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_btree:
195 case BCH_WATERMARK_btree_copygc:
197 case BCH_WATERMARK_copygc:
198 return OPEN_BUCKETS_COUNT / 4;
200 return OPEN_BUCKETS_COUNT / 2;
204 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
206 enum bch_watermark watermark,
207 const struct bch_alloc_v4 *a,
208 struct bucket_alloc_state *s,
211 struct open_bucket *ob;
213 if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
218 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
223 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
224 c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
225 s->skipped_need_journal_commit++;
229 if (bch2_bucket_nocow_is_locked(&c->nocow_locks, POS(ca->dev_idx, bucket))) {
234 spin_lock(&c->freelist_lock);
236 if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(watermark))) {
238 closure_wait(&c->open_buckets_wait, cl);
240 if (!c->blocked_allocate_open_bucket)
241 c->blocked_allocate_open_bucket = local_clock();
243 spin_unlock(&c->freelist_lock);
244 return ERR_PTR(-BCH_ERR_open_buckets_empty);
247 /* Recheck under lock: */
248 if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
249 spin_unlock(&c->freelist_lock);
254 ob = bch2_open_bucket_alloc(c);
256 spin_lock(&ob->lock);
259 ob->sectors_free = ca->mi.bucket_size;
260 ob->dev = ca->dev_idx;
263 spin_unlock(&ob->lock);
265 ca->nr_open_buckets++;
266 bch2_open_bucket_hash_add(c, ob);
268 if (c->blocked_allocate_open_bucket) {
269 bch2_time_stats_update(
270 &c->times[BCH_TIME_blocked_allocate_open_bucket],
271 c->blocked_allocate_open_bucket);
272 c->blocked_allocate_open_bucket = 0;
275 if (c->blocked_allocate) {
276 bch2_time_stats_update(
277 &c->times[BCH_TIME_blocked_allocate],
278 c->blocked_allocate);
279 c->blocked_allocate = 0;
282 spin_unlock(&c->freelist_lock);
286 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
287 enum bch_watermark watermark, u64 free_entry,
288 struct bucket_alloc_state *s,
289 struct bkey_s_c freespace_k,
292 struct bch_fs *c = trans->c;
293 struct btree_iter iter = { NULL };
295 struct open_bucket *ob;
296 struct bch_alloc_v4 a_convert;
297 const struct bch_alloc_v4 *a;
298 u64 b = free_entry & ~(~0ULL << 56);
299 unsigned genbits = free_entry >> 56;
300 struct printbuf buf = PRINTBUF;
303 if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
304 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
306 ca->mi.first_bucket, ca->mi.nbuckets);
307 bch2_bkey_val_to_text(&buf, c, freespace_k);
308 bch2_trans_inconsistent(trans, "%s", buf.buf);
313 k = bch2_bkey_get_iter(trans, &iter,
314 BTREE_ID_alloc, POS(ca->dev_idx, b),
322 a = bch2_alloc_to_v4(k, &a_convert);
324 if (a->data_type != BCH_DATA_free) {
325 if (!test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
330 prt_printf(&buf, "non free bucket in freespace btree\n"
332 bch2_bkey_val_to_text(&buf, c, freespace_k);
333 prt_printf(&buf, "\n ");
334 bch2_bkey_val_to_text(&buf, c, k);
335 bch2_trans_inconsistent(trans, "%s", buf.buf);
340 if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
341 test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
342 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
344 genbits, alloc_freespace_genbits(*a) >> 56);
345 bch2_bkey_val_to_text(&buf, c, freespace_k);
346 prt_printf(&buf, "\n ");
347 bch2_bkey_val_to_text(&buf, c, k);
348 bch2_trans_inconsistent(trans, "%s", buf.buf);
354 if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
355 struct bch_backpointer bp;
356 struct bpos bp_pos = POS_MIN;
358 ret = bch2_get_next_backpointer(trans, POS(ca->dev_idx, b), -1,
360 BTREE_ITER_NOPRESERVE);
366 if (!bkey_eq(bp_pos, POS_MAX)) {
368 * Bucket may have data in it - we don't call
369 * bc2h_trans_inconnsistent() because fsck hasn't
377 ob = __try_alloc_bucket(c, ca, b, watermark, a, s, cl);
379 iter.path->preserve = false;
381 if (iter.trans && iter.path)
382 set_btree_iter_dontneed(&iter);
383 bch2_trans_iter_exit(trans, &iter);
389 * This path is for before the freespace btree is initialized:
391 * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
392 * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
394 static noinline struct open_bucket *
395 bch2_bucket_alloc_early(struct btree_trans *trans,
397 enum bch_watermark watermark,
398 struct bucket_alloc_state *s,
401 struct btree_iter iter;
403 struct open_bucket *ob = NULL;
404 u64 alloc_start = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
405 u64 alloc_cursor = max(alloc_start, READ_ONCE(ca->alloc_cursor));
408 for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, alloc_cursor),
409 BTREE_ITER_SLOTS, k, ret) {
410 struct bch_alloc_v4 a_convert;
411 const struct bch_alloc_v4 *a;
413 if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
416 if (ca->new_fs_bucket_idx &&
417 is_superblock_bucket(ca, k.k->p.offset))
420 a = bch2_alloc_to_v4(k, &a_convert);
422 if (a->data_type != BCH_DATA_free)
427 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, watermark, a, s, cl);
431 bch2_trans_iter_exit(trans, &iter);
433 ca->alloc_cursor = alloc_cursor;
438 if (!ob && alloc_cursor > alloc_start) {
439 alloc_cursor = alloc_start;
446 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
448 enum bch_watermark watermark,
449 struct bucket_alloc_state *s,
452 struct btree_iter iter;
454 struct open_bucket *ob = NULL;
455 u64 alloc_start = max_t(u64, ca->mi.first_bucket, READ_ONCE(ca->alloc_cursor));
456 u64 alloc_cursor = alloc_start;
459 BUG_ON(ca->new_fs_bucket_idx);
461 for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
462 POS(ca->dev_idx, alloc_cursor), 0, k, ret) {
463 if (k.k->p.inode != ca->dev_idx)
466 for (alloc_cursor = max(alloc_cursor, bkey_start_offset(k.k));
467 alloc_cursor < k.k->p.offset;
469 ret = btree_trans_too_many_iters(trans);
477 ob = try_alloc_bucket(trans, ca, watermark,
478 alloc_cursor, s, k, cl);
480 iter.path->preserve = false;
488 bch2_trans_iter_exit(trans, &iter);
490 ca->alloc_cursor = alloc_cursor;
495 if (!ob && alloc_start > ca->mi.first_bucket) {
496 alloc_cursor = alloc_start = ca->mi.first_bucket;
504 * bch_bucket_alloc - allocate a single bucket from a specific device
506 * Returns index of bucket on success, 0 on failure
508 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
510 enum bch_watermark watermark,
512 struct bch_dev_usage *usage)
514 struct bch_fs *c = trans->c;
515 struct open_bucket *ob = NULL;
516 bool freespace = READ_ONCE(ca->mi.freespace_initialized);
518 struct bucket_alloc_state s = { 0 };
519 bool waiting = false;
521 bch2_dev_usage_read_fast(ca, usage);
522 avail = dev_buckets_free(ca, *usage, watermark);
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 ob = likely(freespace)
551 ? bch2_bucket_alloc_freelist(trans, ca, watermark, &s, cl)
552 : bch2_bucket_alloc_early(trans, ca, watermark, &s, cl);
554 if (s.skipped_need_journal_commit * 2 > avail)
555 bch2_journal_flush_async(&c->journal, NULL);
557 if (!ob && freespace && !test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
563 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
566 trace_and_count(c, bucket_alloc, ca,
567 bch2_watermarks[watermark],
569 usage->d[BCH_DATA_free].buckets,
571 bch2_copygc_wait_amount(c),
572 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
576 else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
577 trace_and_count(c, bucket_alloc_fail, ca,
578 bch2_watermarks[watermark],
580 usage->d[BCH_DATA_free].buckets,
582 bch2_copygc_wait_amount(c),
583 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
586 bch2_err_str(PTR_ERR(ob)));
591 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
592 enum bch_watermark watermark,
595 struct bch_dev_usage usage;
596 struct open_bucket *ob;
598 bch2_trans_do(c, NULL, NULL, 0,
599 PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, watermark,
604 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
605 unsigned l, unsigned r)
607 return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
608 (stripe->next_alloc[l] < stripe->next_alloc[r]));
611 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
613 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
614 struct dev_stripe_state *stripe,
615 struct bch_devs_mask *devs)
617 struct dev_alloc_list ret = { .nr = 0 };
620 for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
621 ret.devs[ret.nr++] = i;
623 bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
627 static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
628 struct dev_stripe_state *stripe,
629 struct bch_dev_usage *usage)
631 u64 *v = stripe->next_alloc + ca->dev_idx;
632 u64 free_space = dev_buckets_available(ca, BCH_WATERMARK_normal);
633 u64 free_space_inv = free_space
634 ? div64_u64(1ULL << 48, free_space)
638 if (*v + free_space_inv >= *v)
639 *v += free_space_inv;
643 for (v = stripe->next_alloc;
644 v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
645 *v = *v < scale ? 0 : *v - scale;
648 void bch2_dev_stripe_increment(struct bch_dev *ca,
649 struct dev_stripe_state *stripe)
651 struct bch_dev_usage usage;
653 bch2_dev_usage_read_fast(ca, &usage);
654 bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
657 static int add_new_bucket(struct bch_fs *c,
658 struct open_buckets *ptrs,
659 struct bch_devs_mask *devs_may_alloc,
660 unsigned nr_replicas,
661 unsigned *nr_effective,
664 struct open_bucket *ob)
666 unsigned durability =
667 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
669 BUG_ON(*nr_effective >= nr_replicas);
670 BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
672 __clear_bit(ob->dev, devs_may_alloc->d);
673 *nr_effective += (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)
675 *have_cache |= !durability;
677 ob_push(c, ptrs, ob);
679 if (*nr_effective >= nr_replicas)
686 int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
687 struct open_buckets *ptrs,
688 struct dev_stripe_state *stripe,
689 struct bch_devs_mask *devs_may_alloc,
690 unsigned nr_replicas,
691 unsigned *nr_effective,
694 enum bch_data_type data_type,
695 enum bch_watermark watermark,
698 struct bch_fs *c = trans->c;
699 struct dev_alloc_list devs_sorted =
700 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
703 int ret = -BCH_ERR_insufficient_devices;
706 BUG_ON(*nr_effective >= nr_replicas);
708 for (i = 0; i < devs_sorted.nr; i++) {
709 struct bch_dev_usage usage;
710 struct open_bucket *ob;
712 dev = devs_sorted.devs[i];
715 ca = rcu_dereference(c->devs[dev]);
717 percpu_ref_get(&ca->ref);
723 if (!ca->mi.durability && *have_cache) {
724 percpu_ref_put(&ca->ref);
728 ob = bch2_bucket_alloc_trans(trans, ca, watermark, cl, &usage);
730 bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
731 percpu_ref_put(&ca->ref);
735 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
740 ob->data_type = data_type;
742 if (add_new_bucket(c, ptrs, devs_may_alloc,
743 nr_replicas, nr_effective,
744 have_cache, flags, ob)) {
753 /* Allocate from stripes: */
756 * if we can't allocate a new stripe because there are already too many
757 * partially filled stripes, force allocating from an existing stripe even when
758 * it's to a device we don't want:
761 static int bucket_alloc_from_stripe(struct btree_trans *trans,
762 struct open_buckets *ptrs,
763 struct write_point *wp,
764 struct bch_devs_mask *devs_may_alloc,
766 unsigned nr_replicas,
767 unsigned *nr_effective,
769 enum bch_watermark watermark,
773 struct bch_fs *c = trans->c;
774 struct dev_alloc_list devs_sorted;
775 struct ec_stripe_head *h;
776 struct open_bucket *ob;
784 if (ec_open_bucket(c, ptrs))
787 h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, watermark, cl);
793 devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
795 for (i = 0; i < devs_sorted.nr; i++)
796 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
797 if (!h->s->blocks[ec_idx])
800 ob = c->open_buckets + h->s->blocks[ec_idx];
801 if (ob->dev == devs_sorted.devs[i] &&
802 !test_and_set_bit(ec_idx, h->s->blocks_allocated))
807 ca = bch_dev_bkey_exists(c, ob->dev);
811 ec_stripe_new_get(h->s, STRIPE_REF_io);
813 ret = add_new_bucket(c, ptrs, devs_may_alloc,
814 nr_replicas, nr_effective,
815 have_cache, flags, ob);
817 bch2_ec_stripe_head_put(c, h);
821 /* Sector allocator */
823 static bool want_bucket(struct bch_fs *c,
824 struct write_point *wp,
825 struct bch_devs_mask *devs_may_alloc,
826 bool *have_cache, bool ec,
827 struct open_bucket *ob)
829 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
831 if (!test_bit(ob->dev, devs_may_alloc->d))
834 if (ob->data_type != wp->data_type)
837 if (!ca->mi.durability &&
838 (wp->data_type == BCH_DATA_btree || ec || *have_cache))
841 if (ec != (ob->ec != NULL))
847 static int bucket_alloc_set_writepoint(struct bch_fs *c,
848 struct open_buckets *ptrs,
849 struct write_point *wp,
850 struct bch_devs_mask *devs_may_alloc,
851 unsigned nr_replicas,
852 unsigned *nr_effective,
854 bool ec, unsigned flags)
856 struct open_buckets ptrs_skip = { .nr = 0 };
857 struct open_bucket *ob;
861 open_bucket_for_each(c, &wp->ptrs, ob, i) {
862 if (!ret && want_bucket(c, wp, devs_may_alloc,
864 ret = add_new_bucket(c, ptrs, devs_may_alloc,
865 nr_replicas, nr_effective,
866 have_cache, flags, ob);
868 ob_push(c, &ptrs_skip, ob);
870 wp->ptrs = ptrs_skip;
875 static int bucket_alloc_set_partial(struct bch_fs *c,
876 struct open_buckets *ptrs,
877 struct write_point *wp,
878 struct bch_devs_mask *devs_may_alloc,
879 unsigned nr_replicas,
880 unsigned *nr_effective,
881 bool *have_cache, bool ec,
882 enum bch_watermark watermark,
887 if (!c->open_buckets_partial_nr)
890 spin_lock(&c->freelist_lock);
892 if (!c->open_buckets_partial_nr)
895 for (i = c->open_buckets_partial_nr - 1; i >= 0; --i) {
896 struct open_bucket *ob = c->open_buckets + c->open_buckets_partial[i];
898 if (want_bucket(c, wp, devs_may_alloc, have_cache, ec, ob)) {
899 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
900 struct bch_dev_usage usage;
903 bch2_dev_usage_read_fast(ca, &usage);
904 avail = dev_buckets_free(ca, usage, watermark);
908 array_remove_item(c->open_buckets_partial,
909 c->open_buckets_partial_nr,
911 ob->on_partial_list = false;
913 ret = add_new_bucket(c, ptrs, devs_may_alloc,
914 nr_replicas, nr_effective,
915 have_cache, flags, ob);
921 spin_unlock(&c->freelist_lock);
925 static int __open_bucket_add_buckets(struct btree_trans *trans,
926 struct open_buckets *ptrs,
927 struct write_point *wp,
928 struct bch_devs_list *devs_have,
931 unsigned nr_replicas,
932 unsigned *nr_effective,
934 enum bch_watermark watermark,
938 struct bch_fs *c = trans->c;
939 struct bch_devs_mask devs;
940 struct open_bucket *ob;
941 struct closure *cl = NULL;
945 devs = target_rw_devs(c, wp->data_type, target);
947 /* Don't allocate from devices we already have pointers to: */
948 for (i = 0; i < devs_have->nr; i++)
949 __clear_bit(devs_have->devs[i], devs.d);
951 open_bucket_for_each(c, ptrs, ob, i)
952 __clear_bit(ob->dev, devs.d);
954 if (erasure_code && ec_open_bucket(c, ptrs))
957 ret = bucket_alloc_set_writepoint(c, ptrs, wp, &devs,
958 nr_replicas, nr_effective,
959 have_cache, erasure_code, flags);
963 ret = bucket_alloc_set_partial(c, ptrs, wp, &devs,
964 nr_replicas, nr_effective,
965 have_cache, erasure_code, watermark, flags);
970 ret = bucket_alloc_from_stripe(trans, ptrs, wp, &devs,
972 nr_replicas, nr_effective,
974 watermark, flags, _cl);
978 * Try nonblocking first, so that if one device is full we'll try from
981 ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
982 nr_replicas, nr_effective, have_cache,
983 flags, wp->data_type, watermark, cl);
985 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
986 !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
997 static int open_bucket_add_buckets(struct btree_trans *trans,
998 struct open_buckets *ptrs,
999 struct write_point *wp,
1000 struct bch_devs_list *devs_have,
1002 unsigned erasure_code,
1003 unsigned nr_replicas,
1004 unsigned *nr_effective,
1006 enum bch_watermark watermark,
1013 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1014 devs_have, target, erasure_code,
1015 nr_replicas, nr_effective, have_cache,
1016 watermark, flags, cl);
1017 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1018 bch2_err_matches(ret, BCH_ERR_operation_blocked) ||
1019 bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
1020 bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1022 if (*nr_effective >= nr_replicas)
1026 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1027 devs_have, target, false,
1028 nr_replicas, nr_effective, have_cache,
1029 watermark, flags, cl);
1030 return ret < 0 ? ret : 0;
1033 static bool should_drop_bucket(struct open_bucket *ob, struct bch_fs *c,
1034 struct bch_dev *ca, bool ec)
1037 return ob->ec != NULL;
1039 bool drop = ob->dev == ca->dev_idx;
1040 struct open_bucket *ob2;
1043 if (!drop && ob->ec) {
1044 mutex_lock(&ob->ec->lock);
1045 for (i = 0; i < ob->ec->new_stripe.key.v.nr_blocks; i++) {
1046 if (!ob->ec->blocks[i])
1049 ob2 = c->open_buckets + ob->ec->blocks[i];
1050 drop |= ob2->dev == ca->dev_idx;
1052 mutex_unlock(&ob->ec->lock);
1061 static void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
1062 bool ec, struct write_point *wp)
1064 struct open_buckets ptrs = { .nr = 0 };
1065 struct open_bucket *ob;
1068 mutex_lock(&wp->lock);
1069 open_bucket_for_each(c, &wp->ptrs, ob, i)
1070 if (should_drop_bucket(ob, c, ca, ec))
1071 bch2_open_bucket_put(c, ob);
1073 ob_push(c, &ptrs, ob);
1075 mutex_unlock(&wp->lock);
1078 void bch2_open_buckets_stop(struct bch_fs *c, struct bch_dev *ca,
1083 /* Next, close write points that point to this device... */
1084 for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1085 bch2_writepoint_stop(c, ca, ec, &c->write_points[i]);
1087 bch2_writepoint_stop(c, ca, ec, &c->copygc_write_point);
1088 bch2_writepoint_stop(c, ca, ec, &c->rebalance_write_point);
1089 bch2_writepoint_stop(c, ca, ec, &c->btree_write_point);
1091 mutex_lock(&c->btree_reserve_cache_lock);
1092 while (c->btree_reserve_cache_nr) {
1093 struct btree_alloc *a =
1094 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1096 bch2_open_buckets_put(c, &a->ob);
1098 mutex_unlock(&c->btree_reserve_cache_lock);
1100 spin_lock(&c->freelist_lock);
1102 while (i < c->open_buckets_partial_nr) {
1103 struct open_bucket *ob =
1104 c->open_buckets + c->open_buckets_partial[i];
1106 if (should_drop_bucket(ob, c, ca, ec)) {
1107 --c->open_buckets_partial_nr;
1108 swap(c->open_buckets_partial[i],
1109 c->open_buckets_partial[c->open_buckets_partial_nr]);
1110 ob->on_partial_list = false;
1111 spin_unlock(&c->freelist_lock);
1112 bch2_open_bucket_put(c, ob);
1113 spin_lock(&c->freelist_lock);
1118 spin_unlock(&c->freelist_lock);
1120 bch2_ec_stop_dev(c, ca);
1123 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
1124 unsigned long write_point)
1127 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
1129 return &c->write_points_hash[hash];
1132 static struct write_point *__writepoint_find(struct hlist_head *head,
1133 unsigned long write_point)
1135 struct write_point *wp;
1138 hlist_for_each_entry_rcu(wp, head, node)
1139 if (wp->write_point == write_point)
1147 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1149 u64 stranded = c->write_points_nr * c->bucket_size_max;
1150 u64 free = bch2_fs_usage_read_short(c).free;
1152 return stranded * factor > free;
1155 static bool try_increase_writepoints(struct bch_fs *c)
1157 struct write_point *wp;
1159 if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1160 too_many_writepoints(c, 32))
1163 wp = c->write_points + c->write_points_nr++;
1164 hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1168 static bool try_decrease_writepoints(struct btree_trans *trans, unsigned old_nr)
1170 struct bch_fs *c = trans->c;
1171 struct write_point *wp;
1172 struct open_bucket *ob;
1175 mutex_lock(&c->write_points_hash_lock);
1176 if (c->write_points_nr < old_nr) {
1177 mutex_unlock(&c->write_points_hash_lock);
1181 if (c->write_points_nr == 1 ||
1182 !too_many_writepoints(c, 8)) {
1183 mutex_unlock(&c->write_points_hash_lock);
1187 wp = c->write_points + --c->write_points_nr;
1189 hlist_del_rcu(&wp->node);
1190 mutex_unlock(&c->write_points_hash_lock);
1192 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1193 open_bucket_for_each(c, &wp->ptrs, ob, i)
1194 open_bucket_free_unused(c, ob);
1195 mutex_unlock(&wp->lock);
1199 static struct write_point *writepoint_find(struct btree_trans *trans,
1200 unsigned long write_point)
1202 struct bch_fs *c = trans->c;
1203 struct write_point *wp, *oldest;
1204 struct hlist_head *head;
1206 if (!(write_point & 1UL)) {
1207 wp = (struct write_point *) write_point;
1208 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1212 head = writepoint_hash(c, write_point);
1214 wp = __writepoint_find(head, write_point);
1217 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1218 if (wp->write_point == write_point)
1220 mutex_unlock(&wp->lock);
1223 restart_find_oldest:
1225 for (wp = c->write_points;
1226 wp < c->write_points + c->write_points_nr; wp++)
1227 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1230 bch2_trans_mutex_lock_norelock(trans, &oldest->lock);
1231 bch2_trans_mutex_lock_norelock(trans, &c->write_points_hash_lock);
1232 if (oldest >= c->write_points + c->write_points_nr ||
1233 try_increase_writepoints(c)) {
1234 mutex_unlock(&c->write_points_hash_lock);
1235 mutex_unlock(&oldest->lock);
1236 goto restart_find_oldest;
1239 wp = __writepoint_find(head, write_point);
1240 if (wp && wp != oldest) {
1241 mutex_unlock(&c->write_points_hash_lock);
1242 mutex_unlock(&oldest->lock);
1247 hlist_del_rcu(&wp->node);
1248 wp->write_point = write_point;
1249 hlist_add_head_rcu(&wp->node, head);
1250 mutex_unlock(&c->write_points_hash_lock);
1252 wp->last_used = local_clock();
1257 * Get us an open_bucket we can allocate from, return with it locked:
1259 int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1261 unsigned erasure_code,
1262 struct write_point_specifier write_point,
1263 struct bch_devs_list *devs_have,
1264 unsigned nr_replicas,
1265 unsigned nr_replicas_required,
1266 enum bch_watermark watermark,
1269 struct write_point **wp_ret)
1271 struct bch_fs *c = trans->c;
1272 struct write_point *wp;
1273 struct open_bucket *ob;
1274 struct open_buckets ptrs;
1275 unsigned nr_effective, write_points_nr;
1280 BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
1282 BUG_ON(!nr_replicas || !nr_replicas_required);
1286 write_points_nr = c->write_points_nr;
1289 *wp_ret = wp = writepoint_find(trans, write_point.v);
1291 /* metadata may not allocate on cache devices: */
1292 if (wp->data_type != BCH_DATA_user)
1295 if (target && !(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1296 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1297 target, erasure_code,
1298 nr_replicas, &nr_effective,
1299 &have_cache, watermark,
1302 bch2_err_matches(ret, BCH_ERR_transaction_restart))
1305 /* Don't retry from all devices if we're out of open buckets: */
1306 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1307 goto allocate_blocking;
1310 * Only try to allocate cache (durability = 0 devices) from the
1315 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1317 nr_replicas, &nr_effective,
1318 &have_cache, watermark,
1322 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1323 target, erasure_code,
1324 nr_replicas, &nr_effective,
1325 &have_cache, watermark,
1329 BUG_ON(!ret && nr_effective < nr_replicas);
1331 if (erasure_code && !ec_open_bucket(c, &ptrs))
1332 pr_debug("failed to get ec bucket: ret %u", ret);
1334 if (ret == -BCH_ERR_insufficient_devices &&
1335 nr_effective >= nr_replicas_required)
1341 /* Free buckets we didn't use: */
1342 open_bucket_for_each(c, &wp->ptrs, ob, i)
1343 open_bucket_free_unused(c, ob);
1347 wp->sectors_free = UINT_MAX;
1349 open_bucket_for_each(c, &wp->ptrs, ob, i)
1350 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1352 BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1356 open_bucket_for_each(c, &wp->ptrs, ob, i)
1357 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1358 ob_push(c, &ptrs, ob);
1360 open_bucket_free_unused(c, ob);
1363 mutex_unlock(&wp->lock);
1365 if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1366 try_decrease_writepoints(trans, write_points_nr))
1369 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1370 bch2_err_matches(ret, BCH_ERR_freelist_empty))
1372 ? -BCH_ERR_bucket_alloc_blocked
1373 : -BCH_ERR_ENOSPC_bucket_alloc;
1378 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1380 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1382 return (struct bch_extent_ptr) {
1383 .type = 1 << BCH_EXTENT_ENTRY_ptr,
1386 .offset = bucket_to_sector(ca, ob->bucket) +
1387 ca->mi.bucket_size -
1392 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1393 struct bkey_i *k, unsigned sectors,
1396 bch2_alloc_sectors_append_ptrs_inlined(c, wp, k, sectors, cached);
1400 * Append pointers to the space we just allocated to @k, and mark @sectors space
1401 * as allocated out of @ob
1403 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1405 bch2_alloc_sectors_done_inlined(c, wp);
1408 static inline void writepoint_init(struct write_point *wp,
1409 enum bch_data_type type)
1411 mutex_init(&wp->lock);
1412 wp->data_type = type;
1414 INIT_WORK(&wp->index_update_work, bch2_write_point_do_index_updates);
1415 INIT_LIST_HEAD(&wp->writes);
1416 spin_lock_init(&wp->writes_lock);
1419 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1421 struct open_bucket *ob;
1422 struct write_point *wp;
1424 mutex_init(&c->write_points_hash_lock);
1425 c->write_points_nr = ARRAY_SIZE(c->write_points);
1427 /* open bucket 0 is a sentinal NULL: */
1428 spin_lock_init(&c->open_buckets[0].lock);
1430 for (ob = c->open_buckets + 1;
1431 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1432 spin_lock_init(&ob->lock);
1433 c->open_buckets_nr_free++;
1435 ob->freelist = c->open_buckets_freelist;
1436 c->open_buckets_freelist = ob - c->open_buckets;
1439 writepoint_init(&c->btree_write_point, BCH_DATA_btree);
1440 writepoint_init(&c->rebalance_write_point, BCH_DATA_user);
1441 writepoint_init(&c->copygc_write_point, BCH_DATA_user);
1443 for (wp = c->write_points;
1444 wp < c->write_points + c->write_points_nr; wp++) {
1445 writepoint_init(wp, BCH_DATA_user);
1447 wp->last_used = local_clock();
1448 wp->write_point = (unsigned long) wp;
1449 hlist_add_head_rcu(&wp->node,
1450 writepoint_hash(c, wp->write_point));
1454 static void bch2_open_bucket_to_text(struct printbuf *out, struct bch_fs *c, struct open_bucket *ob)
1456 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1457 unsigned data_type = ob->data_type;
1458 barrier(); /* READ_ONCE() doesn't work on bitfields */
1460 prt_printf(out, "%zu ref %u %s %u:%llu gen %u allocated %u/%u",
1461 ob - c->open_buckets,
1462 atomic_read(&ob->pin),
1463 data_type < BCH_DATA_NR ? bch2_data_types[data_type] : "invalid data type",
1464 ob->dev, ob->bucket, ob->gen,
1465 ca->mi.bucket_size - ob->sectors_free, ca->mi.bucket_size);
1467 prt_printf(out, " ec idx %llu", ob->ec->idx);
1468 if (ob->on_partial_list)
1469 prt_str(out, " partial");
1473 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1475 struct open_bucket *ob;
1479 for (ob = c->open_buckets;
1480 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1482 spin_lock(&ob->lock);
1483 if (ob->valid && !ob->on_partial_list)
1484 bch2_open_bucket_to_text(out, c, ob);
1485 spin_unlock(&ob->lock);
1491 void bch2_open_buckets_partial_to_text(struct printbuf *out, struct bch_fs *c)
1496 spin_lock(&c->freelist_lock);
1498 for (i = 0; i < c->open_buckets_partial_nr; i++)
1499 bch2_open_bucket_to_text(out, c,
1500 c->open_buckets + c->open_buckets_partial[i]);
1502 spin_unlock(&c->freelist_lock);
1506 static const char * const bch2_write_point_states[] = {
1508 WRITE_POINT_STATES()
1513 void bch2_write_points_to_text(struct printbuf *out, struct bch_fs *c)
1515 struct write_point *wp;
1518 for (wp = c->write_points;
1519 wp < c->write_points + ARRAY_SIZE(c->write_points);
1521 prt_printf(out, "%lu: ", wp->write_point);
1522 prt_human_readable_u64(out, wp->sectors_allocated);
1524 prt_printf(out, " last wrote: ");
1525 bch2_pr_time_units(out, sched_clock() - wp->last_used);
1527 for (i = 0; i < WRITE_POINT_STATE_NR; i++) {
1528 prt_printf(out, " %s: ", bch2_write_point_states[i]);
1529 bch2_pr_time_units(out, wp->time[i]);