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, citer;
403 struct bkey_s_c k, ck;
404 struct open_bucket *ob = NULL;
405 u64 first_bucket = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
406 u64 alloc_start = max(first_bucket, READ_ONCE(ca->alloc_cursor));
407 u64 alloc_cursor = alloc_start;
411 * Scan with an uncached iterator to avoid polluting the key cache. An
412 * uncached iter will return a cached key if one exists, but if not
413 * there is no other underlying protection for the associated key cache
414 * slot. To avoid racing bucket allocations, look up the cached key slot
415 * of any likely allocation candidate before attempting to proceed with
416 * the allocation. This provides proper exclusion on the associated
420 for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, alloc_cursor),
421 BTREE_ITER_SLOTS, k, ret) {
422 struct bch_alloc_v4 a_convert;
423 const struct bch_alloc_v4 *a;
425 if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
428 if (ca->new_fs_bucket_idx &&
429 is_superblock_bucket(ca, k.k->p.offset))
432 a = bch2_alloc_to_v4(k, &a_convert);
433 if (a->data_type != BCH_DATA_free)
436 /* now check the cached key to serialize concurrent allocs of the bucket */
437 ck = bch2_bkey_get_iter(trans, &citer, BTREE_ID_alloc, k.k->p, BTREE_ITER_CACHED);
442 a = bch2_alloc_to_v4(ck, &a_convert);
443 if (a->data_type != BCH_DATA_free)
448 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, watermark, a, s, cl);
450 citer.path->preserve = false;
451 bch2_trans_iter_exit(trans, &citer);
455 bch2_trans_iter_exit(trans, &iter);
457 alloc_cursor = iter.pos.offset;
458 ca->alloc_cursor = alloc_cursor;
463 if (!ob && alloc_start > first_bucket) {
464 alloc_cursor = alloc_start = first_bucket;
471 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
473 enum bch_watermark watermark,
474 struct bucket_alloc_state *s,
477 struct btree_iter iter;
479 struct open_bucket *ob = NULL;
480 u64 alloc_start = max_t(u64, ca->mi.first_bucket, READ_ONCE(ca->alloc_cursor));
481 u64 alloc_cursor = alloc_start;
484 BUG_ON(ca->new_fs_bucket_idx);
486 for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
487 POS(ca->dev_idx, alloc_cursor), 0, k, ret) {
488 if (k.k->p.inode != ca->dev_idx)
491 for (alloc_cursor = max(alloc_cursor, bkey_start_offset(k.k));
492 alloc_cursor < k.k->p.offset;
494 ret = btree_trans_too_many_iters(trans);
502 ob = try_alloc_bucket(trans, ca, watermark,
503 alloc_cursor, s, k, cl);
505 iter.path->preserve = false;
513 bch2_trans_iter_exit(trans, &iter);
515 ca->alloc_cursor = alloc_cursor;
520 if (!ob && alloc_start > ca->mi.first_bucket) {
521 alloc_cursor = alloc_start = ca->mi.first_bucket;
529 * bch2_bucket_alloc_trans - allocate a single bucket from a specific device
530 * @trans: transaction object
531 * @ca: device to allocate from
532 * @watermark: how important is this allocation?
533 * @cl: if not NULL, closure to be used to wait if buckets not available
534 * @usage: for secondarily also returning the current device usage
536 * Returns: an open_bucket on success, or an ERR_PTR() on failure.
538 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
540 enum bch_watermark watermark,
542 struct bch_dev_usage *usage)
544 struct bch_fs *c = trans->c;
545 struct open_bucket *ob = NULL;
546 bool freespace = READ_ONCE(ca->mi.freespace_initialized);
548 struct bucket_alloc_state s = { 0 };
549 bool waiting = false;
551 bch2_dev_usage_read_fast(ca, usage);
552 avail = dev_buckets_free(ca, *usage, watermark);
554 if (usage->d[BCH_DATA_need_discard].buckets > avail)
557 if (usage->d[BCH_DATA_need_gc_gens].buckets > avail)
560 if (should_invalidate_buckets(ca, *usage))
561 bch2_do_invalidates(c);
564 if (cl && !waiting) {
565 closure_wait(&c->freelist_wait, cl);
570 if (!c->blocked_allocate)
571 c->blocked_allocate = local_clock();
573 ob = ERR_PTR(-BCH_ERR_freelist_empty);
578 closure_wake_up(&c->freelist_wait);
580 ob = likely(freespace)
581 ? bch2_bucket_alloc_freelist(trans, ca, watermark, &s, cl)
582 : bch2_bucket_alloc_early(trans, ca, watermark, &s, cl);
584 if (s.skipped_need_journal_commit * 2 > avail)
585 bch2_journal_flush_async(&c->journal, NULL);
587 if (!ob && freespace && c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
593 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
596 trace_and_count(c, bucket_alloc, ca,
597 bch2_watermarks[watermark],
599 usage->d[BCH_DATA_free].buckets,
601 bch2_copygc_wait_amount(c),
602 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
606 else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
607 trace_and_count(c, bucket_alloc_fail, ca,
608 bch2_watermarks[watermark],
610 usage->d[BCH_DATA_free].buckets,
612 bch2_copygc_wait_amount(c),
613 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
616 bch2_err_str(PTR_ERR(ob)));
621 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
622 enum bch_watermark watermark,
625 struct bch_dev_usage usage;
626 struct open_bucket *ob;
628 bch2_trans_do(c, NULL, NULL, 0,
629 PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(trans, ca, watermark,
634 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
635 unsigned l, unsigned r)
637 return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
638 (stripe->next_alloc[l] < stripe->next_alloc[r]));
641 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
643 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
644 struct dev_stripe_state *stripe,
645 struct bch_devs_mask *devs)
647 struct dev_alloc_list ret = { .nr = 0 };
650 for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
651 ret.devs[ret.nr++] = i;
653 bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
657 static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
658 struct dev_stripe_state *stripe,
659 struct bch_dev_usage *usage)
661 u64 *v = stripe->next_alloc + ca->dev_idx;
662 u64 free_space = dev_buckets_available(ca, BCH_WATERMARK_normal);
663 u64 free_space_inv = free_space
664 ? div64_u64(1ULL << 48, free_space)
668 if (*v + free_space_inv >= *v)
669 *v += free_space_inv;
673 for (v = stripe->next_alloc;
674 v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
675 *v = *v < scale ? 0 : *v - scale;
678 void bch2_dev_stripe_increment(struct bch_dev *ca,
679 struct dev_stripe_state *stripe)
681 struct bch_dev_usage usage;
683 bch2_dev_usage_read_fast(ca, &usage);
684 bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
687 static int add_new_bucket(struct bch_fs *c,
688 struct open_buckets *ptrs,
689 struct bch_devs_mask *devs_may_alloc,
690 unsigned nr_replicas,
691 unsigned *nr_effective,
694 struct open_bucket *ob)
696 unsigned durability =
697 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
699 BUG_ON(*nr_effective >= nr_replicas);
700 BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
702 __clear_bit(ob->dev, devs_may_alloc->d);
703 *nr_effective += (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)
705 *have_cache |= !durability;
707 ob_push(c, ptrs, ob);
709 if (*nr_effective >= nr_replicas)
716 int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
717 struct open_buckets *ptrs,
718 struct dev_stripe_state *stripe,
719 struct bch_devs_mask *devs_may_alloc,
720 unsigned nr_replicas,
721 unsigned *nr_effective,
724 enum bch_data_type data_type,
725 enum bch_watermark watermark,
728 struct bch_fs *c = trans->c;
729 struct dev_alloc_list devs_sorted =
730 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
733 int ret = -BCH_ERR_insufficient_devices;
736 BUG_ON(*nr_effective >= nr_replicas);
738 for (i = 0; i < devs_sorted.nr; i++) {
739 struct bch_dev_usage usage;
740 struct open_bucket *ob;
742 dev = devs_sorted.devs[i];
745 ca = rcu_dereference(c->devs[dev]);
747 percpu_ref_get(&ca->ref);
753 if (!ca->mi.durability && *have_cache) {
754 percpu_ref_put(&ca->ref);
758 ob = bch2_bucket_alloc_trans(trans, ca, watermark, cl, &usage);
760 bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
761 percpu_ref_put(&ca->ref);
765 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
770 ob->data_type = data_type;
772 if (add_new_bucket(c, ptrs, devs_may_alloc,
773 nr_replicas, nr_effective,
774 have_cache, flags, ob)) {
783 /* Allocate from stripes: */
786 * if we can't allocate a new stripe because there are already too many
787 * partially filled stripes, force allocating from an existing stripe even when
788 * it's to a device we don't want:
791 static int bucket_alloc_from_stripe(struct btree_trans *trans,
792 struct open_buckets *ptrs,
793 struct write_point *wp,
794 struct bch_devs_mask *devs_may_alloc,
796 unsigned nr_replicas,
797 unsigned *nr_effective,
799 enum bch_watermark watermark,
803 struct bch_fs *c = trans->c;
804 struct dev_alloc_list devs_sorted;
805 struct ec_stripe_head *h;
806 struct open_bucket *ob;
813 if (ec_open_bucket(c, ptrs))
816 h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, watermark, cl);
822 devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
824 for (i = 0; i < devs_sorted.nr; i++)
825 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
826 if (!h->s->blocks[ec_idx])
829 ob = c->open_buckets + h->s->blocks[ec_idx];
830 if (ob->dev == devs_sorted.devs[i] &&
831 !test_and_set_bit(ec_idx, h->s->blocks_allocated))
838 ec_stripe_new_get(h->s, STRIPE_REF_io);
840 ret = add_new_bucket(c, ptrs, devs_may_alloc,
841 nr_replicas, nr_effective,
842 have_cache, flags, ob);
844 bch2_ec_stripe_head_put(c, h);
848 /* Sector allocator */
850 static bool want_bucket(struct bch_fs *c,
851 struct write_point *wp,
852 struct bch_devs_mask *devs_may_alloc,
853 bool *have_cache, bool ec,
854 struct open_bucket *ob)
856 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
858 if (!test_bit(ob->dev, devs_may_alloc->d))
861 if (ob->data_type != wp->data_type)
864 if (!ca->mi.durability &&
865 (wp->data_type == BCH_DATA_btree || ec || *have_cache))
868 if (ec != (ob->ec != NULL))
874 static int bucket_alloc_set_writepoint(struct bch_fs *c,
875 struct open_buckets *ptrs,
876 struct write_point *wp,
877 struct bch_devs_mask *devs_may_alloc,
878 unsigned nr_replicas,
879 unsigned *nr_effective,
881 bool ec, unsigned flags)
883 struct open_buckets ptrs_skip = { .nr = 0 };
884 struct open_bucket *ob;
888 open_bucket_for_each(c, &wp->ptrs, ob, i) {
889 if (!ret && want_bucket(c, wp, devs_may_alloc,
891 ret = add_new_bucket(c, ptrs, devs_may_alloc,
892 nr_replicas, nr_effective,
893 have_cache, flags, ob);
895 ob_push(c, &ptrs_skip, ob);
897 wp->ptrs = ptrs_skip;
902 static int bucket_alloc_set_partial(struct bch_fs *c,
903 struct open_buckets *ptrs,
904 struct write_point *wp,
905 struct bch_devs_mask *devs_may_alloc,
906 unsigned nr_replicas,
907 unsigned *nr_effective,
908 bool *have_cache, bool ec,
909 enum bch_watermark watermark,
914 if (!c->open_buckets_partial_nr)
917 spin_lock(&c->freelist_lock);
919 if (!c->open_buckets_partial_nr)
922 for (i = c->open_buckets_partial_nr - 1; i >= 0; --i) {
923 struct open_bucket *ob = c->open_buckets + c->open_buckets_partial[i];
925 if (want_bucket(c, wp, devs_may_alloc, have_cache, ec, ob)) {
926 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
927 struct bch_dev_usage usage;
930 bch2_dev_usage_read_fast(ca, &usage);
931 avail = dev_buckets_free(ca, usage, watermark);
935 array_remove_item(c->open_buckets_partial,
936 c->open_buckets_partial_nr,
938 ob->on_partial_list = false;
940 ret = add_new_bucket(c, ptrs, devs_may_alloc,
941 nr_replicas, nr_effective,
942 have_cache, flags, ob);
948 spin_unlock(&c->freelist_lock);
952 static int __open_bucket_add_buckets(struct btree_trans *trans,
953 struct open_buckets *ptrs,
954 struct write_point *wp,
955 struct bch_devs_list *devs_have,
958 unsigned nr_replicas,
959 unsigned *nr_effective,
961 enum bch_watermark watermark,
965 struct bch_fs *c = trans->c;
966 struct bch_devs_mask devs;
967 struct open_bucket *ob;
968 struct closure *cl = NULL;
972 devs = target_rw_devs(c, wp->data_type, target);
974 /* Don't allocate from devices we already have pointers to: */
975 for (i = 0; i < devs_have->nr; i++)
976 __clear_bit(devs_have->devs[i], devs.d);
978 open_bucket_for_each(c, ptrs, ob, i)
979 __clear_bit(ob->dev, devs.d);
981 if (erasure_code && ec_open_bucket(c, ptrs))
984 ret = bucket_alloc_set_writepoint(c, ptrs, wp, &devs,
985 nr_replicas, nr_effective,
986 have_cache, erasure_code, flags);
990 ret = bucket_alloc_set_partial(c, ptrs, wp, &devs,
991 nr_replicas, nr_effective,
992 have_cache, erasure_code, watermark, flags);
997 ret = bucket_alloc_from_stripe(trans, ptrs, wp, &devs,
999 nr_replicas, nr_effective,
1001 watermark, flags, _cl);
1005 * Try nonblocking first, so that if one device is full we'll try from
1008 ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
1009 nr_replicas, nr_effective, have_cache,
1010 flags, wp->data_type, watermark, cl);
1012 !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
1013 !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
1016 goto retry_blocking;
1023 static int open_bucket_add_buckets(struct btree_trans *trans,
1024 struct open_buckets *ptrs,
1025 struct write_point *wp,
1026 struct bch_devs_list *devs_have,
1028 unsigned erasure_code,
1029 unsigned nr_replicas,
1030 unsigned *nr_effective,
1032 enum bch_watermark watermark,
1039 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1040 devs_have, target, erasure_code,
1041 nr_replicas, nr_effective, have_cache,
1042 watermark, flags, cl);
1043 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1044 bch2_err_matches(ret, BCH_ERR_operation_blocked) ||
1045 bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
1046 bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1048 if (*nr_effective >= nr_replicas)
1052 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1053 devs_have, target, false,
1054 nr_replicas, nr_effective, have_cache,
1055 watermark, flags, cl);
1056 return ret < 0 ? ret : 0;
1060 * should_drop_bucket - check if this is open_bucket should go away
1061 * @ob: open_bucket to predicate on
1062 * @c: filesystem handle
1063 * @ca: if set, we're killing buckets for a particular device
1064 * @ec: if true, we're shutting down erasure coding and killing all ec
1066 * otherwise, return true
1067 * Returns: true if we should kill this open_bucket
1069 * We're killing open_buckets because we're shutting down a device, erasure
1070 * coding, or the entire filesystem - check if this open_bucket matches:
1072 static bool should_drop_bucket(struct open_bucket *ob, struct bch_fs *c,
1073 struct bch_dev *ca, bool ec)
1076 return ob->ec != NULL;
1078 bool drop = ob->dev == ca->dev_idx;
1079 struct open_bucket *ob2;
1082 if (!drop && ob->ec) {
1085 mutex_lock(&ob->ec->lock);
1086 nr_blocks = bkey_i_to_stripe(&ob->ec->new_stripe.key)->v.nr_blocks;
1088 for (i = 0; i < nr_blocks; i++) {
1089 if (!ob->ec->blocks[i])
1092 ob2 = c->open_buckets + ob->ec->blocks[i];
1093 drop |= ob2->dev == ca->dev_idx;
1095 mutex_unlock(&ob->ec->lock);
1104 static void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
1105 bool ec, struct write_point *wp)
1107 struct open_buckets ptrs = { .nr = 0 };
1108 struct open_bucket *ob;
1111 mutex_lock(&wp->lock);
1112 open_bucket_for_each(c, &wp->ptrs, ob, i)
1113 if (should_drop_bucket(ob, c, ca, ec))
1114 bch2_open_bucket_put(c, ob);
1116 ob_push(c, &ptrs, ob);
1118 mutex_unlock(&wp->lock);
1121 void bch2_open_buckets_stop(struct bch_fs *c, struct bch_dev *ca,
1126 /* Next, close write points that point to this device... */
1127 for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1128 bch2_writepoint_stop(c, ca, ec, &c->write_points[i]);
1130 bch2_writepoint_stop(c, ca, ec, &c->copygc_write_point);
1131 bch2_writepoint_stop(c, ca, ec, &c->rebalance_write_point);
1132 bch2_writepoint_stop(c, ca, ec, &c->btree_write_point);
1134 mutex_lock(&c->btree_reserve_cache_lock);
1135 while (c->btree_reserve_cache_nr) {
1136 struct btree_alloc *a =
1137 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1139 bch2_open_buckets_put(c, &a->ob);
1141 mutex_unlock(&c->btree_reserve_cache_lock);
1143 spin_lock(&c->freelist_lock);
1145 while (i < c->open_buckets_partial_nr) {
1146 struct open_bucket *ob =
1147 c->open_buckets + c->open_buckets_partial[i];
1149 if (should_drop_bucket(ob, c, ca, ec)) {
1150 --c->open_buckets_partial_nr;
1151 swap(c->open_buckets_partial[i],
1152 c->open_buckets_partial[c->open_buckets_partial_nr]);
1153 ob->on_partial_list = false;
1154 spin_unlock(&c->freelist_lock);
1155 bch2_open_bucket_put(c, ob);
1156 spin_lock(&c->freelist_lock);
1161 spin_unlock(&c->freelist_lock);
1163 bch2_ec_stop_dev(c, ca);
1166 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
1167 unsigned long write_point)
1170 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
1172 return &c->write_points_hash[hash];
1175 static struct write_point *__writepoint_find(struct hlist_head *head,
1176 unsigned long write_point)
1178 struct write_point *wp;
1181 hlist_for_each_entry_rcu(wp, head, node)
1182 if (wp->write_point == write_point)
1190 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1192 u64 stranded = c->write_points_nr * c->bucket_size_max;
1193 u64 free = bch2_fs_usage_read_short(c).free;
1195 return stranded * factor > free;
1198 static bool try_increase_writepoints(struct bch_fs *c)
1200 struct write_point *wp;
1202 if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1203 too_many_writepoints(c, 32))
1206 wp = c->write_points + c->write_points_nr++;
1207 hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1211 static bool try_decrease_writepoints(struct btree_trans *trans, unsigned old_nr)
1213 struct bch_fs *c = trans->c;
1214 struct write_point *wp;
1215 struct open_bucket *ob;
1218 mutex_lock(&c->write_points_hash_lock);
1219 if (c->write_points_nr < old_nr) {
1220 mutex_unlock(&c->write_points_hash_lock);
1224 if (c->write_points_nr == 1 ||
1225 !too_many_writepoints(c, 8)) {
1226 mutex_unlock(&c->write_points_hash_lock);
1230 wp = c->write_points + --c->write_points_nr;
1232 hlist_del_rcu(&wp->node);
1233 mutex_unlock(&c->write_points_hash_lock);
1235 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1236 open_bucket_for_each(c, &wp->ptrs, ob, i)
1237 open_bucket_free_unused(c, ob);
1239 mutex_unlock(&wp->lock);
1243 static struct write_point *writepoint_find(struct btree_trans *trans,
1244 unsigned long write_point)
1246 struct bch_fs *c = trans->c;
1247 struct write_point *wp, *oldest;
1248 struct hlist_head *head;
1250 if (!(write_point & 1UL)) {
1251 wp = (struct write_point *) write_point;
1252 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1256 head = writepoint_hash(c, write_point);
1258 wp = __writepoint_find(head, write_point);
1261 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1262 if (wp->write_point == write_point)
1264 mutex_unlock(&wp->lock);
1267 restart_find_oldest:
1269 for (wp = c->write_points;
1270 wp < c->write_points + c->write_points_nr; wp++)
1271 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1274 bch2_trans_mutex_lock_norelock(trans, &oldest->lock);
1275 bch2_trans_mutex_lock_norelock(trans, &c->write_points_hash_lock);
1276 if (oldest >= c->write_points + c->write_points_nr ||
1277 try_increase_writepoints(c)) {
1278 mutex_unlock(&c->write_points_hash_lock);
1279 mutex_unlock(&oldest->lock);
1280 goto restart_find_oldest;
1283 wp = __writepoint_find(head, write_point);
1284 if (wp && wp != oldest) {
1285 mutex_unlock(&c->write_points_hash_lock);
1286 mutex_unlock(&oldest->lock);
1291 hlist_del_rcu(&wp->node);
1292 wp->write_point = write_point;
1293 hlist_add_head_rcu(&wp->node, head);
1294 mutex_unlock(&c->write_points_hash_lock);
1296 wp->last_used = local_clock();
1300 static noinline void
1301 deallocate_extra_replicas(struct bch_fs *c,
1302 struct open_buckets *ptrs,
1303 struct open_buckets *ptrs_no_use,
1304 unsigned extra_replicas)
1306 struct open_buckets ptrs2 = { 0 };
1307 struct open_bucket *ob;
1310 open_bucket_for_each(c, ptrs, ob, i) {
1311 unsigned d = bch_dev_bkey_exists(c, ob->dev)->mi.durability;
1313 if (d && d <= extra_replicas) {
1314 extra_replicas -= d;
1315 ob_push(c, ptrs_no_use, ob);
1317 ob_push(c, &ptrs2, ob);
1325 * Get us an open_bucket we can allocate from, return with it locked:
1327 int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1329 unsigned erasure_code,
1330 struct write_point_specifier write_point,
1331 struct bch_devs_list *devs_have,
1332 unsigned nr_replicas,
1333 unsigned nr_replicas_required,
1334 enum bch_watermark watermark,
1337 struct write_point **wp_ret)
1339 struct bch_fs *c = trans->c;
1340 struct write_point *wp;
1341 struct open_bucket *ob;
1342 struct open_buckets ptrs;
1343 unsigned nr_effective, write_points_nr;
1348 if (!IS_ENABLED(CONFIG_BCACHEFS_ERASURE_CODING))
1349 erasure_code = false;
1351 BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
1353 BUG_ON(!nr_replicas || !nr_replicas_required);
1357 write_points_nr = c->write_points_nr;
1360 *wp_ret = wp = writepoint_find(trans, write_point.v);
1362 /* metadata may not allocate on cache devices: */
1363 if (wp->data_type != BCH_DATA_user)
1366 if (target && !(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1367 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1368 target, erasure_code,
1369 nr_replicas, &nr_effective,
1370 &have_cache, watermark,
1373 bch2_err_matches(ret, BCH_ERR_transaction_restart))
1376 /* Don't retry from all devices if we're out of open buckets: */
1377 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1378 goto allocate_blocking;
1381 * Only try to allocate cache (durability = 0 devices) from the
1386 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1388 nr_replicas, &nr_effective,
1389 &have_cache, watermark,
1393 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1394 target, erasure_code,
1395 nr_replicas, &nr_effective,
1396 &have_cache, watermark,
1400 BUG_ON(!ret && nr_effective < nr_replicas);
1402 if (erasure_code && !ec_open_bucket(c, &ptrs))
1403 pr_debug("failed to get ec bucket: ret %u", ret);
1405 if (ret == -BCH_ERR_insufficient_devices &&
1406 nr_effective >= nr_replicas_required)
1412 if (nr_effective > nr_replicas)
1413 deallocate_extra_replicas(c, &ptrs, &wp->ptrs, nr_effective - nr_replicas);
1415 /* Free buckets we didn't use: */
1416 open_bucket_for_each(c, &wp->ptrs, ob, i)
1417 open_bucket_free_unused(c, ob);
1421 wp->sectors_free = UINT_MAX;
1423 open_bucket_for_each(c, &wp->ptrs, ob, i)
1424 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1426 BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1430 open_bucket_for_each(c, &wp->ptrs, ob, i)
1431 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1432 ob_push(c, &ptrs, ob);
1434 open_bucket_free_unused(c, ob);
1437 mutex_unlock(&wp->lock);
1439 if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1440 try_decrease_writepoints(trans, write_points_nr))
1443 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1444 bch2_err_matches(ret, BCH_ERR_freelist_empty))
1446 ? -BCH_ERR_bucket_alloc_blocked
1447 : -BCH_ERR_ENOSPC_bucket_alloc;
1452 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1454 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1456 return (struct bch_extent_ptr) {
1457 .type = 1 << BCH_EXTENT_ENTRY_ptr,
1460 .offset = bucket_to_sector(ca, ob->bucket) +
1461 ca->mi.bucket_size -
1466 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1467 struct bkey_i *k, unsigned sectors,
1470 bch2_alloc_sectors_append_ptrs_inlined(c, wp, k, sectors, cached);
1474 * Append pointers to the space we just allocated to @k, and mark @sectors space
1475 * as allocated out of @ob
1477 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1479 bch2_alloc_sectors_done_inlined(c, wp);
1482 static inline void writepoint_init(struct write_point *wp,
1483 enum bch_data_type type)
1485 mutex_init(&wp->lock);
1486 wp->data_type = type;
1488 INIT_WORK(&wp->index_update_work, bch2_write_point_do_index_updates);
1489 INIT_LIST_HEAD(&wp->writes);
1490 spin_lock_init(&wp->writes_lock);
1493 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1495 struct open_bucket *ob;
1496 struct write_point *wp;
1498 mutex_init(&c->write_points_hash_lock);
1499 c->write_points_nr = ARRAY_SIZE(c->write_points);
1501 /* open bucket 0 is a sentinal NULL: */
1502 spin_lock_init(&c->open_buckets[0].lock);
1504 for (ob = c->open_buckets + 1;
1505 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1506 spin_lock_init(&ob->lock);
1507 c->open_buckets_nr_free++;
1509 ob->freelist = c->open_buckets_freelist;
1510 c->open_buckets_freelist = ob - c->open_buckets;
1513 writepoint_init(&c->btree_write_point, BCH_DATA_btree);
1514 writepoint_init(&c->rebalance_write_point, BCH_DATA_user);
1515 writepoint_init(&c->copygc_write_point, BCH_DATA_user);
1517 for (wp = c->write_points;
1518 wp < c->write_points + c->write_points_nr; wp++) {
1519 writepoint_init(wp, BCH_DATA_user);
1521 wp->last_used = local_clock();
1522 wp->write_point = (unsigned long) wp;
1523 hlist_add_head_rcu(&wp->node,
1524 writepoint_hash(c, wp->write_point));
1528 static void bch2_open_bucket_to_text(struct printbuf *out, struct bch_fs *c, struct open_bucket *ob)
1530 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1531 unsigned data_type = ob->data_type;
1532 barrier(); /* READ_ONCE() doesn't work on bitfields */
1534 prt_printf(out, "%zu ref %u %s %u:%llu gen %u allocated %u/%u",
1535 ob - c->open_buckets,
1536 atomic_read(&ob->pin),
1537 data_type < BCH_DATA_NR ? bch2_data_types[data_type] : "invalid data type",
1538 ob->dev, ob->bucket, ob->gen,
1539 ca->mi.bucket_size - ob->sectors_free, ca->mi.bucket_size);
1541 prt_printf(out, " ec idx %llu", ob->ec->idx);
1542 if (ob->on_partial_list)
1543 prt_str(out, " partial");
1547 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1549 struct open_bucket *ob;
1553 for (ob = c->open_buckets;
1554 ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1556 spin_lock(&ob->lock);
1557 if (ob->valid && !ob->on_partial_list)
1558 bch2_open_bucket_to_text(out, c, ob);
1559 spin_unlock(&ob->lock);
1565 void bch2_open_buckets_partial_to_text(struct printbuf *out, struct bch_fs *c)
1570 spin_lock(&c->freelist_lock);
1572 for (i = 0; i < c->open_buckets_partial_nr; i++)
1573 bch2_open_bucket_to_text(out, c,
1574 c->open_buckets + c->open_buckets_partial[i]);
1576 spin_unlock(&c->freelist_lock);
1580 static const char * const bch2_write_point_states[] = {
1582 WRITE_POINT_STATES()
1587 static void bch2_write_point_to_text(struct printbuf *out, struct bch_fs *c,
1588 struct write_point *wp)
1590 struct open_bucket *ob;
1593 prt_printf(out, "%lu: ", wp->write_point);
1594 prt_human_readable_u64(out, wp->sectors_allocated);
1596 prt_printf(out, " last wrote: ");
1597 bch2_pr_time_units(out, sched_clock() - wp->last_used);
1599 for (i = 0; i < WRITE_POINT_STATE_NR; i++) {
1600 prt_printf(out, " %s: ", bch2_write_point_states[i]);
1601 bch2_pr_time_units(out, wp->time[i]);
1606 printbuf_indent_add(out, 2);
1607 open_bucket_for_each(c, &wp->ptrs, ob, i)
1608 bch2_open_bucket_to_text(out, c, ob);
1609 printbuf_indent_sub(out, 2);
1612 void bch2_write_points_to_text(struct printbuf *out, struct bch_fs *c)
1614 struct write_point *wp;
1616 prt_str(out, "Foreground write points\n");
1617 for (wp = c->write_points;
1618 wp < c->write_points + ARRAY_SIZE(c->write_points);
1620 bch2_write_point_to_text(out, c, wp);
1622 prt_str(out, "Copygc write point\n");
1623 bch2_write_point_to_text(out, c, &c->copygc_write_point);
1625 prt_str(out, "Rebalance write point\n");
1626 bch2_write_point_to_text(out, c, &c->rebalance_write_point);
1628 prt_str(out, "Btree write point\n");
1629 bch2_write_point_to_text(out, c, &c->btree_write_point);