bcachefs: New data structure for buckets waiting on journal commit

Implement a hash table, using cuckoo hashing, for empty buckets that are
waiting on a journal commit before they can be reused.

This replaces the journal_seq field of bucket_mark, and is part of
eventually getting rid of the in memory bucket array.

We may need to make bch2_bucket_needs_journal_commit() lockless, pending
profiling and testing.

Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>

12 files changed, 279 insertions, 124 deletions

diff --git a/fs/bcachefs/Makefile b/fs/bcachefs/Makefile
index a2769a85b029..65eeab56cb4b 100644
--- a/fs/bcachefs/Makefile
+++ b/fs/bcachefs/Makefile
@@ -16,6 +16,7 @@ bcachefs-y		:=	\
 	btree_update_interior.o	\
 	btree_update_leaf.o	\
 	buckets.o		\
+	buckets_waiting_for_journal.o	\
 	chardev.o		\
 	checksum.o		\
 	clock.o			\
diff --git a/fs/bcachefs/alloc_background.c b/fs/bcachefs/alloc_background.c
index ab7d972aac3a..bc5053ebe18f 100644
--- a/fs/bcachefs/alloc_background.c
+++ b/fs/bcachefs/alloc_background.c
@@ -9,6 +9,7 @@
 #include "btree_update_interior.h"
 #include "btree_gc.h"
 #include "buckets.h"
+#include "buckets_waiting_for_journal.h"
 #include "clock.h"
 #include "debug.h"
 #include "ec.h"
@@ -561,8 +562,7 @@ static unsigned bucket_sort_key(struct bucket *g, struct bucket_mark m,
 		 * keys when there's only a small difference, so that we can
 		 * keep sequential buckets together:
 		 */
-		return  (bucket_needs_journal_commit(m, last_seq_ondisk) << 4)|
-			(bucket_gc_gen(g) >> 4);
+		return bucket_gc_gen(g) >> 4;
 	}
 }
 
@@ -611,6 +611,14 @@ static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
 		if (!bch2_can_invalidate_bucket(ca, b, m))
 			continue;
 
+		if (!m.data_type &&
+		    bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
+						     last_seq_ondisk,
+						     ca->dev_idx, b)) {
+			ca->buckets_waiting_on_journal++;
+			continue;
+		}
+
 		if (e.nr && e.bucket + e.nr == b && e.key == key) {
 			e.nr++;
 		} else {
@@ -647,6 +655,7 @@ static size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca)
 
 	ca->inc_gen_needs_gc			= 0;
 	ca->inc_gen_really_needs_gc		= 0;
+	ca->buckets_waiting_on_journal		= 0;
 
 	find_reclaimable_buckets_lru(c, ca);
 
@@ -658,28 +667,6 @@ static size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca)
 	return nr;
 }
 
-/*
- * returns sequence number of most recent journal entry that updated this
- * bucket:
- */
-static u64 bucket_journal_seq(struct bch_fs *c, struct bucket_mark m)
-{
-	if (m.journal_seq_valid) {
-		u64 journal_seq = atomic64_read(&c->journal.seq);
-		u64 bucket_seq	= journal_seq;
-
-		bucket_seq &= ~((u64) U16_MAX);
-		bucket_seq |= m.journal_seq;
-
-		if (bucket_seq > journal_seq)
-			bucket_seq -= 1 << 16;
-
-		return bucket_seq;
-	} else {
-		return 0;
-	}
-}
-
 static int bucket_invalidate_btree(struct btree_trans *trans,
 				   struct bch_dev *ca, u64 b)
 {
@@ -745,9 +732,10 @@ static int bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
 	 * gen in memory here, the incremented gen will be updated in the btree
 	 * by bch2_trans_mark_pointer():
 	 */
-	if (!m.cached_sectors &&
-	    !bucket_needs_journal_commit(m, c->journal.last_seq_ondisk)) {
-		BUG_ON(m.data_type);
+	if (!m.data_type &&
+	    !bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
+					      c->journal.flushed_seq_ondisk,
+					      ca->dev_idx, b)) {
 		bucket_cmpxchg(g, m, m.gen++);
 		*bucket_gen(ca, b) = m.gen;
 		percpu_up_read(&c->mark_lock);
@@ -781,13 +769,6 @@ out:
 
 		if (!top->nr)
 			heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
-
-		/*
-		 * Make sure we flush the last journal entry that updated this
-		 * bucket (i.e. deleting the last reference) before writing to
-		 * this bucket again:
-		 */
-		*journal_seq = max(*journal_seq, bucket_journal_seq(c, m));
 	} else {
 		size_t b2;
 
@@ -954,8 +935,14 @@ static int bch2_allocator_thread(void *arg)
 			gc_count = c->gc_count;
 			nr = find_reclaimable_buckets(c, ca);
 
-			trace_alloc_scan(ca, nr, ca->inc_gen_needs_gc,
-					 ca->inc_gen_really_needs_gc);
+			if (!nr && ca->buckets_waiting_on_journal) {
+				ret = bch2_journal_flush(&c->journal);
+				if (ret)
+					goto stop;
+			} else if (nr < (ca->mi.nbuckets >> 6) &&
+				   ca->buckets_waiting_on_journal >= nr / 2) {
+				bch2_journal_flush_async(&c->journal, NULL);
+			}
 
 			if ((ca->inc_gen_needs_gc >= ALLOC_SCAN_BATCH(ca) ||
 			     ca->inc_gen_really_needs_gc) &&
@@ -963,6 +950,9 @@ static int bch2_allocator_thread(void *arg)
 				atomic_inc(&c->kick_gc);
 				wake_up_process(c->gc_thread);
 			}
+
+			trace_alloc_scan(ca, nr, ca->inc_gen_needs_gc,
+					 ca->inc_gen_really_needs_gc);
 		}
 
 		ret = bch2_invalidate_buckets(c, ca);
diff --git a/fs/bcachefs/bcachefs.h b/fs/bcachefs/bcachefs.h
index 4ebaefd408a4..3d1a6773393c 100644
--- a/fs/bcachefs/bcachefs.h
+++ b/fs/bcachefs/bcachefs.h
@@ -355,6 +355,7 @@ enum bch_time_stats {
 #include "alloc_types.h"
 #include "btree_types.h"
 #include "buckets_types.h"
+#include "buckets_waiting_for_journal_types.h"
 #include "clock_types.h"
 #include "ec_types.h"
 #include "journal_types.h"
@@ -482,6 +483,7 @@ struct bch_dev {
 
 	size_t			inc_gen_needs_gc;
 	size_t			inc_gen_really_needs_gc;
+	size_t			buckets_waiting_on_journal;
 
 	enum allocator_states	allocator_state;
 
@@ -777,6 +779,8 @@ struct bch_fs {
 	struct mutex		write_points_hash_lock;
 	unsigned		write_points_nr;
 
+	struct buckets_waiting_for_journal buckets_waiting_for_journal;
+
 	/* GARBAGE COLLECTION */
 	struct task_struct	*gc_thread;
 	atomic_t		kick_gc;
diff --git a/fs/bcachefs/btree_update_leaf.c b/fs/bcachefs/btree_update_leaf.c
index de33491f2535..24de8604740c 100644
--- a/fs/bcachefs/btree_update_leaf.c
+++ b/fs/bcachefs/btree_update_leaf.c
@@ -396,10 +396,11 @@ static inline void do_btree_insert_one(struct btree_trans *trans,
 	}
 }
 
-static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
+static noinline int bch2_trans_mark_gc(struct btree_trans *trans)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_insert_entry *i;
+	int ret = 0;
 
 	trans_for_each_update(trans, i) {
 		/*
@@ -408,10 +409,15 @@ static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 		 */
 		BUG_ON(i->cached || i->level);
 
-		if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b)))
-			bch2_mark_update(trans, i->path, i->k,
-					 i->flags|BTREE_TRIGGER_GC);
+		if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
+			ret = bch2_mark_update(trans, i->path, i->k,
+					       i->flags|BTREE_TRIGGER_GC);
+			if (ret)
+				break;
+		}
 	}
+
+	return ret;
 }
 
 static inline int
@@ -510,11 +516,17 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 		return BTREE_INSERT_NEED_MARK_REPLICAS;
 
 	trans_for_each_update(trans, i)
-		if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type))
-			bch2_mark_update(trans, i->path, i->k, i->flags);
+		if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
+			ret = bch2_mark_update(trans, i->path, i->k, i->flags);
+			if (ret)
+				return ret;
+		}
 
-	if (unlikely(c->gc_pos.phase))
-		bch2_trans_mark_gc(trans);
+	if (unlikely(c->gc_pos.phase)) {
+		ret = bch2_trans_mark_gc(trans);
+		if  (ret)
+			return ret;
+	}
 
 	trans_for_each_update(trans, i)
 		do_btree_insert_one(trans, i);
diff --git a/fs/bcachefs/buckets.c b/fs/bcachefs/buckets.c
index b80ab1ed22f7..f7a750aff03f 100644
--- a/fs/bcachefs/buckets.c
+++ b/fs/bcachefs/buckets.c
@@ -11,6 +11,7 @@
 #include "btree_gc.h"
 #include "btree_update.h"
 #include "buckets.h"
+#include "buckets_waiting_for_journal.h"
 #include "ec.h"
 #include "error.h"
 #include "inode.h"
@@ -43,43 +44,6 @@ static inline void fs_usage_data_type_to_base(struct bch_fs_usage *fs_usage,
 	}
 }
 
-/*
- * Clear journal_seq_valid for buckets for which it's not needed, to prevent
- * wraparound:
- */
-void bch2_bucket_seq_cleanup(struct bch_fs *c)
-{
-	u64 journal_seq = atomic64_read(&c->journal.seq);
-	u16 last_seq_ondisk = c->journal.flushed_seq_ondisk;
-	struct bch_dev *ca;
-	struct bucket_array *buckets;
-	struct bucket *g;
-	struct bucket_mark m;
-	unsigned i;
-
-	if (journal_seq - c->last_bucket_seq_cleanup <
-	    (1U << (BUCKET_JOURNAL_SEQ_BITS - 2)))
-		return;
-
-	c->last_bucket_seq_cleanup = journal_seq;
-
-	for_each_member_device(ca, c, i) {
-		down_read(&ca->bucket_lock);
-		buckets = bucket_array(ca);
-
-		for_each_bucket(g, buckets) {
-			bucket_cmpxchg(g, m, ({
-				if (!m.journal_seq_valid ||
-				    bucket_needs_journal_commit(m, last_seq_ondisk))
-					break;
-
-				m.journal_seq_valid = 0;
-			}));
-		}
-		up_read(&ca->bucket_lock);
-	}
-}
-
 void bch2_fs_usage_initialize(struct bch_fs *c)
 {
 	struct bch_fs_usage *usage;
@@ -576,16 +540,28 @@ static int bch2_mark_alloc(struct btree_trans *trans,
 		v->journal_seq = cpu_to_le64(new_u.journal_seq);
 	}
 
-	ca = bch_dev_bkey_exists(c, new.k->p.inode);
+	if (old_u.data_type && !new_u.data_type && new_u.journal_seq) {
+		ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
+				c->journal.flushed_seq_ondisk,
+				new_u.dev, new_u.bucket,
+				new_u.journal_seq);
+		if (ret) {
+			bch2_fs_fatal_error(c,
+				"error setting bucket_needs_journal_commit: %i", ret);
+			return ret;
+		}
+	}
+
+	ca = bch_dev_bkey_exists(c, new_u.dev);
 
-	if (new.k->p.offset >= ca->mi.nbuckets)
+	if (new_u.bucket >= ca->mi.nbuckets)
 		return 0;
 
 	percpu_down_read(&c->mark_lock);
 	if (!gc && new_u.gen != old_u.gen)
-		*bucket_gen(ca, new.k->p.offset) = new_u.gen;
+		*bucket_gen(ca, new_u.bucket) = new_u.gen;
 
-	g = __bucket(ca, new.k->p.offset, gc);
+	g = __bucket(ca, new_u.bucket, gc);
 
 	old_m = bucket_cmpxchg(g, m, ({
 		m.gen			= new_u.gen;
@@ -593,11 +569,6 @@ static int bch2_mark_alloc(struct btree_trans *trans,
 		m.dirty_sectors		= new_u.dirty_sectors;
 		m.cached_sectors	= new_u.cached_sectors;
 		m.stripe		= new_u.stripe != 0;
-
-		if (journal_seq) {
-			m.journal_seq_valid	= 1;
-			m.journal_seq		= journal_seq;
-		}
 	}));
 
 	bch2_dev_usage_update(c, ca, old_m, m, journal_seq, gc);
@@ -625,7 +596,7 @@ static int bch2_mark_alloc(struct btree_trans *trans,
 			return ret;
 		}
 
-		trace_invalidate(ca, bucket_to_sector(ca, new.k->p.offset),
+		trace_invalidate(ca, bucket_to_sector(ca, new_u.bucket),
 				 old_m.cached_sectors);
 	}
 
@@ -775,9 +746,10 @@ static int check_bucket_ref(struct bch_fs *c,
 static int mark_stripe_bucket(struct btree_trans *trans,
 			      struct bkey_s_c k,
 			      unsigned ptr_idx,
-			      u64 journal_seq, unsigned flags)
+			      unsigned flags)
 {
 	struct bch_fs *c = trans->c;
+	u64 journal_seq = trans->journal_res.seq;
 	const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 	unsigned nr_data = s->nr_blocks - s->nr_redundant;
 	bool parity = ptr_idx >= nr_data;
@@ -818,11 +790,6 @@ static int mark_stripe_bucket(struct btree_trans *trans,
 		if (data_type)
 			new.data_type		= data_type;
 
-		if (journal_seq) {
-			new.journal_seq_valid	= 1;
-			new.journal_seq		= journal_seq;
-		}
-
 		new.stripe = true;
 	}));
 
@@ -894,11 +861,6 @@ static int bch2_mark_pointer(struct btree_trans *trans,
 
 		new.data_type = bucket_data_type;
 
-		if (journal_seq) {
-			new.journal_seq_valid = 1;
-			new.journal_seq = journal_seq;
-		}
-
 		if (flags & BTREE_TRIGGER_NOATOMIC) {
 			g->_mark = new;
 			break;
@@ -1119,7 +1081,7 @@ static int bch2_mark_stripe(struct btree_trans *trans,
 		memset(m->block_sectors, 0, sizeof(m->block_sectors));
 
 		for (i = 0; i < new_s->nr_blocks; i++) {
-			ret = mark_stripe_bucket(trans, new, i, journal_seq, flags);
+			ret = mark_stripe_bucket(trans, new, i, flags);
 			if (ret)
 				return ret;
 		}
diff --git a/fs/bcachefs/buckets.h b/fs/bcachefs/buckets.h
index 6eeb95068b3b..4b5376684d2c 100644
--- a/fs/bcachefs/buckets.h
+++ b/fs/bcachefs/buckets.h
@@ -159,13 +159,6 @@ static inline bool is_available_bucket(struct bucket_mark mark)
 	return !mark.dirty_sectors && !mark.stripe;
 }
 
-static inline bool bucket_needs_journal_commit(struct bucket_mark m,
-					       u16 last_seq_ondisk)
-{
-	return m.journal_seq_valid &&
-		((s16) m.journal_seq - (s16) last_seq_ondisk > 0);
-}
-
 /* Device usage: */
 
 struct bch_dev_usage bch2_dev_usage_read(struct bch_dev *);
@@ -240,7 +233,6 @@ bch2_fs_usage_read_short(struct bch_fs *);
 
 /* key/bucket marking: */
 
-void bch2_bucket_seq_cleanup(struct bch_fs *);
 void bch2_fs_usage_initialize(struct bch_fs *);
 
 void bch2_mark_alloc_bucket(struct bch_fs *, struct bch_dev *, size_t, bool);
diff --git a/fs/bcachefs/buckets_types.h b/fs/bcachefs/buckets_types.h
index 18bca269b750..24139831226d 100644
--- a/fs/bcachefs/buckets_types.h
+++ b/fs/bcachefs/buckets_types.h
@@ -15,18 +15,9 @@ struct bucket_mark {
 	u8		gen;
 	u8		data_type:3,
 			owned_by_allocator:1,
-			journal_seq_valid:1,
 			stripe:1;
 	u16		dirty_sectors;
 	u16		cached_sectors;
-
-	/*
-	 * low bits of journal sequence number when this bucket was most
-	 * recently modified: if journal_seq_valid is set, this bucket can't be
-	 * reused until the journal sequence number written to disk is >= the
-	 * bucket's journal sequence number:
-	 */
-	u16		journal_seq;
 	};
 	};
 };
diff --git a/fs/bcachefs/buckets_waiting_for_journal.c b/fs/bcachefs/buckets_waiting_for_journal.c
new file mode 100644
index 000000000000..f3774e30b5cd
--- /dev/null
+++ b/fs/bcachefs/buckets_waiting_for_journal.c
@@ -0,0 +1,166 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include "bcachefs.h"
+#include "buckets_waiting_for_journal.h"
+#include <linux/hash.h>
+#include <linux/random.h>
+
+static inline struct bucket_hashed *
+bucket_hash(struct buckets_waiting_for_journal_table *t,
+	    unsigned hash_seed_idx, u64 dev_bucket)
+{
+	return t->d + hash_64(dev_bucket ^ t->hash_seeds[hash_seed_idx], t->bits);
+}
+
+static void bucket_table_init(struct buckets_waiting_for_journal_table *t, size_t bits)
+{
+	unsigned i;
+
+	t->bits = bits;
+	for (i = 0; i < ARRAY_SIZE(t->hash_seeds); i++)
+		get_random_bytes(&t->hash_seeds[i], sizeof(t->hash_seeds[i]));
+	memset(t->d, 0, sizeof(t->d[0]) << t->bits);
+}
+
+bool bch2_bucket_needs_journal_commit(struct buckets_waiting_for_journal *b,
+				      u64 flushed_seq,
+				      unsigned dev, u64 bucket)
+{
+	struct buckets_waiting_for_journal_table *t;
+	u64 dev_bucket = (u64) dev << 56 | bucket;
+	bool ret = false;
+	unsigned i;
+
+	mutex_lock(&b->lock);
+	t = b->t;
+
+	for (i = 0; i < ARRAY_SIZE(t->hash_seeds); i++) {
+		struct bucket_hashed *h = bucket_hash(t, i, dev_bucket);
+
+		if (h->dev_bucket == dev_bucket) {
+			ret = h->journal_seq > flushed_seq;
+			break;
+		}
+	}
+
+	mutex_unlock(&b->lock);
+
+	return ret;
+}
+
+static bool bucket_table_insert(struct buckets_waiting_for_journal_table *t,
+				struct bucket_hashed *new,
+				u64 flushed_seq)
+{
+	struct bucket_hashed *last_evicted = NULL;
+	unsigned tries, i;
+
+	for (tries = 0; tries < 10; tries++) {
+		struct bucket_hashed *old, *victim = NULL;
+
+		for (i = 0; i < ARRAY_SIZE(t->hash_seeds); i++) {
+			old = bucket_hash(t, i, new->dev_bucket);
+
+			if (old->dev_bucket == new->dev_bucket ||
+			    old->journal_seq <= flushed_seq) {
+				*old = *new;
+				return true;
+			}
+
+			if (last_evicted != old)
+				victim = old;
+		}
+
+		/* hashed to same slot 3 times: */
+		if (!victim)
+			break;
+
+		/* Failed to find an empty slot: */
+		swap(*new, *victim);
+		last_evicted = victim;
+	}
+
+	return false;
+}
+
+int bch2_set_bucket_needs_journal_commit(struct buckets_waiting_for_journal *b,
+					 u64 flushed_seq,
+					 unsigned dev, u64 bucket,
+					 u64 journal_seq)
+{
+	struct buckets_waiting_for_journal_table *t, *n;
+	struct bucket_hashed tmp, new = {
+		.dev_bucket	= (u64) dev << 56 | bucket,
+		.journal_seq	= journal_seq,
+	};
+	size_t i, size, new_bits, nr_elements = 1, nr_rehashes = 0;
+	int ret = 0;
+
+	mutex_lock(&b->lock);
+
+	if (likely(bucket_table_insert(b->t, &new, flushed_seq)))
+		goto out;
+
+	t = b->t;
+	size = 1UL << t->bits;
+	for (i = 0; i < size; i++)
+		nr_elements += t->d[i].journal_seq > flushed_seq;
+
+	new_bits = t->bits + (nr_elements * 3 > size);
+
+	n = kvmalloc(sizeof(*n) + (sizeof(n->d[0]) << new_bits), GFP_KERNEL);
+	if (!n) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+retry_rehash:
+	nr_rehashes++;
+	bucket_table_init(n, new_bits);
+
+	tmp = new;
+	BUG_ON(!bucket_table_insert(n, &tmp, flushed_seq));
+
+	for (i = 0; i < 1UL << t->bits; i++) {
+		if (t->d[i].journal_seq <= flushed_seq)
+			continue;
+
+		tmp = t->d[i];
+		if (!bucket_table_insert(n, &tmp, flushed_seq))
+			goto retry_rehash;
+	}
+
+	b->t = n;
+	kvfree(t);
+
+	pr_debug("took %zu rehashes, table at %zu/%zu elements",
+		 nr_rehashes, nr_elements, 1UL << b->t->bits);
+out:
+	mutex_unlock(&b->lock);
+
+	return ret;
+}
+
+void bch2_fs_buckets_waiting_for_journal_exit(struct bch_fs *c)
+{
+	struct buckets_waiting_for_journal *b = &c->buckets_waiting_for_journal;
+
+	kvfree(b->t);
+}
+
+#define INITIAL_TABLE_BITS		3
+
+int bch2_fs_buckets_waiting_for_journal_init(struct bch_fs *c)
+{
+	struct buckets_waiting_for_journal *b = &c->buckets_waiting_for_journal;
+
+	mutex_init(&b->lock);
+
+	b->t = kvmalloc(sizeof(*b->t) +
+			(sizeof(b->t->d[0]) << INITIAL_TABLE_BITS), GFP_KERNEL);
+	if (!b->t)
+		return -ENOMEM;
+
+	bucket_table_init(b->t, INITIAL_TABLE_BITS);
+	return 0;
+}
diff --git a/fs/bcachefs/buckets_waiting_for_journal.h b/fs/bcachefs/buckets_waiting_for_journal.h
new file mode 100644
index 000000000000..d2ae19cbe18c
--- /dev/null
+++ b/fs/bcachefs/buckets_waiting_for_journal.h
@@ -0,0 +1,15 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BUCKETS_WAITING_FOR_JOURNAL_H
+#define _BUCKETS_WAITING_FOR_JOURNAL_H
+
+#include "buckets_waiting_for_journal_types.h"
+
+bool bch2_bucket_needs_journal_commit(struct buckets_waiting_for_journal *,
+				      u64, unsigned, u64);
+int bch2_set_bucket_needs_journal_commit(struct buckets_waiting_for_journal *,
+					 u64, unsigned, u64, u64);
+
+void bch2_fs_buckets_waiting_for_journal_exit(struct bch_fs *);
+int bch2_fs_buckets_waiting_for_journal_init(struct bch_fs *);
+
+#endif /* _BUCKETS_WAITING_FOR_JOURNAL_H */
diff --git a/fs/bcachefs/buckets_waiting_for_journal_types.h b/fs/bcachefs/buckets_waiting_for_journal_types.h
new file mode 100644
index 000000000000..e593db061d81
--- /dev/null
+++ b/fs/bcachefs/buckets_waiting_for_journal_types.h
@@ -0,0 +1,23 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _BUCKETS_WAITING_FOR_JOURNAL_TYPES_H
+#define _BUCKETS_WAITING_FOR_JOURNAL_TYPES_H
+
+#include <linux/siphash.h>
+
+struct bucket_hashed {
+	u64			dev_bucket;
+	u64			journal_seq;
+};
+
+struct buckets_waiting_for_journal_table {
+	unsigned		bits;
+	u64			hash_seeds[3];
+	struct bucket_hashed	d[];
+};
+
+struct buckets_waiting_for_journal {
+	struct mutex		lock;
+	struct buckets_waiting_for_journal_table *t;
+};
+
+#endif /* _BUCKETS_WAITING_FOR_JOURNAL_TYPES_H */
diff --git a/fs/bcachefs/journal_io.c b/fs/bcachefs/journal_io.c
index 4602f581198e..815310e2426f 100644
--- a/fs/bcachefs/journal_io.c
+++ b/fs/bcachefs/journal_io.c
@@ -1666,13 +1666,9 @@ retry_alloc:
 		}
 	}
 
-	bch2_bucket_seq_cleanup(c);
-
 	continue_at(cl, do_journal_write, c->io_complete_wq);
 	return;
 no_io:
-	bch2_bucket_seq_cleanup(c);
-
 	continue_at(cl, journal_write_done, c->io_complete_wq);
 	return;
 err:
diff --git a/fs/bcachefs/super.c b/fs/bcachefs/super.c
index 55bb263a0906..3094eb1e3406 100644
--- a/fs/bcachefs/super.c
+++ b/fs/bcachefs/super.c
@@ -16,6 +16,7 @@
 #include "btree_key_cache.h"
 #include "btree_update_interior.h"
 #include "btree_io.h"
+#include "buckets_waiting_for_journal.h"
 #include "chardev.h"
 #include "checksum.h"
 #include "clock.h"
@@ -475,6 +476,7 @@ static void __bch2_fs_free(struct bch_fs *c)
 	bch2_fs_ec_exit(c);
 	bch2_fs_encryption_exit(c);
 	bch2_fs_io_exit(c);
+	bch2_fs_buckets_waiting_for_journal_exit(c);
 	bch2_fs_btree_interior_update_exit(c);
 	bch2_fs_btree_iter_exit(c);
 	bch2_fs_btree_key_cache_exit(&c->btree_key_cache);
@@ -818,6 +820,7 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	    bch2_fs_btree_key_cache_init(&c->btree_key_cache) ?:
 	    bch2_fs_btree_iter_init(c) ?:
 	    bch2_fs_btree_interior_update_init(c) ?:
+	    bch2_fs_buckets_waiting_for_journal_init(c);
 	    bch2_fs_subvolumes_init(c) ?:
 	    bch2_fs_io_init(c) ?:
 	    bch2_fs_encryption_init(c) ?: