,
985 struct open_buckets *ptrs,
986 struct write_point *wp,
987 struct bch_devs_list *devs_have,
989 unsigned erasure_code,
990 unsigned nr_replicas,
991 unsigned *nr_effective,
993 enum alloc_reserve reserve,
1000 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1001 devs_have, target, erasure_code,
1002 nr_replicas, nr_effective, have_cache,
1003 reserve, flags, cl);
1004 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1005 bch2_err_matches(ret, BCH_ERR_operation_blocked) ||
1006 bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
1007 bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1009 if (*nr_effective >= nr_replicas)
1013 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1014 devs_have, target, false,
1015 nr_replicas, nr_effective, have_cache,
1016 reserve, flags, cl);
1017 return ret < 0 ? ret : 0;
1020 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
1021 struct open_buckets *obs)
1023 struct open_buckets ptrs = { .nr = 0 };
1024 struct open_bucket *ob, *ob2;
1027 open_bucket_for_each(c, obs, ob, i) {
1028 bool drop = !ca || ob->dev == ca->dev_idx;
1030 if (!drop && ob->ec) {
1031 mutex_lock(&ob->ec->lock);
1032 for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
1033 if (!ob->ec->blocks[j])
1036 ob2 = c->open_buckets + ob->ec->blocks[j];
1037 drop |= ob2->dev == ca->dev_idx;
1039 mutex_unlock(&ob->ec->lock);
1043 bch2_open_bucket_put(c, ob);
1045 ob_push(c, &ptrs, ob);
1051 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
1052 struct write_point *wp)
1054 mutex_lock(&wp->lock);
1055 bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
1056 mutex_unlock(&wp->lock);
1059 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
1060 unsigned long write_point)
1063 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
1065 return &c->write_points_hash[hash];
1068 static struct write_point *__writepoint_find(struct hlist_head *head,
1069 unsigned long write_point)
1071 struct write_point *wp;
1074 hlist_for_each_entry_rcu(wp, head, node)
1075 if (wp->write_point == write_point)
1083 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1085 u64 stranded = c->write_points_nr * c->bucket_size_max;
1086 u64 free = bch2_fs_usage_read_short(c).free;
1088 return stranded * factor > free;
1091 static bool try_increase_writepoints(struct bch_fs *c)
1093 struct write_point *wp;
1095 if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1096 too_many_writepoints(c, 32))
1099 wp = c->write_points + c->write_points_nr++;
1100 hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1104 static bool try_decrease_writepoints(struct bch_fs *c,
1107 struct write_point *wp;
1109 mutex_lock(&c->write_points_hash_lock);
1110 if (c->write_points_nr < old_nr) {
1111 mutex_unlock(&c->write_points_hash_lock);
1115 if (c->write_points_nr == 1 ||
1116 !too_many_writepoints(c, 8)) {
1117 mutex_unlock(&c->write_points_hash_lock);
1121 wp = c->write_points + --c->write_points_nr;
1123 hlist_del_rcu(&wp->node);
1124 mutex_unlock(&c->write_points_hash_lock);
1126 bch2_writepoint_stop(c, NULL, wp);
1130 static void bch2_trans_mutex_lock_norelock(struct btree_trans *trans,
1133 if (!mutex_trylock(lock)) {
1134 bch2_trans_unlock(trans);
1139 static struct write_point *writepoint_find(struct btree_trans *trans,
1140 unsigned long write_point)
1142 struct bch_fs *c = trans->c;
1143 struct write_point *wp, *oldest;
1144 struct hlist_head *head;
1146 if (!(write_point & 1UL)) {
1147 wp = (struct write_point *) write_point;
1148 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1152 head = writepoint_hash(c, write_point);
1154 wp = __writepoint_find(head, write_point);
1157 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1158 if (wp->write_point == write_point)
1160 mutex_unlock(&wp->lock);
1163 restart_find_oldest:
1165 for (wp = c->write_points;
1166 wp < c->write_points + c->write_points_nr; wp++)
1167 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1170 bch2_trans_mutex_lock_norelock(trans, &oldest->lock);
1171 bch2_trans_mutex_lock_norelock(trans, &c->write_points_hash_lock);
1172 if (oldest >= c->write_points + c->write_points_nr ||
1173 try_increase_writepoints(c)) {
1174 mutex_unlock(&c->write_points_hash_lock);
1175 mutex_unlock(&oldest->lock);
1176 goto restart_find_oldest;
1179 wp = __writepoint_find(head, write_point);
1180 if (wp && wp != oldest) {
1181 mutex_unlock(&c->write_points_hash_lock);
1182 mutex_unlock(&oldest->lock);
1187 hlist_del_rcu(&wp->node);
1188 wp->write_point = write_point;
1189 hlist_add_head_rcu(&wp->node, head);
1190 mutex_unlock(&c->write_points_hash_lock);
1192 wp->last_used = local_clock();
1197 * Get us an open_bucket we can allocate from, return with it locked:
1199 int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1201 unsigned erasure_code,
1202 struct write_point_specifier write_point,
1203 struct bch_devs_list *devs_have,
1204 unsigned nr_replicas,
1205 unsigned nr_replicas_required,
1206 enum alloc_reserve reserve,
1209 struct write_point **wp_ret)
1211 struct bch_fs *c = trans->c;
1212 struct write_point *wp;
1213 struct open_bucket *ob;
1214 struct open_buckets ptrs;
1215 unsigned nr_effective, write_points_nr;
1220 BUG_ON(!nr_replicas || !nr_replicas_required);
1224 write_points_nr = c->write_points_nr;
1227 *wp_ret = wp = writepoint_find(trans, write_point.v);
1229 /* metadata may not allocate on cache devices: */
1230 if (wp->data_type != BCH_DATA_user)
1233 if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1234 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1235 target, erasure_code,
1236 nr_replicas, &nr_effective,
1237 &have_cache, reserve,
1240 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1241 target, erasure_code,
1242 nr_replicas, &nr_effective,
1243 &have_cache, reserve,
1246 bch2_err_matches(ret, BCH_ERR_transaction_restart))
1249 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1251 nr_replicas, &nr_effective,
1252 &have_cache, reserve,
1256 BUG_ON(!ret && nr_effective < nr_replicas);
1258 if (erasure_code && !ec_open_bucket(c, &ptrs))
1259 pr_debug("failed to get ec bucket: ret %u", ret);
1261 if (ret == -BCH_ERR_insufficient_devices &&
1262 nr_effective >= nr_replicas_required)
1268 /* Free buckets we didn't use: */
1269 open_bucket_for_each(c, &wp->ptrs, ob, i)
1270 open_bucket_free_unused(c, wp, ob);
1274 wp->sectors_free = UINT_MAX;
1276 open_bucket_for_each(c, &wp->ptrs, ob, i)
1277 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1279 BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1283 open_bucket_for_each(c, &wp->ptrs, ob, i)
1284 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1285 ob_push(c, &ptrs, ob);
1287 open_bucket_free_unused(c, wp, ob);
1290 mutex_unlock(&wp->lock);
1292 if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1293 try_decrease_writepoints(c, write_points_nr))
1296 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1297 bch2_err_matches(ret, BCH_ERR_freelist_empty))
1299 ? -BCH_ERR_bucket_alloc_blocked
1300 : -BCH_ERR_ENOSPC_bucket_alloc;
1305 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1307 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1309 return (struct bch_extent_ptr) {
1310 .type = 1 << BCH_EXTENT_ENTRY_ptr,
1313 .offset = bucket_to_sector(ca, ob->bucket) +
1314 ca->mi.bucket_size -
1319 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1320 struct bkey_i *k, unsigned sectors,
1323 bch2_alloc_sectors_append_ptrs_inlined(c, wp, k, sectors, cached);
1327 * Append pointers to the space we just allocated to @k, and mark @sectors space
1328 * as allocated out of @ob
1330 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1332 bch2_alloc_sectors_done_inlined(c, wp);
1335 static inline void writepoint_init(struct write_point *wp,
1336 enum bch_data_type type)
1338 mutex_init(&wp->lock);
1339 wp->data_type = type;
1341 INIT_WORK(&wp->index_update_work, bch2_write_point_do_index_updates);
1342 INIT_LIST_HEAD(&wp->writes);
1343 spin_lock_init(&wp->writes_lock);
1346 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1348 struct open_bucket *ob;
1349 struct write_point *wp;
1351 mutex_init(&c->write_points_hash_lock);
1352 c->write_points_nr = ARRAY_SIZE(c->write_points);
1354 /* open bucket 0 is a sentinal NULL: */
1355 spin_lock_init(&c->open_buckets[0].lock);
1357 for (ob = c->open_buckets + 1;
1358 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1359 spin_lock_init(&ob->lock);
1360 c->open_buckets_nr_free++;
1362 ob->freelist = c->open_buckets_freelist;
1363 c->open_buckets_freelist = ob - c->open_buckets;
1366 writepoint_init(&c->btree_write_point, BCH_DATA_btree);
1367 writepoint_init(&c->rebalance_write_point, BCH_DATA_user);
1368 writepoint_init(&c->copygc_write_point, BCH_DATA_user);
1370 for (wp = c->write_points;
1371 wp < c->write_points + c->write_points_nr; wp++) {
1372 writepoint_init(wp, BCH_DATA_user);
1374 wp->last_used = local_clock();
1375 wp->write_point = (unsigned long) wp;
1376 hlist_add_head_rcu(&wp->node,
1377 writepoint_hash(c, wp->write_point));
1381 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1383 struct open_bucket *ob;
1385 for (ob = c->open_buckets;
1386 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1388 spin_lock(&ob->lock);
1389 if (ob->valid && !ob->on_partial_list) {
1390 prt_printf(out, "%zu ref %u type %s %u:%llu:%u\n",
1391 ob - c->open_buckets,
1392 atomic_read(&ob->pin),
1393 bch2_data_types[ob->data_type],
1394 ob->dev, ob->bucket, ob->gen);
1396 spin_unlock(&ob->lock);
1400 void bch2_open_buckets_partial_to_text(struct printbuf *out, struct bch_fs *c)
1404 spin_lock(&c->freelist_lock);
1405 for (i = 0; i < c->open_buckets_partial_nr; i++) {
1406 struct open_bucket *ob = c->open_buckets + c->open_buckets_partial[i];
1408 prt_printf(out, "%zu ref %u type %s ec %u %u:%llu:%u\n",
1409 ob - c->open_buckets,
1410 atomic_read(&ob->pin),
1411 bch2_data_types[ob->data_type],
1413 ob->dev, ob->bucket, ob->gen);
1415 spin_unlock(&c->freelist_lock);
1418 static const char * const bch2_write_point_states[] = {
1420 WRITE_POINT_STATES()
1425 void bch2_write_points_to_text(struct printbuf *out, struct bch_fs *c)
1427 struct write_point *wp;
1430 for (wp = c->write_points;
1431 wp < c->write_points + ARRAY_SIZE(c->write_points);
1433 prt_printf(out, "%lu: ", wp->write_point);
1434 prt_human_readable_u64(out, wp->sectors_allocated);
1436 prt_printf(out, " last wrote: ");
1437 bch2_pr_time_units(out, sched_clock() - wp->last_used);
1439 for (i = 0; i < WRITE_POINT_STATE_NR; i++) {
1440 prt_printf(out, " %s: ", bch2_write_point_states[i]);
1441 bch2_pr_time_units(out, wp->time[i]);