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"
34 #include <linux/math64.h>
35 #include <linux/rculist.h>
36 #include <linux/rcupdate.h>
38 static void bch2_trans_mutex_lock_norelock(struct btree_trans *trans,
41 if (!mutex_trylock(lock)) {
42 bch2_trans_unlock(trans);
47 const char * const bch2_watermarks[] = {
55 * Open buckets represent a bucket that's currently being allocated from. They
58 * - They track buckets that have been partially allocated, allowing for
59 * sub-bucket sized allocations - they're used by the sector allocator below
61 * - They provide a reference to the buckets they own that mark and sweep GC
62 * can find, until the new allocation has a pointer to it inserted into the
65 * When allocating some space with the sector allocator, the allocation comes
66 * with a reference to an open bucket - the caller is required to put that
67 * reference _after_ doing the index update that makes its allocation reachable.
70 void bch2_reset_alloc_cursors(struct bch_fs *c)
76 for_each_member_device_rcu(ca, c, i, NULL)
81 static void bch2_open_bucket_hash_add(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);
90 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
92 open_bucket_idx_t idx = ob - c->open_buckets;
93 open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
95 while (*slot != idx) {
97 slot = &c->open_buckets[*slot].hash;
104 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
106 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
109 ec_stripe_new_put(c, ob->ec, STRIPE_REF_io);
113 percpu_down_read(&c->mark_lock);
114 spin_lock(&ob->lock);
119 spin_unlock(&ob->lock);
120 percpu_up_read(&c->mark_lock);
122 spin_lock(&c->freelist_lock);
123 bch2_open_bucket_hash_remove(c, ob);
125 ob->freelist = c->open_buckets_freelist;
126 c->open_buckets_freelist = ob - c->open_buckets;
128 c->open_buckets_nr_free++;
129 ca->nr_open_buckets--;
130 spin_unlock(&c->freelist_lock);
132 closure_wake_up(&c->open_buckets_wait);
135 void bch2_open_bucket_write_error(struct bch_fs *c,
136 struct open_buckets *obs,
139 struct open_bucket *ob;
142 open_bucket_for_each(c, obs, ob, i)
143 if (ob->dev == dev && ob->ec)
144 bch2_ec_bucket_cancel(c, ob);
147 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
149 struct open_bucket *ob;
151 BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
153 ob = c->open_buckets + c->open_buckets_freelist;
154 c->open_buckets_freelist = ob->freelist;
155 atomic_set(&ob->pin, 1);
158 c->open_buckets_nr_free--;
162 static void open_bucket_free_unused(struct bch_fs *c, struct open_bucket *ob)
164 BUG_ON(c->open_buckets_partial_nr >=
165 ARRAY_SIZE(c->open_buckets_partial));
167 spin_lock(&c->freelist_lock);
168 ob->on_partial_list = true;
169 c->open_buckets_partial[c->open_buckets_partial_nr++] =
170 ob - c->open_buckets;
171 spin_unlock(&c->freelist_lock);
173 closure_wake_up(&c->open_buckets_wait);
174 closure_wake_up(&c->freelist_wait);
177 /* _only_ for allocating the journal on a new device: */
178 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
180 while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
181 u64 b = ca->new_fs_bucket_idx++;
183 if (!is_superblock_bucket(ca, b) &&
184 (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
191 static inline unsigned open_buckets_reserved(enum bch_watermark watermark)
194 case BCH_WATERMARK_reclaim:
196 case BCH_WATERMARK_btree:
197 case BCH_WATERMARK_btree_copygc:
198 return OPEN_BUCKETS_COUNT / 4;
199 case BCH_WATERMARK_copygc:
200 return OPEN_BUCKETS_COUNT / 3;
202 return OPEN_BUCKETS_COUNT / 2;
206 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
208 enum bch_watermark watermark,
209 const struct bch_alloc_v4 *a,
210 struct bucket_alloc_state *s,
213 struct open_bucket *ob;
215 if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
220 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
225 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
226 c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
227 s->skipped_need_journal_commit++;
231 if (bch2_bucket_nocow_is_locked(&c->nocow_locks, POS(ca->dev_idx, bucket))) {
236 spin_lock(&c->freelist_lock);
238 if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(watermark))) {
240 closure_wait(&c->open_buckets_wait, cl);
242 if (!c->blocked_allocate_open_bucket)
243 c->blocked_allocate_open_bucket = local_clock();
245 spin_unlock(&c->freelist_lock);
246 return ERR_PTR(-BCH_ERR_open_buckets_empty);
249 /* Recheck under lock: */
250 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
251 spin_unlock(&c->freelist_lock);
256 ob = bch2_open_bucket_alloc(c);
258 spin_lock(&ob->lock);
261 ob->sectors_free = ca->mi.bucket_size;
262 ob->dev = ca->dev_idx;
265 spin_unlock(&ob->lock);
267 ca->nr_open_buckets++;
268 bch2_open_bucket_hash_add(c, ob);
270 if (c->blocked_allocate_open_bucket) {
271 bch2_time_stats_update(
272 &c->times[BCH_TIME_blocked_allocate_open_bucket],
273 c->blocked_allocate_open_bucket);
274 c->blocked_allocate_open_bucket = 0;
277 if (c->blocked_allocate) {
278 bch2_time_stats_update(
279 &c->times[BCH_TIME_blocked_allocate],
280 c->blocked_allocate);
281 c->blocked_allocate = 0;
284 spin_unlock(&c->freelist_lock);
288 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
289 enum bch_watermark watermark, u64 free_entry,
290 struct bucket_alloc_state *s,
291 struct bkey_s_c freespace_k,
294 struct bch_fs *c = trans->c;
295 struct btree_iter iter = { NULL };
297 struct open_bucket *ob;
298 struct bch_alloc_v4 a_convert;
299 const struct bch_alloc_v4 *a;
300 u64 b = free_entry & ~(~0ULL << 56);
301 unsigned genbits = free_entry >> 56;
302 struct printbuf buf = PRINTBUF;
305 if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
306 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
308 ca->mi.first_bucket, ca->mi.nbuckets);
309 bch2_bkey_val_to_text(&buf, c, freespace_k);
310 bch2_trans_inconsistent(trans, "%s", buf.buf);
315 k = bch2_bkey_get_iter(trans, &iter,
316 BTREE_ID_alloc, POS(ca->dev_idx, b),
324 a = bch2_alloc_to_v4(k, &a_convert);
326 if (a->data_type != BCH_DATA_free) {
327 if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
332 prt_printf(&buf, "non free bucket in freespace btree\n"
334 bch2_bkey_val_to_text(&buf, c, freespace_k);
335 prt_printf(&buf, "\n ");
336 bch2_bkey_val_to_text(&buf, c, k);
337 bch2_trans_inconsistent(trans, "%s", buf.buf);
342 if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
343 c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
344 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
346 genbits, alloc_freespace_genbits(*a) >> 56);
347 bch2_bkey_val_to_text(&buf, c, freespace_k);
348 prt_printf(&buf, "\n ");
349 bch2_bkey_val_to_text(&buf, c, k);
350 bch2_trans_inconsistent(trans, "%s", buf.buf);
355 if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_extents_to_backpointers) {
356 struct bch_backpointer bp;
357 struct bpos bp_pos = POS_MIN;
359 ret = bch2_get_next_backpointer(trans, POS(ca->dev_idx, b), -1,
361 BTREE_ITER_NOPRESERVE);
367 if (!bkey_eq(bp_pos, POS_MAX)) {
369 * Bucket may have data in it - we don't call
370 * bc2h_trans_inconnsistent() because fsck hasn't
378 ob = __try_alloc_bucket(c, ca, b, watermark, a, s, cl);
380 iter.path->preserve = false;
382 if (iter.trans && iter.path)
383 set_btree_iter_dontneed(&iter);
384 bch2_trans_iter_exit(trans, &iter);
390 * This path is for before the freespace btree is initialized:
392 * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
393 * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
395 static noinline struct open_bucket *
396 bch2_bucket_alloc_early(struct btree_trans *trans,
398 enum bch_watermark watermark,
399 struct bucket_alloc_state *s,
402 struct btree_iter iter;
404 struct open_bucket *ob = NULL;
405 u64 alloc_start = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
406 u64 alloc_cursor = max(alloc_start, READ_ONCE(ca->alloc_cursor));
409 for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, alloc_cursor),
410 BTREE_ITER_SLOTS, k, ret) {
411 struct bch_alloc_v4 a_convert;
412 const struct bch_alloc_v4 *a;
414 if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
417 if (ca->new_fs_bucket_idx &&
418 is_superblock_bucket(ca, k.k->p.offset))
421 a = bch2_alloc_to_v4(k, &a_convert);
423 if (a->data_type != BCH_DATA_free)
428 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, watermark, a, s, cl);
432 bch2_trans_iter_exit(trans, &iter);
434 ca->alloc_cursor = alloc_cursor;
439 if (!ob && alloc_cursor > alloc_start) {
440 alloc_cursor = alloc_start;
447 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
449 enum bch_watermark watermark,
450 struct bucket_alloc_state *s,
453 struct btree_iter iter;
455 struct open_bucket *ob = NULL;
456 u64 alloc_start = max_t(u64, ca->mi.first_bucket, READ_ONCE(ca->alloc_cursor));
457 u64 alloc_cursor = alloc_start;
460 BUG_ON(ca->new_fs_bucket_idx);
462 for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
463 POS(ca->dev_idx, alloc_cursor), 0, k, ret) {
464 if (k.k->p.inode != ca->dev_idx)
467 for (alloc_cursor = max(alloc_cursor, bkey_start_offset(k.k));
468 alloc_cursor < k.k->p.offset;
470 ret = btree_trans_too_many_iters(trans);
478 ob = try_alloc_bucket(trans, ca, watermark,
479 alloc_cursor, s, k, cl);
481 iter.path->preserve = false;
489 bch2_trans_iter_exit(trans, &iter);
491 ca->alloc_cursor = alloc_cursor;
496 if (!ob && alloc_start > ca->mi.first_bucket) {
497 alloc_cursor = alloc_start = ca->mi.first_bucket;
505 * bch_bucket_alloc - allocate a single bucket from a specific device
507 * Returns index of bucket on success, 0 on failure
509 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
511 enum bch_watermark watermark,
513 struct bch_dev_usage *usage)
515 struct bch_fs *c = trans->c;
516 struct open_bucket *ob = NULL;
517 bool freespace = READ_ONCE(ca->mi.freespace_initialized);
519 struct bucket_alloc_state s = { 0 };
520 bool waiting = false;
522 bch2_dev_usage_read_fast(ca, usage);
523 avail = dev_buckets_free(ca, *usage, watermark);
525 if (usage->d[BCH_DATA_need_discard].buckets > avail)
528 if (usage->d[BCH_DATA_need_gc_gens].buckets > avail)
531 if (should_invalidate_buckets(ca, *usage))
532 bch2_do_invalidates(c);
535 if (cl && !waiting) {
536 closure_wait(&c->freelist_wait, cl);
541 if (!c->blocked_allocate)
542 c->blocked_allocate = local_clock();
544 ob = ERR_PTR(-BCH_ERR_freelist_empty);
549 closure_wake_up(&c->freelist_wait);
551 ob = likely(freespace)
552 ? bch2_bucket_alloc_freelist(trans, ca, watermark, &s, cl)
553 : bch2_bucket_alloc_early(trans, ca, watermark, &s, cl);
555 if (s.skipped_need_journal_commit * 2 > avail)
556 bch2_journal_flush_async(&c->journal, NULL);
558 if (!ob && freespace && c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
564 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
567 trace_and_count(c, bucket_alloc, ca,
568 bch2_watermarks[watermark],
570 usage->d[BCH_DATA_free].buckets,
572 bch2_copygc_wait_amount(c),
573 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
577 else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
578 trace_and_count(c, bucket_alloc_fail, ca,
579 bch2_watermarks[watermark],
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 bch2_err_str(PTR_ERR(ob)));
592 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
593 enum bch_watermark watermark,
596 struct bch_dev_usage usage;
597 struct open_bucket *ob;
599 bch2_trans_do(c, NULL, NULL, 0,
600 PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, watermark,
605 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
606 unsigned l, unsigned r)
608 return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
609 (stripe->next_alloc[l] < stripe->next_alloc[r]));
612 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
614 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
615 struct dev_stripe_state *stripe,
616 struct bch_devs_mask *devs)
618 struct dev_alloc_list ret = { .nr = 0 };
621 for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
622 ret.devs[ret.nr++] = i;
624 bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
628 static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
629 struct dev_stripe_state *stripe,
630 struct bch_dev_usage *usage)
632 u64 *v = stripe->next_alloc + ca->dev_idx;
633 u64 free_space = dev_buckets_available(ca, BCH_WATERMARK_normal);
634 u64 free_space_inv = free_space
635 ? div64_u64(1ULL << 48, free_space)
639 if (*v + free_space_inv >= *v)
640 *v += free_space_inv;
644 for (v = stripe->next_alloc;
645 v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
646 *v = *v < scale ? 0 : *v - scale;
649 void bch2_dev_stripe_increment(struct bch_dev *ca,
650 struct dev_stripe_state *stripe)
652 struct bch_dev_usage usage;
654 bch2_dev_usage_read_fast(ca, &usage);
655 bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
658 static int add_new_bucket(struct bch_fs *c,
659 struct open_buckets *ptrs,
660 struct bch_devs_mask *devs_may_alloc,
661 unsigned nr_replicas,
662 unsigned *nr_effective,
665 struct open_bucket *ob)
667 unsigned durability =
668 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
670 BUG_ON(*nr_effective >= nr_replicas);
671 BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
673 __clear_bit(ob->dev, devs_may_alloc->d);
674 *nr_effective += (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)
676 *have_cache |= !durability;
678 ob_push(c, ptrs, ob);
680 if (*nr_effective >= nr_replicas)
687 int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
688 struct open_buckets *ptrs,
689 struct dev_stripe_state *stripe,
690 struct bch_devs_mask *devs_may_alloc,
691 unsigned nr_replicas,
692 unsigned *nr_effective,
695 enum bch_data_type data_type,
696 enum bch_watermark watermark,
699 struct bch_fs *c = trans->c;
700 struct dev_alloc_list devs_sorted =
701 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
704 int ret = -BCH_ERR_insufficient_devices;
707 BUG_ON(*nr_effective >= nr_replicas);
709 for (i = 0; i < devs_sorted.nr; i++) {
710 struct bch_dev_usage usage;
711 struct open_bucket *ob;
713 dev = devs_sorted.devs[i];
716 ca = rcu_dereference(c->devs[dev]);
718 percpu_ref_get(&ca->ref);
724 if (!ca->mi.durability && *have_cache) {
725 percpu_ref_put(&ca->ref);
729 ob = bch2_bucket_alloc_trans(trans, ca, watermark, cl, &usage);
731 bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
732 percpu_ref_put(&ca->ref);
736 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
741 ob->data_type = data_type;
743 if (add_new_bucket(c, ptrs, devs_may_alloc,
744 nr_replicas, nr_effective,
745 have_cache, flags, ob)) {
754 /* Allocate from stripes: */
757 * if we can't allocate a new stripe because there are already too many
758 * partially filled stripes, force allocating from an existing stripe even when
759 * it's to a device we don't want:
762 static int bucket_alloc_from_stripe(struct btree_trans *trans,
763 struct open_buckets *ptrs,
764 struct write_point *wp,
765 struct bch_devs_mask *devs_may_alloc,
767 unsigned nr_replicas,
768 unsigned *nr_effective,
770 enum bch_watermark watermark,
774 struct bch_fs *c = trans->c;
775 struct dev_alloc_list devs_sorted;
776 struct ec_stripe_head *h;
777 struct open_bucket *ob;
785 if (ec_open_bucket(c, ptrs))
788 h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, watermark, cl);
794 devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
796 for (i = 0; i < devs_sorted.nr; i++)
797 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
798 if (!h->s->blocks[ec_idx])
801 ob = c->open_buckets + h->s->blocks[ec_idx];
802 if (ob->dev == devs_sorted.devs[i] &&
803 !test_and_set_bit(ec_idx, h->s->blocks_allocated))
808 ca = bch_dev_bkey_exists(c, ob->dev);
812 ec_stripe_new_get(h->s, STRIPE_REF_io);
814 ret = add_new_bucket(c, ptrs, devs_may_alloc,
815 nr_replicas, nr_effective,
816 have_cache, flags, ob);
818 bch2_ec_stripe_head_put(c, h);
822 /* Sector allocator */
824 static bool want_bucket(struct bch_fs *c,
825 struct write_point *wp,
826 struct bch_devs_mask *devs_may_alloc,
827 bool *have_cache, bool ec,
828 struct open_bucket *ob)
830 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
832 if (!test_bit(ob->dev, devs_may_alloc->d))
835 if (ob->data_type != wp->data_type)
838 if (!ca->mi.durability &&
839 (wp->data_type == BCH_DATA_btree || ec || *have_cache))
842 if (ec != (ob->ec != NULL))
848 static int bucket_alloc_set_writepoint(struct bch_fs *c,
849 struct open_buckets *ptrs,
850 struct write_point *wp,
851 struct bch_devs_mask *devs_may_alloc,
852 unsigned nr_replicas,
853 unsigned *nr_effective,
855 bool ec, unsigned flags)
857 struct open_buckets ptrs_skip = { .nr = 0 };
858 struct open_bucket *ob;
862 open_bucket_for_each(c, &wp->ptrs, ob, i) {
863 if (!ret && want_bucket(c, wp, devs_may_alloc,
865 ret = add_new_bucket(c, ptrs, devs_may_alloc,
866 nr_replicas, nr_effective,
867 have_cache, flags, ob);
869 ob_push(c, &ptrs_skip, ob);
871 wp->ptrs = ptrs_skip;
876 static int bucket_alloc_set_partial(struct bch_fs *c,
877 struct open_buckets *ptrs,
878 struct write_point *wp,
879 struct bch_devs_mask *devs_may_alloc,
880 unsigned nr_replicas,
881 unsigned *nr_effective,
882 bool *have_cache, bool ec,
883 enum bch_watermark watermark,
888 if (!c->open_buckets_partial_nr)
891 spin_lock(&c->freelist_lock);
893 if (!c->open_buckets_partial_nr)
896 for (i = c->open_buckets_partial_nr - 1; i >= 0; --i) {
897 struct open_bucket *ob = c->open_buckets + c->open_buckets_partial[i];
899 if (want_bucket(c, wp, devs_may_alloc, have_cache, ec, ob)) {
900 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
901 struct bch_dev_usage usage;
904 bch2_dev_usage_read_fast(ca, &usage);
905 avail = dev_buckets_free(ca, usage, watermark);
909 array_remove_item(c->open_buckets_partial,
910 c->open_buckets_partial_nr,
912 ob->on_partial_list = false;
914 ret = add_new_bucket(c, ptrs, devs_may_alloc,
915 nr_replicas, nr_effective,
916 have_cache, flags, ob);
922 spin_unlock(&c->freelist_lock);
926 static int __open_bucket_add_buckets(struct btree_trans *trans,
927 struct open_buckets *ptrs,
928 struct write_point *wp,
929 struct bch_devs_list *devs_have,
932 unsigned nr_replicas,
933 unsigned *nr_effective,
935 enum bch_watermark watermark,
939 struct bch_fs *c = trans->c;
940 struct bch_devs_mask devs;
941 struct open_bucket *ob;
942 struct closure *cl = NULL;
946 devs = target_rw_devs(c, wp->data_type, target);
948 /* Don't allocate from devices we already have pointers to: */
949 for (i = 0; i < devs_have->nr; i++)
950 __clear_bit(devs_have->devs[i], devs.d);
952 open_bucket_for_each(c, ptrs, ob, i)
953 __clear_bit(ob->dev, devs.d);
955 if (erasure_code && ec_open_bucket(c, ptrs))
958 ret = bucket_alloc_set_writepoint(c, ptrs, wp, &devs,
959 nr_replicas, nr_effective,
960 have_cache, erasure_code, flags);
964 ret = bucket_alloc_set_partial(c, ptrs, wp, &devs,
965 nr_replicas, nr_effective,
966 have_cache, erasure_code, watermark, flags);
971 ret = bucket_alloc_from_stripe(trans, ptrs, wp, &devs,
973 nr_replicas, nr_effective,
975 watermark, flags, _cl);
979 * Try nonblocking first, so that if one device is full we'll try from
982 ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
983 nr_replicas, nr_effective, have_cache,
984 flags, wp->data_type, watermark, cl);
986 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
987 !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
998 static int open_bucket_add_buckets(struct btree_trans *trans,
999 struct open_buckets *ptrs,
1000 struct write_point *wp,
1001 struct bch_devs_list *devs_have,
1003 unsigned erasure_code,
1004 unsigned nr_replicas,
1005 unsigned *nr_effective,
1007 enum bch_watermark watermark,
1014 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1015 devs_have, target, erasure_code,
1016 nr_replicas, nr_effective, have_cache,
1017 watermark, flags, cl);
1018 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1019 bch2_err_matches(ret, BCH_ERR_operation_blocked) ||
1020 bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
1021 bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1023 if (*nr_effective >= nr_replicas)
1027 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1028 devs_have, target, false,
1029 nr_replicas, nr_effective, have_cache,
1030 watermark, flags, cl);
1031 return ret < 0 ? ret : 0;
1034 static bool should_drop_bucket(struct open_bucket *ob, struct bch_fs *c,
1035 struct bch_dev *ca, bool ec)
1038 return ob->ec != NULL;
1040 bool drop = ob->dev == ca->dev_idx;
1041 struct open_bucket *ob2;
1044 if (!drop && ob->ec) {
1045 mutex_lock(&ob->ec->lock);
1046 for (i = 0; i < ob->ec->new_stripe.key.v.nr_blocks; i++) {
1047 if (!ob->ec->blocks[i])
1050 ob2 = c->open_buckets + ob->ec->blocks[i];
1051 drop |= ob2->dev == ca->dev_idx;
1053 mutex_unlock(&ob->ec->lock);
1062 static void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
1063 bool ec, struct write_point *wp)
1065 struct open_buckets ptrs = { .nr = 0 };
1066 struct open_bucket *ob;
1069 mutex_lock(&wp->lock);
1070 open_bucket_for_each(c, &wp->ptrs, ob, i)
1071 if (should_drop_bucket(ob, c, ca, ec))
1072 bch2_open_bucket_put(c, ob);
1074 ob_push(c, &ptrs, ob);
1076 mutex_unlock(&wp->lock);
1079 void bch2_open_buckets_stop(struct bch_fs *c, struct bch_dev *ca,
1084 /* Next, close write points that point to this device... */
1085 for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1086 bch2_writepoint_stop(c, ca, ec, &c->write_points[i]);
1088 bch2_writepoint_stop(c, ca, ec, &c->copygc_write_point);
1089 bch2_writepoint_stop(c, ca, ec, &c->rebalance_write_point);
1090 bch2_writepoint_stop(c, ca, ec, &c->btree_write_point);
1092 mutex_lock(&c->btree_reserve_cache_lock);
1093 while (c->btree_reserve_cache_nr) {
1094 struct btree_alloc *a =
1095 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1097 bch2_open_buckets_put(c, &a->ob);
1099 mutex_unlock(&c->btree_reserve_cache_lock);
1101 spin_lock(&c->freelist_lock);
1103 while (i < c->open_buckets_partial_nr) {
1104 struct open_bucket *ob =
1105 c->open_buckets + c->open_buckets_partial[i];
1107 if (should_drop_bucket(ob, c, ca, ec)) {
1108 --c->open_buckets_partial_nr;
1109 swap(c->open_buckets_partial[i],
1110 c->open_buckets_partial[c->open_buckets_partial_nr]);
1111 ob->on_partial_list = false;
1112 spin_unlock(&c->freelist_lock);
1113 bch2_open_bucket_put(c, ob);
1114 spin_lock(&c->freelist_lock);
1119 spin_unlock(&c->freelist_lock);
1121 bch2_ec_stop_dev(c, ca);
1124 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
1125 unsigned long write_point)
1128 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
1130 return &c->write_points_hash[hash];
1133 static struct write_point *__writepoint_find(struct hlist_head *head,
1134 unsigned long write_point)
1136 struct write_point *wp;
1139 hlist_for_each_entry_rcu(wp, head, node)
1140 if (wp->write_point == write_point)
1148 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1150 u64 stranded = c->write_points_nr * c->bucket_size_max;
1151 u64 free = bch2_fs_usage_read_short(c).free;
1153 return stranded * factor > free;
1156 static bool try_increase_writepoints(struct bch_fs *c)
1158 struct write_point *wp;
1160 if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1161 too_many_writepoints(c, 32))
1164 wp = c->write_points + c->write_points_nr++;
1165 hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1169 static bool try_decrease_writepoints(struct btree_trans *trans, unsigned old_nr)
1171 struct bch_fs *c = trans->c;
1172 struct write_point *wp;
1173 struct open_bucket *ob;
1176 mutex_lock(&c->write_points_hash_lock);
1177 if (c->write_points_nr < old_nr) {
1178 mutex_unlock(&c->write_points_hash_lock);
1182 if (c->write_points_nr == 1 ||
1183 !too_many_writepoints(c, 8)) {
1184 mutex_unlock(&c->write_points_hash_lock);
1188 wp = c->write_points + --c->write_points_nr;
1190 hlist_del_rcu(&wp->node);
1191 mutex_unlock(&c->write_points_hash_lock);
1193 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1194 open_bucket_for_each(c, &wp->ptrs, ob, i)
1195 open_bucket_free_unused(c, ob);
1197 mutex_unlock(&wp->lock);
1201 static struct write_point *writepoint_find(struct btree_trans *trans,
1202 unsigned long write_point)
1204 struct bch_fs *c = trans->c;
1205 struct write_point *wp, *oldest;
1206 struct hlist_head *head;
1208 if (!(write_point & 1UL)) {
1209 wp = (struct write_point *) write_point;
1210 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1214 head = writepoint_hash(c, write_point);
1216 wp = __writepoint_find(head, write_point);
1219 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1220 if (wp->write_point == write_point)
1222 mutex_unlock(&wp->lock);
1225 restart_find_oldest:
1227 for (wp = c->write_points;
1228 wp < c->write_points + c->write_points_nr; wp++)
1229 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1232 bch2_trans_mutex_lock_norelock(trans, &oldest->lock);
1233 bch2_trans_mutex_lock_norelock(trans, &c->write_points_hash_lock);
1234 if (oldest >= c->write_points + c->write_points_nr ||
1235 try_increase_writepoints(c)) {
1236 mutex_unlock(&c->write_points_hash_lock);
1237 mutex_unlock(&oldest->lock);
1238 goto restart_find_oldest;
1241 wp = __writepoint_find(head, write_point);
1242 if (wp && wp != oldest) {
1243 mutex_unlock(&c->write_points_hash_lock);
1244 mutex_unlock(&oldest->lock);
1249 hlist_del_rcu(&wp->node);
1250 wp->write_point = write_point;
1251 hlist_add_head_rcu(&wp->node, head);
1252 mutex_unlock(&c->write_points_hash_lock);
1254 wp->last_used = local_clock();
1259 * Get us an open_bucket we can allocate from, return with it locked:
1261 int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1263 unsigned erasure_code,
1264 struct write_point_specifier write_point,
1265 struct bch_devs_list *devs_have,
1266 unsigned nr_replicas,
1267 unsigned nr_replicas_required,
1268 enum bch_watermark watermark,
1271 struct write_point **wp_ret)
1273 struct bch_fs *c = trans->c;
1274 struct write_point *wp;
1275 struct open_bucket *ob;
1276 struct open_buckets ptrs;
1277 unsigned nr_effective, write_points_nr;
1282 BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
1284 BUG_ON(!nr_replicas || !nr_replicas_required);
1288 write_points_nr = c->write_points_nr;
1291 *wp_ret = wp = writepoint_find(trans, write_point.v);
1293 /* metadata may not allocate on cache devices: */
1294 if (wp->data_type != BCH_DATA_user)
1297 if (target && !(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1298 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1299 target, erasure_code,
1300 nr_replicas, &nr_effective,
1301 &have_cache, watermark,
1304 bch2_err_matches(ret, BCH_ERR_transaction_restart))
1307 /* Don't retry from all devices if we're out of open buckets: */
1308 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1309 goto allocate_blocking;
1312 * Only try to allocate cache (durability = 0 devices) from the
1317 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1319 nr_replicas, &nr_effective,
1320 &have_cache, watermark,
1324 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1325 target, erasure_code,
1326 nr_replicas, &nr_effective,
1327 &have_cache, watermark,
1331 BUG_ON(!ret && nr_effective < nr_replicas);
1333 if (erasure_code && !ec_open_bucket(c, &ptrs))
1334 pr_debug("failed to get ec bucket: ret %u", ret);
1336 if (ret == -BCH_ERR_insufficient_devices &&
1337 nr_effective >= nr_replicas_required)
1343 /* Free buckets we didn't use: */
1344 open_bucket_for_each(c, &wp->ptrs, ob, i)
1345 open_bucket_free_unused(c, ob);
1349 wp->sectors_free = UINT_MAX;
1351 open_bucket_for_each(c, &wp->ptrs, ob, i)
1352 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1354 BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1358 open_bucket_for_each(c, &wp->ptrs, ob, i)
1359 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1360 ob_push(c, &ptrs, ob);
1362 open_bucket_free_unused(c, ob);
1365 mutex_unlock(&wp->lock);
1367 if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1368 try_decrease_writepoints(trans, write_points_nr))
1371 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1372 bch2_err_matches(ret, BCH_ERR_freelist_empty))
1374 ? -BCH_ERR_bucket_alloc_blocked
1375 : -BCH_ERR_ENOSPC_bucket_alloc;
1380 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1382 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1384 return (struct bch_extent_ptr) {
1385 .type = 1 << BCH_EXTENT_ENTRY_ptr,
1388 .offset = bucket_to_sector(ca, ob->bucket) +
1389 ca->mi.bucket_size -
1394 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1395 struct bkey_i *k, unsigned sectors,
1398 bch2_alloc_sectors_append_ptrs_inlined(c, wp, k, sectors, cached);
1402 * Append pointers to the space we just allocated to @k, and mark @sectors space
1403 * as allocated out of @ob
1405 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1407 bch2_alloc_sectors_done_inlined(c, wp);
1410 static inline void writepoint_init(struct write_point *wp,
1411 enum bch_data_type type)
1413 mutex_init(&wp->lock);
1414 wp->data_type = type;
1416 INIT_WORK(&wp->index_update_work, bch2_write_point_do_index_updates);
1417 INIT_LIST_HEAD(&wp->writes);
1418 spin_lock_init(&wp->writes_lock);
1421 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1423 struct open_bucket *ob;
1424 struct write_point *wp;
1426 mutex_init(&c->write_points_hash_lock);
1427 c->write_points_nr = ARRAY_SIZE(c->write_points);
1429 /* open bucket 0 is a sentinal NULL: */
1430 spin_lock_init(&c->open_buckets[0].lock);
1432 for (ob = c->open_buckets + 1;
1433 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1434 spin_lock_init(&ob->lock);
1435 c->open_buckets_nr_free++;
1437 ob->freelist = c->open_buckets_freelist;
1438 c->open_buckets_freelist = ob - c->open_buckets;
1441 writepoint_init(&c->btree_write_point, BCH_DATA_btree);
1442 writepoint_init(&c->rebalance_write_point, BCH_DATA_user);
1443 writepoint_init(&c->copygc_write_point, BCH_DATA_user);
1445 for (wp = c->write_points;
1446 wp < c->write_points + c->write_points_nr; wp++) {
1447 writepoint_init(wp, BCH_DATA_user);
1449 wp->last_used = local_clock();
1450 wp->write_point = (unsigned long) wp;
1451 hlist_add_head_rcu(&wp->node,
1452 writepoint_hash(c, wp->write_point));
1456 static void bch2_open_bucket_to_text(struct printbuf *out, struct bch_fs *c, struct open_bucket *ob)
1458 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1459 unsigned data_type = ob->data_type;
1460 barrier(); /* READ_ONCE() doesn't work on bitfields */
1462 prt_printf(out, "%zu ref %u %s %u:%llu gen %u allocated %u/%u",
1463 ob - c->open_buckets,
1464 atomic_read(&ob->pin),
1465 data_type < BCH_DATA_NR ? bch2_data_types[data_type] : "invalid data type",
1466 ob->dev, ob->bucket, ob->gen,
1467 ca->mi.bucket_size - ob->sectors_free, ca->mi.bucket_size);
1469 prt_printf(out, " ec idx %llu", ob->ec->idx);
1470 if (ob->on_partial_list)
1471 prt_str(out, " partial");
1475 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1477 struct open_bucket *ob;
1481 for (ob = c->open_buckets;
1482 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1484 spin_lock(&ob->lock);
1485 if (ob->valid && !ob->on_partial_list)
1486 bch2_open_bucket_to_text(out, c, ob);
1487 spin_unlock(&ob->lock);
1493 void bch2_open_buckets_partial_to_text(struct printbuf *out, struct bch_fs *c)
1498 spin_lock(&c->freelist_lock);
1500 for (i = 0; i < c->open_buckets_partial_nr; i++)
1501 bch2_open_bucket_to_text(out, c,
1502 c->open_buckets + c->open_buckets_partial[i]);
1504 spin_unlock(&c->freelist_lock);
1508 static const char * const bch2_write_point_states[] = {
1510 WRITE_POINT_STATES()
1515 void bch2_write_points_to_text(struct printbuf *out, struct bch_fs *c)
1517 struct write_point *wp;
1520 for (wp = c->write_points;
1521 wp < c->write_points + ARRAY_SIZE(c->write_points);
1523 prt_printf(out, "%lu: ", wp->write_point);
1524 prt_human_readable_u64(out, wp->sectors_allocated);
1526 prt_printf(out, " last wrote: ");
1527 bch2_pr_time_units(out, sched_clock() - wp->last_used);
1529 for (i = 0; i < WRITE_POINT_STATE_NR; i++) {
1530 prt_printf(out, " %s: ", bch2_write_point_states[i]);
1531 bch2_pr_time_units(out, wp->time[i]);