bcachefs: KEY_TYPE_alloc_v4

This introduces a new alloc key which doesn't use varints. Soon we'll be
adding backpointers and storing them in alloc keys, which means our
pack/unpack workflow for alloc keys won't really work - we'll need to be
mutating alloc keys in place.

Instead of bch2_alloc_unpack(), we now have bch2_alloc_to_v4() that
converts older types of alloc keys to v4 if needed.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

10 files changed, 379 insertions, 301 deletions

diff --git a/fs/bcachefs/alloc_background.c b/fs/bcachefs/alloc_background.c
index 33b2e4d7da3b..099be1290b4c 100644
--- a/fs/bcachefs/alloc_background.c
+++ b/fs/bcachefs/alloc_background.c
@@ -33,13 +33,27 @@ const char * const bch2_allocator_states[] = {
 	NULL
 };
 
+/* Persistent alloc info: */
+
 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
 	BCH_ALLOC_FIELDS_V1()
 #undef x
 };
 
-/* Persistent alloc info: */
+struct bkey_alloc_unpacked {
+	u64		journal_seq;
+	u64		bucket;
+	u8		dev;
+	u8		gen;
+	u8		oldest_gen;
+	u8		data_type;
+	bool		need_discard:1;
+	bool		need_inc_gen:1;
+#define x(_name, _bits)	u##_bits _name;
+	BCH_ALLOC_FIELDS_V2()
+#undef  x
+};
 
 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
 				     const void **p, unsigned field)
@@ -161,6 +175,8 @@ static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 	out->gen	= a.v->gen;
 	out->oldest_gen	= a.v->oldest_gen;
 	out->data_type	= a.v->data_type;
+	out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
+	out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 	out->journal_seq = le64_to_cpu(a.v->journal_seq);
 
 #define x(_name, _bits)							\
@@ -182,47 +198,7 @@ static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 	return 0;
 }
 
-static void bch2_alloc_pack_v3(struct bkey_alloc_buf *dst,
-			       const struct bkey_alloc_unpacked src)
-{
-	struct bkey_i_alloc_v3 *a = bkey_alloc_v3_init(&dst->k);
-	unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
-	u8 *out = a->v.data;
-	u8 *end = (void *) &dst[1];
-	u8 *last_nonzero_field = out;
-	unsigned bytes;
-
-	a->k.p		= POS(src.dev, src.bucket);
-	a->v.gen	= src.gen;
-	a->v.oldest_gen	= src.oldest_gen;
-	a->v.data_type	= src.data_type;
-	a->v.journal_seq = cpu_to_le64(src.journal_seq);
-
-#define x(_name, _bits)							\
-	nr_fields++;							\
-									\
-	if (src._name) {						\
-		out += bch2_varint_encode_fast(out, src._name);		\
-									\
-		last_nonzero_field = out;				\
-		last_nonzero_fieldnr = nr_fields;			\
-	} else {							\
-		*out++ = 0;						\
-	}
-
-	BCH_ALLOC_FIELDS_V2()
-#undef  x
-	BUG_ON(out > end);
-
-	out = last_nonzero_field;
-	a->v.nr_fields = last_nonzero_fieldnr;
-
-	bytes = (u8 *) out - (u8 *) &a->v;
-	set_bkey_val_bytes(&a->k, bytes);
-	memset_u64s_tail(&a->v, 0, bytes);
-}
-
-struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
+static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 {
 	struct bkey_alloc_unpacked ret = {
 		.dev	= k.k->p.inode,
@@ -245,32 +221,71 @@ struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 	return ret;
 }
 
-struct bkey_alloc_buf *bch2_alloc_pack(struct btree_trans *trans,
-				       const struct bkey_alloc_unpacked src)
+void bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 {
-	struct bkey_alloc_buf *dst;
+	if (k.k->type == KEY_TYPE_alloc_v4) {
+		*out = *bkey_s_c_to_alloc_v4(k).v;
+	} else {
+		struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
+
+		*out = (struct bch_alloc_v4) {
+			.journal_seq		= u.journal_seq,
+			.flags			= u.need_discard,
+			.gen			= u.gen,
+			.oldest_gen		= u.oldest_gen,
+			.data_type		= u.data_type,
+			.stripe_redundancy	= u.stripe_redundancy,
+			.dirty_sectors		= u.dirty_sectors,
+			.cached_sectors		= u.cached_sectors,
+			.io_time[READ]		= u.read_time,
+			.io_time[WRITE]		= u.write_time,
+			.stripe			= u.stripe,
+		};
+	}
+}
 
-	dst = bch2_trans_kmalloc(trans, sizeof(struct bkey_alloc_buf));
-	if (!IS_ERR(dst))
-		bch2_alloc_pack_v3(dst, src);
+struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
+{
+	struct bkey_i_alloc_v4 *ret;
 
-	return dst;
+	if (k.k->type == KEY_TYPE_alloc_v4) {
+		ret = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+		if (!IS_ERR(ret))
+			bkey_reassemble(&ret->k_i, k);
+	} else {
+		ret = bch2_trans_kmalloc(trans, sizeof(*ret));
+		if (!IS_ERR(ret)) {
+			bkey_alloc_v4_init(&ret->k_i);
+			ret->k.p = k.k->p;
+			bch2_alloc_to_v4(k, &ret->v);
+		}
+	}
+	return ret;
 }
 
-int bch2_alloc_write(struct btree_trans *trans, struct btree_iter *iter,
-		     struct bkey_alloc_unpacked *u, unsigned trigger_flags)
+struct bkey_i_alloc_v4 *
+bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
+			      struct bpos pos)
 {
-	struct bkey_alloc_buf *a = bch2_alloc_pack(trans, *u);
+	struct bkey_s_c k;
+	struct bkey_i_alloc_v4 *a;
+	int ret;
 
-	/*
-	 * Without BTREE_UPDATE_NO_KEY_CACHE_COHERENCY, we may end up updating
-	 * the btree instead of the key cache - this can casue the allocator to
-	 * self-deadlock, since updating the btree may require allocating new
-	 * btree nodes:
-	 */
-	return PTR_ERR_OR_ZERO(a) ?:
-		bch2_trans_update(trans, iter, &a->k, trigger_flags|
-				 BTREE_UPDATE_NO_KEY_CACHE_COHERENCY);
+	bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
+			     BTREE_ITER_WITH_UPDATES|
+			     BTREE_ITER_CACHED|
+			     BTREE_ITER_INTENT);
+	k = bch2_btree_iter_peek_slot(iter);
+	ret = bkey_err(k);
+	if (ret) {
+		bch2_trans_iter_exit(trans, iter);
+		return ERR_PTR(ret);
+	}
+
+	a = bch2_alloc_to_v4_mut(trans, k);
+	if (IS_ERR(a))
+		bch2_trans_iter_exit(trans, iter);
+	return a;
 }
 
 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
@@ -316,28 +331,70 @@ const char *bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k)
 const char *bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k)
 {
 	struct bkey_alloc_unpacked u;
+	struct bch_dev *ca;
 
 	if (k.k->p.inode >= c->sb.nr_devices ||
 	    !c->devs[k.k->p.inode])
 		return "invalid device";
 
+	ca = bch_dev_bkey_exists(c, k.k->p.inode);
+
+	if (k.k->p.offset < ca->mi.first_bucket ||
+	    k.k->p.offset >= ca->mi.nbuckets)
+		return "invalid bucket";
+
 	if (bch2_alloc_unpack_v3(&u, k))
 		return "unpack error";
 
 	return NULL;
 }
 
-void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c,
-			   struct bkey_s_c k)
+const char *bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k)
 {
-	struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
+	struct bch_dev *ca;
 
-	pr_buf(out, "gen %u oldest_gen %u data_type %s journal_seq %llu",
-	       u.gen, u.oldest_gen, bch2_data_types[u.data_type],
-	       u.journal_seq);
-#define x(_name, ...)	pr_buf(out, " " #_name " %llu", (u64) u._name);
-	BCH_ALLOC_FIELDS_V2()
-#undef  x
+	if (k.k->p.inode >= c->sb.nr_devices ||
+	    !c->devs[k.k->p.inode])
+		return "invalid device";
+
+	ca = bch_dev_bkey_exists(c, k.k->p.inode);
+
+	if (k.k->p.offset < ca->mi.first_bucket ||
+	    k.k->p.offset >= ca->mi.nbuckets)
+		return "invalid bucket";
+
+	return NULL;
+}
+
+void bch2_alloc_v4_swab(struct bkey_s k)
+{
+	struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
+
+	a->journal_seq		= swab64(a->journal_seq);
+	a->flags		= swab32(a->flags);
+	a->dirty_sectors	= swab32(a->dirty_sectors);
+	a->cached_sectors	= swab32(a->cached_sectors);
+	a->io_time[0]		= swab64(a->io_time[0]);
+	a->io_time[1]		= swab64(a->io_time[1]);
+	a->stripe		= swab32(a->stripe);
+	a->nr_external_backpointers = swab32(a->nr_external_backpointers);
+}
+
+void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
+{
+	struct bch_alloc_v4 a;
+
+	bch2_alloc_to_v4(k, &a);
+
+	pr_buf(out, "gen %u oldest_gen %u data_type %s journal_seq %llu need_discard %llu",
+	       a.gen, a.oldest_gen, bch2_data_types[a.data_type],
+	       a.journal_seq, BCH_ALLOC_V4_NEED_DISCARD(&a));
+	pr_buf(out, " dirty_sectors %u",	a.dirty_sectors);
+	pr_buf(out, " cached_sectors %u",	a.cached_sectors);
+	pr_buf(out, " stripe %u",		a.stripe);
+	pr_buf(out, " stripe_redundancy %u",	a.stripe_redundancy);
+	pr_buf(out, " read_time %llu",		a.io_time[READ]);
+	pr_buf(out, " write_time %llu",		a.io_time[WRITE]);
 }
 
 int bch2_alloc_read(struct bch_fs *c, bool gc, bool metadata_only)
@@ -345,9 +402,9 @@ int bch2_alloc_read(struct bch_fs *c, bool gc, bool metadata_only)
 	struct btree_trans trans;
 	struct btree_iter iter;
 	struct bkey_s_c k;
+	struct bch_alloc_v4 a;
 	struct bch_dev *ca;
 	struct bucket *g;
-	struct bkey_alloc_unpacked u;
 	int ret;
 
 	bch2_trans_init(&trans, c, 0, 0);
@@ -356,28 +413,28 @@ int bch2_alloc_read(struct bch_fs *c, bool gc, bool metadata_only)
 			   BTREE_ITER_PREFETCH, k, ret) {
 		ca = bch_dev_bkey_exists(c, k.k->p.inode);
 		g = __bucket(ca, k.k->p.offset, gc);
-		u = bch2_alloc_unpack(k);
+		bch2_alloc_to_v4(k, &a);
 
 		if (!gc)
-			*bucket_gen(ca, k.k->p.offset) = u.gen;
+			*bucket_gen(ca, k.k->p.offset) = a.gen;
 
-		g->_mark.gen		= u.gen;
-		g->io_time[READ]	= u.read_time;
-		g->io_time[WRITE]	= u.write_time;
-		g->oldest_gen		= !gc ? u.oldest_gen : u.gen;
+		g->_mark.gen		= a.gen;
+		g->io_time[READ]	= a.io_time[READ];
+		g->io_time[WRITE]	= a.io_time[WRITE];
+		g->oldest_gen		= !gc ? a.oldest_gen : a.gen;
 		g->gen_valid		= 1;
 
 		if (!gc ||
 		    (metadata_only &&
-		     (u.data_type == BCH_DATA_user ||
-		      u.data_type == BCH_DATA_cached ||
-		      u.data_type == BCH_DATA_parity))) {
-			g->_mark.data_type	= u.data_type;
-			g->_mark.dirty_sectors	= u.dirty_sectors;
-			g->_mark.cached_sectors	= u.cached_sectors;
-			g->_mark.stripe		= u.stripe != 0;
-			g->stripe		= u.stripe;
-			g->stripe_redundancy	= u.stripe_redundancy;
+		     (a.data_type == BCH_DATA_user ||
+		      a.data_type == BCH_DATA_cached ||
+		      a.data_type == BCH_DATA_parity))) {
+			g->_mark.data_type	= a.data_type;
+			g->_mark.dirty_sectors	= a.dirty_sectors;
+			g->_mark.cached_sectors	= a.cached_sectors;
+			g->_mark.stripe		= a.stripe != 0;
+			g->stripe		= a.stripe;
+			g->stripe_redundancy	= a.stripe_redundancy;
 		}
 
 	}
@@ -398,29 +455,22 @@ int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
 {
 	struct bch_fs *c = trans->c;
 	struct btree_iter iter;
-	struct bkey_s_c k;
-	struct bkey_alloc_unpacked u;
-	u64 *time, now;
+	struct bkey_i_alloc_v4 *a;
+	u64 now;
 	int ret = 0;
 
-	bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(dev, bucket_nr),
-			     BTREE_ITER_CACHED|
-			     BTREE_ITER_INTENT);
-	k = bch2_btree_iter_peek_slot(&iter);
-	ret = bkey_err(k);
+	a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
+	ret = PTR_ERR_OR_ZERO(a);
 	if (ret)
-		goto out;
-
-	u = bch2_alloc_unpack(k);
+		return ret;
 
-	time = rw == READ ? &u.read_time : &u.write_time;
 	now = atomic64_read(&c->io_clock[rw].now);
-	if (*time == now)
+	if (a->v.io_time[rw] == now)
 		goto out;
 
-	*time = now;
+	a->v.io_time[rw] = now;
 
-	ret   = bch2_alloc_write(trans, &iter, &u, 0) ?:
+	ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
 		bch2_trans_commit(trans, NULL, NULL, 0);
 out:
 	bch2_trans_iter_exit(trans, &iter);
@@ -604,7 +654,7 @@ static size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca)
 
 static int bucket_invalidate_btree(struct btree_trans *trans,
 				   struct bch_dev *ca, u64 b,
-				   struct bkey_alloc_unpacked *u)
+				   struct bkey_i_alloc_v4 *a)
 {
 	struct bch_fs *c = trans->c;
 	struct btree_iter iter;
@@ -621,16 +671,19 @@ static int bucket_invalidate_btree(struct btree_trans *trans,
 	if (ret)
 		goto err;
 
-	*u = bch2_alloc_unpack(k);
-	u->gen++;
-	u->data_type		= 0;
-	u->dirty_sectors	= 0;
-	u->cached_sectors	= 0;
-	u->read_time		= atomic64_read(&c->io_clock[READ].now);
-	u->write_time		= atomic64_read(&c->io_clock[WRITE].now);
-
-	ret = bch2_alloc_write(trans, &iter, u,
-			       BTREE_TRIGGER_BUCKET_INVALIDATE);
+	bkey_alloc_v4_init(&a->k_i);
+	a->k.p = iter.pos;
+	bch2_alloc_to_v4(k, &a->v);
+	a->v.gen++;
+	a->v.data_type		= 0;
+	a->v.dirty_sectors	= 0;
+	a->v.cached_sectors	= 0;
+	a->v.io_time[READ]	= atomic64_read(&c->io_clock[READ].now);
+	a->v.io_time[WRITE]	= atomic64_read(&c->io_clock[WRITE].now);
+
+	ret = bch2_trans_update(trans, &iter, &a->k_i,
+			       BTREE_TRIGGER_BUCKET_INVALIDATE|
+			       BTREE_UPDATE_NO_KEY_CACHE_COHERENCY);
 err:
 	bch2_trans_iter_exit(trans, &iter);
 	return ret;
@@ -639,7 +692,7 @@ err:
 static int bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
 				      u64 *journal_seq, unsigned flags)
 {
-	struct bkey_alloc_unpacked u;
+	struct bkey_i_alloc_v4 a;
 	size_t b;
 	u64 commit_seq = 0;
 	int ret = 0;
@@ -671,7 +724,7 @@ static int bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
 			    BTREE_INSERT_NOCHECK_RW|
 			    BTREE_INSERT_NOFAIL|
 			    flags,
-			    bucket_invalidate_btree(&trans, ca, b, &u));
+			    bucket_invalidate_btree(&trans, ca, b, &a));
 
 	if (!ret) {
 		/* remove from alloc_heap: */
@@ -687,14 +740,14 @@ static int bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
 		 * If we invalidating cached data then we need to wait on the
 		 * journal commit:
 		 */
-		if (u.data_type)
+		if (a.v.data_type)
 			*journal_seq = max(*journal_seq, commit_seq);
 
 		/*
 		 * We already waiting on u.alloc_seq when we filtered out
 		 * buckets that need journal commit:
 		 */
-		BUG_ON(*journal_seq > u.journal_seq);
+		BUG_ON(*journal_seq > a.v.journal_seq);
 	} else {
 		size_t b2;
 
diff --git a/fs/bcachefs/alloc_background.h b/fs/bcachefs/alloc_background.h
index 3eaa6d204286..b66c8cf0341e 100644
--- a/fs/bcachefs/alloc_background.h
+++ b/fs/bcachefs/alloc_background.h
@@ -10,48 +10,14 @@
 
 extern const char * const bch2_allocator_states[];
 
-struct bkey_alloc_unpacked {
-	u64		journal_seq;
-	u64		bucket;
-	u8		dev;
-	u8		gen;
-	u8		oldest_gen;
-	u8		data_type;
-#define x(_name, _bits)	u##_bits _name;
-	BCH_ALLOC_FIELDS_V2()
-#undef  x
-};
-
 /* How out of date a pointer gen is allowed to be: */
 #define BUCKET_GC_GEN_MAX	96U
 
-/* returns true if not equal */
-static inline bool bkey_alloc_unpacked_cmp(struct bkey_alloc_unpacked l,
-					   struct bkey_alloc_unpacked r)
-{
-	return  l.gen != r.gen			||
-		l.oldest_gen != r.oldest_gen	||
-		l.data_type != r.data_type
-#define x(_name, ...)	|| l._name != r._name
-	BCH_ALLOC_FIELDS_V2()
-#undef  x
-	;
-}
-
-struct bkey_alloc_buf {
-	struct bkey_i	k;
-	struct bch_alloc_v3 v;
-
-#define x(_name,  _bits)		+ _bits / 8
-	u8		_pad[0 + BCH_ALLOC_FIELDS_V2()];
-#undef  x
-} __attribute__((packed, aligned(8)));
+struct bkey_i_alloc_v4 *
+bch2_trans_start_alloc_update(struct btree_trans *, struct btree_iter *, struct bpos);
 
-struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c);
-struct bkey_alloc_buf *bch2_alloc_pack(struct btree_trans *,
-				       const struct bkey_alloc_unpacked);
-int bch2_alloc_write(struct btree_trans *, struct btree_iter *,
-		     struct bkey_alloc_unpacked *, unsigned);
+void bch2_alloc_to_v4(struct bkey_s_c, struct bch_alloc_v4 *);
+struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *, struct bkey_s_c);
 
 int bch2_bucket_io_time_reset(struct btree_trans *, unsigned, size_t, int);
 
@@ -60,6 +26,8 @@ int bch2_bucket_io_time_reset(struct btree_trans *, unsigned, size_t, int);
 const char *bch2_alloc_v1_invalid(const struct bch_fs *, struct bkey_s_c);
 const char *bch2_alloc_v2_invalid(const struct bch_fs *, struct bkey_s_c);
 const char *bch2_alloc_v3_invalid(const struct bch_fs *, struct bkey_s_c);
+const char *bch2_alloc_v4_invalid(const struct bch_fs *, struct bkey_s_c k);
+void bch2_alloc_v4_swab(struct bkey_s);
 void bch2_alloc_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 
 #define bch2_bkey_ops_alloc (struct bkey_ops) {		\
@@ -80,6 +48,13 @@ void bch2_alloc_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 	.atomic_trigger	= bch2_mark_alloc,		\
 }
 
+#define bch2_bkey_ops_alloc_v4 (struct bkey_ops) {	\
+	.key_invalid	= bch2_alloc_v4_invalid,	\
+	.val_to_text	= bch2_alloc_to_text,		\
+	.swab		= bch2_alloc_v4_swab,		\
+	.atomic_trigger	= bch2_mark_alloc,		\
+}
+
 static inline bool bkey_is_alloc(const struct bkey *k)
 {
 	return  k->type == KEY_TYPE_alloc ||
diff --git a/fs/bcachefs/bcachefs_format.h b/fs/bcachefs/bcachefs_format.h
index 982409ed940e..a640a45a123a 100644
--- a/fs/bcachefs/bcachefs_format.h
+++ b/fs/bcachefs/bcachefs_format.h
@@ -82,6 +82,21 @@
 typedef uuid_t __uuid_t;
 #endif
 
+#define BITMASK(name, type, field, offset, end)				\
+static const unsigned	name##_OFFSET = offset;				\
+static const unsigned	name##_BITS = (end - offset);			\
+									\
+static inline __u64 name(const type *k)					\
+{									\
+	return (k->field >> offset) & ~(~0ULL << (end - offset));	\
+}									\
+									\
+static inline void SET_##name(type *k, __u64 v)				\
+{									\
+	k->field &= ~(~(~0ULL << (end - offset)) << offset);		\
+	k->field |= (v & ~(~0ULL << (end - offset))) << offset;		\
+}
+
 #define LE_BITMASK(_bits, name, type, field, offset, end)		\
 static const unsigned	name##_OFFSET = offset;				\
 static const unsigned	name##_BITS = (end - offset);			\
@@ -353,7 +368,8 @@ static inline void bkey_init(struct bkey *k)
 	x(inode_v2,		23)			\
 	x(alloc_v3,		24)			\
 	x(set,			25)			\
-	x(lru,			26)
+	x(lru,			26)			\
+	x(alloc_v4,		27)
 
 enum bch_bkey_type {
 #define x(name, nr) KEY_TYPE_##name	= nr,
@@ -903,6 +919,30 @@ struct bch_alloc_v3 {
 	__u8			data[];
 } __attribute__((packed, aligned(8)));
 
+struct bch_alloc_v4 {
+	struct bch_val		v;
+	__u64			journal_seq;
+	__u32			flags;
+	__u8			gen;
+	__u8			oldest_gen;
+	__u8			data_type;
+	__u8			stripe_redundancy;
+	__u32			dirty_sectors;
+	__u32			cached_sectors;
+	__u64			io_time[2];
+	__u32			stripe;
+	__u32			nr_external_backpointers;
+	struct bpos		backpointers[0];
+} __attribute__((packed, aligned(8)));
+
+LE32_BITMASK(BCH_ALLOC_V3_NEED_DISCARD,struct bch_alloc_v3, flags,  0,  1)
+LE32_BITMASK(BCH_ALLOC_V3_NEED_INC_GEN,struct bch_alloc_v3, flags,  1,  2)
+
+BITMASK(BCH_ALLOC_V4_NEED_DISCARD,	struct bch_alloc_v4, flags,  0,  1)
+BITMASK(BCH_ALLOC_V4_NEED_INC_GEN,	struct bch_alloc_v4, flags,  1,  2)
+BITMASK(BCH_ALLOC_V4_BACKPOINTERS_START,struct bch_alloc_v4, flags,  2,  8)
+BITMASK(BCH_ALLOC_V4_NR_BACKPOINTERS,	struct bch_alloc_v4, flags,  8,  14)
+
 enum {
 #define x(name, _bits) BCH_ALLOC_FIELD_V1_##name,
 	BCH_ALLOC_FIELDS_V1()
diff --git a/fs/bcachefs/bkey_methods.c b/fs/bcachefs/bkey_methods.c
index 8757218e571b..62774e4f6dbb 100644
--- a/fs/bcachefs/bkey_methods.c
+++ b/fs/bcachefs/bkey_methods.c
@@ -149,7 +149,8 @@ static unsigned bch2_key_types_allowed[] = {
 		(1U << KEY_TYPE_deleted)|
 		(1U << KEY_TYPE_alloc)|
 		(1U << KEY_TYPE_alloc_v2)|
-		(1U << KEY_TYPE_alloc_v3),
+		(1U << KEY_TYPE_alloc_v3)|
+		(1U << KEY_TYPE_alloc_v4),
 	[BKEY_TYPE_quotas] =
 		(1U << KEY_TYPE_deleted)|
 		(1U << KEY_TYPE_quota),
diff --git a/fs/bcachefs/btree_gc.c b/fs/bcachefs/btree_gc.c
index 8ec9c43d98e1..c3d6c62ef062 100644
--- a/fs/bcachefs/btree_gc.c
+++ b/fs/bcachefs/btree_gc.c
@@ -1306,6 +1306,19 @@ static int bch2_gc_start(struct bch_fs *c,
 	return 0;
 }
 
+/* returns true if not equal */
+static inline bool bch2_alloc_v4_cmp(struct bch_alloc_v4 l,
+				     struct bch_alloc_v4 r)
+{
+	return  l.gen != r.gen				||
+		l.oldest_gen != r.oldest_gen		||
+		l.data_type != r.data_type		||
+		l.dirty_sectors	!= r.dirty_sectors	||
+		l.cached_sectors != r.cached_sectors	 ||
+		l.stripe_redundancy != r.stripe_redundancy ||
+		l.stripe != r.stripe;
+}
+
 static int bch2_alloc_write_key(struct btree_trans *trans,
 				struct btree_iter *iter,
 				bool metadata_only)
@@ -1314,8 +1327,8 @@ static int bch2_alloc_write_key(struct btree_trans *trans,
 	struct bch_dev *ca = bch_dev_bkey_exists(c, iter->pos.inode);
 	struct bucket *g;
 	struct bkey_s_c k;
-	struct bkey_alloc_unpacked old_u, new_u, gc_u;
-	struct bkey_alloc_buf *a;
+	struct bkey_i_alloc_v4 *a;
+	struct bch_alloc_v4 old, new, gc;
 	int ret;
 
 	k = bch2_btree_iter_peek_slot(iter);
@@ -1323,60 +1336,61 @@ static int bch2_alloc_write_key(struct btree_trans *trans,
 	if (ret)
 		return ret;
 
-	old_u = new_u = bch2_alloc_unpack(k);
+	bch2_alloc_to_v4(k, &old);
+	new = old;
 
 	percpu_down_read(&c->mark_lock);
 	g	= gc_bucket(ca, iter->pos.offset);
-	gc_u = (struct bkey_alloc_unpacked) {
-		.dev		= iter->pos.inode,
-		.bucket		= iter->pos.offset,
+	gc = (struct bch_alloc_v4) {
 		.gen		= g->mark.gen,
 		.data_type	= g->mark.data_type,
 		.dirty_sectors	= g->mark.dirty_sectors,
 		.cached_sectors	= g->mark.cached_sectors,
-		.read_time	= g->io_time[READ],
-		.write_time	= g->io_time[WRITE],
+		.io_time[READ]	= g->io_time[READ],
+		.io_time[WRITE]	= g->io_time[WRITE],
 		.stripe		= g->stripe,
 		.stripe_redundancy = g->stripe_redundancy,
 	};
 	percpu_up_read(&c->mark_lock);
 
 	if (metadata_only &&
-	    gc_u.data_type != BCH_DATA_sb &&
-	    gc_u.data_type != BCH_DATA_journal &&
-	    gc_u.data_type != BCH_DATA_btree)
+	    gc.data_type != BCH_DATA_sb &&
+	    gc.data_type != BCH_DATA_journal &&
+	    gc.data_type != BCH_DATA_btree)
 		return 0;
 
-	if (gen_after(old_u.gen, gc_u.gen))
+	if (gen_after(old.gen, gc.gen))
 		return 0;
 
 #define copy_bucket_field(_f)						\
-	if (fsck_err_on(new_u._f != gc_u._f, c,				\
+	if (fsck_err_on(new._f != gc._f, c,				\
 			"bucket %llu:%llu gen %u data type %s has wrong " #_f	\
 			": got %u, should be %u",			\
 			iter->pos.inode, iter->pos.offset,		\
-			new_u.gen,					\
-			bch2_data_types[new_u.data_type],		\
-			new_u._f, gc_u._f))				\
-		new_u._f = gc_u._f;					\
+			new.gen,					\
+			bch2_data_types[new.data_type],			\
+			new._f, gc._f))					\
+		new._f = gc._f;						\
 
 	copy_bucket_field(gen);
 	copy_bucket_field(data_type);
-	copy_bucket_field(stripe);
 	copy_bucket_field(dirty_sectors);
 	copy_bucket_field(cached_sectors);
 	copy_bucket_field(stripe_redundancy);
 	copy_bucket_field(stripe);
 #undef copy_bucket_field
 
-	if (!bkey_alloc_unpacked_cmp(old_u, new_u))
+	if (!bch2_alloc_v4_cmp(old, new))
 		return 0;
 
-	a = bch2_alloc_pack(trans, new_u);
-	if (IS_ERR(a))
-		return PTR_ERR(a);
+	a = bch2_alloc_to_v4_mut(trans, k);
+	ret = PTR_ERR_OR_ZERO(a);
+	if (ret)
+		return ret;
+
+	a->v = new;
 
-	ret = bch2_trans_update(trans, iter, &a->k, BTREE_TRIGGER_NORUN);
+	ret = bch2_trans_update(trans, iter, &a->k_i, BTREE_TRIGGER_NORUN);
 fsck_err:
 	return ret;
 }
@@ -1873,7 +1887,8 @@ static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_i
 {
 	struct bch_dev *ca = bch_dev_bkey_exists(trans->c, iter->pos.inode);
 	struct bkey_s_c k;
-	struct bkey_alloc_unpacked u;
+	struct bch_alloc_v4 a;
+	struct bkey_i_alloc_v4 *a_mut;
 	int ret;
 
 	k = bch2_btree_iter_peek_slot(iter);
@@ -1881,14 +1896,19 @@ static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct btree_i
 	if (ret)
 		return ret;
 
-	u = bch2_alloc_unpack(k);
+	bch2_alloc_to_v4(k, &a);
 
-	if (u.oldest_gen == ca->oldest_gen[iter->pos.offset])
+	if (a.oldest_gen == ca->oldest_gen[iter->pos.offset])
 		return 0;
 
-	u.oldest_gen = ca->oldest_gen[iter->pos.offset];
+	a_mut = bch2_alloc_to_v4_mut(trans, k);
+	ret = PTR_ERR_OR_ZERO(a_mut);
+	if (ret)
+		return ret;
+
+	a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
 
-	return bch2_alloc_write(trans, iter, &u, BTREE_TRIGGER_NORUN);
+	return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
 }
 
 int bch2_gc_gens(struct bch_fs *c)
diff --git a/fs/bcachefs/btree_types.h b/fs/bcachefs/btree_types.h
index e848b153ae93..d5be6004071a 100644
--- a/fs/bcachefs/btree_types.h
+++ b/fs/bcachefs/btree_types.h
@@ -678,6 +678,7 @@ enum btree_update_flags {
 	((1U << KEY_TYPE_alloc)|		\
 	 (1U << KEY_TYPE_alloc_v2)|		\
 	 (1U << KEY_TYPE_alloc_v3)|		\
+	 (1U << KEY_TYPE_alloc_v4)|		\
 	 (1U << KEY_TYPE_stripe)|		\
 	 (1U << KEY_TYPE_inode)|		\
 	 (1U << KEY_TYPE_inode_v2)|		\
diff --git a/fs/bcachefs/buckets.c b/fs/bcachefs/buckets.c
index 8eeabb5a66bd..b4252c2f028a 100644
--- a/fs/bcachefs/buckets.c
+++ b/fs/bcachefs/buckets.c
@@ -508,15 +508,14 @@ int bch2_mark_alloc(struct btree_trans *trans,
 	bool gc = flags & BTREE_TRIGGER_GC;
 	u64 journal_seq = trans->journal_res.seq;
 	struct bch_fs *c = trans->c;
-	struct bkey_alloc_unpacked old_u = bch2_alloc_unpack(old);
-	struct bkey_alloc_unpacked new_u = bch2_alloc_unpack(new);
-	struct bch_dev *ca = bch_dev_bkey_exists(c, new_u.dev);
+	struct bch_alloc_v4 old_a, new_a;
+	struct bch_dev *ca = bch_dev_bkey_exists(c, new.k->p.inode);
 	struct bucket *g;
 	struct bucket_mark old_m, m;
 	int ret = 0;
 
-	if (bch2_trans_inconsistent_on(new_u.bucket < ca->mi.first_bucket ||
-				       new_u.bucket >= ca->mi.nbuckets, trans,
+	if (bch2_trans_inconsistent_on(new.k->p.offset < ca->mi.first_bucket ||
+				       new.k->p.offset >= ca->mi.nbuckets, trans,
 				       "alloc key outside range of device's buckets"))
 		return -EIO;
 
@@ -527,11 +526,13 @@ int bch2_mark_alloc(struct btree_trans *trans,
 	    !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 		return 0;
 
+	bch2_alloc_to_v4(old, &old_a);
+	bch2_alloc_to_v4(new, &new_a);
+
 	if ((flags & BTREE_TRIGGER_INSERT) &&
-	    !old_u.data_type != !new_u.data_type &&
-	    new.k->type == KEY_TYPE_alloc_v3) {
-		struct bch_alloc_v3 *v = (struct bch_alloc_v3 *) new.v;
-		u64 old_journal_seq = le64_to_cpu(v->journal_seq);
+	    !old_a.data_type != !new_a.data_type &&
+	    new.k->type == KEY_TYPE_alloc_v4) {
+		struct bch_alloc_v4 *v = (struct bch_alloc_v4 *) new.v;
 
 		BUG_ON(!journal_seq);
 
@@ -540,18 +541,18 @@ int bch2_mark_alloc(struct btree_trans *trans,
 		 * before the bucket became empty again, then the we don't have
 		 * to wait on a journal flush before we can reuse the bucket:
 		 */
-		new_u.journal_seq = !new_u.data_type &&
-			(journal_seq == old_journal_seq ||
-			 bch2_journal_noflush_seq(&c->journal, old_journal_seq))
+		new_a.journal_seq = !new_a.data_type &&
+			(journal_seq == v->journal_seq ||
+			 bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 			? 0 : journal_seq;
-		v->journal_seq = cpu_to_le64(new_u.journal_seq);
+		v->journal_seq = new_a.journal_seq;
 	}
 
-	if (old_u.data_type && !new_u.data_type && new_u.journal_seq) {
+	if (old_a.data_type && !new_a.data_type && new_a.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);
+				new.k->p.inode, new.k->p.offset,
+				new_a.journal_seq);
 		if (ret) {
 			bch2_fs_fatal_error(c,
 				"error setting bucket_needs_journal_commit: %i", ret);
@@ -560,27 +561,27 @@ int bch2_mark_alloc(struct btree_trans *trans,
 	}
 
 	percpu_down_read(&c->mark_lock);
-	if (!gc && new_u.gen != old_u.gen)
-		*bucket_gen(ca, new_u.bucket) = new_u.gen;
+	if (!gc && new_a.gen != old_a.gen)
+		*bucket_gen(ca, new.k->p.offset) = new_a.gen;
 
-	g = __bucket(ca, new_u.bucket, gc);
+	g = __bucket(ca, new.k->p.offset, gc);
 
 	old_m = bucket_cmpxchg(g, m, ({
-		m.gen			= new_u.gen;
-		m.data_type		= new_u.data_type;
-		m.dirty_sectors		= new_u.dirty_sectors;
-		m.cached_sectors	= new_u.cached_sectors;
-		m.stripe		= new_u.stripe != 0;
+		m.gen			= new_a.gen;
+		m.data_type		= new_a.data_type;
+		m.dirty_sectors		= new_a.dirty_sectors;
+		m.cached_sectors	= new_a.cached_sectors;
+		m.stripe		= new_a.stripe != 0;
 	}));
 
 	bch2_dev_usage_update(c, ca, old_m, m, journal_seq, gc);
 
-	g->io_time[READ]	= new_u.read_time;
-	g->io_time[WRITE]	= new_u.write_time;
-	g->oldest_gen		= new_u.oldest_gen;
+	g->io_time[READ]	= new_a.io_time[READ];
+	g->io_time[WRITE]	= new_a.io_time[WRITE];
+	g->oldest_gen		= new_a.oldest_gen;
 	g->gen_valid		= 1;
-	g->stripe		= new_u.stripe;
-	g->stripe_redundancy	= new_u.stripe_redundancy;
+	g->stripe		= new_a.stripe;
+	g->stripe_redundancy	= new_a.stripe_redundancy;
 	percpu_up_read(&c->mark_lock);
 
 	/*
@@ -598,7 +599,7 @@ int bch2_mark_alloc(struct btree_trans *trans,
 			return ret;
 		}
 
-		trace_invalidate(ca, bucket_to_sector(ca, new_u.bucket),
+		trace_invalidate(ca, bucket_to_sector(ca, new.k->p.offset),
 				 old_m.cached_sectors);
 	}
 
@@ -1378,50 +1379,32 @@ need_mark:
 
 /* trans_mark: */
 
-static int bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
-			      const struct bch_extent_ptr *ptr,
-			      struct bkey_alloc_unpacked *u)
-{
-	struct bch_fs *c = trans->c;
-	struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
-	struct bkey_s_c k;
-	int ret;
-
-	bch2_trans_iter_init(trans, iter, BTREE_ID_alloc,
-			     POS(ptr->dev, PTR_BUCKET_NR(ca, ptr)),
-			     BTREE_ITER_WITH_UPDATES|
-			     BTREE_ITER_CACHED|
-			     BTREE_ITER_INTENT);
-	k = bch2_btree_iter_peek_slot(iter);
-	ret = bkey_err(k);
-	if (ret) {
-		bch2_trans_iter_exit(trans, iter);
-		return ret;
-	}
-
-	*u = bch2_alloc_unpack(k);
-	return 0;
-}
-
 static int bch2_trans_mark_pointer(struct btree_trans *trans,
 			struct bkey_s_c k, struct extent_ptr_decoded p,
 			s64 sectors, enum bch_data_type data_type)
 {
 	struct btree_iter iter;
-	struct bkey_alloc_unpacked u;
+	struct bkey_i_alloc_v4 *a;
+	u16 dirty_sectors, cached_sectors;
 	int ret;
 
-	ret = bch2_trans_start_alloc_update(trans, &iter, &p.ptr, &u);
-	if (ret)
-		return ret;
+	a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(trans->c, &p.ptr));
+	if (IS_ERR(a))
+		return PTR_ERR(a);
+
+	dirty_sectors	= a->v.dirty_sectors;
+	cached_sectors	= a->v.cached_sectors;
 
 	ret = __mark_pointer(trans, k, &p.ptr, sectors, data_type,
-			     u.gen, &u.data_type,
-			     &u.dirty_sectors, &u.cached_sectors);
+			     a->v.gen, &a->v.data_type,
+			     &dirty_sectors, &cached_sectors);
 	if (ret)
 		goto out;
 
-	ret = bch2_alloc_write(trans, &iter, &u, 0);
+	a->v.dirty_sectors	= dirty_sectors;
+	a->v.cached_sectors	= cached_sectors;
+
+	ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
 	if (ret)
 		goto out;
 out:
@@ -1554,7 +1537,7 @@ static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
 	struct bch_fs *c = trans->c;
 	const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
 	struct btree_iter iter;
-	struct bkey_alloc_unpacked u;
+	struct bkey_i_alloc_v4 *a;
 	enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
 		? BCH_DATA_parity : 0;
 	s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
@@ -1563,59 +1546,59 @@ static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
 	if (deleting)
 		sectors = -sectors;
 
-	ret = bch2_trans_start_alloc_update(trans, &iter, ptr, &u);
-	if (ret)
-		return ret;
+	a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
+	if (IS_ERR(a))
+		return PTR_ERR(a);
 
 	ret = check_bucket_ref(c, s.s_c, ptr, sectors, data_type,
-			       u.gen, u.data_type,
-			       u.dirty_sectors, u.cached_sectors);
+			       a->v.gen, a->v.data_type,
+			       a->v.dirty_sectors, a->v.cached_sectors);
 	if (ret)
 		goto err;
 
 	if (!deleting) {
-		if (bch2_trans_inconsistent_on(u.stripe ||
-					    u.stripe_redundancy, trans,
+		if (bch2_trans_inconsistent_on(a->v.stripe ||
+					       a->v.stripe_redundancy, trans,
 				"bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
-				iter.pos.inode, iter.pos.offset, u.gen,
-				bch2_data_types[u.data_type],
-				u.dirty_sectors,
-				u.stripe, s.k->p.offset)) {
+				iter.pos.inode, iter.pos.offset, a->v.gen,
+				bch2_data_types[a->v.data_type],
+				a->v.dirty_sectors,
+				a->v.stripe, s.k->p.offset)) {
 			ret = -EIO;
 			goto err;
 		}
 
-		if (bch2_trans_inconsistent_on(data_type && u.dirty_sectors, trans,
+		if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
 				"bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
-				iter.pos.inode, iter.pos.offset, u.gen,
-				bch2_data_types[u.data_type],
-				u.dirty_sectors,
+				iter.pos.inode, iter.pos.offset, a->v.gen,
+				bch2_data_types[a->v.data_type],
+				a->v.dirty_sectors,
 				s.k->p.offset)) {
 			ret = -EIO;
 			goto err;
 		}
 
-		u.stripe		= s.k->p.offset;
-		u.stripe_redundancy	= s.v->nr_redundant;
+		a->v.stripe		= s.k->p.offset;
+		a->v.stripe_redundancy	= s.v->nr_redundant;
 	} else {
-		if (bch2_trans_inconsistent_on(u.stripe != s.k->p.offset ||
-					    u.stripe_redundancy != s.v->nr_redundant, trans,
+		if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
+					       a->v.stripe_redundancy != s.v->nr_redundant, trans,
 				"bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
-				iter.pos.inode, iter.pos.offset, u.gen,
-				s.k->p.offset, u.stripe)) {
+				iter.pos.inode, iter.pos.offset, a->v.gen,
+				s.k->p.offset, a->v.stripe)) {
 			ret = -EIO;
 			goto err;
 		}
 
-		u.stripe		= 0;
-		u.stripe_redundancy	= 0;
+		a->v.stripe		= 0;
+		a->v.stripe_redundancy	= 0;
 	}
 
-	u.dirty_sectors += sectors;
+	a->v.dirty_sectors += sectors;
 	if (data_type)
-		u.data_type = !deleting ? data_type : 0;
+		a->v.data_type = !deleting ? data_type : 0;
 
-	ret = bch2_alloc_write(trans, &iter, &u, 0);
+	ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
 	if (ret)
 		goto err;
 err:
@@ -1845,11 +1828,7 @@ static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
 {
 	struct bch_fs *c = trans->c;
 	struct btree_iter iter;
-	struct bkey_alloc_unpacked u;
-	struct bch_extent_ptr ptr = {
-		.dev = ca->dev_idx,
-		.offset = bucket_to_sector(ca, b),
-	};
+	struct bkey_i_alloc_v4 *a;
 	int ret = 0;
 
 	/*
@@ -1858,26 +1837,27 @@ static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
 	if (b >= ca->mi.nbuckets)
 		return 0;
 
-	ret = bch2_trans_start_alloc_update(trans, &iter, &ptr, &u);
-	if (ret)
-		return ret;
+	a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
+	if (IS_ERR(a))
+		return PTR_ERR(a);
 
-	if (u.data_type && u.data_type != type) {
+	if (a->v.data_type && a->v.data_type != type) {
 		bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 			"bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
 			"while marking %s",
-			iter.pos.inode, iter.pos.offset, u.gen,
-			bch2_data_types[u.data_type],
+			iter.pos.inode, iter.pos.offset, a->v.gen,
+			bch2_data_types[a->v.data_type],
 			bch2_data_types[type],
 			bch2_data_types[type]);
 		ret = -EIO;
 		goto out;
 	}
 
-	u.data_type	= type;
-	u.dirty_sectors	= sectors;
+	a->v.data_type		= type;
+	a->v.dirty_sectors	= sectors;
 
-	ret = bch2_alloc_write(trans, &iter, &u, 0);
+	ret = bch2_trans_update(trans, &iter, &a->k_i,
+				BTREE_UPDATE_NO_KEY_CACHE_COHERENCY);
 	if (ret)
 		goto out;
 out:
diff --git a/fs/bcachefs/buckets.h b/fs/bcachefs/buckets.h
index 90f53e677281..4937d7939c2b 100644
--- a/fs/bcachefs/buckets.h
+++ b/fs/bcachefs/buckets.h
@@ -97,6 +97,14 @@ static inline size_t PTR_BUCKET_NR(const struct bch_dev *ca,
 	return sector_to_bucket(ca, ptr->offset);
 }
 
+static inline struct bpos PTR_BUCKET_POS(const struct bch_fs *c,
+				   const struct bch_extent_ptr *ptr)
+{
+	struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
+
+	return POS(ptr->dev, PTR_BUCKET_NR(ca, ptr));
+}
+
 static inline struct bucket *PTR_GC_BUCKET(struct bch_dev *ca,
 					   const struct bch_extent_ptr *ptr)
 {
diff --git a/fs/bcachefs/buckets_types.h b/fs/bcachefs/buckets_types.h
index 2c73dc60b838..4f7018398385 100644
--- a/fs/bcachefs/buckets_types.h
+++ b/fs/bcachefs/buckets_types.h
@@ -111,7 +111,7 @@ struct copygc_heap_entry {
 	u8			dev;
 	u8			gen;
 	u8			replicas;
-	u16			fragmentation;
+	u32			fragmentation;
 	u32			sectors;
 	u64			offset;
 };
diff --git a/fs/bcachefs/movinggc.c b/fs/bcachefs/movinggc.c
index a54a83d3247b..aecec55eb421 100644
--- a/fs/bcachefs/movinggc.c
+++ b/fs/bcachefs/movinggc.c
@@ -129,7 +129,7 @@ static int walk_buckets_to_copygc(struct bch_fs *c)
 	struct btree_trans trans;
 	struct btree_iter iter;
 	struct bkey_s_c k;
-	struct bkey_alloc_unpacked u;
+	struct bch_alloc_v4 a;
 	int ret;
 
 	bch2_trans_init(&trans, c, 0, 0);
@@ -139,20 +139,20 @@ static int walk_buckets_to_copygc(struct bch_fs *c)
 		struct bch_dev *ca = bch_dev_bkey_exists(c, iter.pos.inode);
 		struct copygc_heap_entry e;
 
-		u = bch2_alloc_unpack(k);
+		bch2_alloc_to_v4(k, &a);
 
-		if (u.data_type != BCH_DATA_user ||
-		    u.dirty_sectors >= ca->mi.bucket_size ||
+		if (a.data_type != BCH_DATA_user ||
+		    a.dirty_sectors >= ca->mi.bucket_size ||
 		    bch2_bucket_is_open(c, iter.pos.inode, iter.pos.offset))
 			continue;
 
 		e = (struct copygc_heap_entry) {
 			.dev		= iter.pos.inode,
-			.gen		= u.gen,
-			.replicas	= 1 + u.stripe_redundancy,
-			.fragmentation	= u.dirty_sectors * (1U << 15)
-				/ ca->mi.bucket_size,
-			.sectors	= u.dirty_sectors,
+			.gen		= a.gen,
+			.replicas	= 1 + a.stripe_redundancy,
+			.fragmentation	= div_u64((u64) a.dirty_sectors * (1ULL << 31),
+						  ca->mi.bucket_size),
+			.sectors	= a.dirty_sectors,
 			.offset		= bucket_to_sector(ca, iter.pos.offset),
 		};
 		heap_add_or_replace(h, e, -fragmentation_cmp, NULL);
@@ -180,7 +180,7 @@ static int check_copygc_was_done(struct bch_fs *c,
 	struct btree_trans trans;
 	struct btree_iter iter;
 	struct bkey_s_c k;
-	struct bkey_alloc_unpacked u;
+	struct bch_alloc_v4 a;
 	struct copygc_heap_entry *i;
 	int ret = 0;
 
@@ -199,10 +199,10 @@ static int check_copygc_was_done(struct bch_fs *c,
 		if (ret)
 			break;
 
-		u = bch2_alloc_unpack(k);
+		bch2_alloc_to_v4(k, &a);
 
-		if (u.gen == i->gen && u.dirty_sectors) {
-			*sectors_not_moved += u.dirty_sectors;
+		if (a.gen == i->gen && a.dirty_sectors) {
+			*sectors_not_moved += a.dirty_sectors;
 			*buckets_not_moved += 1;
 		}
 	}