1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright 2012 Google, Inc.
5 * Foreground allocator code: allocate buckets from freelist, and allocate in
6 * sector granularity from writepoints.
8 * bch2_bucket_alloc() allocates a single bucket from a specific device.
10 * bch2_bucket_alloc_set() allocates one or more buckets from different devices
11 * in a given filesystem.
15 #include "alloc_background.h"
16 #include "alloc_foreground.h"
17 #include "btree_iter.h"
18 #include "btree_update.h"
21 #include "buckets_waiting_for_journal.h"
24 #include "disk_groups.h"
30 #include <linux/math64.h>
31 #include <linux/rculist.h>
32 #include <linux/rcupdate.h>
33 #include <trace/events/bcachefs.h>
35 const char * const bch2_alloc_reserves[] = {
43 * Open buckets represent a bucket that's currently being allocated from. They
46 * - They track buckets that have been partially allocated, allowing for
47 * sub-bucket sized allocations - they're used by the sector allocator below
49 * - They provide a reference to the buckets they own that mark and sweep GC
50 * can find, until the new allocation has a pointer to it inserted into the
53 * When allocating some space with the sector allocator, the allocation comes
54 * with a reference to an open bucket - the caller is required to put that
55 * reference _after_ doing the index update that makes its allocation reachable.
58 static void bch2_open_bucket_hash_add(struct bch_fs *c, struct open_bucket *ob)
60 open_bucket_idx_t idx = ob - c->open_buckets;
61 open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
67 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
69 open_bucket_idx_t idx = ob - c->open_buckets;
70 open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
72 while (*slot != idx) {
74 slot = &c->open_buckets[*slot].hash;
81 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
83 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
86 bch2_ec_bucket_written(c, ob);
90 percpu_down_read(&c->mark_lock);
96 spin_unlock(&ob->lock);
97 percpu_up_read(&c->mark_lock);
99 spin_lock(&c->freelist_lock);
100 bch2_open_bucket_hash_remove(c, ob);
102 ob->freelist = c->open_buckets_freelist;
103 c->open_buckets_freelist = ob - c->open_buckets;
105 c->open_buckets_nr_free++;
106 ca->nr_open_buckets--;
107 spin_unlock(&c->freelist_lock);
109 closure_wake_up(&c->open_buckets_wait);
112 void bch2_open_bucket_write_error(struct bch_fs *c,
113 struct open_buckets *obs,
116 struct open_bucket *ob;
119 open_bucket_for_each(c, obs, ob, i)
120 if (ob->dev == dev && ob->ec)
121 bch2_ec_bucket_cancel(c, ob);
124 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
126 struct open_bucket *ob;
128 BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
130 ob = c->open_buckets + c->open_buckets_freelist;
131 c->open_buckets_freelist = ob->freelist;
132 atomic_set(&ob->pin, 1);
135 c->open_buckets_nr_free--;
139 static void open_bucket_free_unused(struct bch_fs *c,
140 struct write_point *wp,
141 struct open_bucket *ob)
143 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
144 bool may_realloc = wp->data_type == BCH_DATA_user;
146 BUG_ON(ca->open_buckets_partial_nr >
147 ARRAY_SIZE(ca->open_buckets_partial));
149 if (ca->open_buckets_partial_nr <
150 ARRAY_SIZE(ca->open_buckets_partial) &&
152 spin_lock(&c->freelist_lock);
153 ob->on_partial_list = true;
154 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
155 ob - c->open_buckets;
156 spin_unlock(&c->freelist_lock);
158 closure_wake_up(&c->open_buckets_wait);
159 closure_wake_up(&c->freelist_wait);
161 bch2_open_bucket_put(c, ob);
165 /* _only_ for allocating the journal on a new device: */
166 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
168 while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
169 u64 b = ca->new_fs_bucket_idx++;
171 if (!is_superblock_bucket(ca, b) &&
172 (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
179 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
183 case RESERVE_btree_movinggc:
185 case RESERVE_movinggc:
186 return OPEN_BUCKETS_COUNT / 4;
188 return OPEN_BUCKETS_COUNT / 2;
192 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
194 enum alloc_reserve reserve,
195 struct bch_alloc_v4 *a,
197 u64 *skipped_need_journal_commit,
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 (*skipped_need_journal_commit)++;
219 spin_lock(&c->freelist_lock);
221 if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
223 closure_wait(&c->open_buckets_wait, cl);
225 if (!c->blocked_allocate_open_bucket)
226 c->blocked_allocate_open_bucket = local_clock();
228 spin_unlock(&c->freelist_lock);
230 trace_open_bucket_alloc_fail(ca, bch2_alloc_reserves[reserve]);
231 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
234 /* Recheck under lock: */
235 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
236 spin_unlock(&c->freelist_lock);
241 ob = bch2_open_bucket_alloc(c);
243 spin_lock(&ob->lock);
246 ob->sectors_free = ca->mi.bucket_size;
247 ob->alloc_reserve = reserve;
248 ob->dev = ca->dev_idx;
251 spin_unlock(&ob->lock);
253 ca->nr_open_buckets++;
254 bch2_open_bucket_hash_add(c, ob);
256 if (c->blocked_allocate_open_bucket) {
257 bch2_time_stats_update(
258 &c->times[BCH_TIME_blocked_allocate_open_bucket],
259 c->blocked_allocate_open_bucket);
260 c->blocked_allocate_open_bucket = 0;
263 if (c->blocked_allocate) {
264 bch2_time_stats_update(
265 &c->times[BCH_TIME_blocked_allocate],
266 c->blocked_allocate);
267 c->blocked_allocate = 0;
270 spin_unlock(&c->freelist_lock);
272 trace_bucket_alloc(ca, bch2_alloc_reserves[reserve]);
276 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
277 enum alloc_reserve reserve, u64 free_entry,
279 u64 *skipped_need_journal_commit,
283 struct bch_fs *c = trans->c;
284 struct btree_iter iter;
286 struct open_bucket *ob;
287 struct bch_alloc_v4 a;
288 u64 b = free_entry & ~(~0ULL << 56);
289 unsigned genbits = free_entry >> 56;
290 struct printbuf buf = PRINTBUF;
293 bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
294 k = bch2_btree_iter_peek_slot(&iter);
301 bch2_alloc_to_v4(k, &a);
303 if (bch2_fs_inconsistent_on(bucket_state(a) != BUCKET_free, c,
304 "non free bucket in freespace btree (state %s)\n"
306 " at %llu (genbits %u)",
307 bch2_bucket_states[bucket_state(a)],
308 (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
309 free_entry, genbits)) {
314 if (bch2_fs_inconsistent_on(genbits != (alloc_freespace_genbits(a) >> 56), c,
315 "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
317 genbits, alloc_freespace_genbits(a) >> 56,
318 (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
323 if (bch2_fs_inconsistent_on(b < ca->mi.first_bucket || b >= ca->mi.nbuckets, c,
324 "freespace btree has bucket outside allowed range (got %llu, valid %u-%llu)",
325 b, ca->mi.first_bucket, ca->mi.nbuckets)) {
330 ob = __try_alloc_bucket(c, ca, b, reserve, &a,
332 skipped_need_journal_commit,
336 bch2_trans_iter_exit(trans, &iter);
341 static struct open_bucket *try_alloc_partial_bucket(struct bch_fs *c, struct bch_dev *ca,
342 enum alloc_reserve reserve)
344 struct open_bucket *ob;
347 spin_lock(&c->freelist_lock);
349 for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
350 ob = c->open_buckets + ca->open_buckets_partial[i];
352 if (reserve <= ob->alloc_reserve) {
353 array_remove_item(ca->open_buckets_partial,
354 ca->open_buckets_partial_nr,
356 ob->on_partial_list = false;
357 ob->alloc_reserve = reserve;
358 spin_unlock(&c->freelist_lock);
363 spin_unlock(&c->freelist_lock);
368 * This path is for before the freespace btree is initialized:
370 * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
371 * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
373 static noinline struct open_bucket *
374 bch2_bucket_alloc_trans_early(struct btree_trans *trans,
376 enum alloc_reserve reserve,
380 u64 *skipped_need_journal_commit,
384 struct btree_iter iter;
386 struct open_bucket *ob = NULL;
389 *cur_bucket = max_t(u64, *cur_bucket, ca->mi.first_bucket);
390 *cur_bucket = max_t(u64, *cur_bucket, ca->new_fs_bucket_idx);
392 for_each_btree_key(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, *cur_bucket),
393 BTREE_ITER_SLOTS, k, ret) {
394 struct bch_alloc_v4 a;
396 if (bkey_cmp(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
399 if (ca->new_fs_bucket_idx &&
400 is_superblock_bucket(ca, k.k->p.offset))
403 bch2_alloc_to_v4(k, &a);
405 if (bucket_state(a) != BUCKET_free)
410 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, reserve, &a,
412 skipped_need_journal_commit,
418 bch2_trans_iter_exit(trans, &iter);
420 *cur_bucket = iter.pos.offset;
422 return ob ?: ERR_PTR(ret ?: -FREELIST_EMPTY);
425 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
427 enum alloc_reserve reserve,
431 u64 *skipped_need_journal_commit,
435 struct btree_iter iter;
437 struct open_bucket *ob = NULL;
440 if (unlikely(!ca->mi.freespace_initialized))
441 return bch2_bucket_alloc_trans_early(trans, ca, reserve,
445 skipped_need_journal_commit,
449 BUG_ON(ca->new_fs_bucket_idx);
451 for_each_btree_key(trans, iter, BTREE_ID_freespace,
452 POS(ca->dev_idx, *cur_bucket), 0, k, ret) {
453 if (k.k->p.inode != ca->dev_idx)
456 for (*cur_bucket = max(*cur_bucket, bkey_start_offset(k.k));
457 *cur_bucket != k.k->p.offset && !ob;
459 if (btree_trans_too_many_iters(trans)) {
460 ob = ERR_PTR(-EINTR);
466 ob = try_alloc_bucket(trans, ca, reserve,
469 skipped_need_journal_commit,
476 bch2_trans_iter_exit(trans, &iter);
478 return ob ?: ERR_PTR(ret);
482 * bch_bucket_alloc - allocate a single bucket from a specific device
484 * Returns index of bucket on success, 0 on failure
486 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
487 enum alloc_reserve reserve,
488 bool may_alloc_partial,
491 struct open_bucket *ob = NULL;
492 u64 avail = dev_buckets_available(ca, reserve);
494 u64 buckets_seen = 0;
495 u64 skipped_open = 0;
496 u64 skipped_need_journal_commit = 0;
497 u64 skipped_nouse = 0;
500 if (may_alloc_partial) {
501 ob = try_alloc_partial_bucket(c, ca, reserve);
508 closure_wait(&c->freelist_wait, cl);
509 /* recheck after putting ourself on waitlist */
510 avail = dev_buckets_available(ca, reserve);
512 closure_wake_up(&c->freelist_wait);
517 if (!c->blocked_allocate)
518 c->blocked_allocate = local_clock();
520 ob = ERR_PTR(-FREELIST_EMPTY);
524 ret = bch2_trans_do(c, NULL, NULL, 0,
525 PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
529 &skipped_need_journal_commit,
533 if (skipped_need_journal_commit * 2 > avail)
534 bch2_journal_flush_async(&c->journal, NULL);
537 ob = ERR_PTR(ret ?: -FREELIST_EMPTY);
540 trace_bucket_alloc_fail(ca, bch2_alloc_reserves[reserve], avail,
543 skipped_need_journal_commit,
545 cl == NULL, PTR_ERR(ob));
546 atomic_long_inc(&c->bucket_alloc_fail);
552 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
553 unsigned l, unsigned r)
555 return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
556 (stripe->next_alloc[l] < stripe->next_alloc[r]));
559 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
561 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
562 struct dev_stripe_state *stripe,
563 struct bch_devs_mask *devs)
565 struct dev_alloc_list ret = { .nr = 0 };
568 for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
569 ret.devs[ret.nr++] = i;
571 bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
575 void bch2_dev_stripe_increment(struct bch_dev *ca,
576 struct dev_stripe_state *stripe)
578 u64 *v = stripe->next_alloc + ca->dev_idx;
579 u64 free_space = dev_buckets_available(ca, RESERVE_none);
580 u64 free_space_inv = free_space
581 ? div64_u64(1ULL << 48, free_space)
585 if (*v + free_space_inv >= *v)
586 *v += free_space_inv;
590 for (v = stripe->next_alloc;
591 v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
592 *v = *v < scale ? 0 : *v - scale;
595 #define BUCKET_MAY_ALLOC_PARTIAL (1 << 0)
596 #define BUCKET_ALLOC_USE_DURABILITY (1 << 1)
598 static void add_new_bucket(struct bch_fs *c,
599 struct open_buckets *ptrs,
600 struct bch_devs_mask *devs_may_alloc,
601 unsigned *nr_effective,
604 struct open_bucket *ob)
606 unsigned durability =
607 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
609 __clear_bit(ob->dev, devs_may_alloc->d);
610 *nr_effective += (flags & BUCKET_ALLOC_USE_DURABILITY)
612 *have_cache |= !durability;
614 ob_push(c, ptrs, ob);
617 int bch2_bucket_alloc_set(struct bch_fs *c,
618 struct open_buckets *ptrs,
619 struct dev_stripe_state *stripe,
620 struct bch_devs_mask *devs_may_alloc,
621 unsigned nr_replicas,
622 unsigned *nr_effective,
624 enum alloc_reserve reserve,
628 struct dev_alloc_list devs_sorted =
629 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
632 int ret = -INSUFFICIENT_DEVICES;
635 BUG_ON(*nr_effective >= nr_replicas);
637 for (i = 0; i < devs_sorted.nr; i++) {
638 struct open_bucket *ob;
640 dev = devs_sorted.devs[i];
643 ca = rcu_dereference(c->devs[dev]);
645 percpu_ref_get(&ca->ref);
651 if (!ca->mi.durability && *have_cache) {
652 percpu_ref_put(&ca->ref);
656 ob = bch2_bucket_alloc(c, ca, reserve,
657 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
659 bch2_dev_stripe_increment(ca, stripe);
660 percpu_ref_put(&ca->ref);
670 add_new_bucket(c, ptrs, devs_may_alloc,
671 nr_effective, have_cache, flags, ob);
673 if (*nr_effective >= nr_replicas) {
682 /* Allocate from stripes: */
685 * if we can't allocate a new stripe because there are already too many
686 * partially filled stripes, force allocating from an existing stripe even when
687 * it's to a device we don't want:
690 static int bucket_alloc_from_stripe(struct bch_fs *c,
691 struct open_buckets *ptrs,
692 struct write_point *wp,
693 struct bch_devs_mask *devs_may_alloc,
695 unsigned erasure_code,
696 unsigned nr_replicas,
697 unsigned *nr_effective,
702 struct dev_alloc_list devs_sorted;
703 struct ec_stripe_head *h;
704 struct open_bucket *ob;
714 if (ec_open_bucket(c, ptrs))
717 h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
718 wp == &c->copygc_write_point,
725 devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
727 for (i = 0; i < devs_sorted.nr; i++)
728 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
729 if (!h->s->blocks[ec_idx])
732 ob = c->open_buckets + h->s->blocks[ec_idx];
733 if (ob->dev == devs_sorted.devs[i] &&
734 !test_and_set_bit(ec_idx, h->s->blocks_allocated))
739 ca = bch_dev_bkey_exists(c, ob->dev);
744 add_new_bucket(c, ptrs, devs_may_alloc,
745 nr_effective, have_cache, flags, ob);
746 atomic_inc(&h->s->pin);
748 bch2_ec_stripe_head_put(c, h);
752 /* Sector allocator */
754 static void get_buckets_from_writepoint(struct bch_fs *c,
755 struct open_buckets *ptrs,
756 struct write_point *wp,
757 struct bch_devs_mask *devs_may_alloc,
758 unsigned nr_replicas,
759 unsigned *nr_effective,
764 struct open_buckets ptrs_skip = { .nr = 0 };
765 struct open_bucket *ob;
768 open_bucket_for_each(c, &wp->ptrs, ob, i) {
769 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
771 if (*nr_effective < nr_replicas &&
772 test_bit(ob->dev, devs_may_alloc->d) &&
773 (ca->mi.durability ||
774 (wp->data_type == BCH_DATA_user && !*have_cache)) &&
775 (ob->ec || !need_ec)) {
776 add_new_bucket(c, ptrs, devs_may_alloc,
777 nr_effective, have_cache,
780 ob_push(c, &ptrs_skip, ob);
783 wp->ptrs = ptrs_skip;
786 static int open_bucket_add_buckets(struct bch_fs *c,
787 struct open_buckets *ptrs,
788 struct write_point *wp,
789 struct bch_devs_list *devs_have,
791 unsigned erasure_code,
792 unsigned nr_replicas,
793 unsigned *nr_effective,
795 enum alloc_reserve reserve,
799 struct bch_devs_mask devs;
800 struct open_bucket *ob;
801 struct closure *cl = NULL;
806 devs = target_rw_devs(c, wp->data_type, target);
809 /* Don't allocate from devices we already have pointers to: */
810 for (i = 0; i < devs_have->nr; i++)
811 __clear_bit(devs_have->devs[i], devs.d);
813 open_bucket_for_each(c, ptrs, ob, i)
814 __clear_bit(ob->dev, devs.d);
817 if (!ec_open_bucket(c, ptrs)) {
818 get_buckets_from_writepoint(c, ptrs, wp, &devs,
819 nr_replicas, nr_effective,
820 have_cache, flags, true);
821 if (*nr_effective >= nr_replicas)
825 if (!ec_open_bucket(c, ptrs)) {
826 ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
827 target, erasure_code,
828 nr_replicas, nr_effective,
829 have_cache, flags, _cl);
830 if (ret == -FREELIST_EMPTY ||
831 ret == -OPEN_BUCKETS_EMPTY)
833 if (*nr_effective >= nr_replicas)
838 get_buckets_from_writepoint(c, ptrs, wp, &devs,
839 nr_replicas, nr_effective,
840 have_cache, flags, false);
841 if (*nr_effective >= nr_replicas)
846 * Try nonblocking first, so that if one device is full we'll try from
849 ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
850 nr_replicas, nr_effective, have_cache,
852 if (ret && ret != -INSUFFICIENT_DEVICES && !cl && _cl) {
860 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
861 struct open_buckets *obs)
863 struct open_buckets ptrs = { .nr = 0 };
864 struct open_bucket *ob, *ob2;
867 open_bucket_for_each(c, obs, ob, i) {
868 bool drop = !ca || ob->dev == ca->dev_idx;
870 if (!drop && ob->ec) {
871 mutex_lock(&ob->ec->lock);
872 for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
873 if (!ob->ec->blocks[j])
876 ob2 = c->open_buckets + ob->ec->blocks[j];
877 drop |= ob2->dev == ca->dev_idx;
879 mutex_unlock(&ob->ec->lock);
883 bch2_open_bucket_put(c, ob);
885 ob_push(c, &ptrs, ob);
891 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
892 struct write_point *wp)
894 mutex_lock(&wp->lock);
895 bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
896 mutex_unlock(&wp->lock);
899 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
900 unsigned long write_point)
903 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
905 return &c->write_points_hash[hash];
908 static struct write_point *__writepoint_find(struct hlist_head *head,
909 unsigned long write_point)
911 struct write_point *wp;
914 hlist_for_each_entry_rcu(wp, head, node)
915 if (wp->write_point == write_point)
923 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
925 u64 stranded = c->write_points_nr * c->bucket_size_max;
926 u64 free = bch2_fs_usage_read_short(c).free;
928 return stranded * factor > free;
931 static bool try_increase_writepoints(struct bch_fs *c)
933 struct write_point *wp;
935 if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
936 too_many_writepoints(c, 32))
939 wp = c->write_points + c->write_points_nr++;
940 hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
944 static bool try_decrease_writepoints(struct bch_fs *c,
947 struct write_point *wp;
949 mutex_lock(&c->write_points_hash_lock);
950 if (c->write_points_nr < old_nr) {
951 mutex_unlock(&c->write_points_hash_lock);
955 if (c->write_points_nr == 1 ||
956 !too_many_writepoints(c, 8)) {
957 mutex_unlock(&c->write_points_hash_lock);
961 wp = c->write_points + --c->write_points_nr;
963 hlist_del_rcu(&wp->node);
964 mutex_unlock(&c->write_points_hash_lock);
966 bch2_writepoint_stop(c, NULL, wp);
970 static struct write_point *writepoint_find(struct bch_fs *c,
971 unsigned long write_point)
973 struct write_point *wp, *oldest;
974 struct hlist_head *head;
976 if (!(write_point & 1UL)) {
977 wp = (struct write_point *) write_point;
978 mutex_lock(&wp->lock);
982 head = writepoint_hash(c, write_point);
984 wp = __writepoint_find(head, write_point);
987 mutex_lock(&wp->lock);
988 if (wp->write_point == write_point)
990 mutex_unlock(&wp->lock);
995 for (wp = c->write_points;
996 wp < c->write_points + c->write_points_nr; wp++)
997 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1000 mutex_lock(&oldest->lock);
1001 mutex_lock(&c->write_points_hash_lock);
1002 if (oldest >= c->write_points + c->write_points_nr ||
1003 try_increase_writepoints(c)) {
1004 mutex_unlock(&c->write_points_hash_lock);
1005 mutex_unlock(&oldest->lock);
1006 goto restart_find_oldest;
1009 wp = __writepoint_find(head, write_point);
1010 if (wp && wp != oldest) {
1011 mutex_unlock(&c->write_points_hash_lock);
1012 mutex_unlock(&oldest->lock);
1017 hlist_del_rcu(&wp->node);
1018 wp->write_point = write_point;
1019 hlist_add_head_rcu(&wp->node, head);
1020 mutex_unlock(&c->write_points_hash_lock);
1022 wp->last_used = sched_clock();
1027 * Get us an open_bucket we can allocate from, return with it locked:
1029 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
1031 unsigned erasure_code,
1032 struct write_point_specifier write_point,
1033 struct bch_devs_list *devs_have,
1034 unsigned nr_replicas,
1035 unsigned nr_replicas_required,
1036 enum alloc_reserve reserve,
1040 struct write_point *wp;
1041 struct open_bucket *ob;
1042 struct open_buckets ptrs;
1043 unsigned nr_effective, write_points_nr;
1044 unsigned ob_flags = 0;
1049 if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
1050 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
1052 BUG_ON(!nr_replicas || !nr_replicas_required);
1056 write_points_nr = c->write_points_nr;
1059 wp = writepoint_find(c, write_point.v);
1061 if (wp->data_type == BCH_DATA_user)
1062 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
1064 /* metadata may not allocate on cache devices: */
1065 if (wp->data_type != BCH_DATA_user)
1068 if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1069 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1070 target, erasure_code,
1071 nr_replicas, &nr_effective,
1072 &have_cache, reserve,
1075 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1076 target, erasure_code,
1077 nr_replicas, &nr_effective,
1078 &have_cache, reserve,
1083 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1085 nr_replicas, &nr_effective,
1086 &have_cache, reserve,
1090 BUG_ON(!ret && nr_effective < nr_replicas);
1092 if (erasure_code && !ec_open_bucket(c, &ptrs))
1093 pr_debug("failed to get ec bucket: ret %u", ret);
1095 if (ret == -INSUFFICIENT_DEVICES &&
1096 nr_effective >= nr_replicas_required)
1102 /* Free buckets we didn't use: */
1103 open_bucket_for_each(c, &wp->ptrs, ob, i)
1104 open_bucket_free_unused(c, wp, ob);
1108 wp->sectors_free = UINT_MAX;
1110 open_bucket_for_each(c, &wp->ptrs, ob, i)
1111 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1113 BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1117 open_bucket_for_each(c, &wp->ptrs, ob, i)
1118 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1119 ob_push(c, &ptrs, ob);
1121 open_bucket_free_unused(c, wp, ob);
1124 mutex_unlock(&wp->lock);
1126 if (ret == -FREELIST_EMPTY &&
1127 try_decrease_writepoints(c, write_points_nr))
1131 case -OPEN_BUCKETS_EMPTY:
1132 case -FREELIST_EMPTY:
1133 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
1134 case -INSUFFICIENT_DEVICES:
1135 return ERR_PTR(-EROFS);
1137 return ERR_PTR(ret);
1141 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1143 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1145 return (struct bch_extent_ptr) {
1146 .type = 1 << BCH_EXTENT_ENTRY_ptr,
1149 .offset = bucket_to_sector(ca, ob->bucket) +
1150 ca->mi.bucket_size -
1156 * Append pointers to the space we just allocated to @k, and mark @sectors space
1157 * as allocated out of @ob
1159 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1160 struct bkey_i *k, unsigned sectors,
1164 struct open_bucket *ob;
1167 BUG_ON(sectors > wp->sectors_free);
1168 wp->sectors_free -= sectors;
1170 open_bucket_for_each(c, &wp->ptrs, ob, i) {
1171 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1172 struct bch_extent_ptr ptr = bch2_ob_ptr(c, ob);
1174 ptr.cached = cached ||
1175 (!ca->mi.durability &&
1176 wp->data_type == BCH_DATA_user);
1178 bch2_bkey_append_ptr(k, ptr);
1180 BUG_ON(sectors > ob->sectors_free);
1181 ob->sectors_free -= sectors;
1186 * Append pointers to the space we just allocated to @k, and mark @sectors space
1187 * as allocated out of @ob
1189 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1191 struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
1192 struct open_bucket *ob;
1195 open_bucket_for_each(c, &wp->ptrs, ob, i)
1196 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
1199 mutex_unlock(&wp->lock);
1201 bch2_open_buckets_put(c, &ptrs);
1204 static inline void writepoint_init(struct write_point *wp,
1205 enum bch_data_type type)
1207 mutex_init(&wp->lock);
1208 wp->data_type = type;
1211 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1213 struct open_bucket *ob;
1214 struct write_point *wp;
1216 mutex_init(&c->write_points_hash_lock);
1217 c->write_points_nr = ARRAY_SIZE(c->write_points);
1219 /* open bucket 0 is a sentinal NULL: */
1220 spin_lock_init(&c->open_buckets[0].lock);
1222 for (ob = c->open_buckets + 1;
1223 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1224 spin_lock_init(&ob->lock);
1225 c->open_buckets_nr_free++;
1227 ob->freelist = c->open_buckets_freelist;
1228 c->open_buckets_freelist = ob - c->open_buckets;
1231 writepoint_init(&c->btree_write_point, BCH_DATA_btree);
1232 writepoint_init(&c->rebalance_write_point, BCH_DATA_user);
1233 writepoint_init(&c->copygc_write_point, BCH_DATA_user);
1235 for (wp = c->write_points;
1236 wp < c->write_points + c->write_points_nr; wp++) {
1237 writepoint_init(wp, BCH_DATA_user);
1239 wp->last_used = sched_clock();
1240 wp->write_point = (unsigned long) wp;
1241 hlist_add_head_rcu(&wp->node,
1242 writepoint_hash(c, wp->write_point));
1246 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1248 struct open_bucket *ob;
1250 for (ob = c->open_buckets;
1251 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1253 spin_lock(&ob->lock);
1254 if (ob->valid && !ob->on_partial_list) {
1255 pr_buf(out, "%zu ref %u type %s\n",
1256 ob - c->open_buckets,
1257 atomic_read(&ob->pin),
1258 bch2_data_types[ob->data_type]);
1260 spin_unlock(&ob->lock);