Update bcachefs sources to 9017d85854 bcachefs: btree_ptr_v2

35 files changed, 517 insertions, 283 deletions

diff --git a/libbcachefs/alloc_background.c b/libbcachefs/alloc_background.c
index c57df501..b2d1b8f9 100644
--- a/libbcachefs/alloc_background.c
+++ b/libbcachefs/alloc_background.c
@@ -211,33 +211,31 @@ void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c,
 int bch2_alloc_read(struct bch_fs *c, struct journal_keys *journal_keys)
 {
 	struct btree_trans trans;
-	struct btree_iter *iter;
+	struct btree_and_journal_iter iter;
 	struct bkey_s_c k;
 	struct bch_dev *ca;
-	struct journal_key *j;
 	unsigned i;
-	int ret;
+	int ret = 0;
 
 	bch2_trans_init(&trans, c, 0, 0);
 
-	for_each_btree_key(&trans, iter, BTREE_ID_ALLOC, POS_MIN, 0, k, ret)
+	bch2_btree_and_journal_iter_init(&iter, &trans, journal_keys,
+					 BTREE_ID_ALLOC, POS_MIN);
+
+	while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 		bch2_mark_key(c, k, 0, 0, NULL, 0,
 			      BTREE_TRIGGER_ALLOC_READ|
 			      BTREE_TRIGGER_NOATOMIC);
 
+		bch2_btree_and_journal_iter_advance(&iter);
+	}
+
 	ret = bch2_trans_exit(&trans) ?: ret;
 	if (ret) {
 		bch_err(c, "error reading alloc info: %i", ret);
 		return ret;
 	}
 
-	for_each_journal_key(*journal_keys, j)
-		if (j->btree_id == BTREE_ID_ALLOC)
-			bch2_mark_key(c, bkey_i_to_s_c(j->k),
-				      0, 0, NULL, 0,
-				      BTREE_TRIGGER_ALLOC_READ|
-				      BTREE_TRIGGER_NOATOMIC);
-
 	percpu_down_write(&c->mark_lock);
 	bch2_dev_usage_from_buckets(c);
 	percpu_up_write(&c->mark_lock);
diff --git a/libbcachefs/bcachefs_format.h b/libbcachefs/bcachefs_format.h
index 3b5e70a7..bb251fcb 100644
--- a/libbcachefs/bcachefs_format.h
+++ b/libbcachefs/bcachefs_format.h
@@ -339,7 +339,8 @@ static inline void bkey_init(struct bkey *k)
 	x(stripe,		14)			\
 	x(reflink_p,		15)			\
 	x(reflink_v,		16)			\
-	x(inline_data,		17)
+	x(inline_data,		17)			\
+	x(btree_ptr_v2,		18)
 
 enum bch_bkey_type {
 #define x(name, nr) KEY_TYPE_##name	= nr,
@@ -595,6 +596,19 @@ struct bch_btree_ptr {
 	__u64			_data[0];
 } __attribute__((packed, aligned(8)));
 
+struct bch_btree_ptr_v2 {
+	struct bch_val		v;
+
+	__u64			mem_ptr;
+	__le64			seq;
+	__le16			sectors_written;
+	/* In case we ever decide to do variable size btree nodes: */
+	__le16			sectors;
+	struct bpos		min_key;
+	struct bch_extent_ptr	start[0];
+	__u64			_data[0];
+} __attribute__((packed, aligned(8)));
+
 struct bch_extent {
 	struct bch_val		v;
 
@@ -626,7 +640,8 @@ struct bch_reservation {
 
 /* Btree pointers don't carry around checksums: */
 #define BKEY_BTREE_PTR_VAL_U64s_MAX				\
-	((sizeof(struct bch_extent_ptr)) / sizeof(u64) * BCH_REPLICAS_MAX)
+	((sizeof(struct bch_btree_ptr_v2) +			\
+	  sizeof(struct bch_extent_ptr) * BCH_REPLICAS_MAX) / sizeof(u64))
 #define BKEY_BTREE_PTR_U64s_MAX					\
 	(BKEY_U64s + BKEY_BTREE_PTR_VAL_U64s_MAX)
 
@@ -1295,7 +1310,8 @@ LE64_BITMASK(BCH_SB_ERASURE_CODE,	struct bch_sb, flags[3],  0, 16);
 	x(new_siphash,			7)	\
 	x(inline_data,			8)	\
 	x(new_extent_overwrite,		9)	\
-	x(incompressible,		10)
+	x(incompressible,		10)	\
+	x(btree_ptr_v2,			11)
 
 enum bch_sb_feature {
 #define x(f, n) BCH_FEATURE_##f,
diff --git a/libbcachefs/bkey.h b/libbcachefs/bkey.h
index f2d5f300..9106bea9 100644
--- a/libbcachefs/bkey.h
+++ b/libbcachefs/bkey.h
@@ -565,6 +565,7 @@ BKEY_VAL_ACCESSORS(stripe);
 BKEY_VAL_ACCESSORS(reflink_p);
 BKEY_VAL_ACCESSORS(reflink_v);
 BKEY_VAL_ACCESSORS(inline_data);
+BKEY_VAL_ACCESSORS(btree_ptr_v2);
 
 /* byte order helpers */
 
diff --git a/libbcachefs/bkey_methods.c b/libbcachefs/bkey_methods.c
index 320e17d1..c064cf46 100644
--- a/libbcachefs/bkey_methods.c
+++ b/libbcachefs/bkey_methods.c
@@ -202,15 +202,12 @@ void bch2_bkey_val_to_text(struct printbuf *out, struct bch_fs *c,
 	bch2_val_to_text(out, c, k);
 }
 
-void bch2_bkey_swab(const struct bkey_format *f,
-		    struct bkey_packed *k)
+void bch2_bkey_swab_val(struct bkey_s k)
 {
-	const struct bkey_ops *ops = &bch2_bkey_ops[k->type];
-
-	bch2_bkey_swab_key(f, k);
+	const struct bkey_ops *ops = &bch2_bkey_ops[k.k->type];
 
 	if (ops->swab)
-		ops->swab(f, k);
+		ops->swab(k);
 }
 
 bool bch2_bkey_normalize(struct bch_fs *c, struct bkey_s k)
diff --git a/libbcachefs/bkey_methods.h b/libbcachefs/bkey_methods.h
index 8568b65c..d36468b7 100644
--- a/libbcachefs/bkey_methods.h
+++ b/libbcachefs/bkey_methods.h
@@ -29,7 +29,7 @@ struct bkey_ops {
 	void		(*key_debugcheck)(struct bch_fs *, struct bkey_s_c);
 	void		(*val_to_text)(struct printbuf *, struct bch_fs *,
 				       struct bkey_s_c);
-	void		(*swab)(const struct bkey_format *, struct bkey_packed *);
+	void		(*swab)(struct bkey_s);
 	bool		(*key_normalize)(struct bch_fs *, struct bkey_s);
 	enum merge_result (*key_merge)(struct bch_fs *,
 				       struct bkey_s, struct bkey_s);
@@ -51,7 +51,7 @@ void bch2_val_to_text(struct printbuf *, struct bch_fs *,
 void bch2_bkey_val_to_text(struct printbuf *, struct bch_fs *,
 			   struct bkey_s_c);
 
-void bch2_bkey_swab(const struct bkey_format *, struct bkey_packed *);
+void bch2_bkey_swab_val(struct bkey_s);
 
 bool bch2_bkey_normalize(struct bch_fs *, struct bkey_s);
 
diff --git a/libbcachefs/bset.c b/libbcachefs/bset.c
index cf8fa59f..abf87ebd 100644
--- a/libbcachefs/bset.c
+++ b/libbcachefs/bset.c
@@ -1206,7 +1206,8 @@ void bch2_bset_insert(struct btree *b,
 	memcpy_u64s(bkeyp_val(f, where), &insert->v,
 		    bkeyp_val_u64s(f, src));
 
-	bch2_bset_fix_lookup_table(b, t, where, clobber_u64s, src->u64s);
+	if (src->u64s != clobber_u64s)
+		bch2_bset_fix_lookup_table(b, t, where, clobber_u64s, src->u64s);
 
 	bch2_verify_btree_nr_keys(b);
 }
diff --git a/libbcachefs/btree_cache.c b/libbcachefs/btree_cache.c
index 0c737f35..2c9c3c18 100644
--- a/libbcachefs/btree_cache.c
+++ b/libbcachefs/btree_cache.c
@@ -62,13 +62,13 @@ static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
 	const struct btree *b = obj;
 	const u64 *v = arg->key;
 
-	return PTR_HASH(&b->key) == *v ? 0 : 1;
+	return b->hash_val == *v ? 0 : 1;
 }
 
 static const struct rhashtable_params bch_btree_cache_params = {
 	.head_offset	= offsetof(struct btree, hash),
-	.key_offset	= offsetof(struct btree, key.v),
-	.key_len	= sizeof(struct bch_extent_ptr),
+	.key_offset	= offsetof(struct btree, hash_val),
+	.key_len	= sizeof(u64),
 	.obj_cmpfn	= bch2_btree_cache_cmp_fn,
 };
 
@@ -114,11 +114,14 @@ void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
 	rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
 
 	/* Cause future lookups for this node to fail: */
-	PTR_HASH(&b->key) = 0;
+	b->hash_val = 0;
 }
 
 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
 {
+	BUG_ON(b->hash_val);
+	b->hash_val = btree_ptr_hash_val(&b->key);
+
 	return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
 					     bch_btree_cache_params);
 }
@@ -144,8 +147,9 @@ __flatten
 static inline struct btree *btree_cache_find(struct btree_cache *bc,
 				     const struct bkey_i *k)
 {
-	return rhashtable_lookup_fast(&bc->table, &PTR_HASH(k),
-				      bch_btree_cache_params);
+	u64 v = btree_ptr_hash_val(k);
+
+	return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
 }
 
 /*
@@ -199,7 +203,7 @@ static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
 		btree_node_wait_on_io(b);
 	}
 out:
-	if (PTR_HASH(&b->key) && !ret)
+	if (b->hash_val && !ret)
 		trace_btree_node_reap(c, b);
 	return ret;
 out_unlock:
@@ -607,7 +611,7 @@ static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 		/* raced with another fill: */
 
 		/* mark as unhashed... */
-		PTR_HASH(&b->key) = 0;
+		b->hash_val = 0;
 
 		mutex_lock(&bc->lock);
 		list_add(&b->list, &bc->freeable);
@@ -710,7 +714,7 @@ retry:
 		 * free it:
 		 *
 		 * To guard against this, btree nodes are evicted from the cache
-		 * when they're freed - and PTR_HASH() is zeroed out, which we
+		 * when they're freed - and b->hash_val is zeroed out, which we
 		 * check for after we lock the node.
 		 *
 		 * Then, bch2_btree_node_relock() on the parent will fail - because
@@ -723,7 +727,7 @@ retry:
 		if (!btree_node_lock(b, k->k.p, level, iter, lock_type))
 			return ERR_PTR(-EINTR);
 
-		if (unlikely(PTR_HASH(&b->key) != PTR_HASH(k) ||
+		if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
 			     b->level != level ||
 			     race_fault())) {
 			six_unlock_type(&b->lock, lock_type);
diff --git a/libbcachefs/btree_cache.h b/libbcachefs/btree_cache.h
index 83358d6a..d27acd87 100644
--- a/libbcachefs/btree_cache.h
+++ b/libbcachefs/btree_cache.h
@@ -35,13 +35,22 @@ void bch2_fs_btree_cache_exit(struct bch_fs *);
 int bch2_fs_btree_cache_init(struct bch_fs *);
 void bch2_fs_btree_cache_init_early(struct btree_cache *);
 
-#define PTR_HASH(_k)	*((u64 *) &bkey_i_to_btree_ptr_c(_k)->v)
+static inline u64 btree_ptr_hash_val(const struct bkey_i *k)
+{
+	switch (k->k.type) {
+	case KEY_TYPE_btree_ptr:
+		return *((u64 *) bkey_i_to_btree_ptr_c(k)->v.start);
+	case KEY_TYPE_btree_ptr_v2:
+		return bkey_i_to_btree_ptr_v2_c(k)->v.seq;
+	default:
+		return 0;
+	}
+}
 
 /* is btree node in hash table? */
 static inline bool btree_node_hashed(struct btree *b)
 {
-	return b->key.k.type == KEY_TYPE_btree_ptr &&
-		PTR_HASH(&b->key);
+	return b->hash_val != 0;
 }
 
 #define for_each_cached_btree(_b, _c, _tbl, _iter, _pos)		\
diff --git a/libbcachefs/btree_gc.c b/libbcachefs/btree_gc.c
index 05879b66..3705c41f 100644
--- a/libbcachefs/btree_gc.c
+++ b/libbcachefs/btree_gc.c
@@ -124,7 +124,11 @@ static int bch2_gc_mark_key(struct bch_fs *c, struct bkey_s_c k,
 		BUG_ON(journal_seq_verify(c) &&
 		       k.k->version.lo > journal_cur_seq(&c->journal));
 
-		if (k.k->version.lo > atomic64_read(&c->key_version))
+		/* XXX change to fsck check */
+		if (fsck_err_on(k.k->version.lo > atomic64_read(&c->key_version), c,
+				"key version number higher than recorded: %llu > %llu",
+				k.k->version.lo,
+				atomic64_read(&c->key_version)))
 			atomic64_set(&c->key_version, k.k->version.lo);
 
 		if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
diff --git a/libbcachefs/btree_io.c b/libbcachefs/btree_io.c
index 5f1c3183..84fbceea 100644
--- a/libbcachefs/btree_io.c
+++ b/libbcachefs/btree_io.c
@@ -735,6 +735,15 @@ static int validate_bset(struct bch_fs *c, struct btree *b,
 			bch2_bpos_swab(&b->data->max_key);
 		}
 
+		if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
+			struct bch_btree_ptr_v2 *bp =
+				&bkey_i_to_btree_ptr_v2(&b->key)->v;
+
+			btree_err_on(bkey_cmp(b->data->min_key, bp->min_key),
+				     BTREE_ERR_MUST_RETRY, c, b, NULL,
+				     "incorrect min_key");
+		}
+
 		btree_err_on(bkey_cmp(b->data->max_key, b->key.k.p),
 			     BTREE_ERR_MUST_RETRY, c, b, i,
 			     "incorrect max key");
@@ -784,7 +793,7 @@ static int validate_bset(struct bch_fs *c, struct btree *b,
 
 	for (k = i->start;
 	     k != vstruct_last(i);) {
-		struct bkey_s_c u;
+		struct bkey_s u;
 		struct bkey tmp;
 		const char *invalid;
 
@@ -805,21 +814,24 @@ static int validate_bset(struct bch_fs *c, struct btree *b,
 		}
 
 		if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN)
-			bch2_bkey_swab(&b->format, k);
+			bch2_bkey_swab_key(&b->format, k);
 
 		if (!write &&
 		    version < bcachefs_metadata_version_bkey_renumber)
 			bch2_bkey_renumber(btree_node_type(b), k, write);
 
-		u = bkey_disassemble(b, k, &tmp);
+		u = __bkey_disassemble(b, k, &tmp);
 
-		invalid = __bch2_bkey_invalid(c, u, btree_node_type(b)) ?:
-			bch2_bkey_in_btree_node(b, u) ?:
-			(write ? bch2_bkey_val_invalid(c, u) : NULL);
+		if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN)
+			bch2_bkey_swab_val(u);
+
+		invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
+			bch2_bkey_in_btree_node(b, u.s_c) ?:
+			(write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 		if (invalid) {
 			char buf[160];
 
-			bch2_bkey_val_to_text(&PBUF(buf), c, u);
+			bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 			btree_err(BTREE_ERR_FIXABLE, c, b, i,
 				  "invalid bkey:\n%s\n%s", invalid, buf);
 
@@ -895,6 +907,15 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry
 		     BTREE_ERR_MUST_RETRY, c, b, NULL,
 		     "bad btree header");
 
+	if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
+		struct bch_btree_ptr_v2 *bp =
+			&bkey_i_to_btree_ptr_v2(&b->key)->v;
+
+		btree_err_on(b->data->keys.seq != bp->seq,
+			     BTREE_ERR_MUST_RETRY, c, b, NULL,
+			     "got wrong btree node");
+	}
+
 	while (b->written < c->opts.btree_node_size) {
 		unsigned sectors, whiteout_u64s = 0;
 		struct nonce nonce;
@@ -1002,15 +1023,15 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry
 	i = &b->data->keys;
 	for (k = i->start; k != vstruct_last(i);) {
 		struct bkey tmp;
-		struct bkey_s_c u = bkey_disassemble(b, k, &tmp);
-		const char *invalid = bch2_bkey_val_invalid(c, u);
+		struct bkey_s u = __bkey_disassemble(b, k, &tmp);
+		const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
 
 		if (invalid ||
 		    (inject_invalid_keys(c) &&
 		     !bversion_cmp(u.k->version, MAX_VERSION))) {
 			char buf[160];
 
-			bch2_bkey_val_to_text(&PBUF(buf), c, u);
+			bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 			btree_err(BTREE_ERR_FIXABLE, c, b, i,
 				  "invalid bkey %s: %s", buf, invalid);
 
@@ -1023,6 +1044,12 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry
 			continue;
 		}
 
+		if (u.k->type == KEY_TYPE_btree_ptr_v2) {
+			struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
+
+			bp.v->mem_ptr = 0;
+		}
+
 		k = bkey_next_skip_noops(k, vstruct_last(i));
 	}
 
@@ -1252,8 +1279,6 @@ static void bch2_btree_node_write_error(struct bch_fs *c,
 {
 	struct btree *b		= wbio->wbio.bio.bi_private;
 	__BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
-	struct bkey_i_btree_ptr *new_key;
-	struct bkey_s_btree_ptr bp;
 	struct bch_extent_ptr *ptr;
 	struct btree_trans trans;
 	struct btree_iter *iter;
@@ -1279,16 +1304,13 @@ retry:
 
 	bkey_copy(&tmp.k, &b->key);
 
-	new_key = bkey_i_to_btree_ptr(&tmp.k);
-	bp = btree_ptr_i_to_s(new_key);
-
 	bch2_bkey_drop_ptrs(bkey_i_to_s(&tmp.k), ptr,
 		bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
 
-	if (!bch2_bkey_nr_ptrs(bp.s_c))
+	if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&tmp.k)))
 		goto err;
 
-	ret = bch2_btree_node_update_key(c, iter, b, new_key);
+	ret = bch2_btree_node_update_key(c, iter, b, &tmp.k);
 	if (ret == -EINTR)
 		goto retry;
 	if (ret)
diff --git a/libbcachefs/btree_iter.c b/libbcachefs/btree_iter.c
index ea0555b8..c365a2af 100644
--- a/libbcachefs/btree_iter.c
+++ b/libbcachefs/btree_iter.c
@@ -408,7 +408,7 @@ static void __bch2_btree_iter_verify(struct btree_iter *iter,
 	 * For extents, the iterator may have skipped past deleted keys (but not
 	 * whiteouts)
 	 */
-	k = b->level || iter->flags & BTREE_ITER_IS_EXTENTS
+	k = b->level || btree_node_type_is_extents(iter->btree_id)
 		? bch2_btree_node_iter_prev_filter(&tmp, b, KEY_TYPE_discard)
 		: bch2_btree_node_iter_prev_all(&tmp, b);
 	if (k && bkey_iter_pos_cmp(b, k, &pos) >= 0) {
@@ -563,7 +563,7 @@ fixup_done:
 	if (!bch2_btree_node_iter_end(node_iter) &&
 	    iter_current_key_modified &&
 	    (b->level ||
-	     (iter->flags & BTREE_ITER_IS_EXTENTS))) {
+	     btree_node_type_is_extents(iter->btree_id))) {
 		struct bset_tree *t;
 		struct bkey_packed *k, *k2, *p;
 
@@ -1271,6 +1271,29 @@ static unsigned btree_iter_pos_changed(struct btree_iter *iter, int cmp)
 	return l;
 }
 
+void __bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos,
+			       bool strictly_greater)
+{
+	struct bpos old = btree_iter_search_key(iter);
+	unsigned l;
+	int cmp;
+
+	iter->flags &= ~BTREE_ITER_IS_EXTENTS;
+	iter->flags |= strictly_greater ? BTREE_ITER_IS_EXTENTS : 0;
+	iter->pos = new_pos;
+
+	cmp = bkey_cmp(btree_iter_search_key(iter), old);
+	if (!cmp)
+		return;
+
+	l = btree_iter_pos_changed(iter, cmp);
+
+	if (l != iter->level)
+		btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
+	else
+		btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
+}
+
 void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 {
 	int cmp = bkey_cmp(new_pos, iter->pos);
@@ -1709,8 +1732,7 @@ static inline void bch2_btree_iter_init(struct btree_trans *trans,
 	iter->nodes_locked		= 0;
 	iter->nodes_intent_locked	= 0;
 	for (i = 0; i < ARRAY_SIZE(iter->l); i++)
-		iter->l[i].b		= NULL;
-	iter->l[iter->level].b		= BTREE_ITER_NO_NODE_INIT;
+		iter->l[i].b		= BTREE_ITER_NO_NODE_INIT;
 
 	prefetch(c->btree_roots[btree_id].b);
 }
@@ -1729,7 +1751,12 @@ static inline void __bch2_trans_iter_free(struct btree_trans *trans,
 int bch2_trans_iter_put(struct btree_trans *trans,
 			struct btree_iter *iter)
 {
-	int ret = btree_iter_err(iter);
+	int ret;
+
+	if (IS_ERR_OR_NULL(iter))
+		return 0;
+
+	ret = btree_iter_err(iter);
 
 	if (!(trans->iters_touched & (1ULL << iter->idx)) &&
 	    !(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT))
@@ -1742,6 +1769,9 @@ int bch2_trans_iter_put(struct btree_trans *trans,
 int bch2_trans_iter_free(struct btree_trans *trans,
 			 struct btree_iter *iter)
 {
+	if (IS_ERR_OR_NULL(iter))
+		return 0;
+
 	trans->iters_touched &= ~(1ULL << iter->idx);
 
 	return bch2_trans_iter_put(trans, iter);
@@ -1939,7 +1969,8 @@ struct btree_iter *bch2_trans_get_iter(struct btree_trans *trans,
 		__btree_trans_get_iter(trans, btree_id, pos, flags);
 
 	if (!IS_ERR(iter))
-		bch2_btree_iter_set_pos(iter, pos);
+		__bch2_btree_iter_set_pos(iter, pos,
+			btree_node_type_is_extents(btree_id));
 	return iter;
 }
 
@@ -1981,8 +2012,8 @@ struct btree_iter *bch2_trans_copy_iter(struct btree_trans *trans,
 
 	trans->iters_live |= 1ULL << iter->idx;
 	/*
-	 * Don't mark it as touched, we don't need to preserve this iter since
-	 * it's cheap to copy it again:
+	 * We don't need to preserve this iter since it's cheap to copy it
+	 * again - this will cause trans_iter_put() to free it right away:
 	 */
 	trans->iters_touched &= ~(1ULL << iter->idx);
 
diff --git a/libbcachefs/btree_iter.h b/libbcachefs/btree_iter.h
index 96238092..336901f9 100644
--- a/libbcachefs/btree_iter.h
+++ b/libbcachefs/btree_iter.h
@@ -166,6 +166,7 @@ struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *);
 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *);
 
 void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *, struct bpos);
+void __bch2_btree_iter_set_pos(struct btree_iter *, struct bpos, bool);
 void bch2_btree_iter_set_pos(struct btree_iter *, struct bpos);
 
 static inline struct bpos btree_type_successor(enum btree_id id,
diff --git a/libbcachefs/btree_types.h b/libbcachefs/btree_types.h
index b7af88e0..20757d0c 100644
--- a/libbcachefs/btree_types.h
+++ b/libbcachefs/btree_types.h
@@ -64,9 +64,7 @@ struct btree_alloc {
 struct btree {
 	/* Hottest entries first */
 	struct rhash_head	hash;
-
-	/* Key/pointer for this btree node */
-	__BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
+	u64			hash_val;
 
 	struct six_lock		lock;
 
@@ -133,6 +131,9 @@ struct btree {
 #ifdef CONFIG_BCACHEFS_DEBUG
 	bool			*expensive_debug_checks;
 #endif
+
+	/* Key/pointer for this btree node */
+	__BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
 };
 
 struct btree_cache {
diff --git a/libbcachefs/btree_update.h b/libbcachefs/btree_update.h
index 2c34bae6..be4fe818 100644
--- a/libbcachefs/btree_update.h
+++ b/libbcachefs/btree_update.h
@@ -70,7 +70,7 @@ int bch2_btree_delete_range(struct bch_fs *, enum btree_id,
 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *,
 			    __le64, unsigned);
 int bch2_btree_node_update_key(struct bch_fs *, struct btree_iter *,
-			       struct btree *, struct bkey_i_btree_ptr *);
+			       struct btree *, struct bkey_i *);
 
 int bch2_trans_update(struct btree_trans *, struct btree_iter *,
 		      struct bkey_i *, enum btree_trigger_flags);
diff --git a/libbcachefs/btree_update_interior.c b/libbcachefs/btree_update_interior.c
index 748e6356..3cb3a0ff 100644
--- a/libbcachefs/btree_update_interior.c
+++ b/libbcachefs/btree_update_interior.c
@@ -332,7 +332,11 @@ retry:
 		goto retry;
 	}
 
-	bkey_btree_ptr_init(&tmp.k);
+	if (c->sb.features & (1ULL << BCH_FEATURE_btree_ptr_v2))
+		bkey_btree_ptr_v2_init(&tmp.k);
+	else
+		bkey_btree_ptr_init(&tmp.k);
+
 	bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size);
 
 	bch2_open_bucket_get(c, wp, &ob);
@@ -354,14 +358,13 @@ static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned lev
 {
 	struct bch_fs *c = as->c;
 	struct btree *b;
+	int ret;
 
 	BUG_ON(level >= BTREE_MAX_DEPTH);
 	BUG_ON(!as->reserve->nr);
 
 	b = as->reserve->b[--as->reserve->nr];
 
-	BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id));
-
 	set_btree_node_accessed(b);
 	set_btree_node_dirty(b);
 	set_btree_node_need_write(b);
@@ -372,7 +375,16 @@ static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned lev
 	b->data->flags = 0;
 	SET_BTREE_NODE_ID(b->data, as->btree_id);
 	SET_BTREE_NODE_LEVEL(b->data, level);
-	b->data->ptr = bkey_i_to_btree_ptr(&b->key)->v.start[0];
+	b->data->ptr = bch2_bkey_ptrs_c(bkey_i_to_s_c(&b->key)).start->ptr;
+
+	if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
+		struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
+
+		bp->v.mem_ptr		= 0;
+		bp->v.seq		= b->data->keys.seq;
+		bp->v.sectors_written	= 0;
+		bp->v.sectors		= cpu_to_le16(c->opts.btree_node_size);
+	}
 
 	if (c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite))
 		SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
@@ -385,10 +397,26 @@ static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned lev
 
 	btree_node_will_make_reachable(as, b);
 
+	ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
+	BUG_ON(ret);
+
 	trace_btree_node_alloc(c, b);
 	return b;
 }
 
+static void btree_set_min(struct btree *b, struct bpos pos)
+{
+	if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
+		bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
+	b->data->min_key = pos;
+}
+
+static void btree_set_max(struct btree *b, struct bpos pos)
+{
+	b->key.k.p = pos;
+	b->data->max_key = pos;
+}
+
 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 						  struct btree *b,
 						  struct bkey_format format)
@@ -397,11 +425,12 @@ struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
 
 	n = bch2_btree_node_alloc(as, b->level);
 
-	n->data->min_key	= b->data->min_key;
-	n->data->max_key	= b->data->max_key;
-	n->data->format		= format;
 	SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
 
+	btree_set_min(n, b->data->min_key);
+	btree_set_max(n, b->data->max_key);
+
+	n->data->format		= format;
 	btree_node_set_format(n, format);
 
 	bch2_btree_sort_into(as->c, n, b);
@@ -431,10 +460,9 @@ static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
 {
 	struct btree *b = bch2_btree_node_alloc(as, level);
 
-	b->data->min_key = POS_MIN;
-	b->data->max_key = POS_MAX;
+	btree_set_min(b, POS_MIN);
+	btree_set_max(b, POS_MAX);
 	b->data->format = bch2_btree_calc_format(b);
-	b->key.k.p = POS_MAX;
 
 	btree_node_set_format(b, b->data->format);
 	bch2_btree_build_aux_trees(b);
@@ -1263,10 +1291,8 @@ static struct btree *__btree_split_node(struct btree_update *as,
 
 	BUG_ON(!prev);
 
-	n1->key.k.p = bkey_unpack_pos(n1, prev);
-	n1->data->max_key = n1->key.k.p;
-	n2->data->min_key =
-		btree_type_successor(n1->btree_id, n1->key.k.p);
+	btree_set_max(n1, bkey_unpack_pos(n1, prev));
+	btree_set_min(n2, btree_type_successor(n1->btree_id, n1->key.k.p));
 
 	set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
 	set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
@@ -1749,10 +1775,9 @@ retry:
 
 	n = bch2_btree_node_alloc(as, b->level);
 
-	n->data->min_key	= prev->data->min_key;
-	n->data->max_key	= next->data->max_key;
+	btree_set_min(n, prev->data->min_key);
+	btree_set_max(n, next->data->max_key);
 	n->data->format		= new_f;
-	n->key.k.p		= next->key.k.p;
 
 	btree_node_set_format(n, new_f);
 
@@ -1944,7 +1969,7 @@ static void __bch2_btree_node_update_key(struct bch_fs *c,
 					 struct btree_update *as,
 					 struct btree_iter *iter,
 					 struct btree *b, struct btree *new_hash,
-					 struct bkey_i_btree_ptr *new_key)
+					 struct bkey_i *new_key)
 {
 	struct btree *parent;
 	int ret;
@@ -1989,20 +2014,20 @@ static void __bch2_btree_node_update_key(struct bch_fs *c,
 	 */
 	ret = bch2_disk_reservation_add(c, &as->reserve->disk_res,
 			c->opts.btree_node_size *
-			bch2_bkey_nr_ptrs(bkey_i_to_s_c(&new_key->k_i)),
+			bch2_bkey_nr_ptrs(bkey_i_to_s_c(new_key)),
 			BCH_DISK_RESERVATION_NOFAIL);
 	BUG_ON(ret);
 
 	parent = btree_node_parent(iter, b);
 	if (parent) {
 		if (new_hash) {
-			bkey_copy(&new_hash->key, &new_key->k_i);
+			bkey_copy(&new_hash->key, new_key);
 			ret = bch2_btree_node_hash_insert(&c->btree_cache,
 					new_hash, b->level, b->btree_id);
 			BUG_ON(ret);
 		}
 
-		bch2_keylist_add(&as->parent_keys, &new_key->k_i);
+		bch2_keylist_add(&as->parent_keys, new_key);
 		bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
 
 		if (new_hash) {
@@ -2011,12 +2036,12 @@ static void __bch2_btree_node_update_key(struct bch_fs *c,
 
 			bch2_btree_node_hash_remove(&c->btree_cache, b);
 
-			bkey_copy(&b->key, &new_key->k_i);
+			bkey_copy(&b->key, new_key);
 			ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 			BUG_ON(ret);
 			mutex_unlock(&c->btree_cache.lock);
 		} else {
-			bkey_copy(&b->key, &new_key->k_i);
+			bkey_copy(&b->key, new_key);
 		}
 	} else {
 		struct bch_fs_usage *fs_usage;
@@ -2029,11 +2054,11 @@ static void __bch2_btree_node_update_key(struct bch_fs *c,
 		percpu_down_read(&c->mark_lock);
 		fs_usage = bch2_fs_usage_scratch_get(c);
 
-		bch2_mark_key_locked(c, bkey_i_to_s_c(&new_key->k_i),
+		bch2_mark_key_locked(c, bkey_i_to_s_c(new_key),
 			      0, 0, fs_usage, 0,
 			      BTREE_TRIGGER_INSERT);
 		if (gc_visited(c, gc_pos_btree_root(b->btree_id)))
-			bch2_mark_key_locked(c, bkey_i_to_s_c(&new_key->k_i),
+			bch2_mark_key_locked(c, bkey_i_to_s_c(new_key),
 					     0, 0, NULL, 0,
 					     BTREE_TRIGGER_INSERT||
 					     BTREE_TRIGGER_GC);
@@ -2047,16 +2072,16 @@ static void __bch2_btree_node_update_key(struct bch_fs *c,
 		percpu_up_read(&c->mark_lock);
 		mutex_unlock(&c->btree_interior_update_lock);
 
-		if (PTR_HASH(&new_key->k_i) != PTR_HASH(&b->key)) {
+		if (btree_ptr_hash_val(new_key) != b->hash_val) {
 			mutex_lock(&c->btree_cache.lock);
 			bch2_btree_node_hash_remove(&c->btree_cache, b);
 
-			bkey_copy(&b->key, &new_key->k_i);
+			bkey_copy(&b->key, new_key);
 			ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
 			BUG_ON(ret);
 			mutex_unlock(&c->btree_cache.lock);
 		} else {
-			bkey_copy(&b->key, &new_key->k_i);
+			bkey_copy(&b->key, new_key);
 		}
 
 		btree_update_updated_root(as);
@@ -2068,7 +2093,7 @@ static void __bch2_btree_node_update_key(struct bch_fs *c,
 
 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
 			       struct btree *b,
-			       struct bkey_i_btree_ptr *new_key)
+			       struct bkey_i *new_key)
 {
 	struct btree *parent = btree_node_parent(iter, b);
 	struct btree_update *as = NULL;
@@ -2091,8 +2116,11 @@ int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
 		}
 	}
 
-	/* check PTR_HASH() after @b is locked by btree_iter_traverse(): */
-	if (PTR_HASH(&new_key->k_i) != PTR_HASH(&b->key)) {
+	/*
+	 * check btree_ptr_hash_val() after @b is locked by
+	 * btree_iter_traverse():
+	 */
+	if (btree_ptr_hash_val(new_key) != b->hash_val) {
 		/* bch2_btree_reserve_get will unlock */
 		ret = bch2_btree_cache_cannibalize_lock(c, &cl);
 		if (ret) {
@@ -2134,7 +2162,7 @@ int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
 			goto err;
 	}
 
-	ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&new_key->k_i));
+	ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key));
 	if (ret)
 		goto err_free_update;
 
@@ -2193,14 +2221,14 @@ void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
 
 	bkey_btree_ptr_init(&b->key);
 	b->key.k.p = POS_MAX;
-	PTR_HASH(&b->key) = U64_MAX - id;
+	*((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
 
 	bch2_bset_init_first(b, &b->data->keys);
 	bch2_btree_build_aux_trees(b);
 
 	b->data->flags = 0;
-	b->data->min_key = POS_MIN;
-	b->data->max_key = POS_MAX;
+	btree_set_min(b, POS_MIN);
+	btree_set_max(b, POS_MAX);
 	b->data->format = bch2_btree_calc_format(b);
 	btree_node_set_format(b, b->data->format);
 
diff --git a/libbcachefs/btree_update_leaf.c b/libbcachefs/btree_update_leaf.c
index afd2086e..a036c7dd 100644
--- a/libbcachefs/btree_update_leaf.c
+++ b/libbcachefs/btree_update_leaf.c
@@ -53,9 +53,8 @@ bool bch2_btree_bset_insert_key(struct btree_iter *iter,
 				struct btree_node_iter *node_iter,
 				struct bkey_i *insert)
 {
-	const struct bkey_format *f = &b->format;
 	struct bkey_packed *k;
-	unsigned clobber_u64s;
+	unsigned clobber_u64s = 0, new_u64s = 0;
 
 	EBUG_ON(btree_node_just_written(b));
 	EBUG_ON(bset_written(b, btree_bset_last(b)));
@@ -68,30 +67,25 @@ bool bch2_btree_bset_insert_key(struct btree_iter *iter,
 		k = NULL;
 
 	/* @k is the key being overwritten/deleted, if any: */
-
 	EBUG_ON(k && bkey_whiteout(k));
 
+	/* Deleting, but not found? nothing to do: */
+	if (bkey_whiteout(&insert->k) && !k)
+		return false;
+
 	if (bkey_whiteout(&insert->k)) {
 		/* Deleting: */
-
-		/* Not found? Nothing to do: */
-		if (!k)
-			return false;
-
 		btree_account_key_drop(b, k);
 		k->type = KEY_TYPE_deleted;
 
-		if (k->needs_whiteout) {
+		if (k->needs_whiteout)
 			push_whiteout(iter->trans->c, b, k);
-			k->needs_whiteout = false;
-		}
+		k->needs_whiteout = false;
 
 		if (k >= btree_bset_last(b)->start) {
 			clobber_u64s = k->u64s;
-
 			bch2_bset_delete(b, k, clobber_u64s);
-			bch2_btree_node_iter_fix(iter, b, node_iter, k,
-						 clobber_u64s, 0);
+			goto fix_iter;
 		} else {
 			bch2_btree_iter_fix_key_modified(iter, b, k);
 		}
@@ -101,14 +95,6 @@ bool bch2_btree_bset_insert_key(struct btree_iter *iter,
 
 	if (k) {
 		/* Overwriting: */
-		if (!bkey_written(b, k) &&
-		    bkey_val_u64s(&insert->k) == bkeyp_val_u64s(f, k)) {
-			k->type = insert->k.type;
-			memcpy_u64s(bkeyp_val(f, k), &insert->v,
-				    bkey_val_u64s(&insert->k));
-			return true;
-		}
-
 		btree_account_key_drop(b, k);
 		k->type = KEY_TYPE_deleted;
 
@@ -124,11 +110,13 @@ bool bch2_btree_bset_insert_key(struct btree_iter *iter,
 	}
 
 	k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
-	clobber_u64s = 0;
 overwrite:
 	bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
-	bch2_btree_node_iter_fix(iter, b, node_iter, k,
-				 clobber_u64s, k->u64s);
+	new_u64s = k->u64s;
+fix_iter:
+	if (clobber_u64s != new_u64s)
+		bch2_btree_node_iter_fix(iter, b, node_iter, k,
+					 clobber_u64s, new_u64s);
 	return true;
 }
 
@@ -792,7 +780,7 @@ int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 
 	iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 
-	if (iter->flags & BTREE_ITER_IS_EXTENTS) {
+	if (btree_node_type_is_extents(iter->btree_id)) {
 		iter->pos_after_commit = k->k.p;
 		iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
 	}
@@ -910,7 +898,7 @@ retry:
 		 */
 		delete.k.p = iter->pos;
 
-		if (iter->flags & BTREE_ITER_IS_EXTENTS) {
+		if (btree_node_type_is_extents(iter->btree_id)) {
 			unsigned max_sectors =
 				KEY_SIZE_MAX & (~0 << trans->c->block_bits);
 
diff --git a/libbcachefs/buckets.c b/libbcachefs/buckets.c
index 731b9325..b9bc524f 100644
--- a/libbcachefs/buckets.c
+++ b/libbcachefs/buckets.c
@@ -1194,6 +1194,7 @@ int bch2_mark_key_locked(struct bch_fs *c,
 		ret = bch2_mark_alloc(c, k, fs_usage, journal_seq, flags);
 		break;
 	case KEY_TYPE_btree_ptr:
+	case KEY_TYPE_btree_ptr_v2:
 		sectors = !(flags & BTREE_TRIGGER_OVERWRITE)
 			?  c->opts.btree_node_size
 			: -c->opts.btree_node_size;
@@ -1729,6 +1730,7 @@ int bch2_trans_mark_key(struct btree_trans *trans, struct bkey_s_c k,
 
 	switch (k.k->type) {
 	case KEY_TYPE_btree_ptr:
+	case KEY_TYPE_btree_ptr_v2:
 		sectors = !(flags & BTREE_TRIGGER_OVERWRITE)
 			?  c->opts.btree_node_size
 			: -c->opts.btree_node_size;
diff --git a/libbcachefs/buckets.h b/libbcachefs/buckets.h
index 4717a1a6..c1cc63af 100644
--- a/libbcachefs/buckets.h
+++ b/libbcachefs/buckets.h
@@ -97,7 +97,8 @@ static inline struct bucket *PTR_BUCKET(struct bch_dev *ca,
 static inline enum bch_data_type ptr_data_type(const struct bkey *k,
 					       const struct bch_extent_ptr *ptr)
 {
-	if (k->type == KEY_TYPE_btree_ptr)
+	if (k->type == KEY_TYPE_btree_ptr ||
+	    k->type == KEY_TYPE_btree_ptr_v2)
 		return BCH_DATA_BTREE;
 
 	return ptr->cached ? BCH_DATA_CACHED : BCH_DATA_USER;
diff --git a/libbcachefs/ec.c b/libbcachefs/ec.c
index a49d0745..bd813ddf 100644
--- a/libbcachefs/ec.c
+++ b/libbcachefs/ec.c
@@ -1280,9 +1280,8 @@ int bch2_stripes_write(struct bch_fs *c, unsigned flags, bool *wrote)
 int bch2_stripes_read(struct bch_fs *c, struct journal_keys *journal_keys)
 {
 	struct btree_trans trans;
-	struct btree_iter *btree_iter;
-	struct journal_iter journal_iter;
-	struct bkey_s_c btree_k, journal_k;
+	struct btree_and_journal_iter iter;
+	struct bkey_s_c k;
 	int ret;
 
 	ret = bch2_fs_ec_start(c);
@@ -1291,38 +1290,16 @@ int bch2_stripes_read(struct bch_fs *c, struct journal_keys *journal_keys)
 
 	bch2_trans_init(&trans, c, 0, 0);
 
-	btree_iter	= bch2_trans_get_iter(&trans, BTREE_ID_EC, POS_MIN, 0);
-	journal_iter	= bch2_journal_iter_init(journal_keys, BTREE_ID_EC);
+	bch2_btree_and_journal_iter_init(&iter, &trans, journal_keys,
+					 BTREE_ID_EC, POS_MIN);
 
-	btree_k		= bch2_btree_iter_peek(btree_iter);
-	journal_k	= bch2_journal_iter_peek(&journal_iter);
 
-	while (1) {
-		bool btree;
-
-		if (btree_k.k && journal_k.k) {
-			int cmp = bkey_cmp(btree_k.k->p, journal_k.k->p);
-
-			if (!cmp)
-				btree_k = bch2_btree_iter_next(btree_iter);
-			btree = cmp < 0;
-		} else if (btree_k.k) {
-			btree = true;
-		} else if (journal_k.k) {
-			btree = false;
-		} else {
-			break;
-		}
-
-		bch2_mark_key(c, btree ? btree_k : journal_k,
-			      0, 0, NULL, 0,
+	while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
+		bch2_mark_key(c, k, 0, 0, NULL, 0,
 			      BTREE_TRIGGER_ALLOC_READ|
 			      BTREE_TRIGGER_NOATOMIC);
 
-		if (btree)
-			btree_k = bch2_btree_iter_next(btree_iter);
-		else
-			journal_k = bch2_journal_iter_next(&journal_iter);
+		bch2_btree_and_journal_iter_advance(&iter);
 	}
 
 	ret = bch2_trans_exit(&trans) ?: ret;
diff --git a/libbcachefs/ec.h b/libbcachefs/ec.h
index 8d9fbfd1..cf67abd4 100644
--- a/libbcachefs/ec.h
+++ b/libbcachefs/ec.h
@@ -12,6 +12,7 @@ void bch2_stripe_to_text(struct printbuf *, struct bch_fs *,
 #define bch2_bkey_ops_stripe (struct bkey_ops) {	\
 	.key_invalid	= bch2_stripe_invalid,		\
 	.val_to_text	= bch2_stripe_to_text,		\
+	.swab		= bch2_ptr_swab,		\
 }
 
 static inline unsigned stripe_csums_per_device(const struct bch_stripe *s)
diff --git a/libbcachefs/extents.c b/libbcachefs/extents.c
index a19b91f9..10feb856 100644
--- a/libbcachefs/extents.c
+++ b/libbcachefs/extents.c
@@ -749,6 +749,7 @@ void bch2_bkey_append_ptr(struct bkey_i *k,
 
 	switch (k->k.type) {
 	case KEY_TYPE_btree_ptr:
+	case KEY_TYPE_btree_ptr_v2:
 	case KEY_TYPE_extent:
 		EBUG_ON(bkey_val_u64s(&k->k) >= BKEY_EXTENT_VAL_U64s_MAX);
 
@@ -1031,6 +1032,8 @@ const char *bch2_bkey_ptrs_invalid(const struct bch_fs *c, struct bkey_s_c k)
 
 	if (k.k->type == KEY_TYPE_btree_ptr)
 		size_ondisk = c->opts.btree_node_size;
+	if (k.k->type == KEY_TYPE_btree_ptr_v2)
+		size_ondisk = le16_to_cpu(bkey_s_c_to_btree_ptr_v2(k).v->sectors);
 
 	bkey_extent_entry_for_each(ptrs, entry) {
 		if (__extent_entry_type(entry) >= BCH_EXTENT_ENTRY_MAX)
@@ -1079,17 +1082,19 @@ const char *bch2_bkey_ptrs_invalid(const struct bch_fs *c, struct bkey_s_c k)
 	return NULL;
 }
 
-void bch2_ptr_swab(const struct bkey_format *f, struct bkey_packed *k)
+void bch2_ptr_swab(struct bkey_s k)
 {
+	struct bkey_ptrs ptrs = bch2_bkey_ptrs(k);
 	union bch_extent_entry *entry;
-	u64 *d = (u64 *) bkeyp_val(f, k);
-	unsigned i;
+	u64 *d;
 
-	for (i = 0; i < bkeyp_val_u64s(f, k); i++)
-		d[i] = swab64(d[i]);
+	for (d =  (u64 *) ptrs.start;
+	     d != (u64 *) ptrs.end;
+	     d++)
+		*d = swab64(*d);
 
-	for (entry = (union bch_extent_entry *) d;
-	     entry < (union bch_extent_entry *) (d + bkeyp_val_u64s(f, k));
+	for (entry = ptrs.start;
+	     entry < ptrs.end;
 	     entry = extent_entry_next(entry)) {
 		switch (extent_entry_type(entry)) {
 		case BCH_EXTENT_ENTRY_ptr:
diff --git a/libbcachefs/extents.h b/libbcachefs/extents.h
index 0d855417..70b7d702 100644
--- a/libbcachefs/extents.h
+++ b/libbcachefs/extents.h
@@ -225,6 +225,13 @@ static inline struct bkey_ptrs_c bch2_bkey_ptrs_c(struct bkey_s_c k)
 			bkey_val_end(r),
 		};
 	}
+	case KEY_TYPE_btree_ptr_v2: {
+		struct bkey_s_c_btree_ptr_v2 e = bkey_s_c_to_btree_ptr_v2(k);
+		return (struct bkey_ptrs_c) {
+			to_entry(&e.v->start[0]),
+			to_entry(extent_entry_last(e))
+		};
+	}
 	default:
 		return (struct bkey_ptrs_c) { NULL, NULL };
 	}
@@ -372,6 +379,13 @@ void bch2_btree_ptr_to_text(struct printbuf *, struct bch_fs *,
 	.swab		= bch2_ptr_swab,			\
 }
 
+#define bch2_bkey_ops_btree_ptr_v2 (struct bkey_ops) {		\
+	.key_invalid	= bch2_btree_ptr_invalid,		\
+	.key_debugcheck	= bch2_btree_ptr_debugcheck,		\
+	.val_to_text	= bch2_btree_ptr_to_text,		\
+	.swab		= bch2_ptr_swab,			\
+}
+
 /* KEY_TYPE_extent: */
 
 const char *bch2_extent_invalid(const struct bch_fs *, struct bkey_s_c);
@@ -416,6 +430,7 @@ static inline bool bkey_extent_is_direct_data(const struct bkey *k)
 {
 	switch (k->type) {
 	case KEY_TYPE_btree_ptr:
+	case KEY_TYPE_btree_ptr_v2:
 	case KEY_TYPE_extent:
 	case KEY_TYPE_reflink_v:
 		return true;
@@ -532,7 +547,7 @@ void bch2_bkey_ptrs_to_text(struct printbuf *, struct bch_fs *,
 			    struct bkey_s_c);
 const char *bch2_bkey_ptrs_invalid(const struct bch_fs *, struct bkey_s_c);
 
-void bch2_ptr_swab(const struct bkey_format *, struct bkey_packed *);
+void bch2_ptr_swab(struct bkey_s);
 
 /* Generic extent code: */
 
diff --git a/libbcachefs/fs-common.c b/libbcachefs/fs-common.c
index 96f7bbe0..878419d4 100644
--- a/libbcachefs/fs-common.c
+++ b/libbcachefs/fs-common.c
@@ -19,14 +19,15 @@ int bch2_create_trans(struct btree_trans *trans, u64 dir_inum,
 		      struct posix_acl *acl)
 {
 	struct bch_fs *c = trans->c;
-	struct btree_iter *dir_iter;
+	struct btree_iter *dir_iter = NULL;
 	struct bch_hash_info hash = bch2_hash_info_init(c, new_inode);
 	u64 now = bch2_current_time(trans->c);
 	int ret;
 
 	dir_iter = bch2_inode_peek(trans, dir_u, dir_inum, BTREE_ITER_INTENT);
-	if (IS_ERR(dir_iter))
-		return PTR_ERR(dir_iter);
+	ret = PTR_ERR_OR_ZERO(dir_iter);
+	if (ret)
+		goto err;
 
 	bch2_inode_init_late(new_inode, now, uid, gid, mode, rdev, dir_u);
 
@@ -37,20 +38,20 @@ int bch2_create_trans(struct btree_trans *trans, u64 dir_inum,
 				BLOCKDEV_INODE_MAX, 0,
 				&c->unused_inode_hint);
 	if (ret)
-		return ret;
+		goto err;
 
 	if (default_acl) {
 		ret = bch2_set_acl_trans(trans, new_inode, &hash,
 					 default_acl, ACL_TYPE_DEFAULT);
 		if (ret)
-			return ret;
+			goto err;
 	}
 
 	if (acl) {
 		ret = bch2_set_acl_trans(trans, new_inode, &hash,
 					 acl, ACL_TYPE_ACCESS);
 		if (ret)
-			return ret;
+			goto err;
 	}
 
 	if (name) {
@@ -62,48 +63,55 @@ int bch2_create_trans(struct btree_trans *trans, u64 dir_inum,
 
 		ret = bch2_inode_write(trans, dir_iter, dir_u);
 		if (ret)
-			return ret;
+			goto err;
 
 		ret = bch2_dirent_create(trans, dir_inum, &dir_hash,
 					 mode_to_type(new_inode->bi_mode),
 					 name, new_inode->bi_inum,
 					 BCH_HASH_SET_MUST_CREATE);
 		if (ret)
-			return ret;
+			goto err;
 	}
-
-	return 0;
+err:
+	bch2_trans_iter_put(trans, dir_iter);
+	return ret;
 }
 
 int bch2_link_trans(struct btree_trans *trans, u64 dir_inum,
 		    u64 inum, struct bch_inode_unpacked *dir_u,
 		    struct bch_inode_unpacked *inode_u, const struct qstr *name)
 {
-	struct btree_iter *dir_iter, *inode_iter;
+	struct btree_iter *dir_iter = NULL, *inode_iter = NULL;
 	struct bch_hash_info dir_hash;
 	u64 now = bch2_current_time(trans->c);
+	int ret;
 
 	inode_iter = bch2_inode_peek(trans, inode_u, inum, BTREE_ITER_INTENT);
-	if (IS_ERR(inode_iter))
-		return PTR_ERR(inode_iter);
+	ret = PTR_ERR_OR_ZERO(inode_iter);
+	if (ret)
+		goto err;
 
 	inode_u->bi_ctime = now;
 	bch2_inode_nlink_inc(inode_u);
 
 	dir_iter = bch2_inode_peek(trans, dir_u, dir_inum, 0);
-	if (IS_ERR(dir_iter))
-		return PTR_ERR(dir_iter);
+	ret = PTR_ERR_OR_ZERO(dir_iter);
+	if (ret)
+		goto err;
 
 	dir_u->bi_mtime = dir_u->bi_ctime = now;
 
 	dir_hash = bch2_hash_info_init(trans->c, dir_u);
-	bch2_trans_iter_put(trans, dir_iter);
 
-	return bch2_dirent_create(trans, dir_inum, &dir_hash,
+	ret =   bch2_dirent_create(trans, dir_inum, &dir_hash,
 				  mode_to_type(inode_u->bi_mode),
 				  name, inum, BCH_HASH_SET_MUST_CREATE) ?:
 		bch2_inode_write(trans, dir_iter, dir_u) ?:
 		bch2_inode_write(trans, inode_iter, inode_u);
+err:
+	bch2_trans_iter_put(trans, dir_iter);
+	bch2_trans_iter_put(trans, inode_iter);
+	return ret;
 }
 
 int bch2_unlink_trans(struct btree_trans *trans,
@@ -111,39 +119,49 @@ int bch2_unlink_trans(struct btree_trans *trans,
 		      struct bch_inode_unpacked *inode_u,
 		      const struct qstr *name)
 {
-	struct btree_iter *dir_iter, *dirent_iter, *inode_iter;
+	struct btree_iter *dir_iter = NULL, *dirent_iter = NULL,
+			  *inode_iter = NULL;
 	struct bch_hash_info dir_hash;
 	u64 inum, now = bch2_current_time(trans->c);
 	struct bkey_s_c k;
+	int ret;
 
 	dir_iter = bch2_inode_peek(trans, dir_u, dir_inum, BTREE_ITER_INTENT);
-	if (IS_ERR(dir_iter))
-		return PTR_ERR(dir_iter);
+	ret = PTR_ERR_OR_ZERO(dir_iter);
+	if (ret)
+		goto err;
 
 	dir_hash = bch2_hash_info_init(trans->c, dir_u);
 
 	dirent_iter = __bch2_dirent_lookup_trans(trans, dir_inum, &dir_hash,
 						 name, BTREE_ITER_INTENT);
-	if (IS_ERR(dirent_iter))
-		return PTR_ERR(dirent_iter);
+	ret = PTR_ERR_OR_ZERO(dirent_iter);
+	if (ret)
+		goto err;
 
 	k = bch2_btree_iter_peek_slot(dirent_iter);
 	inum = le64_to_cpu(bkey_s_c_to_dirent(k).v->d_inum);
 
 	inode_iter = bch2_inode_peek(trans, inode_u, inum, BTREE_ITER_INTENT);
-	if (IS_ERR(inode_iter))
-		return PTR_ERR(inode_iter);
+	ret = PTR_ERR_OR_ZERO(inode_iter);
+	if (ret)
+		goto err;
 
 	dir_u->bi_mtime = dir_u->bi_ctime = inode_u->bi_ctime = now;
 	dir_u->bi_nlink -= S_ISDIR(inode_u->bi_mode);
 	bch2_inode_nlink_dec(inode_u);
 
-	return  (S_ISDIR(inode_u->bi_mode)
+	ret =   (S_ISDIR(inode_u->bi_mode)
 		 ? bch2_empty_dir_trans(trans, inum)
 		 : 0) ?:
 		bch2_dirent_delete_at(trans, &dir_hash, dirent_iter) ?:
 		bch2_inode_write(trans, dir_iter, dir_u) ?:
 		bch2_inode_write(trans, inode_iter, inode_u);
+err:
+	bch2_trans_iter_put(trans, inode_iter);
+	bch2_trans_iter_put(trans, dirent_iter);
+	bch2_trans_iter_put(trans, dir_iter);
+	return ret;
 }
 
 bool bch2_reinherit_attrs(struct bch_inode_unpacked *dst_u,
@@ -179,24 +197,26 @@ int bch2_rename_trans(struct btree_trans *trans,
 		      const struct qstr *dst_name,
 		      enum bch_rename_mode mode)
 {
-	struct btree_iter *src_dir_iter, *dst_dir_iter = NULL;
-	struct btree_iter *src_inode_iter, *dst_inode_iter = NULL;
+	struct btree_iter *src_dir_iter = NULL, *dst_dir_iter = NULL;
+	struct btree_iter *src_inode_iter = NULL, *dst_inode_iter = NULL;
 	struct bch_hash_info src_hash, dst_hash;
 	u64 src_inode, dst_inode, now = bch2_current_time(trans->c);
 	int ret;
 
 	src_dir_iter = bch2_inode_peek(trans, src_dir_u, src_dir,
 				       BTREE_ITER_INTENT);
-	if (IS_ERR(src_dir_iter))
-		return PTR_ERR(src_dir_iter);
+	ret = PTR_ERR_OR_ZERO(src_dir_iter);
+	if (ret)
+		goto err;
 
 	src_hash = bch2_hash_info_init(trans->c, src_dir_u);
 
 	if (dst_dir != src_dir) {
 		dst_dir_iter = bch2_inode_peek(trans, dst_dir_u, dst_dir,
 					       BTREE_ITER_INTENT);
-		if (IS_ERR(dst_dir_iter))
-			return PTR_ERR(dst_dir_iter);
+		ret = PTR_ERR_OR_ZERO(dst_dir_iter);
+		if (ret)
+			goto err;
 
 		dst_hash = bch2_hash_info_init(trans->c, dst_dir_u);
 	} else {
@@ -211,38 +231,48 @@ int bch2_rename_trans(struct btree_trans *trans,
 				 dst_name, &dst_inode,
 				 mode);
 	if (ret)
-		return ret;
+		goto err;
 
 	src_inode_iter = bch2_inode_peek(trans, src_inode_u, src_inode,
 					 BTREE_ITER_INTENT);
-	if (IS_ERR(src_inode_iter))
-		return PTR_ERR(src_inode_iter);
+	ret = PTR_ERR_OR_ZERO(src_inode_iter);
+	if (ret)
+		goto err;
 
 	if (dst_inode) {
 		dst_inode_iter = bch2_inode_peek(trans, dst_inode_u, dst_inode,
 						 BTREE_ITER_INTENT);
-		if (IS_ERR(dst_inode_iter))
-			return PTR_ERR(dst_inode_iter);
+		ret = PTR_ERR_OR_ZERO(dst_inode_iter);
+		if (ret)
+			goto err;
 	}
 
 	if (mode == BCH_RENAME_OVERWRITE) {
 		if (S_ISDIR(src_inode_u->bi_mode) !=
-		    S_ISDIR(dst_inode_u->bi_mode))
-			return -ENOTDIR;
+		    S_ISDIR(dst_inode_u->bi_mode)) {
+			ret = -ENOTDIR;
+			goto err;
+		}
 
 		if (S_ISDIR(dst_inode_u->bi_mode) &&
-		    bch2_empty_dir_trans(trans, dst_inode))
-			return -ENOTEMPTY;
+		    bch2_empty_dir_trans(trans, dst_inode)) {
+			ret = -ENOTEMPTY;
+			goto err;
+		}
 	}
 
 	if (bch2_reinherit_attrs(src_inode_u, dst_dir_u) &&
-	    S_ISDIR(src_inode_u->bi_mode))
-		return -EXDEV;
+	    S_ISDIR(src_inode_u->bi_mode)) {
+		ret = -EXDEV;
+		goto err;
+	}
 
 	if (mode == BCH_RENAME_EXCHANGE &&
 	    bch2_reinherit_attrs(dst_inode_u, src_dir_u) &&
-	    S_ISDIR(dst_inode_u->bi_mode))
-		return -EXDEV;
+	    S_ISDIR(dst_inode_u->bi_mode)) {
+		ret = -EXDEV;
+		goto err;
+	}
 
 	if (S_ISDIR(src_inode_u->bi_mode)) {
 		src_dir_u->bi_nlink--;
@@ -270,7 +300,7 @@ int bch2_rename_trans(struct btree_trans *trans,
 	if (dst_inode)
 		dst_inode_u->bi_ctime	= now;
 
-	return  bch2_inode_write(trans, src_dir_iter, src_dir_u) ?:
+	ret =   bch2_inode_write(trans, src_dir_iter, src_dir_u) ?:
 		(src_dir != dst_dir
 		 ? bch2_inode_write(trans, dst_dir_iter, dst_dir_u)
 		 : 0 ) ?:
@@ -278,4 +308,10 @@ int bch2_rename_trans(struct btree_trans *trans,
 		(dst_inode
 		 ? bch2_inode_write(trans, dst_inode_iter, dst_inode_u)
 		 : 0 );
+err:
+	bch2_trans_iter_put(trans, dst_inode_iter);
+	bch2_trans_iter_put(trans, src_inode_iter);
+	bch2_trans_iter_put(trans, dst_dir_iter);
+	bch2_trans_iter_put(trans, src_dir_iter);
+	return ret;
 }
diff --git a/libbcachefs/inode.c b/libbcachefs/inode.c
index e811b98d..26171ff7 100644
--- a/libbcachefs/inode.c
+++ b/libbcachefs/inode.c
@@ -362,16 +362,16 @@ int bch2_inode_create(struct btree_trans *trans,
 		      struct bch_inode_unpacked *inode_u,
 		      u64 min, u64 max, u64 *hint)
 {
-	struct bch_fs *c = trans->c;
 	struct bkey_inode_buf *inode_p;
-	struct btree_iter *iter;
+	struct btree_iter *iter = NULL;
+	struct bkey_s_c k;
 	u64 start;
 	int ret;
 
 	if (!max)
 		max = ULLONG_MAX;
 
-	if (c->opts.inodes_32bit)
+	if (trans->c->opts.inodes_32bit)
 		max = min_t(u64, max, U32_MAX);
 
 	start = READ_ONCE(*hint);
@@ -382,48 +382,37 @@ int bch2_inode_create(struct btree_trans *trans,
 	inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 	if (IS_ERR(inode_p))
 		return PTR_ERR(inode_p);
-
-	iter = bch2_trans_get_iter(trans,
-			BTREE_ID_INODES, POS(start, 0),
-			BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
-	if (IS_ERR(iter))
-		return PTR_ERR(iter);
 again:
-	while (1) {
-		struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
-
-		ret = bkey_err(k);
-		if (ret)
-			return ret;
+	for_each_btree_key(trans, iter, BTREE_ID_INODES, POS(start, 0),
+			   BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
+		if (iter->pos.inode > max)
+			break;
 
-		switch (k.k->type) {
-		case KEY_TYPE_inode:
-			/* slot used */
-			if (iter->pos.inode >= max)
-				goto out;
+		if (k.k->type != KEY_TYPE_inode)
+			goto found_slot;
+	}
 
-			bch2_btree_iter_next_slot(iter);
-			break;
+	bch2_trans_iter_put(trans, iter);
 
-		default:
-			*hint			= k.k->p.inode;
-			inode_u->bi_inum	= k.k->p.inode;
-			inode_u->bi_generation	= bkey_generation(k);
+	if (ret)
+		return ret;
 
-			bch2_inode_pack(inode_p, inode_u);
-			bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
-			return 0;
-		}
-	}
-out:
 	if (start != min) {
 		/* Retry from start */
 		start = min;
-		bch2_btree_iter_set_pos(iter, POS(start, 0));
 		goto again;
 	}
 
 	return -ENOSPC;
+found_slot:
+	*hint			= k.k->p.inode;
+	inode_u->bi_inum	= k.k->p.inode;
+	inode_u->bi_generation	= bkey_generation(k);
+
+	bch2_inode_pack(inode_p, inode_u);
+	bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
+	bch2_trans_iter_put(trans, iter);
+	return 0;
 }
 
 int bch2_inode_rm(struct bch_fs *c, u64 inode_nr)
@@ -518,14 +507,13 @@ int bch2_inode_find_by_inum_trans(struct btree_trans *trans, u64 inode_nr,
 	k = bch2_btree_iter_peek_slot(iter);
 	ret = bkey_err(k);
 	if (ret)
-		return ret;
+		goto err;
 
 	ret = k.k->type == KEY_TYPE_inode
 		? bch2_inode_unpack(bkey_s_c_to_inode(k), inode)
 		: -ENOENT;
-
+err:
 	bch2_trans_iter_put(trans, iter);
-
 	return ret;
 }
 
diff --git a/libbcachefs/io.c b/libbcachefs/io.c
index ca83d79e..0d3366c2 100644
--- a/libbcachefs/io.c
+++ b/libbcachefs/io.c
@@ -942,7 +942,7 @@ static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
 
 		if (bch2_csum_type_is_encryption(op->csum_type)) {
 			if (bversion_zero(version)) {
-				version.lo = atomic64_inc_return(&c->key_version) + 1;
+				version.lo = atomic64_inc_return(&c->key_version);
 			} else {
 				crc.nonce = op->nonce;
 				op->nonce += src_len >> 9;
diff --git a/libbcachefs/io.h b/libbcachefs/io.h
index 45c95094..37f7fa61 100644
--- a/libbcachefs/io.h
+++ b/libbcachefs/io.h
@@ -78,6 +78,7 @@ static inline void bch2_write_op_init(struct bch_write_op *op, struct bch_fs *c,
 	op->nr_replicas		= 0;
 	op->nr_replicas_required = c->opts.data_replicas_required;
 	op->alloc_reserve	= RESERVE_NONE;
+	op->incompressible	= 0;
 	op->open_buckets.nr	= 0;
 	op->devs_have.nr	= 0;
 	op->target		= 0;
diff --git a/libbcachefs/io_types.h b/libbcachefs/io_types.h
index 7f7b69b3..684e4c9a 100644
--- a/libbcachefs/io_types.h
+++ b/libbcachefs/io_types.h
@@ -104,7 +104,7 @@ struct bch_write_op {
 	unsigned		compression_type:4;
 	unsigned		nr_replicas:4;
 	unsigned		nr_replicas_required:4;
-	unsigned		alloc_reserve:4;
+	unsigned		alloc_reserve:3;
 	unsigned		incompressible:1;
 
 	struct bch_devs_list	devs_have;
diff --git a/libbcachefs/journal_io.c b/libbcachefs/journal_io.c
index 7112a25d..db722a8a 100644
--- a/libbcachefs/journal_io.c
+++ b/libbcachefs/journal_io.c
@@ -171,8 +171,10 @@ static int journal_validate_key(struct bch_fs *c, struct jset *jset,
 		return 0;
 	}
 
-	if (JSET_BIG_ENDIAN(jset) != CPU_BIG_ENDIAN)
-		bch2_bkey_swab(NULL, bkey_to_packed(k));
+	if (JSET_BIG_ENDIAN(jset) != CPU_BIG_ENDIAN) {
+		bch2_bkey_swab_key(NULL, bkey_to_packed(k));
+		bch2_bkey_swab_val(bkey_i_to_s(k));
+	}
 
 	if (!write &&
 	    version < bcachefs_metadata_version_bkey_renumber)
diff --git a/libbcachefs/migrate.c b/libbcachefs/migrate.c
index 1ef62a18..e26fa160 100644
--- a/libbcachefs/migrate.c
+++ b/libbcachefs/migrate.c
@@ -123,23 +123,21 @@ static int bch2_dev_metadata_drop(struct bch_fs *c, unsigned dev_idx, int flags)
 		for_each_btree_node(&trans, iter, id, POS_MIN,
 				    BTREE_ITER_PREFETCH, b) {
 			__BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
-			struct bkey_i_btree_ptr *new_key;
 retry:
 			if (!bch2_bkey_has_device(bkey_i_to_s_c(&b->key),
 						  dev_idx))
 				continue;
 
 			bkey_copy(&tmp.k, &b->key);
-			new_key = bkey_i_to_btree_ptr(&tmp.k);
 
-			ret = drop_dev_ptrs(c, bkey_i_to_s(&new_key->k_i),
+			ret = drop_dev_ptrs(c, bkey_i_to_s(&tmp.k),
 					    dev_idx, flags, true);
 			if (ret) {
 				bch_err(c, "Cannot drop device without losing data");
 				goto err;
 			}
 
-			ret = bch2_btree_node_update_key(c, iter, b, new_key);
+			ret = bch2_btree_node_update_key(c, iter, b, &tmp.k);
 			if (ret == -EINTR) {
 				b = bch2_btree_iter_peek_node(iter);
 				goto retry;
diff --git a/libbcachefs/recovery.c b/libbcachefs/recovery.c
index 8ecd4abc..c9d12f7c 100644
--- a/libbcachefs/recovery.c
+++ b/libbcachefs/recovery.c
@@ -27,26 +27,15 @@
 
 /* iterate over keys read from the journal: */
 
-struct journal_iter bch2_journal_iter_init(struct journal_keys *keys,
-					   enum btree_id id)
-{
-	return (struct journal_iter) {
-		.keys		= keys,
-		.k		= keys->d,
-		.btree_id	= id,
-	};
-}
-
 struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *iter)
 {
-	while (1) {
-		if (iter->k == iter->keys->d + iter->keys->nr)
-			return bkey_s_c_null;
-
+	while (iter->k) {
 		if (iter->k->btree_id == iter->btree_id)
 			return bkey_i_to_s_c(iter->k->k);
 
 		iter->k++;
+		if (iter->k == iter->keys->d + iter->keys->nr)
+			iter->k = NULL;
 	}
 
 	return bkey_s_c_null;
@@ -54,13 +43,110 @@ struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *iter)
 
 struct bkey_s_c bch2_journal_iter_next(struct journal_iter *iter)
 {
-	if (iter->k == iter->keys->d + iter->keys->nr)
+	if (!iter->k)
 		return bkey_s_c_null;
 
 	iter->k++;
+	if (iter->k == iter->keys->d + iter->keys->nr)
+		iter->k = NULL;
+
 	return bch2_journal_iter_peek(iter);
 }
 
+void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *iter)
+{
+	switch (iter->last) {
+	case none:
+		break;
+	case btree:
+		bch2_btree_iter_next(iter->btree);
+		break;
+	case journal:
+		bch2_journal_iter_next(&iter->journal);
+		break;
+	}
+
+	iter->last = none;
+}
+
+struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *iter)
+{
+	struct bkey_s_c ret;
+
+	while (1) {
+		struct bkey_s_c btree_k		= bch2_btree_iter_peek(iter->btree);
+		struct bkey_s_c journal_k	= bch2_journal_iter_peek(&iter->journal);
+
+		if (btree_k.k && journal_k.k) {
+			int cmp = bkey_cmp(btree_k.k->p, journal_k.k->p);
+
+			if (!cmp)
+				bch2_btree_iter_next(iter->btree);
+
+			iter->last = cmp < 0 ? btree : journal;
+		} else if (btree_k.k) {
+			iter->last = btree;
+		} else if (journal_k.k) {
+			iter->last = journal;
+		} else {
+			iter->last = none;
+			return bkey_s_c_null;
+		}
+
+		ret = iter->last == journal ? journal_k : btree_k;
+		if (!bkey_deleted(ret.k))
+			break;
+
+		bch2_btree_and_journal_iter_advance(iter);
+	}
+
+	return ret;
+}
+
+struct bkey_s_c bch2_btree_and_journal_iter_next(struct btree_and_journal_iter *iter)
+{
+	bch2_btree_and_journal_iter_advance(iter);
+
+	return bch2_btree_and_journal_iter_peek(iter);
+}
+
+struct journal_key *journal_key_search(struct journal_keys *journal_keys,
+				       enum btree_id id, struct bpos pos)
+{
+	size_t l = 0, r = journal_keys->nr, m;
+
+	while (l < r) {
+		m = l + ((r - l) >> 1);
+		if ((cmp_int(id, journal_keys->d[m].btree_id) ?:
+		     bkey_cmp(pos, journal_keys->d[m].k->k.p)) > 0)
+			l = m + 1;
+		else
+			r = m;
+	}
+
+	BUG_ON(l < journal_keys->nr &&
+	       (cmp_int(id, journal_keys->d[l].btree_id) ?:
+		bkey_cmp(pos, journal_keys->d[l].k->k.p)) > 0);
+
+	BUG_ON(l &&
+	       (cmp_int(id, journal_keys->d[l - 1].btree_id) ?:
+		bkey_cmp(pos, journal_keys->d[l - 1].k->k.p)) <= 0);
+
+	return l < journal_keys->nr ? journal_keys->d + l : NULL;
+}
+
+void bch2_btree_and_journal_iter_init(struct btree_and_journal_iter *iter,
+				      struct btree_trans *trans,
+				      struct journal_keys *journal_keys,
+				      enum btree_id id, struct bpos pos)
+{
+	iter->journal.keys	= journal_keys;
+	iter->journal.k		= journal_key_search(journal_keys, id, pos);
+	iter->journal.btree_id	= id;
+
+	iter->btree = bch2_trans_get_iter(trans, id, pos, 0);
+}
+
 /* sort and dedup all keys in the journal: */
 
 static void journal_entries_free(struct list_head *list)
@@ -924,6 +1010,7 @@ int bch2_fs_recovery(struct bch_fs *c)
 		c->disk_sb.sb->version = le16_to_cpu(bcachefs_metadata_version_current);
 		c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_new_siphash;
 		c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_new_extent_overwrite;
+		c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_btree_ptr_v2;
 		write_sb = true;
 	}
 
diff --git a/libbcachefs/recovery.h b/libbcachefs/recovery.h
index 479ea46f..ccd84a8f 100644
--- a/libbcachefs/recovery.h
+++ b/libbcachefs/recovery.h
@@ -24,10 +24,28 @@ struct journal_iter {
 	enum btree_id		btree_id;
 };
 
-struct journal_iter bch2_journal_iter_init(struct journal_keys *,
-					   enum btree_id);
-struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *);
-struct bkey_s_c bch2_journal_iter_next(struct journal_iter *);
+struct btree_and_journal_iter {
+	enum btree_id		btree_id;
+
+	struct btree_iter	*btree;
+	struct journal_iter	journal;
+
+	enum last_key_returned {
+		none,
+		btree,
+		journal,
+	}			last;
+};
+
+void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *);
+struct bkey_s_c bch2_btree_and_journal_iter_peek(struct btree_and_journal_iter *);
+struct bkey_s_c bch2_btree_and_journal_iter_next(struct btree_and_journal_iter *);
+struct journal_key *journal_key_search(struct journal_keys *,
+				       enum btree_id, struct bpos);
+void bch2_btree_and_journal_iter_init(struct btree_and_journal_iter *,
+				      struct btree_trans *,
+				      struct journal_keys *,
+				      enum btree_id, struct bpos);
 
 int bch2_fs_recovery(struct bch_fs *);
 int bch2_fs_initialize(struct bch_fs *);
diff --git a/libbcachefs/reflink.c b/libbcachefs/reflink.c
index 3b8c74ca..d78a3d5f 100644
--- a/libbcachefs/reflink.c
+++ b/libbcachefs/reflink.c
@@ -128,10 +128,9 @@ static int bch2_make_extent_indirect(struct btree_trans *trans,
 
 	bch2_trans_update(trans, extent_iter, &r_p->k_i, 0);
 err:
-	if (!IS_ERR(reflink_iter)) {
+	if (!IS_ERR(reflink_iter))
 		c->reflink_hint = reflink_iter->pos.offset;
-		bch2_trans_iter_put(trans, reflink_iter);
-	}
+	bch2_trans_iter_put(trans, reflink_iter);
 
 	return ret;
 }
diff --git a/libbcachefs/reflink.h b/libbcachefs/reflink.h
index ac23b855..5445c1cf 100644
--- a/libbcachefs/reflink.h
+++ b/libbcachefs/reflink.h
@@ -22,6 +22,7 @@ void bch2_reflink_v_to_text(struct printbuf *, struct bch_fs *,
 #define bch2_bkey_ops_reflink_v (struct bkey_ops) {		\
 	.key_invalid	= bch2_reflink_v_invalid,		\
 	.val_to_text	= bch2_reflink_v_to_text,		\
+	.swab		= bch2_ptr_swab,			\
 }
 
 s64 bch2_remap_range(struct bch_fs *, struct bpos, struct bpos,
diff --git a/libbcachefs/replicas.c b/libbcachefs/replicas.c
index 366888b1..be490857 100644
--- a/libbcachefs/replicas.c
+++ b/libbcachefs/replicas.c
@@ -112,6 +112,7 @@ void bch2_bkey_to_replicas(struct bch_replicas_entry *e,
 
 	switch (k.k->type) {
 	case KEY_TYPE_btree_ptr:
+	case KEY_TYPE_btree_ptr_v2:
 		e->data_type = BCH_DATA_BTREE;
 		extent_to_replicas(k, e);
 		break;
diff --git a/libbcachefs/str_hash.h b/libbcachefs/str_hash.h
index f2779159..19a0d6b2 100644
--- a/libbcachefs/str_hash.h
+++ b/libbcachefs/str_hash.h
@@ -262,10 +262,8 @@ int bch2_hash_set(struct btree_trans *trans,
 	if (!ret)
 		ret = -ENOSPC;
 out:
-	if (!IS_ERR_OR_NULL(slot))
-		bch2_trans_iter_put(trans, slot);
-	if (!IS_ERR_OR_NULL(iter))
-		bch2_trans_iter_put(trans, iter);
+	bch2_trans_iter_put(trans, slot);
+	bch2_trans_iter_put(trans, iter);
 
 	return ret;
 found:
@@ -319,13 +317,16 @@ int bch2_hash_delete(struct btree_trans *trans,
 		     u64 inode, const void *key)
 {
 	struct btree_iter *iter;
+	int ret;
 
 	iter = bch2_hash_lookup(trans, desc, info, inode, key,
 				BTREE_ITER_INTENT);
 	if (IS_ERR(iter))
 		return PTR_ERR(iter);
 
-	return bch2_hash_delete_at(trans, desc, info, iter);
+	ret = bch2_hash_delete_at(trans, desc, info, iter);
+	bch2_trans_iter_put(trans, iter);
+	return ret;
 }
 
 #endif /* _BCACHEFS_STR_HASH_H */