Merge with e4871e8f27 bcachefs: Fix a deadlock on starting an interior btree update

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

60 files changed, 2748 insertions, 2503 deletions

diff --git a/fs/bcachefs/alloc_background.c b/fs/bcachefs/alloc_background.c
index c57df50168e0..b2d1b8f9c9b8 100644
--- a/fs/bcachefs/alloc_background.c
+++ b/fs/bcachefs/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/fs/bcachefs/bcachefs.h b/fs/bcachefs/bcachefs.h
index be2c1ed9fcb2..fa9593764f0c 100644
--- a/fs/bcachefs/bcachefs.h
+++ b/fs/bcachefs/bcachefs.h
@@ -521,6 +521,18 @@ struct journal_seq_blacklist_table {
 	}			entries[0];
 };
 
+struct journal_keys {
+	struct journal_key {
+		enum btree_id	btree_id:8;
+		unsigned	level:8;
+		struct bkey_i	*k;
+		u32		journal_seq;
+		u32		journal_offset;
+	}			*d;
+	size_t			nr;
+	u64			journal_seq_base;
+};
+
 struct bch_fs {
 	struct closure		cl;
 
@@ -608,6 +620,7 @@ struct bch_fs {
 
 	mempool_t		btree_interior_update_pool;
 	struct list_head	btree_interior_update_list;
+	struct list_head	btree_interior_updates_unwritten;
 	struct mutex		btree_interior_update_lock;
 	struct closure_waitlist	btree_interior_update_wait;
 
@@ -719,7 +732,7 @@ struct bch_fs {
 	ZSTD_parameters		zstd_params;
 
 	struct crypto_shash	*sha256;
-	struct crypto_skcipher	*chacha20;
+	struct crypto_sync_skcipher *chacha20;
 	struct crypto_shash	*poly1305;
 
 	atomic64_t		key_version;
@@ -786,6 +799,8 @@ struct bch_fs {
 	mempool_t		btree_bounce_pool;
 
 	struct journal		journal;
+	struct list_head	journal_entries;
+	struct journal_keys	journal_keys;
 
 	u64			last_bucket_seq_cleanup;
 
diff --git a/fs/bcachefs/bcachefs_format.h b/fs/bcachefs/bcachefs_format.h
index f6141fde830b..616863ef77d4 100644
--- a/fs/bcachefs/bcachefs_format.h
+++ b/fs/bcachefs/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)
 
@@ -1141,7 +1156,8 @@ enum bcachefs_metadata_version {
 	bcachefs_metadata_version_min			= 9,
 	bcachefs_metadata_version_new_versioning	= 10,
 	bcachefs_metadata_version_bkey_renumber		= 10,
-	bcachefs_metadata_version_max			= 11,
+	bcachefs_metadata_version_inode_btree_change	= 11,
+	bcachefs_metadata_version_max			= 12,
 };
 
 #define bcachefs_metadata_version_current	(bcachefs_metadata_version_max - 1)
@@ -1294,7 +1310,17 @@ LE64_BITMASK(BCH_SB_ERASURE_CODE,	struct bch_sb, flags[3],  0, 16);
 	x(reflink,			6)	\
 	x(new_siphash,			7)	\
 	x(inline_data,			8)	\
-	x(new_extent_overwrite,		9)
+	x(new_extent_overwrite,		9)	\
+	x(incompressible,		10)	\
+	x(btree_ptr_v2,			11)	\
+	x(extents_above_btree_updates,	12)	\
+	x(btree_updates_journalled,	13)
+
+#define BCH_SB_FEATURES_ALL				\
+	((1ULL << BCH_FEATURE_new_siphash)|		\
+	 (1ULL << BCH_FEATURE_new_extent_overwrite)|	\
+	 (1ULL << BCH_FEATURE_btree_ptr_v2)|		\
+	 (1ULL << BCH_FEATURE_extents_above_btree_updates))
 
 enum bch_sb_feature {
 #define x(f, n) BCH_FEATURE_##f,
@@ -1374,11 +1400,12 @@ enum bch_csum_opts {
 };
 
 #define BCH_COMPRESSION_TYPES()		\
-	x(none,		0)		\
-	x(lz4_old,	1)		\
-	x(gzip,		2)		\
-	x(lz4,		3)		\
-	x(zstd,		4)
+	x(none,			0)	\
+	x(lz4_old,		1)	\
+	x(gzip,			2)	\
+	x(lz4,			3)	\
+	x(zstd,			4)	\
+	x(incompressible,	5)
 
 enum bch_compression_type {
 #define x(t, n) BCH_COMPRESSION_TYPE_##t,
diff --git a/fs/bcachefs/bkey.h b/fs/bcachefs/bkey.h
index f2d5f3009b21..cbcfbd26bc58 100644
--- a/fs/bcachefs/bkey.h
+++ b/fs/bcachefs/bkey.h
@@ -400,9 +400,8 @@ static inline int bch2_compile_bkey_format(const struct bkey_format *format,
 static inline void bkey_reassemble(struct bkey_i *dst,
 				   struct bkey_s_c src)
 {
-	BUG_ON(bkey_packed(src.k));
 	dst->k = *src.k;
-	memcpy_u64s(&dst->v, src.v, bkey_val_u64s(src.k));
+	memcpy_u64s_small(&dst->v, src.v, bkey_val_u64s(src.k));
 }
 
 #define bkey_s_null		((struct bkey_s)   { .k = NULL })
@@ -565,6 +564,7 @@ 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/fs/bcachefs/bkey_methods.c b/fs/bcachefs/bkey_methods.c
index 320e17d108d2..c97e1e9002cb 100644
--- a/fs/bcachefs/bkey_methods.c
+++ b/fs/bcachefs/bkey_methods.c
@@ -134,7 +134,7 @@ const char *bch2_bkey_invalid(struct bch_fs *c, struct bkey_s_c k,
 
 const char *bch2_bkey_in_btree_node(struct btree *b, struct bkey_s_c k)
 {
-	if (bkey_cmp(bkey_start_pos(k.k), b->data->min_key) < 0)
+	if (bkey_cmp(k.k->p, b->data->min_key) < 0)
 		return "key before start of btree node";
 
 	if (bkey_cmp(k.k->p, b->data->max_key) > 0)
@@ -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)
@@ -276,3 +273,59 @@ void bch2_bkey_renumber(enum btree_node_type btree_node_type,
 			break;
 		}
 }
+
+void __bch2_bkey_compat(unsigned level, enum btree_id btree_id,
+			unsigned version, unsigned big_endian,
+			int write,
+			struct bkey_format *f,
+			struct bkey_packed *k)
+{
+	const struct bkey_ops *ops;
+	struct bkey uk;
+	struct bkey_s u;
+
+	if (big_endian != CPU_BIG_ENDIAN)
+		bch2_bkey_swab_key(f, k);
+
+	if (version < bcachefs_metadata_version_bkey_renumber)
+		bch2_bkey_renumber(__btree_node_type(level, btree_id), k, write);
+
+	if (version < bcachefs_metadata_version_inode_btree_change &&
+	    btree_id == BTREE_ID_INODES) {
+		if (!bkey_packed(k)) {
+			struct bkey_i *u = packed_to_bkey(k);
+			swap(u->k.p.inode, u->k.p.offset);
+		} else if (f->bits_per_field[BKEY_FIELD_INODE] &&
+			   f->bits_per_field[BKEY_FIELD_OFFSET]) {
+			struct bkey_format tmp = *f, *in = f, *out = &tmp;
+
+			swap(tmp.bits_per_field[BKEY_FIELD_INODE],
+			     tmp.bits_per_field[BKEY_FIELD_OFFSET]);
+			swap(tmp.field_offset[BKEY_FIELD_INODE],
+			     tmp.field_offset[BKEY_FIELD_OFFSET]);
+
+			if (!write)
+				swap(in, out);
+
+			uk = __bch2_bkey_unpack_key(in, k);
+			swap(uk.p.inode, uk.p.offset);
+			BUG_ON(!bch2_bkey_pack_key(k, &uk, out));
+		}
+	}
+
+	if (!bkey_packed(k)) {
+		u = bkey_i_to_s(packed_to_bkey(k));
+	} else {
+		uk = __bch2_bkey_unpack_key(f, k);
+		u.k = &uk;
+		u.v = bkeyp_val(f, k);
+	}
+
+	if (big_endian != CPU_BIG_ENDIAN)
+		bch2_bkey_swab_val(u);
+
+	ops = &bch2_bkey_ops[k->type];
+
+	if (ops->compat)
+		ops->compat(btree_id, version, big_endian, write, u);
+}
diff --git a/fs/bcachefs/bkey_methods.h b/fs/bcachefs/bkey_methods.h
index 8568b65c1ed2..0bca725ae3b8 100644
--- a/fs/bcachefs/bkey_methods.h
+++ b/fs/bcachefs/bkey_methods.h
@@ -29,10 +29,13 @@ 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);
+	void		(*compat)(enum btree_id id, unsigned version,
+				  unsigned big_endian, int write,
+				  struct bkey_s);
 };
 
 const char *bch2_bkey_val_invalid(struct bch_fs *, struct bkey_s_c);
@@ -51,7 +54,7 @@ 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);
 
@@ -60,4 +63,20 @@ enum merge_result bch2_bkey_merge(struct bch_fs *,
 
 void bch2_bkey_renumber(enum btree_node_type, struct bkey_packed *, int);
 
+void __bch2_bkey_compat(unsigned, enum btree_id, unsigned, unsigned,
+			int, struct bkey_format *, struct bkey_packed *);
+
+static inline void bch2_bkey_compat(unsigned level, enum btree_id btree_id,
+			       unsigned version, unsigned big_endian,
+			       int write,
+			       struct bkey_format *f,
+			       struct bkey_packed *k)
+{
+	if (version < bcachefs_metadata_version_current ||
+	    big_endian != CPU_BIG_ENDIAN)
+		__bch2_bkey_compat(level, btree_id, version,
+				   big_endian, write, f, k);
+
+}
+
 #endif /* _BCACHEFS_BKEY_METHODS_H */
diff --git a/fs/bcachefs/bkey_sort.c b/fs/bcachefs/bkey_sort.c
index 7cbb57042af1..839e78d1dc35 100644
--- a/fs/bcachefs/bkey_sort.c
+++ b/fs/bcachefs/bkey_sort.c
@@ -130,44 +130,21 @@ bch2_key_sort_fix_overlapping(struct bch_fs *c, struct bset *dst,
 	return nr;
 }
 
-static void extent_sort_advance_prev(struct bkey_format *f,
-				     struct btree_nr_keys *nr,
-				     struct bkey_packed *start,
-				     struct bkey_packed **prev)
-{
-	if (*prev) {
-		bch2_bkey_pack(*prev, (void *) *prev, f);
-
-		btree_keys_account_key_add(nr, 0, *prev);
-		*prev = bkey_next(*prev);
-	} else {
-		*prev = start;
-	}
-}
-
 static void extent_sort_append(struct bch_fs *c,
 			       struct bkey_format *f,
 			       struct btree_nr_keys *nr,
-			       struct bkey_packed *start,
-			       struct bkey_packed **prev,
+			       struct bkey_packed **out,
 			       struct bkey_s k)
 {
-	if (bkey_whiteout(k.k))
-		return;
-
-	/*
-	 * prev is always unpacked, for key merging - until right before we
-	 * advance it:
-	 */
+	if (!bkey_whiteout(k.k)) {
+		if (!bch2_bkey_pack_key(*out, k.k, f))
+			memcpy_u64s_small(*out, k.k, BKEY_U64s);
 
-	if (*prev &&
-	    bch2_bkey_merge(c, bkey_i_to_s((void *) *prev), k) ==
-	    BCH_MERGE_MERGE)
-		return;
+		memcpy_u64s_small(bkeyp_val(f, *out), k.v, bkey_val_u64s(k.k));
 
-	extent_sort_advance_prev(f, nr, start, prev);
-
-	bkey_reassemble((void *) *prev, k.s_c);
+		btree_keys_account_key_add(nr, 0, *out);
+		*out = bkey_next(*out);
+	}
 }
 
 /* Sort + repack in a new format: */
@@ -201,7 +178,7 @@ bch2_sort_repack(struct bset *dst, struct btree *src,
 	return nr;
 }
 
-/* Sort, repack, and merge: */
+/* Sort, repack, and call bch2_bkey_normalize() to drop stale pointers: */
 struct btree_nr_keys
 bch2_sort_repack_merge(struct bch_fs *c,
 		       struct bset *dst, struct btree *src,
@@ -209,7 +186,7 @@ bch2_sort_repack_merge(struct bch_fs *c,
 		       struct bkey_format *out_f,
 		       bool filter_whiteouts)
 {
-	struct bkey_packed *prev = NULL, *k_packed;
+	struct bkey_packed *out = vstruct_last(dst), *k_packed;
 	struct bkey_on_stack k;
 	struct btree_nr_keys nr;
 
@@ -234,13 +211,10 @@ bch2_sort_repack_merge(struct bch_fs *c,
 		    bch2_bkey_normalize(c, bkey_i_to_s(k.k)))
 			continue;
 
-		extent_sort_append(c, out_f, &nr, vstruct_last(dst),
-				   &prev, bkey_i_to_s(k.k));
+		extent_sort_append(c, out_f, &nr, &out, bkey_i_to_s(k.k));
 	}
 
-	extent_sort_advance_prev(out_f, &nr, vstruct_last(dst), &prev);
-
-	dst->u64s = cpu_to_le16((u64 *) prev - dst->_data);
+	dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
 	bkey_on_stack_exit(&k, c);
 	return nr;
 }
@@ -311,6 +285,25 @@ static inline int extent_sort_fix_overlapping_cmp(struct btree *b,
 		cmp_int((unsigned long) r, (unsigned long) l);
 }
 
+/*
+ * The algorithm in extent_sort_fix_overlapping() relies on keys in the same
+ * bset being ordered by start offset - but 0 size whiteouts (which are always
+ * KEY_TYPE_deleted) break this ordering, so we need to skip over them:
+ */
+static void extent_iter_advance(struct sort_iter *iter, unsigned idx)
+{
+	struct sort_iter_set *i = iter->data + idx;
+
+	do {
+		i->k = bkey_next_skip_noops(i->k, i->end);
+	} while (i->k != i->end && bkey_deleted(i->k));
+
+	if (i->k == i->end)
+		array_remove_item(iter->data, iter->used, idx);
+	else
+		__sort_iter_sift(iter, idx, extent_sort_fix_overlapping_cmp);
+}
+
 struct btree_nr_keys
 bch2_extent_sort_fix_overlapping(struct bch_fs *c, struct bset *dst,
 				 struct sort_iter *iter)
@@ -318,24 +311,31 @@ bch2_extent_sort_fix_overlapping(struct bch_fs *c, struct bset *dst,
 	struct btree *b = iter->b;
 	struct bkey_format *f = &b->format;
 	struct sort_iter_set *_l = iter->data, *_r = iter->data + 1;
-	struct bkey_packed *prev = NULL;
+	struct bkey_packed *out = dst->start;
 	struct bkey l_unpacked, r_unpacked;
 	struct bkey_s l, r;
 	struct btree_nr_keys nr;
 	struct bkey_on_stack split;
+	unsigned i;
 
 	memset(&nr, 0, sizeof(nr));
 	bkey_on_stack_init(&split);
 
 	sort_iter_sort(iter, extent_sort_fix_overlapping_cmp);
+	for (i = 0; i < iter->used;) {
+		if (bkey_deleted(iter->data[i].k))
+			__sort_iter_advance(iter, i,
+					    extent_sort_fix_overlapping_cmp);
+		else
+			i++;
+	}
 
 	while (!sort_iter_end(iter)) {
 		l = __bkey_disassemble(b, _l->k, &l_unpacked);
 
 		if (iter->used == 1) {
-			extent_sort_append(c, f, &nr, dst->start, &prev, l);
-			sort_iter_advance(iter,
-					  extent_sort_fix_overlapping_cmp);
+			extent_sort_append(c, f, &nr, &out, l);
+			extent_iter_advance(iter, 0);
 			continue;
 		}
 
@@ -343,16 +343,14 @@ bch2_extent_sort_fix_overlapping(struct bch_fs *c, struct bset *dst,
 
 		/* If current key and next key don't overlap, just append */
 		if (bkey_cmp(l.k->p, bkey_start_pos(r.k)) <= 0) {
-			extent_sort_append(c, f, &nr, dst->start, &prev, l);
-			sort_iter_advance(iter,
-					  extent_sort_fix_overlapping_cmp);
+			extent_sort_append(c, f, &nr, &out, l);
+			extent_iter_advance(iter, 0);
 			continue;
 		}
 
 		/* Skip 0 size keys */
 		if (!r.k->size) {
-			__sort_iter_advance(iter, 1,
-					    extent_sort_fix_overlapping_cmp);
+			extent_iter_advance(iter, 1);
 			continue;
 		}
 
@@ -369,8 +367,7 @@ bch2_extent_sort_fix_overlapping(struct bch_fs *c, struct bset *dst,
 		if (_l->k > _r->k) {
 			/* l wins, trim r */
 			if (bkey_cmp(l.k->p, r.k->p) >= 0) {
-				__sort_iter_advance(iter, 1,
-					 extent_sort_fix_overlapping_cmp);
+				extent_iter_advance(iter, 1);
 			} else {
 				bch2_cut_front_s(l.k->p, r);
 				extent_save(b, _r->k, r.k);
@@ -391,17 +388,15 @@ bch2_extent_sort_fix_overlapping(struct bch_fs *c, struct bset *dst,
 			__sort_iter_sift(iter, 0,
 					 extent_sort_fix_overlapping_cmp);
 
-			extent_sort_append(c, f, &nr, dst->start,
-					   &prev, bkey_i_to_s(split.k));
+			extent_sort_append(c, f, &nr, &out,
+					   bkey_i_to_s(split.k));
 		} else {
 			bch2_cut_back_s(bkey_start_pos(r.k), l);
 			extent_save(b, _l->k, l.k);
 		}
 	}
 
-	extent_sort_advance_prev(f, &nr, dst->start, &prev);
-
-	dst->u64s = cpu_to_le16((u64 *) prev - dst->_data);
+	dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
 
 	bkey_on_stack_exit(&split, c);
 	return nr;
diff --git a/fs/bcachefs/bset.c b/fs/bcachefs/bset.c
index cf8fa59fada1..6360b2e8cf73 100644
--- a/fs/bcachefs/bset.c
+++ b/fs/bcachefs/bset.c
@@ -79,8 +79,8 @@ void bch2_dump_bset(struct btree *b, struct bset *i, unsigned set)
 		_n = bkey_next_skip_noops(_k, vstruct_last(i));
 
 		bch2_bkey_to_text(&PBUF(buf), &k);
-		printk(KERN_ERR "block %u key %5u: %s\n", set,
-		       __btree_node_key_to_offset(b, _k), buf);
+		printk(KERN_ERR "block %u key %5zu: %s\n", set,
+		       _k->_data - i->_data, buf);
 
 		if (_n == vstruct_last(i))
 			continue;
@@ -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);
 }
@@ -1681,7 +1682,8 @@ struct bkey_packed *bch2_btree_node_iter_prev_all(struct btree_node_iter *iter,
 	struct bset_tree *t;
 	unsigned end = 0;
 
-	bch2_btree_node_iter_verify(iter, b);
+	if (btree_keys_expensive_checks(b))
+		bch2_btree_node_iter_verify(iter, b);
 
 	for_each_bset(b, t) {
 		k = bch2_bkey_prev_all(b, t,
@@ -1716,7 +1718,8 @@ found:
 	iter->data[0].k = __btree_node_key_to_offset(b, prev);
 	iter->data[0].end = end;
 
-	bch2_btree_node_iter_verify(iter, b);
+	if (btree_keys_expensive_checks(b))
+		bch2_btree_node_iter_verify(iter, b);
 	return prev;
 }
 
diff --git a/fs/bcachefs/btree_cache.c b/fs/bcachefs/btree_cache.c
index 0c737f35f430..c12f8a6b5205 100644
--- a/fs/bcachefs/btree_cache.c
+++ b/fs/bcachefs/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:
@@ -584,6 +588,7 @@ err:
 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 				struct btree_iter *iter,
 				const struct bkey_i *k,
+				enum btree_id btree_id,
 				unsigned level,
 				enum six_lock_type lock_type,
 				bool sync)
@@ -591,23 +596,24 @@ static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 	struct btree_cache *bc = &c->btree_cache;
 	struct btree *b;
 
+	BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
 	/*
 	 * Parent node must be locked, else we could read in a btree node that's
 	 * been freed:
 	 */
-	BUG_ON(!btree_node_locked(iter, level + 1));
-	BUG_ON(level >= BTREE_MAX_DEPTH);
+	if (iter && !bch2_btree_node_relock(iter, level + 1))
+		return ERR_PTR(-EINTR);
 
 	b = bch2_btree_node_mem_alloc(c);
 	if (IS_ERR(b))
 		return b;
 
 	bkey_copy(&b->key, k);
-	if (bch2_btree_node_hash_insert(bc, b, level, iter->btree_id)) {
+	if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
 		/* 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);
@@ -619,15 +625,11 @@ static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
 	}
 
 	/*
-	 * If the btree node wasn't cached, we can't drop our lock on
-	 * the parent until after it's added to the cache - because
-	 * otherwise we could race with a btree_split() freeing the node
-	 * we're trying to lock.
+	 * Unlock before doing IO:
 	 *
-	 * But the deadlock described below doesn't exist in this case,
-	 * so it's safe to not drop the parent lock until here:
+	 * XXX: ideally should be dropping all btree node locks here
 	 */
-	if (btree_node_read_locked(iter, level + 1))
+	if (iter && btree_node_read_locked(iter, level + 1))
 		btree_node_unlock(iter, level + 1);
 
 	bch2_btree_node_read(c, b, sync);
@@ -662,16 +664,11 @@ struct btree *bch2_btree_node_get(struct bch_fs *c, struct btree_iter *iter,
 	struct btree *b;
 	struct bset_tree *t;
 
-	/*
-	 * XXX: locking optimization
-	 *
-	 * we can make the locking looser here - caller can drop lock on parent
-	 * node before locking child node (and potentially blocking): we just
-	 * have to have bch2_btree_node_fill() call relock on the parent and
-	 * return -EINTR if that fails
-	 */
-	EBUG_ON(!btree_node_locked(iter, level + 1));
 	EBUG_ON(level >= BTREE_MAX_DEPTH);
+
+	b = btree_node_mem_ptr(k);
+	if (b)
+		goto lock_node;
 retry:
 	b = btree_cache_find(bc, k);
 	if (unlikely(!b)) {
@@ -680,7 +677,8 @@ retry:
 		 * else we could read in a btree node from disk that's been
 		 * freed:
 		 */
-		b = bch2_btree_node_fill(c, iter, k, level, lock_type, true);
+		b = bch2_btree_node_fill(c, iter, k, iter->btree_id,
+					 level, lock_type, true);
 
 		/* We raced and found the btree node in the cache */
 		if (!b)
@@ -689,6 +687,7 @@ retry:
 		if (IS_ERR(b))
 			return b;
 	} else {
+lock_node:
 		/*
 		 * There's a potential deadlock with splits and insertions into
 		 * interior nodes we have to avoid:
@@ -710,7 +709,7 @@ 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 +722,7 @@ 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);
@@ -735,6 +734,7 @@ retry:
 		}
 	}
 
+	/* XXX: waiting on IO with btree locks held: */
 	wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
 		       TASK_UNINTERRUPTIBLE);
 
@@ -749,7 +749,7 @@ retry:
 	}
 
 	/* avoid atomic set bit if it's not needed: */
-	if (btree_node_accessed(b))
+	if (!btree_node_accessed(b))
 		set_btree_node_accessed(b);
 
 	if (unlikely(btree_node_read_error(b))) {
@@ -764,6 +764,74 @@ retry:
 	return b;
 }
 
+struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
+					 const struct bkey_i *k,
+					 enum btree_id btree_id,
+					 unsigned level)
+{
+	struct btree_cache *bc = &c->btree_cache;
+	struct btree *b;
+	struct bset_tree *t;
+
+	EBUG_ON(level >= BTREE_MAX_DEPTH);
+
+	b = btree_node_mem_ptr(k);
+	if (b)
+		goto lock_node;
+retry:
+	b = btree_cache_find(bc, k);
+	if (unlikely(!b)) {
+		b = bch2_btree_node_fill(c, NULL, k, btree_id,
+					 level, SIX_LOCK_read, true);
+
+		/* We raced and found the btree node in the cache */
+		if (!b)
+			goto retry;
+
+		if (IS_ERR(b))
+			return b;
+	} else {
+lock_node:
+		six_lock_read(&b->lock);
+
+		if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
+			     b->btree_id != btree_id ||
+			     b->level != level)) {
+			six_unlock_read(&b->lock);
+			goto retry;
+		}
+	}
+
+	/* XXX: waiting on IO with btree locks held: */
+	wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
+		       TASK_UNINTERRUPTIBLE);
+
+	prefetch(b->aux_data);
+
+	for_each_bset(b, t) {
+		void *p = (u64 *) b->aux_data + t->aux_data_offset;
+
+		prefetch(p + L1_CACHE_BYTES * 0);
+		prefetch(p + L1_CACHE_BYTES * 1);
+		prefetch(p + L1_CACHE_BYTES * 2);
+	}
+
+	/* avoid atomic set bit if it's not needed: */
+	if (!btree_node_accessed(b))
+		set_btree_node_accessed(b);
+
+	if (unlikely(btree_node_read_error(b))) {
+		six_unlock_read(&b->lock);
+		return ERR_PTR(-EIO);
+	}
+
+	EBUG_ON(b->btree_id != btree_id ||
+		BTREE_NODE_LEVEL(b->data) != level ||
+		bkey_cmp(b->data->max_key, k->k.p));
+
+	return b;
+}
+
 struct btree *bch2_btree_node_get_sibling(struct bch_fs *c,
 					  struct btree_iter *iter,
 					  struct btree *b,
@@ -855,8 +923,7 @@ out:
 		if (sib != btree_prev_sib)
 			swap(n1, n2);
 
-		BUG_ON(bkey_cmp(btree_type_successor(n1->btree_id,
-						     n1->key.k.p),
+		BUG_ON(bkey_cmp(bkey_successor(n1->key.k.p),
 				n2->data->min_key));
 	}
 
@@ -878,7 +945,8 @@ void bch2_btree_node_prefetch(struct bch_fs *c, struct btree_iter *iter,
 	if (b)
 		return;
 
-	bch2_btree_node_fill(c, iter, k, level, SIX_LOCK_read, false);
+	bch2_btree_node_fill(c, iter, k, iter->btree_id,
+			     level, SIX_LOCK_read, false);
 }
 
 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
diff --git a/fs/bcachefs/btree_cache.h b/fs/bcachefs/btree_cache.h
index 83358d6a4df8..132cc95a4c02 100644
--- a/fs/bcachefs/btree_cache.h
+++ b/fs/bcachefs/btree_cache.h
@@ -25,6 +25,9 @@ struct btree *bch2_btree_node_get(struct bch_fs *, struct btree_iter *,
 				  const struct bkey_i *, unsigned,
 				  enum six_lock_type);
 
+struct btree *bch2_btree_node_get_noiter(struct bch_fs *, const struct bkey_i *,
+					 enum btree_id, unsigned);
+
 struct btree *bch2_btree_node_get_sibling(struct bch_fs *, struct btree_iter *,
 				struct btree *, enum btree_node_sibling);
 
@@ -35,13 +38,29 @@ 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;
+	}
+}
+
+static inline struct btree *btree_node_mem_ptr(const struct bkey_i *k)
+{
+	return k->k.type == KEY_TYPE_btree_ptr_v2
+		? (void *)(unsigned long)bkey_i_to_btree_ptr_v2_c(k)->v.mem_ptr
+		: NULL;
+}
 
 /* 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/fs/bcachefs/btree_gc.c b/fs/bcachefs/btree_gc.c
index 05879b66d6af..cef8e148f784 100644
--- a/fs/bcachefs/btree_gc.c
+++ b/fs/bcachefs/btree_gc.c
@@ -47,65 +47,42 @@ static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
 	__gc_pos_set(c, new_pos);
 }
 
-/* range_checks - for validating min/max pos of each btree node: */
-
-struct range_checks {
-	struct range_level {
-		struct bpos	min;
-		struct bpos	max;
-	}			l[BTREE_MAX_DEPTH];
-	unsigned		depth;
-};
-
-static void btree_node_range_checks_init(struct range_checks *r, unsigned depth)
+static int bch2_gc_check_topology(struct bch_fs *c,
+				  struct bkey_s_c k,
+				  struct bpos *expected_start,
+				  struct bpos expected_end,
+				  bool is_last)
 {
-	unsigned i;
-
-	for (i = 0; i < BTREE_MAX_DEPTH; i++)
-		r->l[i].min = r->l[i].max = POS_MIN;
-	r->depth = depth;
-}
-
-static void btree_node_range_checks(struct bch_fs *c, struct btree *b,
-				    struct range_checks *r)
-{
-	struct range_level *l = &r->l[b->level];
-
-	struct bpos expected_min = bkey_cmp(l->min, l->max)
-		? btree_type_successor(b->btree_id, l->max)
-		: l->max;
-
-	bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, expected_min), c,
-		"btree node has incorrect min key: %llu:%llu != %llu:%llu",
-		b->data->min_key.inode,
-		b->data->min_key.offset,
-		expected_min.inode,
-		expected_min.offset);
-
-	l->max = b->data->max_key;
+	int ret = 0;
 
-	if (b->level > r->depth) {
-		l = &r->l[b->level - 1];
+	if (k.k->type == KEY_TYPE_btree_ptr_v2) {
+		struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
 
-		bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, l->min), c,
-			"btree node min doesn't match min of child nodes: %llu:%llu != %llu:%llu",
-			b->data->min_key.inode,
-			b->data->min_key.offset,
-			l->min.inode,
-			l->min.offset);
+		if (fsck_err_on(bkey_cmp(*expected_start, bp.v->min_key), c,
+				"btree node with incorrect min_key: got %llu:%llu, should be %llu:%llu",
+				bp.v->min_key.inode,
+				bp.v->min_key.offset,
+				expected_start->inode,
+				expected_start->offset)) {
+			BUG();
+		}
+	}
 
-		bch2_fs_inconsistent_on(bkey_cmp(b->data->max_key, l->max), c,
-			"btree node max doesn't match max of child nodes: %llu:%llu != %llu:%llu",
-			b->data->max_key.inode,
-			b->data->max_key.offset,
-			l->max.inode,
-			l->max.offset);
-
-		if (bkey_cmp(b->data->max_key, POS_MAX))
-			l->min = l->max =
-				btree_type_successor(b->btree_id,
-						     b->data->max_key);
+	*expected_start = bkey_cmp(k.k->p, POS_MAX)
+		? bkey_successor(k.k->p)
+		: k.k->p;
+
+	if (fsck_err_on(is_last &&
+			bkey_cmp(k.k->p, expected_end), c,
+			"btree node with incorrect max_key: got %llu:%llu, should be %llu:%llu",
+			k.k->p.inode,
+			k.k->p.offset,
+			expected_end.inode,
+			expected_end.offset)) {
+		BUG();
 	}
+fsck_err:
+	return ret;
 }
 
 /* marking of btree keys/nodes: */
@@ -124,7 +101,11 @@ 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) ||
@@ -180,9 +161,10 @@ fsck_err:
 	return ret;
 }
 
-static int btree_gc_mark_node(struct bch_fs *c, struct btree *b,
-			      u8 *max_stale, bool initial)
+static int btree_gc_mark_node(struct bch_fs *c, struct btree *b, u8 *max_stale,
+			      bool initial)
 {
+	struct bpos next_node_start = b->data->min_key;
 	struct btree_node_iter iter;
 	struct bkey unpacked;
 	struct bkey_s_c k;
@@ -193,13 +175,25 @@ static int btree_gc_mark_node(struct bch_fs *c, struct btree *b,
 	if (!btree_node_type_needs_gc(btree_node_type(b)))
 		return 0;
 
-	for_each_btree_node_key_unpack(b, k, &iter,
-				       &unpacked) {
+	bch2_btree_node_iter_init_from_start(&iter, b);
+
+	while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 		bch2_bkey_debugcheck(c, b, k);
 
 		ret = bch2_gc_mark_key(c, k, max_stale, initial);
 		if (ret)
 			break;
+
+		bch2_btree_node_iter_advance(&iter, b);
+
+		if (b->level) {
+			ret = bch2_gc_check_topology(c, k,
+					&next_node_start,
+					b->data->max_key,
+					bch2_btree_node_iter_end(&iter));
+			if (ret)
+				break;
+		}
 	}
 
 	return ret;
@@ -211,7 +205,6 @@ static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
 	struct btree_trans trans;
 	struct btree_iter *iter;
 	struct btree *b;
-	struct range_checks r;
 	unsigned depth = metadata_only			? 1
 		: expensive_debug_checks(c)		? 0
 		: !btree_node_type_needs_gc(btree_id)	? 1
@@ -223,12 +216,8 @@ static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
 
 	gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 
-	btree_node_range_checks_init(&r, depth);
-
 	__for_each_btree_node(&trans, iter, btree_id, POS_MIN,
 			      0, depth, BTREE_ITER_PREFETCH, b) {
-		btree_node_range_checks(c, b, &r);
-
 		bch2_verify_btree_nr_keys(b);
 
 		gc_pos_set(c, gc_pos_btree_node(b));
@@ -269,40 +258,116 @@ static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
 	return ret;
 }
 
-static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
+static int bch2_gc_btree_init_recurse(struct bch_fs *c, struct btree *b,
+				      struct journal_keys *journal_keys,
+				      unsigned target_depth)
 {
-	return  (int) btree_id_to_gc_phase(l) -
-		(int) btree_id_to_gc_phase(r);
+	struct btree_and_journal_iter iter;
+	struct bkey_s_c k;
+	struct bpos next_node_start = b->data->min_key;
+	u8 max_stale = 0;
+	int ret = 0;
+
+	bch2_btree_and_journal_iter_init_node_iter(&iter, journal_keys, b);
+
+	while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
+		bch2_bkey_debugcheck(c, b, k);
+
+		BUG_ON(bkey_cmp(k.k->p, b->data->min_key) < 0);
+		BUG_ON(bkey_cmp(k.k->p, b->data->max_key) > 0);
+
+		ret = bch2_gc_mark_key(c, k, &max_stale, true);
+		if (ret)
+			break;
+
+		if (b->level) {
+			struct btree *child;
+			BKEY_PADDED(k) tmp;
+
+			bkey_reassemble(&tmp.k, k);
+			k = bkey_i_to_s_c(&tmp.k);
+
+			bch2_btree_and_journal_iter_advance(&iter);
+
+			ret = bch2_gc_check_topology(c, k,
+					&next_node_start,
+					b->data->max_key,
+					!bch2_btree_and_journal_iter_peek(&iter).k);
+			if (ret)
+				break;
+
+			if (b->level > target_depth) {
+				child = bch2_btree_node_get_noiter(c, &tmp.k,
+							b->btree_id, b->level - 1);
+				ret = PTR_ERR_OR_ZERO(child);
+				if (ret)
+					break;
+
+				ret = bch2_gc_btree_init_recurse(c, child,
+						journal_keys, target_depth);
+				six_unlock_read(&child->lock);
+
+				if (ret)
+					break;
+			}
+		} else {
+			bch2_btree_and_journal_iter_advance(&iter);
+		}
+	}
+
+	return ret;
 }
 
-static int mark_journal_key(struct bch_fs *c, enum btree_id id,
-			    struct bkey_i *insert)
+static int bch2_gc_btree_init(struct bch_fs *c,
+			      struct journal_keys *journal_keys,
+			      enum btree_id btree_id,
+			      bool metadata_only)
 {
-	struct btree_trans trans;
-	struct btree_iter *iter;
-	struct bkey_s_c k;
-	u8 max_stale;
+	struct btree *b;
+	unsigned target_depth = metadata_only		? 1
+		: expensive_debug_checks(c)		? 0
+		: !btree_node_type_needs_gc(btree_id)	? 1
+		: 0;
+	u8 max_stale = 0;
 	int ret = 0;
 
-	ret = bch2_gc_mark_key(c, bkey_i_to_s_c(insert), &max_stale, true);
-	if (ret)
-		return ret;
+	b = c->btree_roots[btree_id].b;
 
-	bch2_trans_init(&trans, c, 0, 0);
+	if (btree_node_fake(b))
+		return 0;
 
-	for_each_btree_key(&trans, iter, id, bkey_start_pos(&insert->k),
-			   BTREE_ITER_SLOTS, k, ret) {
-		percpu_down_read(&c->mark_lock);
-		ret = bch2_mark_overwrite(&trans, iter, k, insert, NULL,
-					 BTREE_TRIGGER_GC|
-					 BTREE_TRIGGER_NOATOMIC);
-		percpu_up_read(&c->mark_lock);
+	six_lock_read(&b->lock);
+	if (fsck_err_on(bkey_cmp(b->data->min_key, POS_MIN), c,
+			"btree root with incorrect min_key: %llu:%llu",
+			b->data->min_key.inode,
+			b->data->min_key.offset)) {
+		BUG();
+	}
 
-		if (!ret)
-			break;
+	if (fsck_err_on(bkey_cmp(b->data->max_key, POS_MAX), c,
+			"btree root with incorrect min_key: %llu:%llu",
+			b->data->max_key.inode,
+			b->data->max_key.offset)) {
+		BUG();
 	}
 
-	return bch2_trans_exit(&trans) ?: ret;
+	if (b->level >= target_depth)
+		ret = bch2_gc_btree_init_recurse(c, b,
+					journal_keys, target_depth);
+
+	if (!ret)
+		ret = bch2_gc_mark_key(c, bkey_i_to_s_c(&b->key),
+				       &max_stale, true);
+fsck_err:
+	six_unlock_read(&b->lock);
+
+	return ret;
+}
+
+static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
+{
+	return  (int) btree_id_to_gc_phase(l) -
+		(int) btree_id_to_gc_phase(r);
 }
 
 static int bch2_gc_btrees(struct bch_fs *c, struct journal_keys *journal_keys,
@@ -317,24 +382,12 @@ static int bch2_gc_btrees(struct bch_fs *c, struct journal_keys *journal_keys,
 
 	for (i = 0; i < BTREE_ID_NR; i++) {
 		enum btree_id id = ids[i];
-		enum btree_node_type type = __btree_node_type(0, id);
-
-		int ret = bch2_gc_btree(c, id, initial, metadata_only);
+		int ret = initial
+			? bch2_gc_btree_init(c, journal_keys,
+					     id, metadata_only)
+			: bch2_gc_btree(c, id, initial, metadata_only);
 		if (ret)
 			return ret;
-
-		if (journal_keys && !metadata_only &&
-		    btree_node_type_needs_gc(type)) {
-			struct journal_key *j;
-			int ret;
-
-			for_each_journal_key(*journal_keys, j)
-				if (j->btree_id == id) {
-					ret = mark_journal_key(c, id, j->k);
-					if (ret)
-						return ret;
-				}
-		}
 	}
 
 	return 0;
@@ -880,7 +933,7 @@ static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
 		return;
 	}
 
-	as = bch2_btree_update_start(c, iter->btree_id,
+	as = bch2_btree_update_start(iter->trans, iter->btree_id,
 			btree_update_reserve_required(c, parent) + nr_old_nodes,
 			BTREE_INSERT_NOFAIL|
 			BTREE_INSERT_USE_RESERVE,
@@ -951,9 +1004,7 @@ static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
 			n1->key.k.p = n1->data->max_key =
 				bkey_unpack_pos(n1, last);
 
-			n2->data->min_key =
-				btree_type_successor(iter->btree_id,
-						     n1->data->max_key);
+			n2->data->min_key = bkey_successor(n1->data->max_key);
 
 			memcpy_u64s(vstruct_last(s1),
 				    s2->start, u64s);
diff --git a/fs/bcachefs/btree_io.c b/fs/bcachefs/btree_io.c
index c126985b6ef5..ac8b98861aae 100644
--- a/fs/bcachefs/btree_io.c
+++ b/fs/bcachefs/btree_io.c
@@ -19,6 +19,7 @@
 #include "journal_seq_blacklist.h"
 #include "super-io.h"
 
+#include <linux/sched/mm.h>
 #include <trace/events/bcachefs.h>
 
 static void verify_no_dups(struct btree *b,
@@ -68,17 +69,19 @@ static void btree_bounce_free(struct bch_fs *c, unsigned order,
 static void *btree_bounce_alloc(struct bch_fs *c, unsigned order,
 				bool *used_mempool)
 {
+	unsigned flags = memalloc_nofs_save();
 	void *p;
 
 	BUG_ON(order > btree_page_order(c));
 
 	*used_mempool = false;
 	p = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOWAIT, order);
-	if (p)
-		return p;
-
-	*used_mempool = true;
-	return mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
+	if (!p) {
+		*used_mempool = true;
+		p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
+	}
+	memalloc_nofs_restore(flags);
+	return p;
 }
 
 static void sort_bkey_ptrs(const struct btree *bt,
@@ -617,7 +620,7 @@ static void bset_encrypt(struct bch_fs *c, struct bset *i, unsigned offset)
 		bch2_encrypt(c, BSET_CSUM_TYPE(i), nonce, &bn->flags,
 			     bytes);
 
-		nonce = nonce_add(nonce, round_up(bytes, CHACHA20_BLOCK_SIZE));
+		nonce = nonce_add(nonce, round_up(bytes, CHACHA_BLOCK_SIZE));
 	}
 
 	bch2_encrypt(c, BSET_CSUM_TYPE(i), nonce, i->_data,
@@ -706,76 +709,107 @@ out:									\
 
 static int validate_bset(struct bch_fs *c, struct btree *b,
 			 struct bset *i, unsigned sectors,
-			 unsigned *whiteout_u64s, int write,
-			 bool have_retry)
+			 int write, bool have_retry)
 {
-	struct bkey_packed *k, *prev = NULL;
-	struct bpos prev_pos	= POS_MIN;
-	struct bpos prev_data	= POS_MIN;
-	bool seen_non_whiteout = false;
-	unsigned version;
+	unsigned version = le16_to_cpu(i->version);
 	const char *err;
 	int ret = 0;
 
-	if (i == &b->data->keys) {
+	btree_err_on((version != BCH_BSET_VERSION_OLD &&
+		      version < bcachefs_metadata_version_min) ||
+		     version >= bcachefs_metadata_version_max,
+		     BTREE_ERR_FATAL, c, b, i,
+		     "unsupported bset version");
+
+	if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
+			 BTREE_ERR_FIXABLE, c, b, i,
+			 "bset past end of btree node")) {
+		i->u64s = 0;
+		return 0;
+	}
+
+	btree_err_on(b->written && !i->u64s,
+		     BTREE_ERR_FIXABLE, c, b, i,
+		     "empty bset");
+
+	if (!b->written) {
+		struct btree_node *bn =
+			container_of(i, struct btree_node, keys);
 		/* These indicate that we read the wrong btree node: */
-		btree_err_on(BTREE_NODE_ID(b->data) != b->btree_id,
+		btree_err_on(BTREE_NODE_ID(bn) != b->btree_id,
 			     BTREE_ERR_MUST_RETRY, c, b, i,
 			     "incorrect btree id");
 
-		btree_err_on(BTREE_NODE_LEVEL(b->data) != b->level,
+		btree_err_on(BTREE_NODE_LEVEL(bn) != b->level,
 			     BTREE_ERR_MUST_RETRY, c, b, i,
 			     "incorrect level");
 
 		if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN) {
-			u64 *p = (u64 *) &b->data->ptr;
+			u64 *p = (u64 *) &bn->ptr;
 
 			*p = swab64(*p);
-			bch2_bpos_swab(&b->data->min_key);
-			bch2_bpos_swab(&b->data->max_key);
 		}
 
-		btree_err_on(bkey_cmp(b->data->max_key, b->key.k.p),
+		if (!write)
+			compat_btree_node(b->level, b->btree_id, version,
+					  BSET_BIG_ENDIAN(i), write, bn);
+
+		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: got %llu:%llu should be %llu:%llu",
+				     b->data->min_key.inode,
+				     b->data->min_key.offset,
+				     bp->min_key.inode,
+				     bp->min_key.offset);
+		}
+
+		btree_err_on(bkey_cmp(bn->max_key, b->key.k.p),
 			     BTREE_ERR_MUST_RETRY, c, b, i,
 			     "incorrect max key");
 
+		if (write)
+			compat_btree_node(b->level, b->btree_id, version,
+					  BSET_BIG_ENDIAN(i), write, bn);
+
 		/* XXX: ideally we would be validating min_key too */
 #if 0
 		/*
 		 * not correct anymore, due to btree node write error
 		 * handling
 		 *
-		 * need to add b->data->seq to btree keys and verify
+		 * need to add bn->seq to btree keys and verify
 		 * against that
 		 */
 		btree_err_on(!extent_contains_ptr(bkey_i_to_s_c_extent(&b->key),
-						  b->data->ptr),
+						  bn->ptr),
 			     BTREE_ERR_FATAL, c, b, i,
 			     "incorrect backpointer");
 #endif
-		err = bch2_bkey_format_validate(&b->data->format);
+		err = bch2_bkey_format_validate(&bn->format);
 		btree_err_on(err,
 			     BTREE_ERR_FATAL, c, b, i,
 			     "invalid bkey format: %s", err);
-	}
 
-	version = le16_to_cpu(i->version);
-	btree_err_on((version != BCH_BSET_VERSION_OLD &&
-		      version < bcachefs_metadata_version_min) ||
-		     version >= bcachefs_metadata_version_max,
-		     BTREE_ERR_FATAL, c, b, i,
-		     "unsupported bset version");
-
-	if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
-			 BTREE_ERR_FIXABLE, c, b, i,
-			 "bset past end of btree node")) {
-		i->u64s = 0;
-		return 0;
+		compat_bformat(b->level, b->btree_id, version,
+			       BSET_BIG_ENDIAN(i), write,
+			       &bn->format);
 	}
+fsck_err:
+	return ret;
+}
 
-	btree_err_on(b->written && !i->u64s,
-		     BTREE_ERR_FIXABLE, c, b, i,
-		     "empty bset");
+static int validate_bset_keys(struct bch_fs *c, struct btree *b,
+			 struct bset *i, unsigned *whiteout_u64s,
+			 int write, bool have_retry)
+{
+	unsigned version = le16_to_cpu(i->version);
+	struct bkey_packed *k, *prev = NULL;
+	bool seen_non_whiteout = false;
+	int ret = 0;
 
 	if (!BSET_SEPARATE_WHITEOUTS(i)) {
 		seen_non_whiteout = true;
@@ -784,7 +818,7 @@ 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;
 
@@ -804,22 +838,21 @@ static int validate_bset(struct bch_fs *c, struct btree *b,
 			continue;
 		}
 
-		if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN)
-			bch2_bkey_swab(&b->format, k);
-
-		if (!write &&
-		    version < bcachefs_metadata_version_bkey_renumber)
-			bch2_bkey_renumber(btree_node_type(b), k, write);
+		/* XXX: validate k->u64s */
+		if (!write)
+			bch2_bkey_compat(b->level, b->btree_id, version,
+				    BSET_BIG_ENDIAN(i), write,
+				    &b->format, k);
 
-		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);
+		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);
 
@@ -829,9 +862,10 @@ static int validate_bset(struct bch_fs *c, struct btree *b,
 			continue;
 		}
 
-		if (write &&
-		    version < bcachefs_metadata_version_bkey_renumber)
-			bch2_bkey_renumber(btree_node_type(b), k, write);
+		if (write)
+			bch2_bkey_compat(b->level, b->btree_id, version,
+				    BSET_BIG_ENDIAN(i), write,
+				    &b->format, k);
 
 		/*
 		 * with the separate whiteouts thing (used for extents), the
@@ -841,29 +875,27 @@ static int validate_bset(struct bch_fs *c, struct btree *b,
 
 		if (!seen_non_whiteout &&
 		    (!bkey_whiteout(k) ||
-		     (bkey_cmp(prev_pos, bkey_start_pos(u.k)) > 0))) {
+		     (prev && bkey_iter_cmp(b, prev, k) > 0))) {
 			*whiteout_u64s = k->_data - i->_data;
 			seen_non_whiteout = true;
-		} else if (bkey_cmp(prev_data, bkey_start_pos(u.k)) > 0 ||
-			   bkey_cmp(prev_pos, u.k->p) > 0) {
+		} else if (prev && bkey_iter_cmp(b, prev, k) > 0) {
+			char buf1[80];
+			char buf2[80];
+			struct bkey up = bkey_unpack_key(b, prev);
+
+			bch2_bkey_to_text(&PBUF(buf1), &up);
+			bch2_bkey_to_text(&PBUF(buf2), u.k);
+
+			bch2_dump_bset(b, i, 0);
 			btree_err(BTREE_ERR_FATAL, c, b, i,
-				  "keys out of order: %llu:%llu > %llu:%llu",
-				  prev_pos.inode,
-				  prev_pos.offset,
-				  u.k->p.inode,
-				  bkey_start_offset(u.k));
+				  "keys out of order: %s > %s",
+				  buf1, buf2);
 			/* XXX: repair this */
 		}
 
-		if (!bkey_deleted(u.k))
-			prev_data = u.k->p;
-		prev_pos = u.k->p;
-
 		prev = k;
 		k = bkey_next_skip_noops(k, vstruct_last(i));
 	}
-
-	SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 fsck_err:
 	return ret;
 }
@@ -895,6 +927,15 @@ 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;
@@ -922,8 +963,6 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry
 				set_btree_node_old_extent_overwrite(b);
 
 			sectors = vstruct_sectors(b->data, c->block_bits);
-
-			btree_node_set_format(b, b->data->format);
 		} else {
 			bne = write_block(b);
 			i = &bne->keys;
@@ -947,11 +986,21 @@ int bch2_btree_node_read_done(struct bch_fs *c, struct btree *b, bool have_retry
 			sectors = vstruct_sectors(bne, c->block_bits);
 		}
 
-		ret = validate_bset(c, b, i, sectors, &whiteout_u64s,
+		ret = validate_bset(c, b, i, sectors,
 				    READ, have_retry);
 		if (ret)
 			goto fsck_err;
 
+		if (!b->written)
+			btree_node_set_format(b, b->data->format);
+
+		ret = validate_bset_keys(c, b, i, &whiteout_u64s,
+				    READ, have_retry);
+		if (ret)
+			goto fsck_err;
+
+		SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
+
 		b->written += sectors;
 
 		blacklisted = bch2_journal_seq_is_blacklisted(c,
@@ -1002,15 +1051,15 @@ 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 +1072,12 @@ 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));
 	}
 
@@ -1236,7 +1291,6 @@ void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
 		closure_put(&((struct btree_update *) new)->cl);
 
 	bch2_journal_pin_drop(&c->journal, &w->journal);
-	closure_wake_up(&w->wait);
 }
 
 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
@@ -1252,8 +1306,6 @@ 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 +1331,13 @@ 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)
@@ -1394,7 +1443,8 @@ static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
 	if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_BTREE))
 		return -1;
 
-	ret = validate_bset(c, b, i, sectors, &whiteout_u64s, WRITE, false);
+	ret = validate_bset(c, b, i, sectors, WRITE, false) ?:
+		validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false);
 	if (ret)
 		bch2_inconsistent_error(c);
 
@@ -1544,8 +1594,7 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
 		validate_before_checksum = true;
 
 	/* validate_bset will be modifying: */
-	if (le16_to_cpu(i->version) <
-	    bcachefs_metadata_version_bkey_renumber)
+	if (le16_to_cpu(i->version) < bcachefs_metadata_version_max)
 		validate_before_checksum = true;
 
 	/* if we're going to be encrypting, check metadata validity first: */
@@ -1598,9 +1647,6 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
 	wbio->wbio.bio.bi_end_io	= btree_node_write_endio;
 	wbio->wbio.bio.bi_private	= b;
 
-	if (b->level || !b->written)
-		wbio->wbio.bio.bi_opf |= REQ_FUA;
-
 	bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
 
 	/*
@@ -1625,6 +1671,7 @@ void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
 
 	b->written += sectors_to_write;
 
+	/* XXX: submitting IO with btree locks held: */
 	bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_BTREE, &k.key);
 	return;
 err:
@@ -1773,12 +1820,11 @@ ssize_t bch2_dirty_btree_nodes_print(struct bch_fs *c, char *buf)
 	rcu_read_lock();
 	for_each_cached_btree(b, c, tbl, i, pos) {
 		unsigned long flags = READ_ONCE(b->flags);
-		unsigned idx = (flags & (1 << BTREE_NODE_write_idx)) != 0;
 
 		if (!(flags & (1 << BTREE_NODE_dirty)))
 			continue;
 
-		pr_buf(&out, "%p d %u n %u l %u w %u b %u r %u:%lu c %u p %u\n",
+		pr_buf(&out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
 		       b,
 		       (flags & (1 << BTREE_NODE_dirty)) != 0,
 		       (flags & (1 << BTREE_NODE_need_write)) != 0,
@@ -1786,9 +1832,7 @@ ssize_t bch2_dirty_btree_nodes_print(struct bch_fs *c, char *buf)
 		       b->written,
 		       !list_empty_careful(&b->write_blocked),
 		       b->will_make_reachable != 0,
-		       b->will_make_reachable & 1,
-		       b->writes[ idx].wait.list.first != NULL,
-		       b->writes[!idx].wait.list.first != NULL);
+		       b->will_make_reachable & 1);
 	}
 	rcu_read_unlock();
 
diff --git a/fs/bcachefs/btree_io.h b/fs/bcachefs/btree_io.h
index e90e89eee273..337d2bdd29e8 100644
--- a/fs/bcachefs/btree_io.h
+++ b/fs/bcachefs/btree_io.h
@@ -2,6 +2,7 @@
 #ifndef _BCACHEFS_BTREE_IO_H
 #define _BCACHEFS_BTREE_IO_H
 
+#include "bkey_methods.h"
 #include "bset.h"
 #include "btree_locking.h"
 #include "extents.h"
@@ -102,19 +103,20 @@ bool bch2_btree_post_write_cleanup(struct bch_fs *, struct btree *);
 void bch2_btree_node_write(struct bch_fs *, struct btree *,
 			  enum six_lock_type);
 
-static inline void btree_node_write_if_need(struct bch_fs *c, struct btree *b)
+static inline void btree_node_write_if_need(struct bch_fs *c, struct btree *b,
+					    enum six_lock_type lock_held)
 {
 	while (b->written &&
 	       btree_node_need_write(b) &&
 	       btree_node_may_write(b)) {
 		if (!btree_node_write_in_flight(b)) {
-			bch2_btree_node_write(c, b, SIX_LOCK_read);
+			bch2_btree_node_write(c, b, lock_held);
 			break;
 		}
 
-		six_unlock_read(&b->lock);
+		six_unlock_type(&b->lock, lock_held);
 		btree_node_wait_on_io(b);
-		btree_node_lock_type(c, b, SIX_LOCK_read);
+		btree_node_lock_type(c, b, lock_held);
 	}
 }
 
@@ -131,7 +133,7 @@ do {									\
 		new |= (1 << BTREE_NODE_need_write);			\
 	} while ((v = cmpxchg(&(_b)->flags, old, new)) != old);		\
 									\
-	btree_node_write_if_need(_c, _b);				\
+	btree_node_write_if_need(_c, _b, SIX_LOCK_read);		\
 } while (0)
 
 void bch2_btree_flush_all_reads(struct bch_fs *);
@@ -139,4 +141,50 @@ void bch2_btree_flush_all_writes(struct bch_fs *);
 void bch2_btree_verify_flushed(struct bch_fs *);
 ssize_t bch2_dirty_btree_nodes_print(struct bch_fs *, char *);
 
+static inline void compat_bformat(unsigned level, enum btree_id btree_id,
+				 unsigned version, unsigned big_endian,
+				 int write, struct bkey_format *f)
+{
+	if (version < bcachefs_metadata_version_inode_btree_change &&
+	    btree_id == BTREE_ID_INODES) {
+		swap(f->bits_per_field[BKEY_FIELD_INODE],
+		     f->bits_per_field[BKEY_FIELD_OFFSET]);
+		swap(f->field_offset[BKEY_FIELD_INODE],
+		     f->field_offset[BKEY_FIELD_OFFSET]);
+	}
+}
+
+static inline void compat_bpos(unsigned level, enum btree_id btree_id,
+			       unsigned version, unsigned big_endian,
+			       int write, struct bpos *p)
+{
+	if (big_endian != CPU_BIG_ENDIAN)
+		bch2_bpos_swab(p);
+
+	if (version < bcachefs_metadata_version_inode_btree_change &&
+	    btree_id == BTREE_ID_INODES)
+		swap(p->inode, p->offset);
+}
+
+static inline void compat_btree_node(unsigned level, enum btree_id btree_id,
+				     unsigned version, unsigned big_endian,
+				     int write,
+				     struct btree_node *bn)
+{
+	if (version < bcachefs_metadata_version_inode_btree_change &&
+	    btree_node_type_is_extents(btree_id) &&
+	    bkey_cmp(bn->min_key, POS_MIN) &&
+	    write)
+		bn->min_key = bkey_predecessor(bn->min_key);
+
+	compat_bpos(level, btree_id, version, big_endian, write, &bn->min_key);
+	compat_bpos(level, btree_id, version, big_endian, write, &bn->max_key);
+
+	if (version < bcachefs_metadata_version_inode_btree_change &&
+	    btree_node_type_is_extents(btree_id) &&
+	    bkey_cmp(bn->min_key, POS_MIN) &&
+	    !write)
+		bn->min_key = bkey_successor(bn->min_key);
+}
+
 #endif /* _BCACHEFS_BTREE_IO_H */
diff --git a/fs/bcachefs/btree_iter.c b/fs/bcachefs/btree_iter.c
index ea0555b806f0..5528ba0f1d44 100644
--- a/fs/bcachefs/btree_iter.c
+++ b/fs/bcachefs/btree_iter.c
@@ -5,6 +5,7 @@
 #include "btree_cache.h"
 #include "btree_iter.h"
 #include "btree_locking.h"
+#include "btree_update.h"
 #include "debug.h"
 #include "extents.h"
 
@@ -35,6 +36,26 @@ static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
 	return pos;
 }
 
+static inline bool btree_iter_pos_before_node(struct btree_iter *iter,
+					      struct btree *b)
+{
+	return bkey_cmp(btree_iter_search_key(iter), b->data->min_key) < 0;
+}
+
+static inline bool btree_iter_pos_after_node(struct btree_iter *iter,
+					     struct btree *b)
+{
+	return bkey_cmp(b->key.k.p, btree_iter_search_key(iter)) < 0;
+}
+
+static inline bool btree_iter_pos_in_node(struct btree_iter *iter,
+					  struct btree *b)
+{
+	return iter->btree_id == b->btree_id &&
+		!btree_iter_pos_before_node(iter, b) &&
+		!btree_iter_pos_after_node(iter, b);
+}
+
 /* Btree node locking: */
 
 void bch2_btree_node_unlock_write(struct btree *b, struct btree_iter *iter)
@@ -241,7 +262,7 @@ bool __bch2_btree_node_lock(struct btree *b, struct bpos pos,
 /* Btree iterator locking: */
 
 #ifdef CONFIG_BCACHEFS_DEBUG
-void bch2_btree_iter_verify_locks(struct btree_iter *iter)
+static void bch2_btree_iter_verify_locks(struct btree_iter *iter)
 {
 	unsigned l;
 
@@ -262,6 +283,8 @@ void bch2_btree_trans_verify_locks(struct btree_trans *trans)
 	trans_for_each_iter(trans, iter)
 		bch2_btree_iter_verify_locks(iter);
 }
+#else
+static inline void bch2_btree_iter_verify_locks(struct btree_iter *iter) {}
 #endif
 
 __flatten
@@ -385,21 +408,43 @@ void bch2_trans_unlock(struct btree_trans *trans)
 
 #ifdef CONFIG_BCACHEFS_DEBUG
 
-static void __bch2_btree_iter_verify(struct btree_iter *iter,
-				     struct btree *b)
+static void bch2_btree_iter_verify_level(struct btree_iter *iter,
+					 unsigned level)
 {
 	struct bpos pos = btree_iter_search_key(iter);
-	struct btree_iter_level *l = &iter->l[b->level];
+	struct btree_iter_level *l = &iter->l[level];
 	struct btree_node_iter tmp = l->iter;
-	struct bkey_packed *k;
+	bool locked = btree_node_locked(iter, level);
+	struct bkey_packed *p, *k;
+	char buf1[100], buf2[100];
+	const char *msg;
 
 	if (!debug_check_iterators(iter->trans->c))
 		return;
 
-	if (iter->uptodate > BTREE_ITER_NEED_PEEK)
+	BUG_ON(iter->level < iter->min_depth);
+
+	if (!btree_iter_node(iter, level))
 		return;
 
-	bch2_btree_node_iter_verify(&l->iter, b);
+	if (!bch2_btree_node_relock(iter, level))
+		return;
+
+	/*
+	 * Ideally this invariant would always be true, and hopefully in the
+	 * future it will be, but for now set_pos_same_leaf() breaks it:
+	 */
+	BUG_ON(iter->uptodate < BTREE_ITER_NEED_TRAVERSE &&
+	       !btree_iter_pos_in_node(iter, l->b));
+
+	/*
+	 * node iterators don't use leaf node iterator:
+	 */
+	if (btree_iter_type(iter) == BTREE_ITER_NODES &&
+	    level <= iter->min_depth)
+		goto unlock;
+
+	bch2_btree_node_iter_verify(&l->iter, l->b);
 
 	/*
 	 * For interior nodes, the iterator will have skipped past
@@ -408,46 +453,73 @@ 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
-		? 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) {
-		char buf[100];
-		struct bkey uk = bkey_unpack_key(b, k);
+	p = level || btree_node_type_is_extents(iter->btree_id)
+		? bch2_btree_node_iter_prev_filter(&tmp, l->b, KEY_TYPE_discard)
+		: bch2_btree_node_iter_prev_all(&tmp, l->b);
+	k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
 
-		bch2_bkey_to_text(&PBUF(buf), &uk);
-		panic("iterator should be before prev key:\n%s\n%llu:%llu\n",
-		      buf, iter->pos.inode, iter->pos.offset);
+	if (p && bkey_iter_pos_cmp(l->b, p, &pos) >= 0) {
+		msg = "before";
+		goto err;
 	}
 
-	k = bch2_btree_node_iter_peek_all(&l->iter, b);
-	if (k && bkey_iter_pos_cmp(b, k, &pos) < 0) {
-		char buf[100];
-		struct bkey uk = bkey_unpack_key(b, k);
+	if (k && bkey_iter_pos_cmp(l->b, k, &pos) < 0) {
+		msg = "after";
+		goto err;
+	}
+unlock:
+	if (!locked)
+		btree_node_unlock(iter, level);
+	return;
+err:
+	strcpy(buf1, "(none)");
+	strcpy(buf2, "(none)");
+
+	if (p) {
+		struct bkey uk = bkey_unpack_key(l->b, p);
+		bch2_bkey_to_text(&PBUF(buf1), &uk);
+	}
 
-		bch2_bkey_to_text(&PBUF(buf), &uk);
-		panic("iter should be after current key:\n"
-		      "iter pos %llu:%llu\n"
-		      "cur key  %s\n",
-		      iter->pos.inode, iter->pos.offset, buf);
+	if (k) {
+		struct bkey uk = bkey_unpack_key(l->b, k);
+		bch2_bkey_to_text(&PBUF(buf2), &uk);
 	}
+
+	panic("iterator should be %s key at level %u:\n"
+	      "iter pos %s %llu:%llu\n"
+	      "prev key %s\n"
+	      "cur  key %s\n",
+	      msg, level,
+	      iter->flags & BTREE_ITER_IS_EXTENTS ? ">" : "=>",
+	      iter->pos.inode, iter->pos.offset,
+	      buf1, buf2);
 }
 
-void bch2_btree_iter_verify(struct btree_iter *iter, struct btree *b)
+static void bch2_btree_iter_verify(struct btree_iter *iter)
 {
-	struct btree_iter *linked;
+	unsigned i;
 
-	if (!debug_check_iterators(iter->trans->c))
+	bch2_btree_trans_verify_locks(iter->trans);
+
+	for (i = 0; i < BTREE_MAX_DEPTH; i++)
+		bch2_btree_iter_verify_level(iter, i);
+}
+
+void bch2_btree_trans_verify_iters(struct btree_trans *trans, struct btree *b)
+{
+	struct btree_iter *iter;
+
+	if (!debug_check_iterators(trans->c))
 		return;
 
-	trans_for_each_iter_with_node(iter->trans, b, linked)
-		__bch2_btree_iter_verify(linked, b);
+	trans_for_each_iter_with_node(trans, b, iter)
+		bch2_btree_iter_verify_level(iter, b->level);
 }
 
 #else
 
-static inline void __bch2_btree_iter_verify(struct btree_iter *iter,
-					    struct btree *b) {}
+static inline void bch2_btree_iter_verify_level(struct btree_iter *iter, unsigned l) {}
+static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
 
 #endif
 
@@ -492,7 +564,7 @@ void bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
 
 	trans_for_each_iter_with_node(iter->trans, b, linked) {
 		__bch2_btree_iter_fix_key_modified(linked, b, where);
-		__bch2_btree_iter_verify(linked, b);
+		bch2_btree_iter_verify_level(linked, b->level);
 	}
 }
 
@@ -563,7 +635,7 @@ 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;
 
@@ -591,19 +663,8 @@ fixup_done:
 
 	if (!b->level &&
 	    node_iter == &iter->l[0].iter &&
-	    iter_current_key_modified) {
-		struct bkey_packed *k =
-			bch2_btree_node_iter_peek_all(node_iter, b);
-
-		if (likely(k)) {
-			bkey_disassemble(b, k, &iter->k);
-		} else {
-			/* XXX: for extents, calculate size of hole? */
-			iter->k.type = KEY_TYPE_deleted;
-		}
-
+	    iter_current_key_modified)
 		btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
-	}
 }
 
 void bch2_btree_node_iter_fix(struct btree_iter *iter,
@@ -619,14 +680,16 @@ void bch2_btree_node_iter_fix(struct btree_iter *iter,
 	if (node_iter != &iter->l[b->level].iter) {
 		__bch2_btree_node_iter_fix(iter, b, node_iter, t,
 					   where, clobber_u64s, new_u64s);
-		bch2_btree_node_iter_verify(node_iter, b);
+
+		if (debug_check_iterators(iter->trans->c))
+			bch2_btree_node_iter_verify(node_iter, b);
 	}
 
 	trans_for_each_iter_with_node(iter->trans, b, linked) {
 		__bch2_btree_node_iter_fix(linked, b,
 					   &linked->l[b->level].iter, t,
 					   where, clobber_u64s, new_u64s);
-		__bch2_btree_iter_verify(linked, b);
+		bch2_btree_iter_verify_level(linked, b->level);
 	}
 }
 
@@ -736,26 +799,6 @@ static void btree_iter_verify_new_node(struct btree_iter *iter, struct btree *b)
 		btree_node_unlock(iter, b->level + 1);
 }
 
-static inline bool btree_iter_pos_before_node(struct btree_iter *iter,
-					      struct btree *b)
-{
-	return bkey_cmp(iter->pos, b->data->min_key) < 0;
-}
-
-static inline bool btree_iter_pos_after_node(struct btree_iter *iter,
-					     struct btree *b)
-{
-	return bkey_cmp(b->key.k.p, btree_iter_search_key(iter)) < 0;
-}
-
-static inline bool btree_iter_pos_in_node(struct btree_iter *iter,
-					  struct btree *b)
-{
-	return iter->btree_id == b->btree_id &&
-		!btree_iter_pos_before_node(iter, b) &&
-		!btree_iter_pos_after_node(iter, b);
-}
-
 static inline void __btree_iter_init(struct btree_iter *iter,
 				     unsigned level)
 {
@@ -912,6 +955,27 @@ static void btree_iter_prefetch(struct btree_iter *iter)
 		btree_node_unlock(iter, iter->level);
 }
 
+static noinline void btree_node_mem_ptr_set(struct btree_iter *iter,
+					    unsigned plevel, struct btree *b)
+{
+	struct btree_iter_level *l = &iter->l[plevel];
+	bool locked = btree_node_locked(iter, plevel);
+	struct bkey_packed *k;
+	struct bch_btree_ptr_v2 *bp;
+
+	if (!bch2_btree_node_relock(iter, plevel))
+		return;
+
+	k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
+	BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
+
+	bp = (void *) bkeyp_val(&l->b->format, k);
+	bp->mem_ptr = (unsigned long)b;
+
+	if (!locked)
+		btree_node_unlock(iter, plevel);
+}
+
 static __always_inline int btree_iter_down(struct btree_iter *iter)
 {
 	struct bch_fs *c = iter->trans->c;
@@ -933,6 +997,10 @@ static __always_inline int btree_iter_down(struct btree_iter *iter)
 	mark_btree_node_locked(iter, level, lock_type);
 	btree_iter_node_set(iter, b);
 
+	if (tmp.k.k.type == KEY_TYPE_btree_ptr_v2 &&
+	    unlikely(b != btree_node_mem_ptr(&tmp.k)))
+		btree_node_mem_ptr_set(iter, level + 1, b);
+
 	if (iter->flags & BTREE_ITER_PREFETCH)
 		btree_iter_prefetch(iter);
 
@@ -1000,7 +1068,14 @@ retry_all:
 			goto retry_all;
 	}
 
-	ret = hweight64(trans->iters_live) > 1 ? -EINTR : 0;
+	if (hweight64(trans->iters_live) > 1)
+		ret = -EINTR;
+	else
+		trans_for_each_iter(trans, iter)
+			if (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) {
+				ret = -EINTR;
+				break;
+			}
 out:
 	bch2_btree_cache_cannibalize_unlock(c);
 	return ret;
@@ -1107,9 +1182,7 @@ static int btree_iter_traverse_one(struct btree_iter *iter)
 
 	iter->uptodate = BTREE_ITER_NEED_PEEK;
 
-	bch2_btree_trans_verify_locks(iter->trans);
-	if (btree_iter_node(iter, iter->level))
-		__bch2_btree_iter_verify(iter, iter->l[iter->level].b);
+	bch2_btree_iter_verify(iter);
 	return 0;
 }
 
@@ -1129,12 +1202,14 @@ static inline void bch2_btree_iter_checks(struct btree_iter *iter,
 					  enum btree_iter_type type)
 {
 	EBUG_ON(iter->btree_id >= BTREE_ID_NR);
-	EBUG_ON(!!(iter->flags & BTREE_ITER_IS_EXTENTS) !=
-		(btree_node_type_is_extents(iter->btree_id) &&
-		 type != BTREE_ITER_NODES));
 	EBUG_ON(btree_iter_type(iter) != type);
 
-	bch2_btree_trans_verify_locks(iter->trans);
+	BUG_ON(type == BTREE_ITER_KEYS &&
+	       (bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
+		bkey_cmp(iter->pos, iter->k.p) > 0));
+
+	bch2_btree_iter_verify_locks(iter);
+	bch2_btree_iter_verify_level(iter, iter->level);
 }
 
 /* Iterate across nodes (leaf and interior nodes) */
@@ -1162,10 +1237,12 @@ struct btree *bch2_btree_iter_peek_node(struct btree_iter *iter)
 	iter->pos = b->key.k.p;
 	iter->uptodate = BTREE_ITER_UPTODATE;
 
+	bch2_btree_iter_verify(iter);
+
 	return b;
 }
 
-struct btree *bch2_btree_iter_next_node(struct btree_iter *iter, unsigned depth)
+struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
 {
 	struct btree *b;
 	int ret;
@@ -1207,11 +1284,8 @@ struct btree *bch2_btree_iter_next_node(struct btree_iter *iter, unsigned depth)
 		if (btree_node_read_locked(iter, iter->level))
 			btree_node_unlock(iter, iter->level);
 
-		/* ick: */
-		iter->pos	= iter->btree_id == BTREE_ID_INODES
-			? btree_type_successor(iter->btree_id, iter->pos)
-			: bkey_successor(iter->pos);
-		iter->level	= depth;
+		iter->pos	= bkey_successor(iter->pos);
+		iter->level	= iter->min_depth;
 
 		btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
 		ret = bch2_btree_iter_traverse(iter);
@@ -1224,6 +1298,8 @@ struct btree *bch2_btree_iter_next_node(struct btree_iter *iter, unsigned depth)
 	iter->pos = b->key.k.p;
 	iter->uptodate = BTREE_ITER_UPTODATE;
 
+	bch2_btree_iter_verify(iter);
+
 	return b;
 }
 
@@ -1238,7 +1314,8 @@ void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *iter, struct bpos new_
 	EBUG_ON(!btree_node_locked(iter, 0));
 	EBUG_ON(bkey_cmp(new_pos, l->b->key.k.p) > 0);
 
-	iter->pos = new_pos;
+	bkey_init(&iter->k);
+	iter->k.p = iter->pos = new_pos;
 	btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
 
 	btree_iter_advance_to_pos(iter, l, -1);
@@ -1248,9 +1325,14 @@ void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *iter, struct bpos new_
 		btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
 }
 
-static unsigned btree_iter_pos_changed(struct btree_iter *iter, int cmp)
+static void btree_iter_pos_changed(struct btree_iter *iter, int cmp)
 {
-	unsigned l = btree_iter_up_until_good_node(iter, cmp);
+	unsigned l = iter->level;
+
+	if (!cmp)
+		goto out;
+
+	l = btree_iter_up_until_good_node(iter, cmp);
 
 	if (btree_iter_node(iter, l)) {
 		/*
@@ -1267,64 +1349,81 @@ static unsigned btree_iter_pos_changed(struct btree_iter *iter, int cmp)
 		if (btree_lock_want(iter, l) == BTREE_NODE_UNLOCKED)
 			btree_node_unlock(iter, l);
 	}
+out:
+	if (l != iter->level)
+		btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
+	else
+		btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
+}
 
-	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);
+	int cmp;
+
+	iter->flags &= ~BTREE_ITER_IS_EXTENTS;
+	iter->flags |= strictly_greater ? BTREE_ITER_IS_EXTENTS : 0;
+
+	bkey_init(&iter->k);
+	iter->k.p = iter->pos = new_pos;
+
+	cmp = bkey_cmp(btree_iter_search_key(iter), old);
+
+	btree_iter_pos_changed(iter, cmp);
 }
 
 void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 {
 	int cmp = bkey_cmp(new_pos, iter->pos);
-	unsigned l;
-
-	if (!cmp)
-		return;
-
-	iter->pos = new_pos;
 
-	l = btree_iter_pos_changed(iter, cmp);
+	bkey_init(&iter->k);
+	iter->k.p = iter->pos = new_pos;
 
-	if (l != iter->level)
-		btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
-	else
-		btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
+	btree_iter_pos_changed(iter, cmp);
 }
 
 static inline bool btree_iter_set_pos_to_next_leaf(struct btree_iter *iter)
 {
 	struct btree_iter_level *l = &iter->l[0];
+	bool ret;
 
-	iter->pos	= l->b->key.k.p;
-	iter->uptodate	= BTREE_ITER_NEED_TRAVERSE;
+	bkey_init(&iter->k);
+	iter->k.p = iter->pos = l->b->key.k.p;
 
-	if (!bkey_cmp(iter->pos, POS_MAX)) {
-		bkey_init(&iter->k);
-		iter->k.p	= POS_MAX;
-		return false;
-	}
+	ret = bkey_cmp(iter->pos, POS_MAX) != 0;
+	if (ret && !(iter->flags & BTREE_ITER_IS_EXTENTS))
+		iter->k.p = iter->pos = bkey_successor(iter->pos);
 
-	iter->pos = btree_type_successor(iter->btree_id, iter->pos);
 	btree_iter_pos_changed(iter, 1);
-	return true;
+	return ret;
 }
 
 static inline bool btree_iter_set_pos_to_prev_leaf(struct btree_iter *iter)
 {
 	struct btree_iter_level *l = &iter->l[0];
+	bool ret;
 
-	iter->pos	= l->b->data->min_key;
+	bkey_init(&iter->k);
+	iter->k.p = iter->pos = l->b->data->min_key;
 	iter->uptodate	= BTREE_ITER_NEED_TRAVERSE;
 
-	if (!bkey_cmp(iter->pos, POS_MIN)) {
-		bkey_init(&iter->k);
-		iter->k.p	= POS_MIN;
-		return false;
+	ret = bkey_cmp(iter->pos, POS_MIN) != 0;
+	if (ret) {
+		iter->k.p = iter->pos = bkey_predecessor(iter->pos);
+
+		if (iter->flags & BTREE_ITER_IS_EXTENTS)
+			iter->k.p = iter->pos = bkey_predecessor(iter->pos);
 	}
 
-	iter->pos = btree_type_predecessor(iter->btree_id, iter->pos);
 	btree_iter_pos_changed(iter, -1);
-	return true;
+	return ret;
 }
 
+/**
+ * btree_iter_peek_uptodate - given an iterator that is uptodate, return the key
+ * it currently points to
+ */
 static inline struct bkey_s_c btree_iter_peek_uptodate(struct btree_iter *iter)
 {
 	struct btree_iter_level *l = &iter->l[0];
@@ -1361,7 +1460,8 @@ struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
 
 	bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
 
-	if (iter->uptodate == BTREE_ITER_UPTODATE)
+	if (iter->uptodate == BTREE_ITER_UPTODATE &&
+	    !bkey_deleted(&iter->k))
 		return btree_iter_peek_uptodate(iter);
 
 	while (1) {
@@ -1386,6 +1486,8 @@ struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
 		iter->pos = bkey_start_pos(k.k);
 
 	iter->uptodate = BTREE_ITER_UPTODATE;
+
+	bch2_btree_iter_verify_level(iter, 0);
 	return k;
 }
 
@@ -1395,52 +1497,101 @@ struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
  */
 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
 {
+	if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
+		return bkey_s_c_null;
+
+	bch2_btree_iter_set_pos(iter,
+		(iter->flags & BTREE_ITER_IS_EXTENTS)
+		? iter->k.p
+		: bkey_successor(iter->k.p));
+
+	return bch2_btree_iter_peek(iter);
+}
+
+static struct bkey_s_c __btree_trans_updates_peek(struct btree_iter *iter)
+{
+	struct bpos pos = btree_iter_search_key(iter);
+	struct btree_trans *trans = iter->trans;
+	struct btree_insert_entry *i;
+
+	trans_for_each_update2(trans, i)
+		if ((cmp_int(iter->btree_id,	i->iter->btree_id) ?:
+		     bkey_cmp(pos,		i->k->k.p)) <= 0)
+			break;
+
+	return i < trans->updates2 + trans->nr_updates2 &&
+		iter->btree_id == i->iter->btree_id
+		? bkey_i_to_s_c(i->k)
+		: bkey_s_c_null;
+}
+
+static struct bkey_s_c __bch2_btree_iter_peek_with_updates(struct btree_iter *iter)
+{
 	struct btree_iter_level *l = &iter->l[0];
-	struct bkey_packed *p;
+	struct bkey_s_c k = __btree_iter_peek(iter, l);
+	struct bkey_s_c u = __btree_trans_updates_peek(iter);
+
+	if (k.k && (!u.k || bkey_cmp(k.k->p, u.k->p) < 0))
+		return k;
+	if (u.k && bkey_cmp(u.k->p, l->b->key.k.p) <= 0) {
+		iter->k = *u.k;
+		return u;
+	}
+	return bkey_s_c_null;
+}
+
+struct bkey_s_c bch2_btree_iter_peek_with_updates(struct btree_iter *iter)
+{
 	struct bkey_s_c k;
+	int ret;
 
 	bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
 
-	if (unlikely(iter->uptodate != BTREE_ITER_UPTODATE)) {
-		if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
-			return bkey_s_c_null;
+	while (1) {
+		ret = bch2_btree_iter_traverse(iter);
+		if (unlikely(ret))
+			return bkey_s_c_err(ret);
 
-		/*
-		 * XXX: when we just need to relock we should be able to avoid
-		 * calling traverse, but we need to kill BTREE_ITER_NEED_PEEK
-		 * for that to work
-		 */
-		iter->uptodate	= BTREE_ITER_NEED_TRAVERSE;
+		k = __bch2_btree_iter_peek_with_updates(iter);
 
-		bch2_btree_iter_set_pos(iter,
-			btree_type_successor(iter->btree_id, iter->k.p));
+		if (k.k && bkey_deleted(k.k)) {
+			bch2_btree_iter_set_pos(iter,
+				(iter->flags & BTREE_ITER_IS_EXTENTS)
+				? iter->k.p
+				: bkey_successor(iter->k.p));
+			continue;
+		}
 
-		return bch2_btree_iter_peek(iter);
-	}
+		if (likely(k.k))
+			break;
 
-	if (unlikely(bkey_deleted(&iter->k))) {
-		/*
-		 * we're currently pointed at a hole, because previously we were
-		 * iterating over slots:
-		 */
-		return bch2_btree_iter_peek(iter);
+		if (!btree_iter_set_pos_to_next_leaf(iter))
+			return bkey_s_c_null;
 	}
 
-	do {
-		bch2_btree_node_iter_advance(&l->iter, l->b);
-		p = bch2_btree_node_iter_peek_all(&l->iter, l->b);
-	} while (likely(p) && bkey_whiteout(p));
+	/*
+	 * iter->pos should always be equal to the key we just
+	 * returned - except extents can straddle iter->pos:
+	 */
+	if (!(iter->flags & BTREE_ITER_IS_EXTENTS) ||
+	    bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
+		iter->pos = bkey_start_pos(k.k);
 
-	if (unlikely(!p))
-		return btree_iter_set_pos_to_next_leaf(iter)
-			? bch2_btree_iter_peek(iter)
-			: bkey_s_c_null;
+	iter->uptodate = BTREE_ITER_UPTODATE;
+	return k;
+}
 
-	k = __btree_iter_unpack(iter, l, &iter->k, p);
+struct bkey_s_c bch2_btree_iter_next_with_updates(struct btree_iter *iter)
+{
+	if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
+		return bkey_s_c_null;
 
-	EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) < 0);
-	iter->pos = bkey_start_pos(k.k);
-	return k;
+	bch2_btree_iter_set_pos(iter,
+		(iter->flags & BTREE_ITER_IS_EXTENTS)
+		? iter->k.p
+		: bkey_successor(iter->k.p));
+
+	return bch2_btree_iter_peek_with_updates(iter);
 }
 
 /**
@@ -1449,13 +1600,15 @@ struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
  */
 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
 {
+	struct bpos pos = iter->pos;
 	struct btree_iter_level *l = &iter->l[0];
 	struct bkey_s_c k;
 	int ret;
 
 	bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
 
-	if (iter->uptodate == BTREE_ITER_UPTODATE)
+	if (iter->uptodate == BTREE_ITER_UPTODATE &&
+	    !bkey_deleted(&iter->k))
 		return btree_iter_peek_uptodate(iter);
 
 	while (1) {
@@ -1464,8 +1617,7 @@ struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
 			return bkey_s_c_err(ret);
 
 		k = __btree_iter_peek(iter, l);
-		if (!k.k ||
-		    bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
+		if (!k.k || bkey_cmp(bkey_start_pos(k.k), pos) > 0)
 			k = __btree_iter_prev(iter, l);
 
 		if (likely(k.k))
@@ -1475,7 +1627,7 @@ struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
 			return bkey_s_c_null;
 	}
 
-	EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0);
+	EBUG_ON(bkey_cmp(bkey_start_pos(k.k), pos) > 0);
 	iter->pos	= bkey_start_pos(k.k);
 	iter->uptodate	= BTREE_ITER_UPTODATE;
 	return k;
@@ -1487,33 +1639,16 @@ struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
  */
 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
 {
-	struct btree_iter_level *l = &iter->l[0];
-	struct bkey_s_c k;
+	struct bpos pos = bkey_start_pos(&iter->k);
 
 	bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
 
-	if (unlikely(iter->uptodate != BTREE_ITER_UPTODATE)) {
-		/*
-		 * XXX: when we just need to relock we should be able to avoid
-		 * calling traverse, but we need to kill BTREE_ITER_NEED_PEEK
-		 * for that to work
-		 */
-		iter->pos	= btree_type_predecessor(iter->btree_id,
-							 iter->pos);
-		iter->uptodate	= BTREE_ITER_NEED_TRAVERSE;
-
-		return bch2_btree_iter_peek_prev(iter);
-	}
+	if (unlikely(!bkey_cmp(pos, POS_MIN)))
+		return bkey_s_c_null;
 
-	k = __btree_iter_prev(iter, l);
-	if (unlikely(!k.k))
-		return btree_iter_set_pos_to_prev_leaf(iter)
-			? bch2_btree_iter_peek(iter)
-			: bkey_s_c_null;
+	bch2_btree_iter_set_pos(iter, bkey_predecessor(pos));
 
-	EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) >= 0);
-	iter->pos	= bkey_start_pos(k.k);
-	return k;
+	return bch2_btree_iter_peek_prev(iter);
 }
 
 static inline struct bkey_s_c
@@ -1525,8 +1660,17 @@ __bch2_btree_iter_peek_slot_extents(struct btree_iter *iter)
 	struct bkey n;
 	int ret;
 
-recheck:
-	btree_iter_advance_to_pos(iter, l, -1);
+	/* keys & holes can't span inode numbers: */
+	if (iter->pos.offset == KEY_OFFSET_MAX) {
+		if (iter->pos.inode == KEY_INODE_MAX)
+			return bkey_s_c_null;
+
+		bch2_btree_iter_set_pos(iter, bkey_successor(iter->pos));
+
+		ret = bch2_btree_iter_traverse(iter);
+		if (unlikely(ret))
+			return bkey_s_c_err(ret);
+	}
 
 	/*
 	 * iterator is now at the correct position for inserting at iter->pos,
@@ -1540,47 +1684,17 @@ recheck:
 
 	if (k.k && bkey_cmp(bkey_start_pos(k.k), iter->pos) <= 0) {
 		/*
-		 * If there wasn't actually a hole, want the iterator to be
-		 * pointed at the key we found:
-		 *
-		 * XXX: actually, we shouldn't be changing the iterator here:
-		 * the iterator needs to be correct for inserting at iter->pos,
-		 * and there may be whiteouts between iter->pos and what this
-		 * iterator points at:
+		 * We're not setting iter->uptodate because the node iterator
+		 * doesn't necessarily point at the key we're returning:
 		 */
-		l->iter = node_iter;
 
 		EBUG_ON(bkey_cmp(k.k->p, iter->pos) <= 0);
-		iter->uptodate = BTREE_ITER_UPTODATE;
-
-		__bch2_btree_iter_verify(iter, l->b);
+		bch2_btree_iter_verify_level(iter, 0);
 		return k;
 	}
 
-	/*
-	 * If we got to the end of the node, check if we need to traverse to the
-	 * next node:
-	 */
-	if (unlikely(!k.k && btree_iter_pos_after_node(iter, l->b))) {
-		btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
-		ret = bch2_btree_iter_traverse(iter);
-		if (unlikely(ret))
-			return bkey_s_c_err(ret);
-
-		goto recheck;
-	}
-
 	/* hole */
 
-	/* holes can't span inode numbers: */
-	if (iter->pos.offset == KEY_OFFSET_MAX) {
-		if (iter->pos.inode == KEY_INODE_MAX)
-			return bkey_s_c_null;
-
-		iter->pos = bkey_successor(iter->pos);
-		goto recheck;
-	}
-
 	if (!k.k)
 		k.k = &l->b->key.k;
 
@@ -1598,42 +1712,33 @@ recheck:
 	iter->k	= n;
 	iter->uptodate = BTREE_ITER_UPTODATE;
 
-	__bch2_btree_iter_verify(iter, l->b);
+	bch2_btree_iter_verify_level(iter, 0);
 	return (struct bkey_s_c) { &iter->k, NULL };
 }
 
-static inline struct bkey_s_c
-__bch2_btree_iter_peek_slot(struct btree_iter *iter)
+struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
 {
 	struct btree_iter_level *l = &iter->l[0];
 	struct bkey_s_c k;
 	int ret;
 
+	bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
+
+	if (iter->uptodate == BTREE_ITER_UPTODATE)
+		return btree_iter_peek_uptodate(iter);
+
+	ret = bch2_btree_iter_traverse(iter);
+	if (unlikely(ret))
+		return bkey_s_c_err(ret);
+
 	if (iter->flags & BTREE_ITER_IS_EXTENTS)
 		return __bch2_btree_iter_peek_slot_extents(iter);
 
-recheck:
-	while ((k = __btree_iter_peek_all(iter, l, &iter->k)).k &&
-	       bkey_deleted(k.k) &&
-	       bkey_cmp(k.k->p, iter->pos) == 0)
-		bch2_btree_node_iter_advance(&l->iter, l->b);
-
-	/*
-	 * If we got to the end of the node, check if we need to traverse to the
-	 * next node:
-	 */
-	if (unlikely(!k.k && btree_iter_pos_after_node(iter, l->b))) {
-		btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
-		ret = bch2_btree_iter_traverse(iter);
-		if (unlikely(ret))
-			return bkey_s_c_err(ret);
+	k = __btree_iter_peek_all(iter, l, &iter->k);
 
-		goto recheck;
-	}
+	EBUG_ON(k.k && bkey_deleted(k.k) && bkey_cmp(k.k->p, iter->pos) == 0);
 
-	if (!k.k ||
-	    bkey_deleted(k.k) ||
-	    bkey_cmp(iter->pos, k.k->p)) {
+	if (!k.k || bkey_cmp(iter->pos, k.k->p)) {
 		/* hole */
 		bkey_init(&iter->k);
 		iter->k.p = iter->pos;
@@ -1641,49 +1746,21 @@ recheck:
 	}
 
 	iter->uptodate = BTREE_ITER_UPTODATE;
-	__bch2_btree_iter_verify(iter, l->b);
+	bch2_btree_iter_verify_level(iter, 0);
 	return k;
 }
 
-struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
-{
-	int ret;
-
-	bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
-
-	if (iter->uptodate == BTREE_ITER_UPTODATE)
-		return btree_iter_peek_uptodate(iter);
-
-	ret = bch2_btree_iter_traverse(iter);
-	if (unlikely(ret))
-		return bkey_s_c_err(ret);
-
-	return __bch2_btree_iter_peek_slot(iter);
-}
-
 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
 {
-	bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
-
-	iter->pos = btree_type_successor(iter->btree_id, iter->k.p);
-
-	if (unlikely(iter->uptodate != BTREE_ITER_UPTODATE)) {
-		/*
-		 * XXX: when we just need to relock we should be able to avoid
-		 * calling traverse, but we need to kill BTREE_ITER_NEED_PEEK
-		 * for that to work
-		 */
-		btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
-
-		return bch2_btree_iter_peek_slot(iter);
-	}
+	if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
+		return bkey_s_c_null;
 
-	if (!bkey_deleted(&iter->k))
-		bch2_btree_node_iter_advance(&iter->l[0].iter, iter->l[0].b);
+	bch2_btree_iter_set_pos(iter,
+		(iter->flags & BTREE_ITER_IS_EXTENTS)
+		? iter->k.p
+		: bkey_successor(iter->k.p));
 
-	btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
-
-	return __bch2_btree_iter_peek_slot(iter);
+	return bch2_btree_iter_peek_slot(iter);
 }
 
 static inline void bch2_btree_iter_init(struct btree_trans *trans,
@@ -1705,12 +1782,12 @@ static inline void bch2_btree_iter_init(struct btree_trans *trans,
 	iter->uptodate			= BTREE_ITER_NEED_TRAVERSE;
 	iter->btree_id			= btree_id;
 	iter->level			= 0;
+	iter->min_depth			= 0;
 	iter->locks_want		= flags & BTREE_ITER_INTENT ? 1 : 0;
 	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 +1806,14 @@ 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;
+
+	BUG_ON(trans->iters + iter->idx != iter);
+
+	ret = btree_iter_err(iter);
 
 	if (!(trans->iters_touched & (1ULL << iter->idx)) &&
 	    !(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT))
@@ -1742,6 +1826,9 @@ 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);
@@ -1750,7 +1837,7 @@ int bch2_trans_iter_free(struct btree_trans *trans,
 static int bch2_trans_realloc_iters(struct btree_trans *trans,
 				    unsigned new_size)
 {
-	void *new_iters, *new_updates;
+	void *p, *new_iters, *new_updates, *new_updates2;
 	size_t iters_bytes;
 	size_t updates_bytes;
 
@@ -1768,21 +1855,27 @@ static int bch2_trans_realloc_iters(struct btree_trans *trans,
 	iters_bytes	= sizeof(struct btree_iter) * new_size;
 	updates_bytes	= sizeof(struct btree_insert_entry) * new_size;
 
-	new_iters = kmalloc(iters_bytes + updates_bytes, GFP_NOFS);
-	if (new_iters)
+	p = kmalloc(iters_bytes +
+		    updates_bytes +
+		    updates_bytes, GFP_NOFS);
+	if (p)
 		goto success;
 
-	new_iters = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS);
+	p = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS);
 	new_size = BTREE_ITER_MAX;
 
 	trans->used_mempool = true;
 success:
-	new_updates	= new_iters + iters_bytes;
+	new_iters	= p; p += iters_bytes;
+	new_updates	= p; p += updates_bytes;
+	new_updates2	= p; p += updates_bytes;
 
 	memcpy(new_iters, trans->iters,
 	       sizeof(struct btree_iter) * trans->nr_iters);
 	memcpy(new_updates, trans->updates,
 	       sizeof(struct btree_insert_entry) * trans->nr_updates);
+	memcpy(new_updates2, trans->updates2,
+	       sizeof(struct btree_insert_entry) * trans->nr_updates2);
 
 	if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 		memset(trans->iters, POISON_FREE,
@@ -1794,6 +1887,7 @@ success:
 
 	trans->iters		= new_iters;
 	trans->updates		= new_updates;
+	trans->updates2		= new_updates2;
 	trans->size		= new_size;
 
 	if (trans->iters_live) {
@@ -1818,13 +1912,14 @@ static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *trans)
 			struct btree_iter *iter;
 
 			trans_for_each_iter(trans, iter) {
-				pr_err("iter: btree %s pos %llu:%llu%s%s%s",
+				pr_err("iter: btree %s pos %llu:%llu%s%s%s %pf",
 				       bch2_btree_ids[iter->btree_id],
 				       iter->pos.inode,
 				       iter->pos.offset,
 				       (trans->iters_live & (1ULL << iter->idx)) ? " live" : "",
 				       (trans->iters_touched & (1ULL << iter->idx)) ? " touched" : "",
-				       iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT ? " keep" : "");
+				       iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT ? " keep" : "",
+				       (void *) iter->ip_allocated);
 			}
 
 			panic("trans iter oveflow\n");
@@ -1931,15 +2026,16 @@ static struct btree_iter *__btree_trans_get_iter(struct btree_trans *trans,
 	return iter;
 }
 
-struct btree_iter *bch2_trans_get_iter(struct btree_trans *trans,
-				       enum btree_id btree_id,
-				       struct bpos pos, unsigned flags)
+struct btree_iter *__bch2_trans_get_iter(struct btree_trans *trans,
+					 enum btree_id btree_id,
+					 struct bpos pos, unsigned flags)
 {
 	struct btree_iter *iter =
 		__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;
 }
 
@@ -1960,6 +2056,7 @@ struct btree_iter *bch2_trans_get_node_iter(struct btree_trans *trans,
 
 	iter->locks_want = locks_want;
 	iter->level	= depth;
+	iter->min_depth	= depth;
 
 	for (i = 0; i < ARRAY_SIZE(iter->l); i++)
 		iter->l[i].b		= NULL;
@@ -1968,7 +2065,7 @@ struct btree_iter *bch2_trans_get_node_iter(struct btree_trans *trans,
 	return iter;
 }
 
-struct btree_iter *bch2_trans_copy_iter(struct btree_trans *trans,
+struct btree_iter *__bch2_trans_copy_iter(struct btree_trans *trans,
 					struct btree_iter *src)
 {
 	struct btree_iter *iter;
@@ -1981,8 +2078,8 @@ 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);
 
@@ -2049,16 +2146,12 @@ void bch2_trans_reset(struct btree_trans *trans, unsigned flags)
 
 	bch2_trans_unlink_iters(trans);
 
-	if (flags & TRANS_RESET_ITERS)
-		trans->iters_live = 0;
-
 	trans->iters_touched &= trans->iters_live;
 
 	trans->need_reset		= 0;
 	trans->nr_updates		= 0;
-
-	if (flags & TRANS_RESET_MEM)
-		trans->mem_top		= 0;
+	trans->nr_updates2		= 0;
+	trans->mem_top			= 0;
 
 	if (trans->fs_usage_deltas) {
 		trans->fs_usage_deltas->used = 0;
@@ -2077,11 +2170,18 @@ void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
 {
 	memset(trans, 0, offsetof(struct btree_trans, iters_onstack));
 
+	/*
+	 * reallocating iterators currently completely breaks
+	 * bch2_trans_iter_put():
+	 */
+	expected_nr_iters = BTREE_ITER_MAX;
+
 	trans->c		= c;
 	trans->ip		= _RET_IP_;
 	trans->size		= ARRAY_SIZE(trans->iters_onstack);
 	trans->iters		= trans->iters_onstack;
 	trans->updates		= trans->updates_onstack;
+	trans->updates2		= trans->updates2_onstack;
 	trans->fs_usage_deltas	= NULL;
 
 	if (expected_nr_iters > trans->size)
@@ -2119,5 +2219,5 @@ int bch2_fs_btree_iter_init(struct bch_fs *c)
 	return mempool_init_kmalloc_pool(&c->btree_iters_pool, 1,
 			sizeof(struct btree_iter) * nr +
 			sizeof(struct btree_insert_entry) * nr +
-			sizeof(u8) * nr);
+			sizeof(struct btree_insert_entry) * nr);
 }
diff --git a/fs/bcachefs/btree_iter.h b/fs/bcachefs/btree_iter.h
index 962380925511..6456787a8f77 100644
--- a/fs/bcachefs/btree_iter.h
+++ b/fs/bcachefs/btree_iter.h
@@ -96,11 +96,11 @@ __trans_next_iter_with_node(struct btree_trans *trans, struct btree *b,
 						 (_iter)->idx + 1))
 
 #ifdef CONFIG_BCACHEFS_DEBUG
-void bch2_btree_iter_verify(struct btree_iter *, struct btree *);
+void bch2_btree_trans_verify_iters(struct btree_trans *, struct btree *);
 void bch2_btree_trans_verify_locks(struct btree_trans *);
 #else
-static inline void bch2_btree_iter_verify(struct btree_iter *iter,
-					  struct btree *b) {}
+static inline void bch2_btree_trans_verify_iters(struct btree_trans *trans,
+						 struct btree *b) {}
 static inline void bch2_btree_trans_verify_locks(struct btree_trans *iter) {}
 #endif
 
@@ -154,11 +154,14 @@ bch2_btree_iter_traverse(struct btree_iter *iter)
 int bch2_btree_iter_traverse_all(struct btree_trans *);
 
 struct btree *bch2_btree_iter_peek_node(struct btree_iter *);
-struct btree *bch2_btree_iter_next_node(struct btree_iter *, unsigned);
+struct btree *bch2_btree_iter_next_node(struct btree_iter *);
 
 struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *);
 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *);
 
+struct bkey_s_c bch2_btree_iter_peek_with_updates(struct btree_iter *);
+struct bkey_s_c bch2_btree_iter_next_with_updates(struct btree_iter *);
+
 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *);
 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *);
 
@@ -166,41 +169,14 @@ 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,
-					       struct bpos pos)
-{
-	if (id == BTREE_ID_INODES) {
-		pos.inode++;
-		pos.offset = 0;
-	} else if (!btree_node_type_is_extents(id)) {
-		pos = bkey_successor(pos);
-	}
-
-	return pos;
-}
-
-static inline struct bpos btree_type_predecessor(enum btree_id id,
-					       struct bpos pos)
-{
-	if (id == BTREE_ID_INODES) {
-		--pos.inode;
-		pos.offset = 0;
-	} else {
-		pos = bkey_predecessor(pos);
-	}
-
-	return pos;
-}
-
 static inline int __btree_iter_cmp(enum btree_id id,
 				   struct bpos pos,
 				   const struct btree_iter *r)
 {
-	if (id != r->btree_id)
-		return id < r->btree_id ? -1 : 1;
-	return bkey_cmp(pos, r->pos);
+	return cmp_int(id, r->btree_id) ?: bkey_cmp(pos, r->pos);
 }
 
 static inline int btree_iter_cmp(const struct btree_iter *l,
@@ -230,7 +206,7 @@ static inline int bch2_trans_cond_resched(struct btree_trans *trans)
 				_start, _locks_want, _depth, _flags),	\
 	     _b = bch2_btree_iter_peek_node(_iter);			\
 	     (_b);							\
-	     (_b) = bch2_btree_iter_next_node(_iter, _depth))
+	     (_b) = bch2_btree_iter_next_node(_iter))
 
 #define for_each_btree_node(_trans, _iter, _btree_id, _start,		\
 			    _flags, _b)					\
@@ -281,23 +257,46 @@ int bch2_trans_iter_free(struct btree_trans *, struct btree_iter *);
 
 void bch2_trans_unlink_iters(struct btree_trans *);
 
-struct btree_iter *bch2_trans_get_iter(struct btree_trans *, enum btree_id,
-				       struct bpos, unsigned);
-struct btree_iter *bch2_trans_copy_iter(struct btree_trans *,
+struct btree_iter *__bch2_trans_get_iter(struct btree_trans *, enum btree_id,
+					 struct bpos, unsigned);
+
+static inline struct btree_iter *
+bch2_trans_get_iter(struct btree_trans *trans, enum btree_id btree_id,
+		    struct bpos pos, unsigned flags)
+{
+	struct btree_iter *iter =
+		__bch2_trans_get_iter(trans, btree_id, pos, flags);
+
+	if (!IS_ERR(iter))
+		iter->ip_allocated = _THIS_IP_;
+	return iter;
+}
+
+struct btree_iter *__bch2_trans_copy_iter(struct btree_trans *,
 					struct btree_iter *);
+static inline struct btree_iter *
+bch2_trans_copy_iter(struct btree_trans *trans, struct btree_iter *src)
+{
+	struct btree_iter *iter =
+		__bch2_trans_copy_iter(trans, src);
+
+	if (!IS_ERR(iter))
+		iter->ip_allocated = _THIS_IP_;
+	return iter;
+
+}
+
 struct btree_iter *bch2_trans_get_node_iter(struct btree_trans *,
 				enum btree_id, struct bpos,
 				unsigned, unsigned, unsigned);
 
-#define TRANS_RESET_ITERS		(1 << 0)
-#define TRANS_RESET_MEM			(1 << 1)
-#define TRANS_RESET_NOTRAVERSE		(1 << 2)
+#define TRANS_RESET_NOTRAVERSE		(1 << 0)
 
 void bch2_trans_reset(struct btree_trans *, unsigned);
 
 static inline void bch2_trans_begin(struct btree_trans *trans)
 {
-	return bch2_trans_reset(trans, TRANS_RESET_ITERS|TRANS_RESET_MEM);
+	return bch2_trans_reset(trans, 0);
 }
 
 void *bch2_trans_kmalloc(struct btree_trans *, size_t);
diff --git a/fs/bcachefs/btree_types.h b/fs/bcachefs/btree_types.h
index b7af88e05837..732cdc35aa7c 100644
--- a/fs/bcachefs/btree_types.h
+++ b/fs/bcachefs/btree_types.h
@@ -53,7 +53,6 @@ struct bset_tree {
 
 struct btree_write {
 	struct journal_entry_pin	journal;
-	struct closure_waitlist		wait;
 };
 
 struct btree_alloc {
@@ -64,9 +63,7 @@ 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 +130,9 @@ 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 {
@@ -234,9 +234,10 @@ struct btree_iter {
 	u16			flags;
 	u8			idx;
 
-	enum btree_iter_uptodate uptodate:4;
 	enum btree_id		btree_id:4;
+	enum btree_iter_uptodate uptodate:4;
 	unsigned		level:4,
+				min_depth:4,
 				locks_want:4,
 				nodes_locked:4,
 				nodes_intent_locked:4;
@@ -252,6 +253,7 @@ struct btree_iter {
 	 * bch2_btree_iter_next_slot() can correctly advance pos.
 	 */
 	struct bkey		k;
+	unsigned long		ip_allocated;
 };
 
 static inline enum btree_iter_type btree_iter_type(struct btree_iter *iter)
@@ -259,6 +261,11 @@ static inline enum btree_iter_type btree_iter_type(struct btree_iter *iter)
 	return iter->flags & BTREE_ITER_TYPE;
 }
 
+static inline struct btree_iter_level *iter_l(struct btree_iter *iter)
+{
+	return iter->l + iter->level;
+}
+
 struct btree_insert_entry {
 	unsigned		trigger_flags;
 	unsigned		trans_triggers_run:1;
@@ -266,7 +273,11 @@ struct btree_insert_entry {
 	struct btree_iter	*iter;
 };
 
+#ifndef CONFIG_LOCKDEP
 #define BTREE_ITER_MAX		64
+#else
+#define BTREE_ITER_MAX		32
+#endif
 
 struct btree_trans {
 	struct bch_fs		*c;
@@ -278,6 +289,7 @@ struct btree_trans {
 
 	u8			nr_iters;
 	u8			nr_updates;
+	u8			nr_updates2;
 	u8			size;
 	unsigned		used_mempool:1;
 	unsigned		error:1;
@@ -290,6 +302,7 @@ struct btree_trans {
 
 	struct btree_iter	*iters;
 	struct btree_insert_entry *updates;
+	struct btree_insert_entry *updates2;
 
 	/* update path: */
 	struct journal_res	journal_res;
@@ -303,6 +316,7 @@ struct btree_trans {
 
 	struct btree_iter	iters_onstack[2];
 	struct btree_insert_entry updates_onstack[2];
+	struct btree_insert_entry updates2_onstack[2];
 };
 
 #define BTREE_FLAG(flag)						\
@@ -534,8 +548,6 @@ static inline bool btree_node_type_needs_gc(enum btree_node_type type)
 struct btree_root {
 	struct btree		*b;
 
-	struct btree_update	*as;
-
 	/* On disk root - see async splits: */
 	__BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
 	u8			level;
diff --git a/fs/bcachefs/btree_update.h b/fs/bcachefs/btree_update.h
index 2c34bae64281..11f7d02de622 100644
--- a/fs/bcachefs/btree_update.h
+++ b/fs/bcachefs/btree_update.h
@@ -12,8 +12,7 @@ void bch2_btree_node_lock_for_insert(struct bch_fs *, struct btree *,
 				     struct btree_iter *);
 bool bch2_btree_bset_insert_key(struct btree_iter *, struct btree *,
 				struct btree_node_iter *, struct bkey_i *);
-void bch2_btree_journal_key(struct btree_trans *, struct btree_iter *,
-			    struct bkey_i *);
+void bch2_btree_add_journal_pin(struct bch_fs *, struct btree *, u64);
 
 enum btree_insert_flags {
 	__BTREE_INSERT_NOUNLOCK,
@@ -59,6 +58,7 @@ enum btree_insert_flags {
 
 int bch2_btree_delete_at(struct btree_trans *, struct btree_iter *, unsigned);
 
+int __bch2_btree_insert(struct btree_trans *, enum btree_id, struct bkey_i *);
 int bch2_btree_insert(struct bch_fs *, enum btree_id, struct bkey_i *,
 		     struct disk_reservation *, u64 *, int flags);
 
@@ -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);
@@ -98,17 +98,17 @@ static inline int bch2_trans_commit(struct btree_trans *trans,
 	return __bch2_trans_commit(trans);
 }
 
-#define __bch2_trans_do(_trans, _disk_res, _journal_seq,		\
-			_flags,	_reset_flags, _do)			\
+#define __bch2_trans_do(_trans, _disk_res, _journal_seq, _flags, _do)	\
 ({									\
 	int _ret;							\
 									\
-	do {								\
-		bch2_trans_reset(_trans, _reset_flags);			\
-									\
+	while (1) {							\
 		_ret = (_do) ?:	bch2_trans_commit(_trans, (_disk_res),	\
 					(_journal_seq), (_flags));	\
-	} while (_ret == -EINTR);					\
+		if (_ret != -EINTR)					\
+			break;						\
+		bch2_trans_reset(_trans, 0);				\
+	}								\
 									\
 	_ret;								\
 })
@@ -120,7 +120,7 @@ static inline int bch2_trans_commit(struct btree_trans *trans,
 									\
 	bch2_trans_init(&trans, (_c), 0, 0);				\
 	_ret = __bch2_trans_do(&trans, _disk_res, _journal_seq, _flags,	\
-			       TRANS_RESET_MEM|TRANS_RESET_ITERS, _do);	\
+			       _do);					\
 	_ret2 = bch2_trans_exit(&trans);				\
 									\
 	_ret ?: _ret2;							\
@@ -131,4 +131,9 @@ static inline int bch2_trans_commit(struct btree_trans *trans,
 	     (_i) < (_trans)->updates + (_trans)->nr_updates;		\
 	     (_i)++)
 
+#define trans_for_each_update2(_trans, _i)				\
+	for ((_i) = (_trans)->updates2;					\
+	     (_i) < (_trans)->updates2 + (_trans)->nr_updates2;		\
+	     (_i)++)
+
 #endif /* _BCACHEFS_BTREE_UPDATE_H */
diff --git a/fs/bcachefs/btree_update_interior.c b/fs/bcachefs/btree_update_interior.c
index 748e6356f3d6..82b66a667e35 100644
--- a/fs/bcachefs/btree_update_interior.c
+++ b/fs/bcachefs/btree_update_interior.c
@@ -24,47 +24,42 @@
 static void btree_node_will_make_reachable(struct btree_update *,
 					   struct btree *);
 static void btree_update_drop_new_node(struct bch_fs *, struct btree *);
-static void bch2_btree_set_root_ondisk(struct bch_fs *, struct btree *, int);
 
 /* Debug code: */
 
+/*
+ * Verify that child nodes correctly span parent node's range:
+ */
 static void btree_node_interior_verify(struct btree *b)
 {
+#ifdef CONFIG_BCACHEFS_DEBUG
+	struct bpos next_node = b->data->min_key;
 	struct btree_node_iter iter;
-	struct bkey_packed *k;
+	struct bkey_s_c k;
+	struct bkey_s_c_btree_ptr_v2 bp;
+	struct bkey unpacked;
 
 	BUG_ON(!b->level);
 
-	bch2_btree_node_iter_init(&iter, b, &b->key.k.p);
-#if 1
-	BUG_ON(!(k = bch2_btree_node_iter_peek(&iter, b)) ||
-	       bkey_cmp_left_packed(b, k, &b->key.k.p));
+	bch2_btree_node_iter_init_from_start(&iter, b);
 
-	BUG_ON((bch2_btree_node_iter_advance(&iter, b),
-		!bch2_btree_node_iter_end(&iter)));
-#else
-	const char *msg;
+	while (1) {
+		k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
+		if (k.k->type != KEY_TYPE_btree_ptr_v2)
+			break;
+		bp = bkey_s_c_to_btree_ptr_v2(k);
 
-	msg = "not found";
-	k = bch2_btree_node_iter_peek(&iter, b);
-	if (!k)
-		goto err;
+		BUG_ON(bkey_cmp(next_node, bp.v->min_key));
 
-	msg = "isn't what it should be";
-	if (bkey_cmp_left_packed(b, k, &b->key.k.p))
-		goto err;
+		bch2_btree_node_iter_advance(&iter, b);
 
-	bch2_btree_node_iter_advance(&iter, b);
+		if (bch2_btree_node_iter_end(&iter)) {
+			BUG_ON(bkey_cmp(k.k->p, b->key.k.p));
+			break;
+		}
 
-	msg = "isn't last key";
-	if (!bch2_btree_node_iter_end(&iter))
-		goto err;
-	return;
-err:
-	bch2_dump_btree_node(b);
-	printk(KERN_ERR "last key %llu:%llu %s\n", b->key.k.p.inode,
-	       b->key.k.p.offset, msg);
-	BUG();
+		next_node = bkey_successor(k.k->p);
+	}
 #endif
 }
 
@@ -260,16 +255,17 @@ void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
 }
 
 static void bch2_btree_node_free_ondisk(struct bch_fs *c,
-					struct pending_btree_node_free *pending)
+			struct pending_btree_node_free *pending,
+			u64 journal_seq)
 {
 	BUG_ON(!pending->index_update_done);
 
 	bch2_mark_key(c, bkey_i_to_s_c(&pending->key),
-		      0, 0, NULL, 0, BTREE_TRIGGER_OVERWRITE);
+		      0, 0, NULL, journal_seq, BTREE_TRIGGER_OVERWRITE);
 
 	if (gc_visited(c, gc_phase(GC_PHASE_PENDING_DELETE)))
 		bch2_mark_key(c, bkey_i_to_s_c(&pending->key),
-			      0, 0, NULL, 0,
+			      0, 0, NULL, journal_seq,
 			      BTREE_TRIGGER_OVERWRITE|
 			      BTREE_TRIGGER_GC);
 }
@@ -332,7 +328,11 @@ 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);
@@ -354,25 +354,36 @@ 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);
 
 	bch2_bset_init_first(b, &b->data->keys);
+	b->level	= level;
+	b->btree_id	= as->btree_id;
+
 	memset(&b->nr, 0, sizeof(b->nr));
 	b->data->magic = cpu_to_le64(bset_magic(c));
 	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 +396,26 @@ 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 +424,12 @@ 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 +459,9 @@ 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);
@@ -550,43 +577,47 @@ err_free:
 
 /* Asynchronous interior node update machinery */
 
-static void bch2_btree_update_free(struct btree_update *as)
+static void __bch2_btree_update_free(struct btree_update *as)
 {
 	struct bch_fs *c = as->c;
 
+	bch2_journal_preres_put(&c->journal, &as->journal_preres);
+
+	bch2_journal_pin_drop(&c->journal, &as->journal);
 	bch2_journal_pin_flush(&c->journal, &as->journal);
 
-	BUG_ON(as->nr_new_nodes);
-	BUG_ON(as->nr_pending);
+	BUG_ON((as->nr_new_nodes || as->nr_pending) &&
+	       !bch2_journal_error(&c->journal));;
 
 	if (as->reserve)
 		bch2_btree_reserve_put(c, as->reserve);
 
-	mutex_lock(&c->btree_interior_update_lock);
 	list_del(&as->list);
 
 	closure_debug_destroy(&as->cl);
 	mempool_free(as, &c->btree_interior_update_pool);
 
 	closure_wake_up(&c->btree_interior_update_wait);
-	mutex_unlock(&c->btree_interior_update_lock);
 }
 
-static void btree_update_nodes_reachable(struct closure *cl)
+static void bch2_btree_update_free(struct btree_update *as)
 {
-	struct btree_update *as = container_of(cl, struct btree_update, cl);
 	struct bch_fs *c = as->c;
 
-	bch2_journal_pin_drop(&c->journal, &as->journal);
-
 	mutex_lock(&c->btree_interior_update_lock);
+	__bch2_btree_update_free(as);
+	mutex_unlock(&c->btree_interior_update_lock);
+}
+
+static void btree_update_nodes_reachable(struct btree_update *as, u64 seq)
+{
+	struct bch_fs *c = as->c;
 
 	while (as->nr_new_nodes) {
 		struct btree *b = as->new_nodes[--as->nr_new_nodes];
 
 		BUG_ON(b->will_make_reachable != (unsigned long) as);
 		b->will_make_reachable = 0;
-		mutex_unlock(&c->btree_interior_update_lock);
 
 		/*
 		 * b->will_make_reachable prevented it from being written, so
@@ -595,150 +626,128 @@ static void btree_update_nodes_reachable(struct closure *cl)
 		btree_node_lock_type(c, b, SIX_LOCK_read);
 		bch2_btree_node_write_cond(c, b, btree_node_need_write(b));
 		six_unlock_read(&b->lock);
-		mutex_lock(&c->btree_interior_update_lock);
 	}
 
 	while (as->nr_pending)
-		bch2_btree_node_free_ondisk(c, &as->pending[--as->nr_pending]);
-
-	mutex_unlock(&c->btree_interior_update_lock);
-
-	closure_wake_up(&as->wait);
-
-	bch2_btree_update_free(as);
-}
-
-static void btree_update_wait_on_journal(struct closure *cl)
-{
-	struct btree_update *as = container_of(cl, struct btree_update, cl);
-	struct bch_fs *c = as->c;
-	int ret;
-
-	ret = bch2_journal_open_seq_async(&c->journal, as->journal_seq, cl);
-	if (ret == -EAGAIN) {
-		continue_at(cl, btree_update_wait_on_journal, system_wq);
-		return;
-	}
-	if (ret < 0)
-		goto err;
-
-	bch2_journal_flush_seq_async(&c->journal, as->journal_seq, cl);
-err:
-	continue_at(cl, btree_update_nodes_reachable, system_wq);
+		bch2_btree_node_free_ondisk(c, &as->pending[--as->nr_pending],
+					    seq);
 }
 
 static void btree_update_nodes_written(struct closure *cl)
 {
 	struct btree_update *as = container_of(cl, struct btree_update, cl);
+	struct journal_res res = { 0 };
 	struct bch_fs *c = as->c;
 	struct btree *b;
+	struct bset *i;
+	int ret;
 
 	/*
 	 * We did an update to a parent node where the pointers we added pointed
 	 * to child nodes that weren't written yet: now, the child nodes have
 	 * been written so we can write out the update to the interior node.
 	 */
-retry:
 	mutex_lock(&c->btree_interior_update_lock);
 	as->nodes_written = true;
+again:
+	as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
+				      struct btree_update, unwritten_list);
+	if (!as || !as->nodes_written) {
+		mutex_unlock(&c->btree_interior_update_lock);
+		return;
+	}
+
+	b = as->b;
+	if (b && !six_trylock_intent(&b->lock)) {
+		mutex_unlock(&c->btree_interior_update_lock);
+		btree_node_lock_type(c, b, SIX_LOCK_intent);
+		six_unlock_intent(&b->lock);
+		mutex_lock(&c->btree_interior_update_lock);
+		goto again;
+	}
+
+	list_del(&as->unwritten_list);
+
+	ret = bch2_journal_res_get(&c->journal, &res, as->journal_u64s,
+				   JOURNAL_RES_GET_RESERVED);
+	if (ret) {
+		BUG_ON(!bch2_journal_error(&c->journal));
+		/* can't unblock btree writes */
+		goto free_update;
+	}
+
+	{
+		struct journal_buf *buf = &c->journal.buf[res.idx];
+		struct jset_entry *entry = vstruct_idx(buf->data, res.offset);
+
+		res.offset	+= as->journal_u64s;
+		res.u64s	-= as->journal_u64s;
+		memcpy_u64s(entry, as->journal_entries, as->journal_u64s);
+	}
 
 	switch (as->mode) {
 	case BTREE_INTERIOR_NO_UPDATE:
 		BUG();
 	case BTREE_INTERIOR_UPDATING_NODE:
-		/* The usual case: */
-		b = READ_ONCE(as->b);
-
-		if (!six_trylock_read(&b->lock)) {
-			mutex_unlock(&c->btree_interior_update_lock);
-			btree_node_lock_type(c, b, SIX_LOCK_read);
-			six_unlock_read(&b->lock);
-			goto retry;
-		}
-
-		BUG_ON(!btree_node_dirty(b));
-		closure_wait(&btree_current_write(b)->wait, cl);
+		/* @b is the node we did the final insert into: */
+		BUG_ON(!res.ref);
 
+		six_lock_write(&b->lock);
 		list_del(&as->write_blocked_list);
 
-		/*
-		 * for flush_held_btree_writes() waiting on updates to flush or
-		 * nodes to be writeable:
-		 */
-		closure_wake_up(&c->btree_interior_update_wait);
-		mutex_unlock(&c->btree_interior_update_lock);
+		i = btree_bset_last(b);
+		i->journal_seq = cpu_to_le64(
+			max(res.seq,
+			    le64_to_cpu(i->journal_seq)));
 
-		/*
-		 * b->write_blocked prevented it from being written, so
-		 * write it now if it needs to be written:
-		 */
-		bch2_btree_node_write_cond(c, b, true);
-		six_unlock_read(&b->lock);
+		bch2_btree_add_journal_pin(c, b, res.seq);
+		six_unlock_write(&b->lock);
 		break;
 
 	case BTREE_INTERIOR_UPDATING_AS:
-		/*
-		 * The btree node we originally updated has been freed and is
-		 * being rewritten - so we need to write anything here, we just
-		 * need to signal to that btree_update that it's ok to make the
-		 * new replacement node visible:
-		 */
-		closure_put(&as->parent_as->cl);
-
-		/*
-		 * and then we have to wait on that btree_update to finish:
-		 */
-		closure_wait(&as->parent_as->wait, cl);
-		mutex_unlock(&c->btree_interior_update_lock);
+		BUG_ON(b);
 		break;
 
-	case BTREE_INTERIOR_UPDATING_ROOT:
-		/* b is the new btree root: */
-		b = READ_ONCE(as->b);
-
-		if (!six_trylock_read(&b->lock)) {
-			mutex_unlock(&c->btree_interior_update_lock);
-			btree_node_lock_type(c, b, SIX_LOCK_read);
-			six_unlock_read(&b->lock);
-			goto retry;
-		}
+	case BTREE_INTERIOR_UPDATING_ROOT: {
+		struct btree_root *r = &c->btree_roots[as->btree_id];
 
-		BUG_ON(c->btree_roots[b->btree_id].as != as);
-		c->btree_roots[b->btree_id].as = NULL;
+		BUG_ON(b);
 
-		bch2_btree_set_root_ondisk(c, b, WRITE);
+		mutex_lock(&c->btree_root_lock);
+		bkey_copy(&r->key, as->parent_keys.keys);
+		r->level = as->level;
+		r->alive = true;
+		c->btree_roots_dirty = true;
+		mutex_unlock(&c->btree_root_lock);
+		break;
+	}
+	}
 
-		/*
-		 * We don't have to wait anything anything here (before
-		 * btree_update_nodes_reachable frees the old nodes
-		 * ondisk) - we've ensured that the very next journal write will
-		 * have the pointer to the new root, and before the allocator
-		 * can reuse the old nodes it'll have to do a journal commit:
-		 */
-		six_unlock_read(&b->lock);
-		mutex_unlock(&c->btree_interior_update_lock);
+	bch2_journal_pin_drop(&c->journal, &as->journal);
 
+	bch2_journal_res_put(&c->journal, &res);
+	bch2_journal_preres_put(&c->journal, &as->journal_preres);
+free_update:
+	/* Do btree write after dropping journal res: */
+	if (b) {
 		/*
-		 * Bit of funny circularity going on here we have to break:
-		 *
-		 * We have to drop our journal pin before writing the journal
-		 * entry that points to the new btree root: else, we could
-		 * deadlock if the journal currently happens to be full.
-		 *
-		 * This mean we're dropping the journal pin _before_ the new
-		 * nodes are technically reachable - but this is safe, because
-		 * after the bch2_btree_set_root_ondisk() call above they will
-		 * be reachable as of the very next journal write:
+		 * b->write_blocked prevented it from being written, so
+		 * write it now if it needs to be written:
 		 */
-		bch2_journal_pin_drop(&c->journal, &as->journal);
-
-		as->journal_seq = bch2_journal_last_unwritten_seq(&c->journal);
-
-		btree_update_wait_on_journal(cl);
-		return;
+		btree_node_write_if_need(c, b, SIX_LOCK_intent);
+		six_unlock_intent(&b->lock);
 	}
 
-	continue_at(cl, btree_update_nodes_reachable, system_wq);
+	if (!ret)
+		btree_update_nodes_reachable(as, res.seq);
+
+	__bch2_btree_update_free(as);
+	/*
+	 * for flush_held_btree_writes() waiting on updates to flush or
+	 * nodes to be writeable:
+	 */
+	closure_wake_up(&c->btree_interior_update_wait);
+	goto again;
 }
 
 /*
@@ -750,52 +759,17 @@ static void btree_update_updated_node(struct btree_update *as, struct btree *b)
 	struct bch_fs *c = as->c;
 
 	mutex_lock(&c->btree_interior_update_lock);
+	list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 
 	BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
 	BUG_ON(!btree_node_dirty(b));
 
-	as->mode = BTREE_INTERIOR_UPDATING_NODE;
-	as->b = b;
+	as->mode	= BTREE_INTERIOR_UPDATING_NODE;
+	as->b		= b;
+	as->level	= b->level;
 	list_add(&as->write_blocked_list, &b->write_blocked);
 
 	mutex_unlock(&c->btree_interior_update_lock);
-
-	/*
-	 * In general, when you're staging things in a journal that will later
-	 * be written elsewhere, and you also want to guarantee ordering: that
-	 * is, if you have updates a, b, c, after a crash you should never see c
-	 * and not a or b - there's a problem:
-	 *
-	 * If the final destination of the update(s) (i.e. btree node) can be
-	 * written/flushed _before_ the relevant journal entry - oops, that
-	 * breaks ordering, since the various leaf nodes can be written in any
-	 * order.
-	 *
-	 * Normally we use bset->journal_seq to deal with this - if during
-	 * recovery we find a btree node write that's newer than the newest
-	 * journal entry, we just ignore it - we don't need it, anything we're
-	 * supposed to have (that we reported as completed via fsync()) will
-	 * still be in the journal, and as far as the state of the journal is
-	 * concerned that btree node write never happened.
-	 *
-	 * That breaks when we're rewriting/splitting/merging nodes, since we're
-	 * mixing btree node writes that haven't happened yet with previously
-	 * written data that has been reported as completed to the journal.
-	 *
-	 * Thus, before making the new nodes reachable, we have to wait the
-	 * newest journal sequence number we have data for to be written (if it
-	 * hasn't been yet).
-	 */
-	bch2_journal_wait_on_seq(&c->journal, as->journal_seq, &as->cl);
-}
-
-static void interior_update_flush(struct journal *j,
-			struct journal_entry_pin *pin, u64 seq)
-{
-	struct btree_update *as =
-		container_of(pin, struct btree_update, journal);
-
-	bch2_journal_flush_seq_async(j, as->journal_seq, NULL);
 }
 
 static void btree_update_reparent(struct btree_update *as,
@@ -803,10 +777,10 @@ static void btree_update_reparent(struct btree_update *as,
 {
 	struct bch_fs *c = as->c;
 
+	lockdep_assert_held(&c->btree_interior_update_lock);
+
 	child->b = NULL;
 	child->mode = BTREE_INTERIOR_UPDATING_AS;
-	child->parent_as = as;
-	closure_get(&as->cl);
 
 	/*
 	 * When we write a new btree root, we have to drop our journal pin
@@ -817,45 +791,24 @@ static void btree_update_reparent(struct btree_update *as,
 	 * just transfer the journal pin to the new interior update so
 	 * btree_update_nodes_written() can drop it.
 	 */
-	bch2_journal_pin_add_if_older(&c->journal, &child->journal,
-				      &as->journal, interior_update_flush);
+	bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
 	bch2_journal_pin_drop(&c->journal, &child->journal);
-
-	as->journal_seq = max(as->journal_seq, child->journal_seq);
 }
 
-static void btree_update_updated_root(struct btree_update *as)
+static void btree_update_updated_root(struct btree_update *as, struct btree *b)
 {
 	struct bch_fs *c = as->c;
-	struct btree_root *r = &c->btree_roots[as->btree_id];
-
-	mutex_lock(&c->btree_interior_update_lock);
 
 	BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
+	BUG_ON(!bch2_keylist_empty(&as->parent_keys));
 
-	/*
-	 * Old root might not be persistent yet - if so, redirect its
-	 * btree_update operation to point to us:
-	 */
-	if (r->as)
-		btree_update_reparent(as, r->as);
-
-	as->mode = BTREE_INTERIOR_UPDATING_ROOT;
-	as->b = r->b;
-	r->as = as;
+	mutex_lock(&c->btree_interior_update_lock);
+	list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
 
+	as->mode	= BTREE_INTERIOR_UPDATING_ROOT;
+	as->level	= b->level;
+	bch2_keylist_add(&as->parent_keys, &b->key);
 	mutex_unlock(&c->btree_interior_update_lock);
-
-	/*
-	 * When we're rewriting nodes and updating interior nodes, there's an
-	 * issue with updates that haven't been written in the journal getting
-	 * mixed together with older data - see btree_update_updated_node()
-	 * for the explanation.
-	 *
-	 * However, this doesn't affect us when we're writing a new btree root -
-	 * because to make that new root reachable we have to write out a new
-	 * journal entry, which must necessarily be newer than as->journal_seq.
-	 */
 }
 
 static void btree_node_will_make_reachable(struct btree_update *as,
@@ -932,10 +885,8 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 					       struct btree *b)
 {
 	struct bch_fs *c = as->c;
-	struct closure *cl, *cl_n;
 	struct btree_update *p, *n;
 	struct btree_write *w;
-	struct bset_tree *t;
 
 	set_btree_node_dying(b);
 
@@ -944,18 +895,6 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 
 	btree_interior_update_add_node_reference(as, b);
 
-	/*
-	 * Does this node have data that hasn't been written in the journal?
-	 *
-	 * If so, we have to wait for the corresponding journal entry to be
-	 * written before making the new nodes reachable - we can't just carry
-	 * over the bset->journal_seq tracking, since we'll be mixing those keys
-	 * in with keys that aren't in the journal anymore:
-	 */
-	for_each_bset(b, t)
-		as->journal_seq = max(as->journal_seq,
-				      le64_to_cpu(bset(b, t)->journal_seq));
-
 	mutex_lock(&c->btree_interior_update_lock);
 
 	/*
@@ -979,16 +918,6 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 
 	clear_btree_node_dirty(b);
 	clear_btree_node_need_write(b);
-	w = btree_current_write(b);
-
-	/*
-	 * Does this node have any btree_update operations waiting on this node
-	 * to be written?
-	 *
-	 * If so, wake them up when this btree_update operation is reachable:
-	 */
-	llist_for_each_entry_safe(cl, cl_n, llist_del_all(&w->wait.list), list)
-		llist_add(&cl->list, &as->wait.list);
 
 	/*
 	 * Does this node have unwritten data that has a pin on the journal?
@@ -998,13 +927,12 @@ void bch2_btree_interior_update_will_free_node(struct btree_update *as,
 	 * oldest pin of any of the nodes we're freeing. We'll release the pin
 	 * when the new nodes are persistent and reachable on disk:
 	 */
-	bch2_journal_pin_add_if_older(&c->journal, &w->journal,
-				      &as->journal, interior_update_flush);
+	w = btree_current_write(b);
+	bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 	bch2_journal_pin_drop(&c->journal, &w->journal);
 
 	w = btree_prev_write(b);
-	bch2_journal_pin_add_if_older(&c->journal, &w->journal,
-				      &as->journal, interior_update_flush);
+	bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
 	bch2_journal_pin_drop(&c->journal, &w->journal);
 
 	mutex_unlock(&c->btree_interior_update_lock);
@@ -1021,12 +949,33 @@ void bch2_btree_update_done(struct btree_update *as)
 }
 
 struct btree_update *
-bch2_btree_update_start(struct bch_fs *c, enum btree_id id,
+bch2_btree_update_start(struct btree_trans *trans, enum btree_id id,
 			unsigned nr_nodes, unsigned flags,
 			struct closure *cl)
 {
+	struct bch_fs *c = trans->c;
+	struct journal_preres journal_preres = { 0 };
 	struct btree_reserve *reserve;
 	struct btree_update *as;
+	int ret;
+
+	ret = bch2_journal_preres_get(&c->journal, &journal_preres,
+				      BTREE_UPDATE_JOURNAL_RES,
+				      JOURNAL_RES_GET_NONBLOCK);
+	if (ret == -EAGAIN) {
+		bch2_trans_unlock(trans);
+
+		ret = bch2_journal_preres_get(&c->journal, &journal_preres,
+					      BTREE_UPDATE_JOURNAL_RES,
+					      JOURNAL_RES_GET_NONBLOCK);
+		if (ret)
+			return ERR_PTR(ret);
+
+		if (!bch2_trans_relock(trans)) {
+			bch2_journal_preres_put(&c->journal, &journal_preres);
+			return ERR_PTR(-EINTR);
+		}
+	}
 
 	reserve = bch2_btree_reserve_get(c, nr_nodes, flags, cl);
 	if (IS_ERR(reserve))
@@ -1040,6 +989,7 @@ bch2_btree_update_start(struct bch_fs *c, enum btree_id id,
 	as->btree_id	= id;
 	as->reserve	= reserve;
 	INIT_LIST_HEAD(&as->write_blocked_list);
+	as->journal_preres = journal_preres;
 
 	bch2_keylist_init(&as->parent_keys, as->inline_keys);
 
@@ -1102,22 +1052,6 @@ static void bch2_btree_set_root_inmem(struct btree_update *as, struct btree *b)
 	mutex_unlock(&c->btree_interior_update_lock);
 }
 
-static void bch2_btree_set_root_ondisk(struct bch_fs *c, struct btree *b, int rw)
-{
-	struct btree_root *r = &c->btree_roots[b->btree_id];
-
-	mutex_lock(&c->btree_root_lock);
-
-	BUG_ON(b != r->b);
-	bkey_copy(&r->key, &b->key);
-	r->level = b->level;
-	r->alive = true;
-	if (rw == WRITE)
-		c->btree_roots_dirty = true;
-
-	mutex_unlock(&c->btree_root_lock);
-}
-
 /**
  * bch_btree_set_root - update the root in memory and on disk
  *
@@ -1150,7 +1084,7 @@ static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
 
 	bch2_btree_set_root_inmem(as, b);
 
-	btree_update_updated_root(as);
+	btree_update_updated_root(as, b);
 
 	/*
 	 * Unlock old root after new root is visible:
@@ -1171,10 +1105,21 @@ static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b
 {
 	struct bch_fs *c = as->c;
 	struct bch_fs_usage *fs_usage;
+	struct jset_entry *entry;
 	struct bkey_packed *k;
 	struct bkey tmp;
 
-	BUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(c, b));
+	BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
+	       ARRAY_SIZE(as->journal_entries));
+
+	entry = (void *) &as->journal_entries[as->journal_u64s];
+	memset(entry, 0, sizeof(*entry));
+	entry->u64s	= cpu_to_le16(insert->k.u64s);
+	entry->type	= BCH_JSET_ENTRY_btree_keys;
+	entry->btree_id = b->btree_id;
+	entry->level	= b->level;
+	memcpy_u64s_small(entry->_data, insert, insert->k.u64s);
+	as->journal_u64s += jset_u64s(insert->k.u64s);
 
 	mutex_lock(&c->btree_interior_update_lock);
 	percpu_down_read(&c->mark_lock);
@@ -1263,10 +1208,8 @@ 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, bkey_successor(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));
@@ -1325,6 +1268,14 @@ static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
 	struct bkey_packed *src, *dst, *n;
 	struct bset *i;
 
+	/*
+	 * XXX
+	 *
+	 * these updates must be journalled
+	 *
+	 * oops
+	 */
+
 	BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
 
 	bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
@@ -1332,11 +1283,6 @@ static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
 	while (!bch2_keylist_empty(keys)) {
 		k = bch2_keylist_front(keys);
 
-		BUG_ON(bch_keylist_u64s(keys) >
-		       bch_btree_keys_u64s_remaining(as->c, b));
-		BUG_ON(bkey_cmp(k->k.p, b->data->min_key) < 0);
-		BUG_ON(bkey_cmp(k->k.p, b->data->max_key) > 0);
-
 		bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
 		bch2_keylist_pop_front(keys);
 	}
@@ -1422,7 +1368,8 @@ static void btree_split(struct btree_update *as, struct btree *b,
 		bch2_btree_build_aux_trees(n1);
 		six_unlock_write(&n1->lock);
 
-		bch2_keylist_add(&as->parent_keys, &n1->key);
+		if (parent)
+			bch2_keylist_add(&as->parent_keys, &n1->key);
 	}
 
 	bch2_btree_node_write(c, n1, SIX_LOCK_intent);
@@ -1496,19 +1443,15 @@ bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
 	       (bkey_cmp_packed(b, k, &insert->k) >= 0))
 		;
 
-	while (!bch2_keylist_empty(keys)) {
-		insert = bch2_keylist_front(keys);
-
+	for_each_keylist_key(keys, insert)
 		bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
-		bch2_keylist_pop_front(keys);
-	}
 
 	btree_update_updated_node(as, b);
 
 	trans_for_each_iter_with_node(iter->trans, b, linked)
 		bch2_btree_node_iter_peek(&linked->l[b->level].iter, b);
 
-	bch2_btree_iter_verify(iter, b);
+	bch2_btree_trans_verify_iters(iter->trans, b);
 }
 
 /**
@@ -1581,7 +1524,7 @@ int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
 			  unsigned flags)
 {
 	struct btree_trans *trans = iter->trans;
-	struct btree *b = iter->l[0].b;
+	struct btree *b = iter_l(iter)->b;
 	struct btree_update *as;
 	struct closure cl;
 	int ret = 0;
@@ -1620,7 +1563,7 @@ int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
 		goto out;
 	}
 
-	as = bch2_btree_update_start(c, iter->btree_id,
+	as = bch2_btree_update_start(trans, iter->btree_id,
 		btree_update_reserve_required(c, b), flags,
 		!(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
 	if (IS_ERR(as)) {
@@ -1732,7 +1675,7 @@ retry:
 		goto err_unlock;
 	}
 
-	as = bch2_btree_update_start(c, iter->btree_id,
+	as = bch2_btree_update_start(trans, iter->btree_id,
 			 btree_update_reserve_required(c, parent) + 1,
 			 BTREE_INSERT_NOFAIL|
 			 BTREE_INSERT_USE_RESERVE,
@@ -1749,10 +1692,9 @@ 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);
 
@@ -1779,7 +1721,7 @@ retry:
 
 	bch2_btree_iter_node_replace(iter, n);
 
-	bch2_btree_iter_verify(iter, n);
+	bch2_btree_trans_verify_iters(trans, n);
 
 	bch2_btree_node_free_inmem(c, b, iter);
 	bch2_btree_node_free_inmem(c, m, iter);
@@ -1846,7 +1788,7 @@ static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
 	struct btree *n, *parent = btree_node_parent(iter, b);
 	struct btree_update *as;
 
-	as = bch2_btree_update_start(c, iter->btree_id,
+	as = bch2_btree_update_start(iter->trans, iter->btree_id,
 		(parent
 		 ? btree_update_reserve_required(c, parent)
 		 : 0) + 1,
@@ -1944,7 +1886,7 @@ 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 +1931,20 @@ 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 +1953,12 @@ 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 +1971,11 @@ 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);
@@ -2047,19 +1989,19 @@ 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);
+		btree_update_updated_root(as, b);
 		bch2_btree_node_unlock_write(b, iter);
 	}
 
@@ -2068,7 +2010,7 @@ 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 +2033,11 @@ 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) {
@@ -2110,7 +2055,7 @@ int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
 		new_hash = bch2_btree_node_mem_alloc(c);
 	}
 
-	as = bch2_btree_update_start(c, iter->btree_id,
+	as = bch2_btree_update_start(iter->trans, iter->btree_id,
 		parent ? btree_update_reserve_required(c, parent) : 0,
 		BTREE_INSERT_NOFAIL|
 		BTREE_INSERT_USE_RESERVE|
@@ -2134,7 +2079,7 @@ 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 +2138,14 @@ 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/fs/bcachefs/btree_update_interior.h b/fs/bcachefs/btree_update_interior.h
index 2d8e0b7f3aaf..2fddf5d31eb9 100644
--- a/fs/bcachefs/btree_update_interior.h
+++ b/fs/bcachefs/btree_update_interior.h
@@ -32,6 +32,9 @@ struct pending_btree_node_free {
 	__BKEY_PADDED(key, BKEY_BTREE_PTR_VAL_U64s_MAX);
 };
 
+#define BTREE_UPDATE_JOURNAL_RES		\
+	((BKEY_BTREE_PTR_U64s_MAX + 1) * (BTREE_MAX_DEPTH - 1) * 2)
+
 /*
  * Tracks an in progress split/rewrite of a btree node and the update to the
  * parent node:
@@ -55,6 +58,7 @@ struct btree_update {
 	struct bch_fs			*c;
 
 	struct list_head		list;
+	struct list_head		unwritten_list;
 
 	/* What kind of update are we doing? */
 	enum {
@@ -68,8 +72,10 @@ struct btree_update {
 	unsigned			nodes_written:1;
 
 	enum btree_id			btree_id;
+	u8				level;
 
 	struct btree_reserve		*reserve;
+	struct journal_preres		journal_preres;
 
 	/*
 	 * BTREE_INTERIOR_UPDATING_NODE:
@@ -83,18 +89,6 @@ struct btree_update {
 	struct list_head		write_blocked_list;
 
 	/*
-	 * BTREE_INTERIOR_UPDATING_AS: btree node we updated was freed, so now
-	 * we're now blocking another btree_update
-	 * @parent_as - btree_update that's waiting on our nodes to finish
-	 * writing, before it can make new nodes visible on disk
-	 * @wait - list of child btree_updates that are waiting on this
-	 * btree_update to make all the new nodes visible before they can free
-	 * their old btree nodes
-	 */
-	struct btree_update		*parent_as;
-	struct closure_waitlist		wait;
-
-	/*
 	 * We may be freeing nodes that were dirty, and thus had journal entries
 	 * pinned: we need to transfer the oldest of those pins to the
 	 * btree_update operation, and release it when the new node(s)
@@ -102,8 +96,6 @@ struct btree_update {
 	 */
 	struct journal_entry_pin	journal;
 
-	u64				journal_seq;
-
 	/*
 	 * Nodes being freed:
 	 * Protected by c->btree_node_pending_free_lock
@@ -115,6 +107,9 @@ struct btree_update {
 	struct btree			*new_nodes[BTREE_MAX_DEPTH * 2 + GC_MERGE_NODES];
 	unsigned			nr_new_nodes;
 
+	unsigned			journal_u64s;
+	u64				journal_entries[BTREE_UPDATE_JOURNAL_RES];
+
 	/* Only here to reduce stack usage on recursive splits: */
 	struct keylist			parent_keys;
 	/*
@@ -139,7 +134,7 @@ struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *,
 
 void bch2_btree_update_done(struct btree_update *);
 struct btree_update *
-bch2_btree_update_start(struct bch_fs *, enum btree_id, unsigned,
+bch2_btree_update_start(struct btree_trans *, enum btree_id, unsigned,
 			unsigned, struct closure *);
 
 void bch2_btree_interior_update_will_free_node(struct btree_update *,
@@ -302,18 +297,23 @@ static inline struct btree_node_entry *want_new_bset(struct bch_fs *c,
 }
 
 static inline void push_whiteout(struct bch_fs *c, struct btree *b,
-				 struct bkey_packed *k)
+				 struct bpos pos)
 {
-	unsigned u64s = bkeyp_key_u64s(&b->format, k);
-	struct bkey_packed *dst;
+	struct bkey_packed k;
+
+	BUG_ON(bch_btree_keys_u64s_remaining(c, b) < BKEY_U64s);
+
+	if (!bkey_pack_pos(&k, pos, b)) {
+		struct bkey *u = (void *) &k;
+
+		bkey_init(u);
+		u->p = pos;
+	}
 
-	BUG_ON(u64s > bch_btree_keys_u64s_remaining(c, b));
+	k.needs_whiteout = true;
 
-	b->whiteout_u64s += bkeyp_key_u64s(&b->format, k);
-	dst = unwritten_whiteouts_start(c, b);
-	memcpy_u64s(dst, k, u64s);
-	dst->u64s = u64s;
-	dst->type = KEY_TYPE_deleted;
+	b->whiteout_u64s += k.u64s;
+	bkey_copy(unwritten_whiteouts_start(c, b), &k);
 }
 
 /*
diff --git a/fs/bcachefs/btree_update_leaf.c b/fs/bcachefs/btree_update_leaf.c
index afd2086edeff..7faf98fd2f64 100644
--- a/fs/bcachefs/btree_update_leaf.c
+++ b/fs/bcachefs/btree_update_leaf.c
@@ -23,11 +23,10 @@
 static inline bool same_leaf_as_prev(struct btree_trans *trans,
 				     struct btree_insert_entry *i)
 {
-	return i != trans->updates &&
-		i[0].iter->l[0].b == i[-1].iter->l[0].b;
+	return i != trans->updates2 &&
+		iter_l(i[0].iter)->b == iter_l(i[-1].iter)->b;
 }
 
-
 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
 					    struct btree_iter *iter)
 {
@@ -53,45 +52,45 @@ 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)));
 	EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
-	EBUG_ON(bkey_cmp(bkey_start_pos(&insert->k), b->data->min_key) < 0 ||
-		bkey_cmp(insert->k.p, b->data->max_key) > 0);
+	EBUG_ON(bkey_cmp(b->data->min_key, POS_MIN) &&
+		bkey_cmp(bkey_start_pos(&insert->k),
+			 bkey_predecessor(b->data->min_key)) < 0);
+	EBUG_ON(bkey_cmp(insert->k.p, b->data->min_key) < 0);
+	EBUG_ON(bkey_cmp(insert->k.p, b->data->max_key) > 0);
+	EBUG_ON(insert->k.u64s >
+		bch_btree_keys_u64s_remaining(iter->trans->c, b));
+	EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
 
 	k = bch2_btree_node_iter_peek_all(node_iter, b);
 	if (k && bkey_cmp_packed(b, k, &insert->k))
 		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) {
-			push_whiteout(iter->trans->c, b, k);
-			k->needs_whiteout = false;
-		}
+		if (k->needs_whiteout)
+			push_whiteout(iter->trans->c, b, insert->k.p);
+		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 +100,6 @@ 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 +115,13 @@ 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;
 }
 
@@ -155,6 +148,17 @@ static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin,
 	return __btree_node_flush(j, pin, 1, seq);
 }
 
+inline void bch2_btree_add_journal_pin(struct bch_fs *c,
+				       struct btree *b, u64 seq)
+{
+	struct btree_write *w = btree_current_write(b);
+
+	bch2_journal_pin_add(&c->journal, seq, &w->journal,
+			     btree_node_write_idx(b) == 0
+			     ? btree_node_flush0
+			     : btree_node_flush1);
+}
+
 static inline void __btree_journal_key(struct btree_trans *trans,
 				       enum btree_id btree_id,
 				       struct bkey_i *insert)
@@ -176,16 +180,14 @@ static inline void __btree_journal_key(struct btree_trans *trans,
 		*trans->journal_seq = seq;
 }
 
-void bch2_btree_journal_key(struct btree_trans *trans,
-			   struct btree_iter *iter,
-			   struct bkey_i *insert)
+static void bch2_btree_journal_key(struct btree_trans *trans,
+				   struct btree_iter *iter,
+				   struct bkey_i *insert)
 {
 	struct bch_fs *c = trans->c;
 	struct journal *j = &c->journal;
-	struct btree *b = iter->l[0].b;
-	struct btree_write *w = btree_current_write(b);
+	struct btree *b = iter_l(iter)->b;
 
-	EBUG_ON(iter->level || b->level);
 	EBUG_ON(trans->journal_res.ref !=
 		!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 
@@ -195,35 +197,15 @@ void bch2_btree_journal_key(struct btree_trans *trans,
 			cpu_to_le64(trans->journal_res.seq);
 	}
 
-	if (unlikely(!journal_pin_active(&w->journal))) {
-		u64 seq = likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
+	bch2_btree_add_journal_pin(c, b,
+		likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
 			? trans->journal_res.seq
-			: j->replay_journal_seq;
-
-		bch2_journal_pin_add(j, seq, &w->journal,
-				     btree_node_write_idx(b) == 0
-				     ? btree_node_flush0
-				     : btree_node_flush1);
-	}
+			: j->replay_journal_seq);
 
 	if (unlikely(!btree_node_dirty(b)))
 		set_btree_node_dirty(b);
 }
 
-static void bch2_insert_fixup_key(struct btree_trans *trans,
-				  struct btree_iter *iter,
-				  struct bkey_i *insert)
-{
-	struct btree_iter_level *l = &iter->l[0];
-
-	EBUG_ON(iter->level);
-	EBUG_ON(insert->k.u64s >
-		bch_btree_keys_u64s_remaining(trans->c, l->b));
-
-	if (likely(bch2_btree_bset_insert_key(iter, l->b, &l->iter, insert)))
-		bch2_btree_journal_key(trans, iter, insert);
-}
-
 /**
  * btree_insert_key - insert a key one key into a leaf node
  */
@@ -232,7 +214,7 @@ static void btree_insert_key_leaf(struct btree_trans *trans,
 				  struct bkey_i *insert)
 {
 	struct bch_fs *c = trans->c;
-	struct btree *b = iter->l[0].b;
+	struct btree *b = iter_l(iter)->b;
 	struct bset_tree *t = bset_tree_last(b);
 	int old_u64s = bset_u64s(t);
 	int old_live_u64s = b->nr.live_u64s;
@@ -240,10 +222,8 @@ static void btree_insert_key_leaf(struct btree_trans *trans,
 
 	insert->k.needs_whiteout = false;
 
-	if (!btree_node_is_extents(b))
-		bch2_insert_fixup_key(trans, iter, insert);
-	else
-		bch2_insert_fixup_extent(trans, iter, insert);
+	if (likely(bch2_btree_bset_insert_key(iter, b, &iter_l(iter)->iter, insert)))
+		bch2_btree_journal_key(trans, iter, insert);
 
 	live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 	u64s_added = (int) bset_u64s(t) - old_u64s;
@@ -268,14 +248,10 @@ static inline void btree_insert_entry_checks(struct btree_trans *trans,
 {
 	struct bch_fs *c = trans->c;
 
-	BUG_ON(iter->level);
-	BUG_ON(bkey_cmp(bkey_start_pos(&insert->k), iter->pos));
-	EBUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
-		bkey_cmp(insert->k.p, iter->l[0].b->key.k.p) > 0);
-
+	BUG_ON(bkey_cmp(insert->k.p, iter->pos));
 	BUG_ON(debug_check_bkeys(c) &&
-	       !bkey_deleted(&insert->k) &&
-	       bch2_bkey_invalid(c, bkey_i_to_s_c(insert), iter->btree_id));
+	       bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
+				 __btree_node_type(iter->level, iter->btree_id)));
 }
 
 static noinline int
@@ -321,15 +297,22 @@ btree_key_can_insert(struct btree_trans *trans,
 		     unsigned *u64s)
 {
 	struct bch_fs *c = trans->c;
-	struct btree *b = iter->l[0].b;
+	struct btree *b = iter_l(iter)->b;
 	static enum btree_insert_ret ret;
 
 	if (unlikely(btree_node_fake(b)))
 		return BTREE_INSERT_BTREE_NODE_FULL;
 
-	ret = !btree_node_is_extents(b)
+	/*
+	 * old bch2_extent_sort_fix_overlapping() algorithm won't work with new
+	 * style extent updates:
+	 */
+	if (unlikely(btree_node_old_extent_overwrite(b)))
+		return BTREE_INSERT_BTREE_NODE_FULL;
+
+	ret = !(iter->flags & BTREE_ITER_IS_EXTENTS)
 		? BTREE_INSERT_OK
-		: bch2_extent_can_insert(trans, iter, insert, u64s);
+		: bch2_extent_can_insert(trans, iter, insert);
 	if (ret)
 		return ret;
 
@@ -369,7 +352,7 @@ static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 	struct btree_insert_entry *i;
 
 	trans_for_each_update(trans, i)
-		if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
+		if (gc_visited(c, gc_pos_btree_node(iter_l(i->iter)->b)))
 			bch2_mark_update(trans, i->iter, i->k, NULL,
 					 i->trigger_flags|BTREE_TRIGGER_GC);
 }
@@ -398,7 +381,7 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 
 	prefetch(&trans->c->journal.flags);
 
-	trans_for_each_update(trans, i) {
+	trans_for_each_update2(trans, i) {
 		/* Multiple inserts might go to same leaf: */
 		if (!same_leaf_as_prev(trans, i))
 			u64s = 0;
@@ -437,10 +420,10 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 
 	if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 		if (journal_seq_verify(c))
-			trans_for_each_update(trans, i)
+			trans_for_each_update2(trans, i)
 				i->k->k.version.lo = trans->journal_res.seq;
 		else if (inject_invalid_keys(c))
-			trans_for_each_update(trans, i)
+			trans_for_each_update2(trans, i)
 				i->k->k.version = MAX_VERSION;
 	}
 
@@ -463,7 +446,7 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 	if (unlikely(c->gc_pos.phase))
 		bch2_trans_mark_gc(trans);
 
-	trans_for_each_update(trans, i)
+	trans_for_each_update2(trans, i)
 		do_btree_insert_one(trans, i->iter, i->k);
 err:
 	if (marking) {
@@ -484,8 +467,8 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans,
 	struct btree_iter *iter;
 	int ret;
 
-	trans_for_each_update(trans, i)
-		BUG_ON(!btree_node_intent_locked(i->iter, 0));
+	trans_for_each_update2(trans, i)
+		BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level));
 
 	ret = bch2_journal_preres_get(&trans->c->journal,
 			&trans->journal_preres, trans->journal_preres_u64s,
@@ -512,20 +495,20 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans,
 	}
 
 	if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
-		trans_for_each_update(trans, i)
+		trans_for_each_update2(trans, i)
 			btree_insert_entry_checks(trans, i->iter, i->k);
 	bch2_btree_trans_verify_locks(trans);
 
-	trans_for_each_update(trans, i)
+	trans_for_each_update2(trans, i)
 		if (!same_leaf_as_prev(trans, i))
 			bch2_btree_node_lock_for_insert(trans->c,
-						i->iter->l[0].b, i->iter);
+					iter_l(i->iter)->b, i->iter);
 
 	ret = bch2_trans_commit_write_locked(trans, stopped_at);
 
-	trans_for_each_update(trans, i)
+	trans_for_each_update2(trans, i)
 		if (!same_leaf_as_prev(trans, i))
-			bch2_btree_node_unlock_write_inlined(i->iter->l[0].b,
+			bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 							     i->iter);
 
 	/*
@@ -540,14 +523,14 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans,
 	if (trans->flags & BTREE_INSERT_NOUNLOCK)
 		trans->nounlock = true;
 
-	trans_for_each_update(trans, i)
+	trans_for_each_update2(trans, i)
 		if (!same_leaf_as_prev(trans, i))
 			bch2_foreground_maybe_merge(trans->c, i->iter,
 						    0, trans->flags);
 
 	trans->nounlock = false;
 
-	trans_for_each_update(trans, i)
+	trans_for_each_update2(trans, i)
 		bch2_btree_iter_downgrade(i->iter);
 
 	return 0;
@@ -670,6 +653,135 @@ bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 	return 0;
 }
 
+static void bch2_trans_update2(struct btree_trans *trans,
+			       struct btree_iter *iter,
+			       struct bkey_i *insert)
+{
+	struct btree_insert_entry *i, n = (struct btree_insert_entry) {
+		.iter = iter, .k = insert
+	};
+
+	btree_insert_entry_checks(trans, n.iter, n.k);
+
+	BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
+
+	EBUG_ON(trans->nr_updates2 >= trans->nr_iters);
+
+	iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
+
+	trans_for_each_update2(trans, i) {
+		if (btree_iter_cmp(n.iter, i->iter) == 0) {
+			*i = n;
+			return;
+		}
+
+		if (btree_iter_cmp(n.iter, i->iter) <= 0)
+			break;
+	}
+
+	array_insert_item(trans->updates2, trans->nr_updates2,
+			  i - trans->updates2, n);
+}
+
+static int extent_update_to_keys(struct btree_trans *trans,
+				 struct btree_iter *orig_iter,
+				 struct bkey_i *insert)
+{
+	struct btree_iter *iter;
+
+	if (bkey_deleted(&insert->k))
+		return 0;
+
+	iter = bch2_trans_copy_iter(trans, orig_iter);
+	if (IS_ERR(iter))
+		return PTR_ERR(iter);
+
+	iter->flags |= BTREE_ITER_INTENT;
+	__bch2_btree_iter_set_pos(iter, insert->k.p, false);
+	bch2_trans_update2(trans, iter, insert);
+	bch2_trans_iter_put(trans, iter);
+	return 0;
+}
+
+static int extent_handle_overwrites(struct btree_trans *trans,
+				    enum btree_id btree_id,
+				    struct bpos start, struct bpos end)
+{
+	struct btree_iter *iter = NULL, *update_iter;
+	struct bkey_i *update;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT);
+	ret = PTR_ERR_OR_ZERO(iter);
+	if (ret)
+		return ret;
+
+	k = bch2_btree_iter_peek_with_updates(iter);
+
+	while (k.k && !(ret = bkey_err(k))) {
+		if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0)
+			break;
+
+		if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
+			update_iter = bch2_trans_copy_iter(trans, iter);
+			if ((ret = PTR_ERR_OR_ZERO(update_iter)))
+				goto err;
+
+			update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+			if ((ret = PTR_ERR_OR_ZERO(update)))
+				goto err;
+
+			bkey_reassemble(update, k);
+			bch2_cut_back(start, update);
+
+			__bch2_btree_iter_set_pos(update_iter, update->k.p, false);
+			bch2_trans_update2(trans, update_iter, update);
+			bch2_trans_iter_put(trans, update_iter);
+		}
+
+		if (bkey_cmp(k.k->p, end) > 0) {
+			update_iter = bch2_trans_copy_iter(trans, iter);
+			if ((ret = PTR_ERR_OR_ZERO(update_iter)))
+				goto err;
+
+			update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+			if ((ret = PTR_ERR_OR_ZERO(update)))
+				goto err;
+
+			bkey_reassemble(update, k);
+			bch2_cut_front(end, update);
+
+			__bch2_btree_iter_set_pos(update_iter, update->k.p, false);
+			bch2_trans_update2(trans, update_iter, update);
+			bch2_trans_iter_put(trans, update_iter);
+		} else {
+			update_iter = bch2_trans_copy_iter(trans, iter);
+			if ((ret = PTR_ERR_OR_ZERO(update_iter)))
+				goto err;
+
+			update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
+			if ((ret = PTR_ERR_OR_ZERO(update)))
+				goto err;
+
+			update->k = *k.k;
+			set_bkey_val_u64s(&update->k, 0);
+			update->k.type = KEY_TYPE_deleted;
+			update->k.size = 0;
+
+			__bch2_btree_iter_set_pos(update_iter, update->k.p, false);
+			bch2_trans_update2(trans, update_iter, update);
+			bch2_trans_iter_put(trans, update_iter);
+		}
+
+		k = bch2_btree_iter_next_with_updates(iter);
+	}
+err:
+	if (!IS_ERR_OR_NULL(iter))
+		bch2_trans_iter_put(trans, iter);
+	return ret;
+}
+
 int __bch2_trans_commit(struct btree_trans *trans)
 {
 	struct btree_insert_entry *i = NULL;
@@ -739,7 +851,36 @@ int __bch2_trans_commit(struct btree_trans *trans)
 		}
 	} while (trans_trigger_run);
 
+	/* Turn extents updates into keys: */
+	trans_for_each_update(trans, i)
+		if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
+			struct bpos start = bkey_start_pos(&i->k->k);
+
+			while (i + 1 < trans->updates + trans->nr_updates &&
+			       i[0].iter->btree_id == i[1].iter->btree_id &&
+			       !bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)))
+				i++;
+
+			ret = extent_handle_overwrites(trans, i->iter->btree_id,
+						       start, i->k->k.p);
+			if (ret)
+				goto out;
+		}
+
 	trans_for_each_update(trans, i) {
+		if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
+			ret = extent_update_to_keys(trans, i->iter, i->k);
+			if (ret)
+				goto out;
+		} else {
+			bch2_trans_update2(trans, i->iter, i->k);
+		}
+	}
+
+	trans_for_each_update2(trans, i) {
+		BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK);
+		BUG_ON(i->iter->locks_want < 1);
+
 		u64s = jset_u64s(i->k->k.u64s);
 		if (0)
 			trans->journal_preres_u64s += u64s;
@@ -770,7 +911,7 @@ out:
 	if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 		percpu_ref_put(&trans->c->writes);
 out_noupdates:
-	bch2_trans_reset(trans, TRANS_RESET_MEM|TRANS_RESET_NOTRAVERSE);
+	bch2_trans_reset(trans, !ret ? TRANS_RESET_NOTRAVERSE : 0);
 
 	return ret;
 err:
@@ -788,11 +929,14 @@ int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 		.trigger_flags = flags, .iter = iter, .k = k
 	};
 
-	EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&k->k)));
+	EBUG_ON(bkey_cmp(iter->pos,
+			 (iter->flags & BTREE_ITER_IS_EXTENTS)
+			 ? bkey_start_pos(&k->k)
+			 : k->k.p));
 
 	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;
 	}
@@ -851,18 +995,21 @@ int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 	return 0;
 }
 
-static int __bch2_btree_insert(struct btree_trans *trans,
-			       enum btree_id id, struct bkey_i *k)
+int __bch2_btree_insert(struct btree_trans *trans,
+			enum btree_id id, struct bkey_i *k)
 {
 	struct btree_iter *iter;
+	int ret;
 
 	iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
 				   BTREE_ITER_INTENT);
 	if (IS_ERR(iter))
 		return PTR_ERR(iter);
 
-	bch2_trans_update(trans, iter, k, 0);
-	return 0;
+	ret   = bch2_btree_iter_traverse(iter) ?:
+		bch2_trans_update(trans, iter, k, 0);
+	bch2_trans_iter_put(trans, iter);
+	return ret;
 }
 
 /**
@@ -894,7 +1041,7 @@ retry:
 	       bkey_cmp(iter->pos, end) < 0) {
 		struct bkey_i delete;
 
-		bch2_trans_reset(trans, TRANS_RESET_MEM);
+		bch2_trans_begin(trans);
 
 		bkey_init(&delete.k);
 
@@ -910,7 +1057,7 @@ 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/fs/bcachefs/buckets.c b/fs/bcachefs/buckets.c
index 731b93255876..2e1df04c760d 100644
--- a/fs/bcachefs/buckets.c
+++ b/fs/bcachefs/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;
@@ -1253,21 +1254,21 @@ inline int bch2_mark_overwrite(struct btree_trans *trans,
 			       struct bkey_s_c old,
 			       struct bkey_i *new,
 			       struct bch_fs_usage *fs_usage,
-			       unsigned flags)
+			       unsigned flags,
+			       bool is_extents)
 {
 	struct bch_fs		*c = trans->c;
-	struct btree		*b = iter->l[0].b;
 	unsigned		offset = 0;
-	s64			sectors = 0;
+	s64			sectors = -((s64) old.k->size);
 
 	flags |= BTREE_TRIGGER_OVERWRITE;
 
-	if (btree_node_is_extents(b)
+	if (is_extents
 	    ? bkey_cmp(new->k.p, bkey_start_pos(old.k)) <= 0
 	    : bkey_cmp(new->k.p, old.k->p))
 		return 0;
 
-	if (btree_node_is_extents(b)) {
+	if (is_extents) {
 		switch (bch2_extent_overlap(&new->k, old.k)) {
 		case BCH_EXTENT_OVERLAP_ALL:
 			offset = 0;
@@ -1334,13 +1335,13 @@ int bch2_mark_update(struct btree_trans *trans,
 	    !bkey_deleted(&insert->k))
 		return 0;
 
-	while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
-						      KEY_TYPE_discard))) {
+	while ((_k = bch2_btree_node_iter_peek(&node_iter, b))) {
 		struct bkey		unpacked;
 		struct bkey_s_c		k = bkey_disassemble(b, _k, &unpacked);
 
 		ret = bch2_mark_overwrite(trans, iter, k, insert,
-					  fs_usage, flags);
+					  fs_usage, flags,
+					  btree_node_type_is_extents(iter->btree_id));
 		if (ret <= 0)
 			break;
 
@@ -1380,8 +1381,7 @@ void bch2_trans_fs_usage_apply(struct btree_trans *trans,
 		pr_err("overlapping with");
 
 		node_iter = iter->l[0].iter;
-		while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
-							KEY_TYPE_discard))) {
+		while ((_k = bch2_btree_node_iter_peek(&node_iter, b))) {
 			struct bkey		unpacked;
 			struct bkey_s_c		k;
 
@@ -1443,8 +1443,7 @@ static int bch2_trans_mark_pointer(struct btree_trans *trans,
 	struct bkey_s_c k;
 	struct bkey_alloc_unpacked u;
 	struct bkey_i_alloc *a;
-	u16 *dst_sectors;
-	bool overflow;
+	u16 *dst_sectors, orig_sectors;
 	int ret;
 
 	ret = trans_get_key(trans, BTREE_ID_ALLOC,
@@ -1501,13 +1500,12 @@ static int bch2_trans_mark_pointer(struct btree_trans *trans,
 	dst_sectors = !p.ptr.cached
 		? &u.dirty_sectors
 		: &u.cached_sectors;
+	orig_sectors = *dst_sectors;
 
-	overflow = checked_add(*dst_sectors, sectors);
-
-	if (overflow) {
+	if (checked_add(*dst_sectors, sectors)) {
 		bch2_fs_inconsistent(c,
 			"bucket sector count overflow: %u + %lli > U16_MAX",
-			*dst_sectors, sectors);
+			orig_sectors, sectors);
 		/* return an error indicating that we need full fsck */
 		ret = -EIO;
 		goto out;
@@ -1672,8 +1670,7 @@ static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
 	     k.k->p.offset > idx + sectors))
 		goto out;
 
-	bch2_btree_iter_set_pos(iter, bkey_start_pos(k.k));
-	BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
+	sectors = k.k->p.offset - idx;
 
 	r_v = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 	ret = PTR_ERR_OR_ZERO(r_v);
@@ -1690,9 +1687,12 @@ static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
 		set_bkey_val_u64s(&r_v->k, 0);
 	}
 
+	bch2_btree_iter_set_pos(iter, bkey_start_pos(k.k));
+	BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
+
 	bch2_trans_update(trans, iter, &r_v->k_i, 0);
 out:
-	ret = k.k->p.offset - idx;
+	ret = sectors;
 err:
 	bch2_trans_iter_put(trans, iter);
 	return ret;
@@ -1729,6 +1729,7 @@ 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;
@@ -1792,8 +1793,7 @@ int bch2_trans_mark_update(struct btree_trans *trans,
 	if (unlikely(flags & BTREE_TRIGGER_NOOVERWRITES))
 		return 0;
 
-	while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
-						      KEY_TYPE_discard))) {
+	while ((_k = bch2_btree_node_iter_peek(&node_iter, b))) {
 		struct bkey		unpacked;
 		struct bkey_s_c		k;
 		unsigned		offset = 0;
diff --git a/fs/bcachefs/buckets.h b/fs/bcachefs/buckets.h
index 4717a1a6f568..765650ce9d0a 100644
--- a/fs/bcachefs/buckets.h
+++ b/fs/bcachefs/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;
@@ -267,7 +268,7 @@ int bch2_fs_usage_apply(struct bch_fs *, struct bch_fs_usage *,
 
 int bch2_mark_overwrite(struct btree_trans *, struct btree_iter *,
 			struct bkey_s_c, struct bkey_i *,
-			struct bch_fs_usage *, unsigned);
+			struct bch_fs_usage *, unsigned, bool);
 int bch2_mark_update(struct btree_trans *, struct btree_iter *,
 		     struct bkey_i *, struct bch_fs_usage *, unsigned);
 
diff --git a/fs/bcachefs/checksum.c b/fs/bcachefs/checksum.c
index ad6993b7565a..6f1afa4a3119 100644
--- a/fs/bcachefs/checksum.c
+++ b/fs/bcachefs/checksum.c
@@ -10,7 +10,7 @@
 #include <linux/random.h>
 #include <linux/scatterlist.h>
 #include <crypto/algapi.h>
-#include <crypto/chacha20.h>
+#include <crypto/chacha.h>
 #include <crypto/hash.h>
 #include <crypto/poly1305.h>
 #include <keys/user-type.h>
@@ -67,21 +67,21 @@ static u64 bch2_checksum_update(unsigned type, u64 crc, const void *data, size_t
 	}
 }
 
-static inline void do_encrypt_sg(struct crypto_skcipher *tfm,
+static inline void do_encrypt_sg(struct crypto_sync_skcipher *tfm,
 				 struct nonce nonce,
 				 struct scatterlist *sg, size_t len)
 {
-	SKCIPHER_REQUEST_ON_STACK(req, tfm);
+	SYNC_SKCIPHER_REQUEST_ON_STACK(req, tfm);
 	int ret;
 
-	skcipher_request_set_tfm(req, tfm);
+	skcipher_request_set_sync_tfm(req, tfm);
 	skcipher_request_set_crypt(req, sg, sg, len, nonce.d);
 
 	ret = crypto_skcipher_encrypt(req);
 	BUG_ON(ret);
 }
 
-static inline void do_encrypt(struct crypto_skcipher *tfm,
+static inline void do_encrypt(struct crypto_sync_skcipher *tfm,
 			      struct nonce nonce,
 			      void *buf, size_t len)
 {
@@ -94,8 +94,8 @@ static inline void do_encrypt(struct crypto_skcipher *tfm,
 int bch2_chacha_encrypt_key(struct bch_key *key, struct nonce nonce,
 			    void *buf, size_t len)
 {
-	struct crypto_skcipher *chacha20 =
-		crypto_alloc_skcipher("chacha20", 0, 0);
+	struct crypto_sync_skcipher *chacha20 =
+		crypto_alloc_sync_skcipher("chacha20", 0, 0);
 	int ret;
 
 	if (!chacha20) {
@@ -103,7 +103,8 @@ int bch2_chacha_encrypt_key(struct bch_key *key, struct nonce nonce,
 		return PTR_ERR(chacha20);
 	}
 
-	ret = crypto_skcipher_setkey(chacha20, (void *) key, sizeof(*key));
+	ret = crypto_skcipher_setkey(&chacha20->base,
+				     (void *) key, sizeof(*key));
 	if (ret) {
 		pr_err("crypto_skcipher_setkey() error: %i", ret);
 		goto err;
@@ -111,7 +112,7 @@ int bch2_chacha_encrypt_key(struct bch_key *key, struct nonce nonce,
 
 	do_encrypt(chacha20, nonce, buf, len);
 err:
-	crypto_free_skcipher(chacha20);
+	crypto_free_sync_skcipher(chacha20);
 	return ret;
 }
 
@@ -198,7 +199,7 @@ static struct bch_csum __bch2_checksum_bio(struct bch_fs *c, unsigned type,
 			kunmap_atomic(p);
 		}
 #else
-		__bio_for_each_contig_segment(bv, bio, *iter, *iter)
+		__bio_for_each_bvec(bv, bio, *iter, *iter)
 			crc = bch2_checksum_update(type, crc,
 				page_address(bv.bv_page) + bv.bv_offset,
 				bv.bv_len);
@@ -223,7 +224,7 @@ static struct bch_csum __bch2_checksum_bio(struct bch_fs *c, unsigned type,
 			kunmap_atomic(p);
 		}
 #else
-		__bio_for_each_contig_segment(bv, bio, *iter, *iter)
+		__bio_for_each_bvec(bv, bio, *iter, *iter)
 			crypto_shash_update(desc,
 				page_address(bv.bv_page) + bv.bv_offset,
 				bv.bv_len);
@@ -324,7 +325,7 @@ int bch2_rechecksum_bio(struct bch_fs *c, struct bio *bio,
 
 	BUG_ON(len_a + len_b > bio_sectors(bio));
 	BUG_ON(crc_old.uncompressed_size != bio_sectors(bio));
-	BUG_ON(crc_old.compression_type);
+	BUG_ON(crc_is_compressed(crc_old));
 	BUG_ON(bch2_csum_type_is_encryption(crc_old.csum_type) !=
 	       bch2_csum_type_is_encryption(new_csum_type));
 
@@ -353,6 +354,7 @@ int bch2_rechecksum_bio(struct bch_fs *c, struct bio *bio,
 		if (i->crc)
 			*i->crc = (struct bch_extent_crc_unpacked) {
 				.csum_type		= i->csum_type,
+				.compression_type	= crc_old.compression_type,
 				.compressed_size	= i->len,
 				.uncompressed_size	= i->len,
 				.offset			= 0,
@@ -461,7 +463,7 @@ err:
 static int bch2_alloc_ciphers(struct bch_fs *c)
 {
 	if (!c->chacha20)
-		c->chacha20 = crypto_alloc_skcipher("chacha20", 0, 0);
+		c->chacha20 = crypto_alloc_sync_skcipher("chacha20", 0, 0);
 	if (IS_ERR(c->chacha20)) {
 		bch_err(c, "error requesting chacha20 module: %li",
 			PTR_ERR(c->chacha20));
@@ -544,7 +546,7 @@ int bch2_enable_encryption(struct bch_fs *c, bool keyed)
 			goto err;
 	}
 
-	ret = crypto_skcipher_setkey(c->chacha20,
+	ret = crypto_skcipher_setkey(&c->chacha20->base,
 			(void *) &key.key, sizeof(key.key));
 	if (ret)
 		goto err;
@@ -572,7 +574,7 @@ void bch2_fs_encryption_exit(struct bch_fs *c)
 	if (!IS_ERR_OR_NULL(c->poly1305))
 		crypto_free_shash(c->poly1305);
 	if (!IS_ERR_OR_NULL(c->chacha20))
-		crypto_free_skcipher(c->chacha20);
+		crypto_free_sync_skcipher(c->chacha20);
 	if (!IS_ERR_OR_NULL(c->sha256))
 		crypto_free_shash(c->sha256);
 }
@@ -604,7 +606,7 @@ int bch2_fs_encryption_init(struct bch_fs *c)
 	if (ret)
 		goto out;
 
-	ret = crypto_skcipher_setkey(c->chacha20,
+	ret = crypto_skcipher_setkey(&c->chacha20->base,
 			(void *) &key.key, sizeof(key.key));
 	if (ret)
 		goto out;
diff --git a/fs/bcachefs/checksum.h b/fs/bcachefs/checksum.h
index 1963cbfaaa05..24dee8039d57 100644
--- a/fs/bcachefs/checksum.h
+++ b/fs/bcachefs/checksum.h
@@ -7,7 +7,7 @@
 #include "super-io.h"
 
 #include <linux/crc64.h>
-#include <crypto/chacha20.h>
+#include <crypto/chacha.h>
 
 static inline bool bch2_checksum_mergeable(unsigned type)
 {
@@ -138,9 +138,9 @@ static inline bool bch2_crc_cmp(struct bch_csum l, struct bch_csum r)
 /* for skipping ahead and encrypting/decrypting at an offset: */
 static inline struct nonce nonce_add(struct nonce nonce, unsigned offset)
 {
-	EBUG_ON(offset & (CHACHA20_BLOCK_SIZE - 1));
+	EBUG_ON(offset & (CHACHA_BLOCK_SIZE - 1));
 
-	le32_add_cpu(&nonce.d[0], offset / CHACHA20_BLOCK_SIZE);
+	le32_add_cpu(&nonce.d[0], offset / CHACHA_BLOCK_SIZE);
 	return nonce;
 }
 
@@ -155,13 +155,16 @@ static inline struct nonce null_nonce(void)
 static inline struct nonce extent_nonce(struct bversion version,
 					struct bch_extent_crc_unpacked crc)
 {
-	unsigned size = crc.compression_type ? crc.uncompressed_size : 0;
+	unsigned compression_type = crc_is_compressed(crc)
+		? crc.compression_type
+		: 0;
+	unsigned size = compression_type ? crc.uncompressed_size : 0;
 	struct nonce nonce = (struct nonce) {{
 		[0] = cpu_to_le32(size << 22),
 		[1] = cpu_to_le32(version.lo),
 		[2] = cpu_to_le32(version.lo >> 32),
 		[3] = cpu_to_le32(version.hi|
-				  (crc.compression_type << 24))^BCH_NONCE_EXTENT,
+				  (compression_type << 24))^BCH_NONCE_EXTENT,
 	}};
 
 	return nonce_add(nonce, crc.nonce << 9);
diff --git a/fs/bcachefs/compress.c b/fs/bcachefs/compress.c
index b6b4ec48dccc..0713286d7999 100644
--- a/fs/bcachefs/compress.c
+++ b/fs/bcachefs/compress.c
@@ -17,7 +17,6 @@ struct bbuf {
 		BB_NONE,
 		BB_VMAP,
 		BB_KMALLOC,
-		BB_VMALLOC,
 		BB_MEMPOOL,
 	}		type;
 	int		rw;
@@ -33,17 +32,7 @@ static struct bbuf __bounce_alloc(struct bch_fs *c, unsigned size, int rw)
 	if (b)
 		return (struct bbuf) { .b = b, .type = BB_KMALLOC, .rw = rw };
 
-	b = mempool_alloc(&c->compression_bounce[rw], GFP_NOWAIT);
-	b = b ? page_address(b) : NULL;
-	if (b)
-		return (struct bbuf) { .b = b, .type = BB_MEMPOOL, .rw = rw };
-
-	b = vmalloc(size);
-	if (b)
-		return (struct bbuf) { .b = b, .type = BB_VMALLOC, .rw = rw };
-
 	b = mempool_alloc(&c->compression_bounce[rw], GFP_NOIO);
-	b = b ? page_address(b) : NULL;
 	if (b)
 		return (struct bbuf) { .b = b, .type = BB_MEMPOOL, .rw = rw };
 
@@ -66,7 +55,7 @@ static struct bbuf __bio_map_or_bounce(struct bch_fs *c, struct bio *bio,
 	BUG_ON(bvec_iter_sectors(start) > c->sb.encoded_extent_max);
 
 #ifndef CONFIG_HIGHMEM
-	__bio_for_each_contig_segment(bv, bio, iter, start) {
+	__bio_for_each_bvec(bv, bio, iter, start) {
 		if (bv.bv_len == start.bi_size)
 			return (struct bbuf) {
 				.b = page_address(bv.bv_page) + bv.bv_offset,
@@ -129,12 +118,8 @@ static void bio_unmap_or_unbounce(struct bch_fs *c, struct bbuf buf)
 	case BB_KMALLOC:
 		kfree(buf.b);
 		break;
-	case BB_VMALLOC:
-		vfree(buf.b);
-		break;
 	case BB_MEMPOOL:
-		mempool_free(virt_to_page(buf.b),
-			     &c->compression_bounce[buf.rw]);
+		mempool_free(buf.b, &c->compression_bounce[buf.rw]);
 		break;
 	}
 }
@@ -434,7 +419,7 @@ out:
 	bio_unmap_or_unbounce(c, dst_data);
 	return compression_type;
 err:
-	compression_type = 0;
+	compression_type = BCH_COMPRESSION_TYPE_incompressible;
 	goto out;
 }
 
@@ -561,15 +546,15 @@ static int __bch2_fs_compress_init(struct bch_fs *c, u64 features)
 have_compressed:
 
 	if (!mempool_initialized(&c->compression_bounce[READ])) {
-		ret = mempool_init_page_pool(&c->compression_bounce[READ],
-					     1, order);
+		ret = mempool_init_kvpmalloc_pool(&c->compression_bounce[READ],
+						  1, order);
 		if (ret)
 			goto out;
 	}
 
 	if (!mempool_initialized(&c->compression_bounce[WRITE])) {
-		ret = mempool_init_page_pool(&c->compression_bounce[WRITE],
-					     1, order);
+		ret = mempool_init_kvpmalloc_pool(&c->compression_bounce[WRITE],
+						  1, order);
 		if (ret)
 			goto out;
 	}
diff --git a/fs/bcachefs/dirent.c b/fs/bcachefs/dirent.c
index 623b6c3eda95..ae5c9fd8d9f7 100644
--- a/fs/bcachefs/dirent.c
+++ b/fs/bcachefs/dirent.c
@@ -169,12 +169,12 @@ int bch2_dirent_rename(struct btree_trans *trans,
 		       const struct qstr *dst_name, u64 *dst_inum,
 		       enum bch_rename_mode mode)
 {
-	struct btree_iter *src_iter, *dst_iter;
+	struct btree_iter *src_iter = NULL, *dst_iter = NULL;
 	struct bkey_s_c old_src, old_dst;
 	struct bkey_i_dirent *new_src = NULL, *new_dst = NULL;
 	struct bpos dst_pos =
 		POS(dst_dir, bch2_dirent_hash(dst_hash, dst_name));
-	int ret;
+	int ret = 0;
 
 	*src_inum = *dst_inum = 0;
 
@@ -191,8 +191,10 @@ int bch2_dirent_rename(struct btree_trans *trans,
 		: bch2_hash_lookup(trans, bch2_dirent_hash_desc,
 				   dst_hash, dst_dir, dst_name,
 				   BTREE_ITER_INTENT);
-	if (IS_ERR(dst_iter))
-		return PTR_ERR(dst_iter);
+	ret = PTR_ERR_OR_ZERO(dst_iter);
+	if (ret)
+		goto out;
+
 	old_dst = bch2_btree_iter_peek_slot(dst_iter);
 
 	if (mode != BCH_RENAME)
@@ -202,15 +204,18 @@ int bch2_dirent_rename(struct btree_trans *trans,
 	src_iter = bch2_hash_lookup(trans, bch2_dirent_hash_desc,
 				    src_hash, src_dir, src_name,
 				    BTREE_ITER_INTENT);
-	if (IS_ERR(src_iter))
-		return PTR_ERR(src_iter);
+	ret = PTR_ERR_OR_ZERO(src_iter);
+	if (ret)
+		goto out;
+
 	old_src = bch2_btree_iter_peek_slot(src_iter);
 	*src_inum = le64_to_cpu(bkey_s_c_to_dirent(old_src).v->d_inum);
 
 	/* Create new dst key: */
 	new_dst = dirent_create_key(trans, 0, dst_name, 0);
-	if (IS_ERR(new_dst))
-		return PTR_ERR(new_dst);
+	ret = PTR_ERR_OR_ZERO(new_dst);
+	if (ret)
+		goto out;
 
 	dirent_copy_target(new_dst, bkey_s_c_to_dirent(old_src));
 	new_dst->k.p = dst_iter->pos;
@@ -218,15 +223,18 @@ int bch2_dirent_rename(struct btree_trans *trans,
 	/* Create new src key: */
 	if (mode == BCH_RENAME_EXCHANGE) {
 		new_src = dirent_create_key(trans, 0, src_name, 0);
-		if (IS_ERR(new_src))
-			return PTR_ERR(new_src);
+		ret = PTR_ERR_OR_ZERO(new_src);
+		if (ret)
+			goto out;
 
 		dirent_copy_target(new_src, bkey_s_c_to_dirent(old_dst));
 		new_src->k.p = src_iter->pos;
 	} else {
 		new_src = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
-		if (IS_ERR(new_src))
-			return PTR_ERR(new_src);
+		ret = PTR_ERR_OR_ZERO(new_src);
+		if (ret)
+			goto out;
+
 		bkey_init(&new_src->k);
 		new_src->k.p = src_iter->pos;
 
@@ -247,7 +255,7 @@ int bch2_dirent_rename(struct btree_trans *trans,
 				new_dst->k.p = src_iter->pos;
 				bch2_trans_update(trans, src_iter,
 						  &new_dst->k_i, 0);
-				return 0;
+				goto out;
 			} else {
 				/* If we're overwriting, we can't insert new_dst
 				 * at a different slot because it has to
@@ -261,7 +269,7 @@ int bch2_dirent_rename(struct btree_trans *trans,
 			ret = bch2_hash_needs_whiteout(trans, bch2_dirent_hash_desc,
 						       src_hash, src_iter);
 			if (ret < 0)
-				return ret;
+				goto out;
 
 			if (ret)
 				new_src->k.type = KEY_TYPE_whiteout;
@@ -270,7 +278,10 @@ int bch2_dirent_rename(struct btree_trans *trans,
 
 	bch2_trans_update(trans, src_iter, &new_src->k_i, 0);
 	bch2_trans_update(trans, dst_iter, &new_dst->k_i, 0);
-	return 0;
+out:
+	bch2_trans_iter_put(trans, src_iter);
+	bch2_trans_iter_put(trans, dst_iter);
+	return ret;
 }
 
 int bch2_dirent_delete_at(struct btree_trans *trans,
@@ -331,9 +342,7 @@ int bch2_empty_dir_trans(struct btree_trans *trans, u64 dir_inum)
 			break;
 		}
 	}
-
-	if (!IS_ERR(iter))
-		bch2_trans_iter_put(trans, iter);
+	bch2_trans_iter_put(trans, iter);
 
 	return ret;
 }
diff --git a/fs/bcachefs/ec.c b/fs/bcachefs/ec.c
index a49d0745c720..933945b65925 100644
--- a/fs/bcachefs/ec.c
+++ b/fs/bcachefs/ec.c
@@ -741,6 +741,8 @@ found_slot:
 	ret = bch2_trans_commit(&trans, NULL, NULL,
 				BTREE_INSERT_NOFAIL);
 err:
+	bch2_trans_iter_put(&trans, iter);
+
 	if (ret == -EINTR)
 		goto retry;
 
@@ -802,8 +804,6 @@ static int ec_stripe_update_ptrs(struct bch_fs *c,
 			continue;
 		}
 
-		bch2_btree_iter_set_pos(iter, bkey_start_pos(k.k));
-
 		dev = s->key.v.ptrs[idx].dev;
 
 		bkey_on_stack_reassemble(&sk, c, k);
@@ -818,6 +818,7 @@ static int ec_stripe_update_ptrs(struct bch_fs *c,
 
 		extent_stripe_ptr_add(e, s, ec_ptr, idx);
 
+		bch2_btree_iter_set_pos(iter, bkey_start_pos(&sk.k->k));
 		bch2_trans_update(&trans, iter, sk.k, 0);
 
 		ret = bch2_trans_commit(&trans, NULL, NULL,
@@ -1201,8 +1202,7 @@ static int __bch2_stripe_write_key(struct btree_trans *trans,
 				   struct btree_iter *iter,
 				   struct stripe *m,
 				   size_t idx,
-				   struct bkey_i_stripe *new_key,
-				   unsigned flags)
+				   struct bkey_i_stripe *new_key)
 {
 	struct bch_fs *c = trans->c;
 	struct bkey_s_c k;
@@ -1231,9 +1231,7 @@ static int __bch2_stripe_write_key(struct btree_trans *trans,
 	spin_unlock(&c->ec_stripes_heap_lock);
 
 	bch2_trans_update(trans, iter, &new_key->k_i, 0);
-
-	return bch2_trans_commit(trans, NULL, NULL,
-				 BTREE_INSERT_NOFAIL|flags);
+	return 0;
 }
 
 int bch2_stripes_write(struct bch_fs *c, unsigned flags, bool *wrote)
@@ -1257,12 +1255,10 @@ int bch2_stripes_write(struct bch_fs *c, unsigned flags, bool *wrote)
 		if (!m->dirty)
 			continue;
 
-		do {
-			bch2_trans_reset(&trans, TRANS_RESET_MEM);
-
-			ret = __bch2_stripe_write_key(&trans, iter, m,
-					giter.pos, new_key, flags);
-		} while (ret == -EINTR);
+		ret = __bch2_trans_do(&trans, NULL, NULL,
+				      BTREE_INSERT_NOFAIL|flags,
+			__bch2_stripe_write_key(&trans, iter, m,
+					giter.pos, new_key));
 
 		if (ret)
 			break;
@@ -1280,9 +1276,8 @@ 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 +1286,16 @@ 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/fs/bcachefs/ec.h b/fs/bcachefs/ec.h
index 8d9fbfd19f66..cf67abd48490 100644
--- a/fs/bcachefs/ec.h
+++ b/fs/bcachefs/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/fs/bcachefs/extent_update.c b/fs/bcachefs/extent_update.c
index 846d77dc2530..2a7d913bdda3 100644
--- a/fs/bcachefs/extent_update.c
+++ b/fs/bcachefs/extent_update.c
@@ -39,6 +39,16 @@ static int count_iters_for_insert(struct btree_trans *trans,
 {
 	int ret = 0;
 
+	/*
+	 * The extent update path requires an _additional_ iterator for each
+	 * extent we're inserting and overwriting:
+	 */
+	*nr_iters += 1;
+	if (*nr_iters >= max_iters) {
+		*end = bpos_min(*end, k.k->p);
+		ret = 1;
+	}
+
 	switch (k.k->type) {
 	case KEY_TYPE_extent:
 	case KEY_TYPE_reflink_v:
@@ -105,7 +115,9 @@ int bch2_extent_atomic_end(struct btree_iter *iter,
 	b = iter->l[0].b;
 	node_iter = iter->l[0].iter;
 
-	BUG_ON(bkey_cmp(bkey_start_pos(&insert->k), b->data->min_key) < 0);
+	BUG_ON(bkey_cmp(b->data->min_key, POS_MIN) &&
+	       bkey_cmp(bkey_start_pos(&insert->k),
+			bkey_predecessor(b->data->min_key)) < 0);
 
 	*end = bpos_min(insert->k.p, b->key.k.p);
 
@@ -114,8 +126,7 @@ int bch2_extent_atomic_end(struct btree_iter *iter,
 	if (ret < 0)
 		return ret;
 
-	while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
-						      KEY_TYPE_discard))) {
+	while ((_k = bch2_btree_node_iter_peek(&node_iter, b))) {
 		struct bkey	unpacked;
 		struct bkey_s_c	k = bkey_disassemble(b, _k, &unpacked);
 		unsigned offset = 0;
@@ -167,402 +178,39 @@ int bch2_extent_is_atomic(struct bkey_i *k, struct btree_iter *iter)
 enum btree_insert_ret
 bch2_extent_can_insert(struct btree_trans *trans,
 		       struct btree_iter *iter,
-		       struct bkey_i *insert,
-		       unsigned *u64s)
+		       struct bkey_i *insert)
 {
 	struct btree_iter_level *l = &iter->l[0];
 	struct btree_node_iter node_iter = l->iter;
 	struct bkey_packed *_k;
+	struct bkey_s_c k;
 	struct bkey unpacked;
 	int sectors;
 
-	while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, l->b,
-						      KEY_TYPE_discard))) {
-		struct bkey_s_c k = bkey_disassemble(l->b, _k, &unpacked);
-		enum bch_extent_overlap overlap =
-			bch2_extent_overlap(&insert->k, k.k);
-
-		if (bkey_cmp(bkey_start_pos(k.k), insert->k.p) >= 0)
-			break;
-
-		overlap = bch2_extent_overlap(&insert->k, k.k);
-
-		/*
-		 * If we're overwriting an existing extent, we may need to emit
-		 * a whiteout - unless we're inserting a new extent at the same
-		 * position:
-		 */
-		if (k.k->needs_whiteout &&
-		    (!bkey_whiteout(&insert->k) ||
-		     bkey_cmp(k.k->p, insert->k.p)))
-			*u64s += BKEY_U64s;
-
-		/*
-		 * If we're partially overwriting an existing extent which has
-		 * been written out to disk, we'll need to emit a new version of
-		 * that extent:
-		 */
-		if (bkey_written(l->b, _k) &&
-		    overlap != BCH_EXTENT_OVERLAP_ALL)
-			*u64s += _k->u64s;
-
-		/* And we may be splitting an existing extent: */
-		if (overlap == BCH_EXTENT_OVERLAP_MIDDLE)
-			*u64s += _k->u64s;
-
-		if (overlap == BCH_EXTENT_OVERLAP_MIDDLE &&
-		    (sectors = bch2_bkey_sectors_compressed(k))) {
-			int flags = trans->flags & BTREE_INSERT_NOFAIL
-				? BCH_DISK_RESERVATION_NOFAIL : 0;
-
-			switch (bch2_disk_reservation_add(trans->c,
-					trans->disk_res,
-					sectors, flags)) {
-			case 0:
-				break;
-			case -ENOSPC:
-				return BTREE_INSERT_ENOSPC;
-			default:
-				BUG();
-			}
-		}
-
-		if (overlap == BCH_EXTENT_OVERLAP_FRONT ||
-		    overlap == BCH_EXTENT_OVERLAP_MIDDLE)
-			break;
-
-		bch2_btree_node_iter_advance(&node_iter, l->b);
-	}
-
-	return BTREE_INSERT_OK;
-}
-
-static void verify_extent_nonoverlapping(struct bch_fs *c,
-					 struct btree *b,
-					 struct btree_node_iter *_iter,
-					 struct bkey_i *insert)
-{
-#ifdef CONFIG_BCACHEFS_DEBUG
-	struct btree_node_iter iter;
-	struct bkey_packed *k;
-	struct bkey uk;
-
-	if (!expensive_debug_checks(c))
-		return;
-
-	iter = *_iter;
-	k = bch2_btree_node_iter_prev_filter(&iter, b, KEY_TYPE_discard);
-	BUG_ON(k &&
-	       (uk = bkey_unpack_key(b, k),
-		bkey_cmp(uk.p, bkey_start_pos(&insert->k)) > 0));
-
-	iter = *_iter;
-	k = bch2_btree_node_iter_peek_filter(&iter, b, KEY_TYPE_discard);
-#if 0
-	BUG_ON(k &&
-	       (uk = bkey_unpack_key(b, k),
-		bkey_cmp(insert->k.p, bkey_start_pos(&uk))) > 0);
-#else
-	if (k &&
-	    (uk = bkey_unpack_key(b, k),
-	     bkey_cmp(insert->k.p, bkey_start_pos(&uk))) > 0) {
-		char buf1[100];
-		char buf2[100];
-
-		bch2_bkey_to_text(&PBUF(buf1), &insert->k);
-		bch2_bkey_to_text(&PBUF(buf2), &uk);
-
-		bch2_dump_btree_node(b);
-		panic("insert > next :\n"
-		      "insert %s\n"
-		      "next   %s\n",
-		      buf1, buf2);
-	}
-#endif
-
-#endif
-}
-
-static void extent_bset_insert(struct bch_fs *c, struct btree_iter *iter,
-			       struct bkey_i *insert)
-{
-	struct btree_iter_level *l = &iter->l[0];
-	struct bkey_packed *k =
-		bch2_btree_node_iter_bset_pos(&l->iter, l->b, bset_tree_last(l->b));
-
-	BUG_ON(insert->k.u64s > bch_btree_keys_u64s_remaining(c, l->b));
-
-	EBUG_ON(bkey_deleted(&insert->k) || !insert->k.size);
-	verify_extent_nonoverlapping(c, l->b, &l->iter, insert);
-
-	if (debug_check_bkeys(c))
-		bch2_bkey_debugcheck(c, l->b, bkey_i_to_s_c(insert));
-
-	bch2_bset_insert(l->b, &l->iter, k, insert, 0);
-	bch2_btree_node_iter_fix(iter, l->b, &l->iter, k, 0, k->u64s);
-}
-
-static void pack_push_whiteout(struct bch_fs *c, struct btree *b,
-			       struct bpos pos)
-{
-	struct bkey_packed k;
-
-	if (!bkey_pack_pos(&k, pos, b)) {
-		struct bkey_i tmp;
+	_k = bch2_btree_node_iter_peek(&node_iter, l->b);
+	if (!_k)
+		return BTREE_INSERT_OK;
 
-		bkey_init(&tmp.k);
-		tmp.k.p = pos;
-		bkey_copy(&k, &tmp);
-	}
+	k = bkey_disassemble(l->b, _k, &unpacked);
 
-	k.needs_whiteout = true;
-	push_whiteout(c, b, &k);
-}
+	/* Check if we're splitting a compressed extent: */
 
-static void
-extent_drop(struct bch_fs *c, struct btree_iter *iter,
-	    struct bkey_packed *_k, struct bkey_s k)
-{
-	struct btree_iter_level *l = &iter->l[0];
-
-	if (!bkey_whiteout(k.k))
-		btree_account_key_drop(l->b, _k);
-
-	k.k->size = 0;
-	k.k->type = KEY_TYPE_deleted;
-
-	if (!btree_node_old_extent_overwrite(l->b) &&
-	    k.k->needs_whiteout) {
-		pack_push_whiteout(c, l->b, k.k->p);
-		k.k->needs_whiteout = false;
-	}
-
-	if (_k >= btree_bset_last(l->b)->start) {
-		unsigned u64s = _k->u64s;
-
-		bch2_bset_delete(l->b, _k, _k->u64s);
-		bch2_btree_node_iter_fix(iter, l->b, &l->iter, _k, u64s, 0);
-	} else {
-		extent_save(l->b, _k, k.k);
-		bch2_btree_iter_fix_key_modified(iter, l->b, _k);
-	}
-}
-
-static void
-extent_squash(struct bch_fs *c, struct btree_iter *iter,
-	      struct bkey_i *insert,
-	      struct bkey_packed *_k, struct bkey_s k,
-	      enum bch_extent_overlap overlap)
-{
-	struct btree_iter_level *l = &iter->l[0];
-	struct bkey_on_stack tmp, split;
+	if (bkey_cmp(bkey_start_pos(&insert->k), bkey_start_pos(k.k)) > 0 &&
+	    bkey_cmp(insert->k.p, k.k->p) < 0 &&
+	    (sectors = bch2_bkey_sectors_compressed(k))) {
+		int flags = trans->flags & BTREE_INSERT_NOFAIL
+			? BCH_DISK_RESERVATION_NOFAIL : 0;
 
-	bkey_on_stack_init(&tmp);
-	bkey_on_stack_init(&split);
-
-	if (!btree_node_old_extent_overwrite(l->b)) {
-		if (!bkey_whiteout(&insert->k) &&
-		    !bkey_cmp(k.k->p, insert->k.p)) {
-			insert->k.needs_whiteout = k.k->needs_whiteout;
-			k.k->needs_whiteout = false;
-		}
-	} else {
-		insert->k.needs_whiteout |= k.k->needs_whiteout;
-	}
-
-	switch (overlap) {
-	case BCH_EXTENT_OVERLAP_FRONT:
-		if (bkey_written(l->b, _k)) {
-			bkey_on_stack_reassemble(&tmp, c, k.s_c);
-			bch2_cut_front(insert->k.p, tmp.k);
-
-			/*
-			 * needs_whiteout was propagated to new version of @k,
-			 * @tmp:
-			 */
-			if (!btree_node_old_extent_overwrite(l->b))
-				k.k->needs_whiteout = false;
-
-			extent_drop(c, iter, _k, k);
-			extent_bset_insert(c, iter, tmp.k);
-		} else {
-			btree_keys_account_val_delta(l->b, _k,
-				bch2_cut_front_s(insert->k.p, k));
-
-			extent_save(l->b, _k, k.k);
-			/*
-			 * No need to call bset_fix_invalidated_key, start of
-			 * extent changed but extents are indexed by where they
-			 * end
-			 */
-			bch2_btree_iter_fix_key_modified(iter, l->b, _k);
-		}
-		break;
-	case BCH_EXTENT_OVERLAP_BACK:
-		if (bkey_written(l->b, _k)) {
-			bkey_on_stack_reassemble(&tmp, c, k.s_c);
-			bch2_cut_back(bkey_start_pos(&insert->k), tmp.k);
-
-			/*
-			 * @tmp has different position than @k, needs_whiteout
-			 * should not be propagated:
-			 */
-			if (!btree_node_old_extent_overwrite(l->b))
-				tmp.k->k.needs_whiteout = false;
-
-			extent_drop(c, iter, _k, k);
-			extent_bset_insert(c, iter, tmp.k);
-		} else {
-			/*
-			 * position of @k is changing, emit a whiteout if
-			 * needs_whiteout is set:
-			 */
-			if (!btree_node_old_extent_overwrite(l->b) &&
-			    k.k->needs_whiteout) {
-				pack_push_whiteout(c, l->b, k.k->p);
-				k.k->needs_whiteout = false;
-			}
-
-			btree_keys_account_val_delta(l->b, _k,
-				bch2_cut_back_s(bkey_start_pos(&insert->k), k));
-			extent_save(l->b, _k, k.k);
-
-			bch2_bset_fix_invalidated_key(l->b, _k);
-			bch2_btree_node_iter_fix(iter, l->b, &l->iter,
-						 _k, _k->u64s, _k->u64s);
-		}
-		break;
-	case BCH_EXTENT_OVERLAP_ALL:
-		extent_drop(c, iter, _k, k);
-		break;
-	case BCH_EXTENT_OVERLAP_MIDDLE:
-		bkey_on_stack_reassemble(&split, c, k.s_c);
-		bch2_cut_back(bkey_start_pos(&insert->k), split.k);
-
-		if (!btree_node_old_extent_overwrite(l->b))
-			split.k->k.needs_whiteout = false;
-
-		/* this is identical to BCH_EXTENT_OVERLAP_FRONT: */
-		if (bkey_written(l->b, _k)) {
-			bkey_on_stack_reassemble(&tmp, c, k.s_c);
-			bch2_cut_front(insert->k.p, tmp.k);
-
-			if (!btree_node_old_extent_overwrite(l->b))
-				k.k->needs_whiteout = false;
-
-			extent_drop(c, iter, _k, k);
-			extent_bset_insert(c, iter, tmp.k);
-		} else {
-			btree_keys_account_val_delta(l->b, _k,
-				bch2_cut_front_s(insert->k.p, k));
-
-			extent_save(l->b, _k, k.k);
-			bch2_btree_iter_fix_key_modified(iter, l->b, _k);
-		}
-
-		extent_bset_insert(c, iter, split.k);
-		break;
-	}
-
-	bkey_on_stack_exit(&split, c);
-	bkey_on_stack_exit(&tmp, c);
-}
-
-/**
- * bch_extent_insert_fixup - insert a new extent and deal with overlaps
- *
- * this may result in not actually doing the insert, or inserting some subset
- * of the insert key. For cmpxchg operations this is where that logic lives.
- *
- * All subsets of @insert that need to be inserted are inserted using
- * bch2_btree_insert_and_journal(). If @b or @res fills up, this function
- * returns false, setting @iter->pos for the prefix of @insert that actually got
- * inserted.
- *
- * BSET INVARIANTS: this function is responsible for maintaining all the
- * invariants for bsets of extents in memory. things get really hairy with 0
- * size extents
- *
- * within one bset:
- *
- * bkey_start_pos(bkey_next(k)) >= k
- * or bkey_start_offset(bkey_next(k)) >= k->offset
- *
- * i.e. strict ordering, no overlapping extents.
- *
- * multiple bsets (i.e. full btree node):
- *
- * ∀ k, j
- *   k.size != 0 ∧ j.size != 0 →
- *     ¬ (k > bkey_start_pos(j) ∧ k < j)
- *
- * i.e. no two overlapping keys _of nonzero size_
- *
- * We can't realistically maintain this invariant for zero size keys because of
- * the key merging done in bch2_btree_insert_key() - for two mergeable keys k, j
- * there may be another 0 size key between them in another bset, and it will
- * thus overlap with the merged key.
- *
- * In addition, the end of iter->pos indicates how much has been processed.
- * If the end of iter->pos is not the same as the end of insert, then
- * key insertion needs to continue/be retried.
- */
-void bch2_insert_fixup_extent(struct btree_trans *trans,
-			      struct btree_iter *iter,
-			      struct bkey_i *insert)
-{
-	struct bch_fs *c = trans->c;
-	struct btree_iter_level *l = &iter->l[0];
-	struct btree_node_iter node_iter = l->iter;
-	bool do_update		= !bkey_whiteout(&insert->k);
-	struct bkey_packed *_k;
-	struct bkey unpacked;
-
-	EBUG_ON(iter->level);
-	EBUG_ON(!insert->k.size);
-	EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&insert->k)));
-
-	while ((_k = bch2_btree_node_iter_peek_filter(&l->iter, l->b,
-						      KEY_TYPE_discard))) {
-		struct bkey_s k = __bkey_disassemble(l->b, _k, &unpacked);
-		enum bch_extent_overlap overlap =
-			bch2_extent_overlap(&insert->k, k.k);
-
-		if (bkey_cmp(bkey_start_pos(k.k), insert->k.p) >= 0)
+		switch (bch2_disk_reservation_add(trans->c, trans->disk_res,
+						  sectors, flags)) {
+		case 0:
 			break;
-
-		if (!bkey_whiteout(k.k))
-			do_update = true;
-
-		if (!do_update) {
-			struct bpos cur_end = bpos_min(insert->k.p, k.k->p);
-
-			bch2_cut_front(cur_end, insert);
-			bch2_btree_iter_set_pos_same_leaf(iter, cur_end);
-		} else {
-			extent_squash(c, iter, insert, _k, k, overlap);
+		case -ENOSPC:
+			return BTREE_INSERT_ENOSPC;
+		default:
+			BUG();
 		}
-
-		node_iter = l->iter;
-
-		if (overlap == BCH_EXTENT_OVERLAP_FRONT ||
-		    overlap == BCH_EXTENT_OVERLAP_MIDDLE)
-			break;
 	}
 
-	l->iter = node_iter;
-	bch2_btree_iter_set_pos_same_leaf(iter, insert->k.p);
-
-	if (do_update) {
-		if (insert->k.type == KEY_TYPE_deleted)
-			insert->k.type = KEY_TYPE_discard;
-
-		if (!bkey_whiteout(&insert->k) ||
-		    btree_node_old_extent_overwrite(l->b))
-			extent_bset_insert(c, iter, insert);
-
-		bch2_btree_journal_key(trans, iter, insert);
-	}
-
-	bch2_cut_front(insert->k.p, insert);
+	return BTREE_INSERT_OK;
 }
diff --git a/fs/bcachefs/extent_update.h b/fs/bcachefs/extent_update.h
index e9dc8091ba3f..38dc084627d2 100644
--- a/fs/bcachefs/extent_update.h
+++ b/fs/bcachefs/extent_update.h
@@ -11,9 +11,6 @@ int bch2_extent_is_atomic(struct bkey_i *, struct btree_iter *);
 
 enum btree_insert_ret
 bch2_extent_can_insert(struct btree_trans *, struct btree_iter *,
-		       struct bkey_i *, unsigned *);
-void bch2_insert_fixup_extent(struct btree_trans *,
-			      struct btree_iter *,
-			      struct bkey_i *);
+		       struct bkey_i *);
 
 #endif /* _BCACHEFS_EXTENT_UPDATE_H */
diff --git a/fs/bcachefs/extents.c b/fs/bcachefs/extents.c
index c4b0b9e15a8f..3f66457d2272 100644
--- a/fs/bcachefs/extents.c
+++ b/fs/bcachefs/extents.c
@@ -9,6 +9,7 @@
 #include "bcachefs.h"
 #include "bkey_methods.h"
 #include "btree_gc.h"
+#include "btree_io.h"
 #include "btree_iter.h"
 #include "buckets.h"
 #include "checksum.h"
@@ -214,6 +215,37 @@ void bch2_btree_ptr_to_text(struct printbuf *out, struct bch_fs *c,
 	bch2_bkey_ptrs_to_text(out, c, k);
 }
 
+void bch2_btree_ptr_v2_to_text(struct printbuf *out, struct bch_fs *c,
+			    struct bkey_s_c k)
+{
+	struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
+
+	pr_buf(out, "seq %llu sectors %u written %u min_key ",
+	       le64_to_cpu(bp.v->seq),
+	       le16_to_cpu(bp.v->sectors),
+	       le16_to_cpu(bp.v->sectors_written));
+
+	bch2_bpos_to_text(out, bp.v->min_key);
+	pr_buf(out, " ");
+	bch2_bkey_ptrs_to_text(out, c, k);
+}
+
+void bch2_btree_ptr_v2_compat(enum btree_id btree_id, unsigned version,
+			      unsigned big_endian, int write,
+			      struct bkey_s k)
+{
+	struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(k);
+
+	compat_bpos(0, btree_id, version, big_endian, write, &bp.v->min_key);
+
+	if (version < bcachefs_metadata_version_inode_btree_change &&
+	    btree_node_type_is_extents(btree_id) &&
+	    bkey_cmp(bp.v->min_key, POS_MIN))
+		bp.v->min_key = write
+			? bkey_predecessor(bp.v->min_key)
+			: bkey_successor(bp.v->min_key);
+}
+
 /* KEY_TYPE_extent: */
 
 const char *bch2_extent_invalid(const struct bch_fs *c, struct bkey_s_c k)
@@ -337,7 +369,7 @@ enum merge_result bch2_extent_merge(struct bch_fs *c,
 			if (!bch2_checksum_mergeable(crc_l.csum_type))
 				return BCH_MERGE_NOMERGE;
 
-			if (crc_l.compression_type)
+			if (crc_is_compressed(crc_l))
 				return BCH_MERGE_NOMERGE;
 
 			if (crc_l.csum_type &&
@@ -345,7 +377,7 @@ enum merge_result bch2_extent_merge(struct bch_fs *c,
 			    crc_r.uncompressed_size > c->sb.encoded_extent_max)
 				return BCH_MERGE_NOMERGE;
 
-			if (crc_l.uncompressed_size + crc_r.uncompressed_size - 1 >
+			if (crc_l.uncompressed_size + crc_r.uncompressed_size >
 			    bch2_crc_field_size_max[extent_entry_type(en_l)])
 				return BCH_MERGE_NOMERGE;
 
@@ -448,7 +480,7 @@ static inline bool bch2_crc_unpacked_cmp(struct bch_extent_crc_unpacked l,
 static inline bool can_narrow_crc(struct bch_extent_crc_unpacked u,
 				  struct bch_extent_crc_unpacked n)
 {
-	return !u.compression_type &&
+	return !crc_is_compressed(u) &&
 		u.csum_type &&
 		u.uncompressed_size > u.live_size &&
 		bch2_csum_type_is_encryption(u.csum_type) ==
@@ -492,7 +524,7 @@ bool bch2_bkey_narrow_crcs(struct bkey_i *k, struct bch_extent_crc_unpacked n)
 	/* Find a checksum entry that covers only live data: */
 	if (!n.csum_type) {
 		bkey_for_each_crc(&k->k, ptrs, u, i)
-			if (!u.compression_type &&
+			if (!crc_is_compressed(u) &&
 			    u.csum_type &&
 			    u.live_size == u.uncompressed_size) {
 				n = u;
@@ -501,7 +533,7 @@ bool bch2_bkey_narrow_crcs(struct bkey_i *k, struct bch_extent_crc_unpacked n)
 		return false;
 	}
 found:
-	BUG_ON(n.compression_type);
+	BUG_ON(crc_is_compressed(n));
 	BUG_ON(n.offset);
 	BUG_ON(n.live_size != k->k.size);
 
@@ -563,15 +595,15 @@ void bch2_extent_crc_append(struct bkey_i *k,
 	enum bch_extent_entry_type type;
 
 	if (bch_crc_bytes[new.csum_type]	<= 4 &&
-	    new.uncompressed_size - 1		<= CRC32_SIZE_MAX &&
+	    new.uncompressed_size		<= CRC32_SIZE_MAX &&
 	    new.nonce				<= CRC32_NONCE_MAX)
 		type = BCH_EXTENT_ENTRY_crc32;
 	else if (bch_crc_bytes[new.csum_type]	<= 10 &&
-		   new.uncompressed_size - 1	<= CRC64_SIZE_MAX &&
+		   new.uncompressed_size	<= CRC64_SIZE_MAX &&
 		   new.nonce			<= CRC64_NONCE_MAX)
 		type = BCH_EXTENT_ENTRY_crc64;
 	else if (bch_crc_bytes[new.csum_type]	<= 16 &&
-		   new.uncompressed_size - 1	<= CRC128_SIZE_MAX &&
+		   new.uncompressed_size	<= CRC128_SIZE_MAX &&
 		   new.nonce			<= CRC128_NONCE_MAX)
 		type = BCH_EXTENT_ENTRY_crc128;
 	else
@@ -610,8 +642,7 @@ unsigned bch2_bkey_nr_ptrs_fully_allocated(struct bkey_s_c k)
 		struct extent_ptr_decoded p;
 
 		bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
-			ret += !p.ptr.cached &&
-				p.crc.compression_type == BCH_COMPRESSION_TYPE_none;
+			ret += !p.ptr.cached && !crc_is_compressed(p.crc);
 	}
 
 	return ret;
@@ -625,13 +656,24 @@ unsigned bch2_bkey_sectors_compressed(struct bkey_s_c k)
 	unsigned ret = 0;
 
 	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
-		if (!p.ptr.cached &&
-		    p.crc.compression_type != BCH_COMPRESSION_TYPE_none)
+		if (!p.ptr.cached && crc_is_compressed(p.crc))
 			ret += p.crc.compressed_size;
 
 	return ret;
 }
 
+bool bch2_bkey_is_incompressible(struct bkey_s_c k)
+{
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
+	struct bch_extent_crc_unpacked crc;
+
+	bkey_for_each_crc(k.k, ptrs, crc, entry)
+		if (crc.compression_type == BCH_COMPRESSION_TYPE_incompressible)
+			return true;
+	return false;
+}
+
 bool bch2_check_range_allocated(struct bch_fs *c, struct bpos pos, u64 size,
 				unsigned nr_replicas)
 {
@@ -739,6 +781,7 @@ 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);
 
@@ -1021,6 +1064,8 @@ 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)
@@ -1069,17 +1114,19 @@ 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/fs/bcachefs/extents.h b/fs/bcachefs/extents.h
index 7c5a41e6d79d..29b15365d19c 100644
--- a/fs/bcachefs/extents.h
+++ b/fs/bcachefs/extents.h
@@ -175,6 +175,12 @@ bch2_extent_crc_unpack(const struct bkey *k, const union bch_extent_crc *crc)
 #undef common_fields
 }
 
+static inline bool crc_is_compressed(struct bch_extent_crc_unpacked crc)
+{
+	return (crc.compression_type != BCH_COMPRESSION_TYPE_none &&
+		crc.compression_type != BCH_COMPRESSION_TYPE_incompressible);
+}
+
 /* bkey_ptrs: generically over any key type that has ptrs */
 
 struct bkey_ptrs_c {
@@ -219,6 +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 };
 	}
@@ -359,6 +372,11 @@ void bch2_btree_ptr_debugcheck(struct bch_fs *, struct bkey_s_c);
 void bch2_btree_ptr_to_text(struct printbuf *, struct bch_fs *,
 			    struct bkey_s_c);
 
+void bch2_btree_ptr_v2_to_text(struct printbuf *, struct bch_fs *,
+			    struct bkey_s_c);
+void bch2_btree_ptr_v2_compat(enum btree_id, unsigned, unsigned,
+			      int, struct bkey_s);
+
 #define bch2_bkey_ops_btree_ptr (struct bkey_ops) {		\
 	.key_invalid	= bch2_btree_ptr_invalid,		\
 	.key_debugcheck	= bch2_btree_ptr_debugcheck,		\
@@ -366,6 +384,14 @@ 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_v2_to_text,		\
+	.swab		= bch2_ptr_swab,			\
+	.compat		= bch2_btree_ptr_v2_compat,		\
+}
+
 /* KEY_TYPE_extent: */
 
 const char *bch2_extent_invalid(const struct bch_fs *, struct bkey_s_c);
@@ -410,6 +436,7 @@ 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;
@@ -483,6 +510,7 @@ static inline struct bch_devs_list bch2_bkey_cached_devs(struct bkey_s_c k)
 unsigned bch2_bkey_nr_ptrs(struct bkey_s_c);
 unsigned bch2_bkey_nr_ptrs_allocated(struct bkey_s_c);
 unsigned bch2_bkey_nr_ptrs_fully_allocated(struct bkey_s_c);
+bool bch2_bkey_is_incompressible(struct bkey_s_c);
 unsigned bch2_bkey_sectors_compressed(struct bkey_s_c);
 bool bch2_check_range_allocated(struct bch_fs *, struct bpos, u64, unsigned);
 unsigned bch2_bkey_durability(struct bch_fs *, struct bkey_s_c);
@@ -525,7 +553,7 @@ 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/fs/bcachefs/fs-common.c b/fs/bcachefs/fs-common.c
index 96f7bbe0a3ed..878419d40992 100644
--- a/fs/bcachefs/fs-common.c
+++ b/fs/bcachefs/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/fs/bcachefs/fs-io.c b/fs/bcachefs/fs-io.c
index c0f8cd8942e4..0aa3afade4ea 100644
--- a/fs/bcachefs/fs-io.c
+++ b/fs/bcachefs/fs-io.c
@@ -602,7 +602,7 @@ int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 	EBUG_ON(!PageLocked(page));
 	EBUG_ON(!PageLocked(newpage));
 
-	ret = migrate_page_move_mapping(mapping, newpage, page, NULL, mode, 0);
+	ret = migrate_page_move_mapping(mapping, newpage, page, 0);
 	if (ret != MIGRATEPAGE_SUCCESS)
 		return ret;
 
@@ -627,10 +627,10 @@ int bch2_migrate_page(struct address_space *mapping, struct page *newpage,
 
 static void bch2_readpages_end_io(struct bio *bio)
 {
+	struct bvec_iter_all iter;
 	struct bio_vec *bv;
-	unsigned i;
 
-	bio_for_each_segment_all(bv, bio, i) {
+	bio_for_each_segment_all(bv, bio, iter) {
 		struct page *page = bv->bv_page;
 
 		if (!bio->bi_status) {
@@ -782,11 +782,8 @@ static void readpage_bio_extend(struct readpages_iter *iter,
 			if (!get_more)
 				break;
 
-			rcu_read_lock();
-			page = radix_tree_lookup(&iter->mapping->i_pages, page_offset);
-			rcu_read_unlock();
-
-			if (page && !radix_tree_exceptional_entry(page))
+			page = xa_load(&iter->mapping->i_pages, page_offset);
+			if (page && !xa_is_value(page))
 				break;
 
 			page = __page_cache_alloc(readahead_gfp_mask(iter->mapping));
@@ -1037,32 +1034,33 @@ static void bch2_writepage_io_done(struct closure *cl)
 					struct bch_writepage_io, cl);
 	struct bch_fs *c = io->op.c;
 	struct bio *bio = &io->op.wbio.bio;
+	struct bvec_iter_all iter;
 	struct bio_vec *bvec;
-	unsigned i, j;
+	unsigned i;
 
 	if (io->op.error) {
-		bio_for_each_segment_all(bvec, bio, i) {
+		bio_for_each_segment_all(bvec, bio, iter) {
 			struct bch_page_state *s;
 
 			SetPageError(bvec->bv_page);
-			mapping_set_error(io->inode->v.i_mapping, -EIO);
+			mapping_set_error(bvec->bv_page->mapping, -EIO);
 
 			s = __bch2_page_state(bvec->bv_page);
 			spin_lock(&s->lock);
-			for (j = 0; j < PAGE_SECTORS; j++)
-				s->s[j].nr_replicas = 0;
+			for (i = 0; i < PAGE_SECTORS; i++)
+				s->s[i].nr_replicas = 0;
 			spin_unlock(&s->lock);
 		}
 	}
 
 	if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
-		bio_for_each_segment_all(bvec, bio, i) {
+		bio_for_each_segment_all(bvec, bio, iter) {
 			struct bch_page_state *s;
 
 			s = __bch2_page_state(bvec->bv_page);
 			spin_lock(&s->lock);
-			for (j = 0; j < PAGE_SECTORS; j++)
-				s->s[j].nr_replicas = 0;
+			for (i = 0; i < PAGE_SECTORS; i++)
+				s->s[i].nr_replicas = 0;
 			spin_unlock(&s->lock);
 		}
 	}
@@ -1086,7 +1084,7 @@ static void bch2_writepage_io_done(struct closure *cl)
 	 */
 	i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
 
-	bio_for_each_segment_all(bvec, bio, i) {
+	bio_for_each_segment_all(bvec, bio, iter) {
 		struct bch_page_state *s = __bch2_page_state(bvec->bv_page);
 
 		if (atomic_dec_and_test(&s->write_count))
@@ -1240,7 +1238,7 @@ do_io:
 
 		if (w->io &&
 		    (w->io->op.res.nr_replicas != nr_replicas_this_write ||
-		     bio_full(&w->io->op.wbio.bio) ||
+		     bio_full(&w->io->op.wbio.bio, PAGE_SIZE) ||
 		     w->io->op.wbio.bio.bi_iter.bi_size >= (256U << 20) ||
 		     bio_end_sector(&w->io->op.wbio.bio) != sector))
 			bch2_writepage_do_io(w);
@@ -1805,10 +1803,11 @@ static long bch2_dio_write_loop(struct dio_write *dio)
 	struct address_space *mapping = req->ki_filp->f_mapping;
 	struct bch_inode_info *inode = file_bch_inode(req->ki_filp);
 	struct bio *bio = &dio->op.wbio.bio;
+	struct bvec_iter_all iter;
 	struct bio_vec *bv;
-	unsigned i, unaligned;
+	unsigned unaligned;
 	u64 new_i_size;
-	bool sync;
+	bool sync = dio->sync;
 	long ret;
 
 	if (dio->loop)
@@ -1838,7 +1837,7 @@ static long bch2_dio_write_loop(struct dio_write *dio)
 			 * bio_iov_iter_get_pages was only able to get <
 			 * blocksize worth of pages:
 			 */
-			bio_for_each_segment_all(bv, bio, i)
+			bio_for_each_segment_all(bv, bio, iter)
 				put_page(bv->bv_page);
 			ret = -EFAULT;
 			goto err;
@@ -1856,7 +1855,7 @@ static long bch2_dio_write_loop(struct dio_write *dio)
 				iov = kmalloc(dio->iter.nr_segs * sizeof(*iov),
 					      GFP_KERNEL);
 				if (unlikely(!iov)) {
-					dio->sync = true;
+					dio->sync = sync = true;
 					goto do_io;
 				}
 
@@ -1870,7 +1869,7 @@ do_io:
 		dio->loop = true;
 		closure_call(&dio->op.cl, bch2_write, NULL, NULL);
 
-		if (dio->sync)
+		if (sync)
 			wait_for_completion(&dio->done);
 		else
 			return -EIOCBQUEUED;
@@ -1886,7 +1885,7 @@ loop:
 			i_size_write(&inode->v, new_i_size);
 		spin_unlock(&inode->v.i_lock);
 
-		bio_for_each_segment_all(bv, bio, i)
+		bio_for_each_segment_all(bv, bio, iter)
 			put_page(bv->bv_page);
 		if (!dio->iter.count || dio->op.error)
 			break;
@@ -1904,7 +1903,6 @@ err:
 	if (dio->free_iov)
 		kfree(dio->iter.iov);
 
-	sync = dio->sync;
 	bio_put(bio);
 
 	/* inode->i_dio_count is our ref on inode and thus bch_fs */
@@ -2514,10 +2512,8 @@ reassemble:
 		bkey_on_stack_reassemble(&copy, c, k);
 
 		if (insert &&
-		    bkey_cmp(bkey_start_pos(k.k), move_pos) < 0) {
+		    bkey_cmp(bkey_start_pos(k.k), move_pos) < 0)
 			bch2_cut_front(move_pos, copy.k);
-			bch2_btree_iter_set_pos(src, bkey_start_pos(&copy.k->k));
-		}
 
 		copy.k->k.p.offset += shift >> 9;
 		bch2_btree_iter_set_pos(dst, bkey_start_pos(&copy.k->k));
@@ -2537,8 +2533,9 @@ reassemble:
 		}
 
 		bkey_init(&delete.k);
-		delete.k.p = src->pos;
-		bch2_key_resize(&delete.k, copy.k->k.size);
+		delete.k.p = copy.k->k.p;
+		delete.k.size = copy.k->k.size;
+		delete.k.p.offset -= shift >> 9;
 
 		next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
 
@@ -2559,6 +2556,8 @@ reassemble:
 			BUG_ON(ret);
 		}
 
+		bch2_btree_iter_set_pos(src, bkey_start_pos(&delete.k));
+
 		ret =   bch2_trans_update(&trans, src, &delete, trigger_flags) ?:
 			bch2_trans_update(&trans, dst, copy.k, trigger_flags) ?:
 			bch2_trans_commit(&trans, &disk_res,
@@ -2649,7 +2648,7 @@ static long bchfs_fallocate(struct bch_inode_info *inode, int mode,
 		struct bkey_i_reservation reservation;
 		struct bkey_s_c k;
 
-		bch2_trans_reset(&trans, TRANS_RESET_MEM);
+		bch2_trans_begin(&trans);
 
 		k = bch2_btree_iter_peek_slot(iter);
 		if ((ret = bkey_err(k)))
@@ -2823,235 +2822,6 @@ static void mark_range_unallocated(struct bch_inode_info *inode,
 	} while (index <= end_index);
 }
 
-static int generic_access_check_limits(struct file *file, loff_t pos,
-				       loff_t *count)
-{
-	struct inode *inode = file->f_mapping->host;
-	loff_t max_size = inode->i_sb->s_maxbytes;
-
-	if (!(file->f_flags & O_LARGEFILE))
-		max_size = MAX_NON_LFS;
-
-	if (unlikely(pos >= max_size))
-		return -EFBIG;
-	*count = min(*count, max_size - pos);
-	return 0;
-}
-
-static int generic_write_check_limits(struct file *file, loff_t pos,
-				      loff_t *count)
-{
-	loff_t limit = rlimit(RLIMIT_FSIZE);
-
-	if (limit != RLIM_INFINITY) {
-		if (pos >= limit) {
-			send_sig(SIGXFSZ, current, 0);
-			return -EFBIG;
-		}
-		*count = min(*count, limit - pos);
-	}
-
-	return generic_access_check_limits(file, pos, count);
-}
-
-static int generic_remap_checks(struct file *file_in, loff_t pos_in,
-			 struct file *file_out, loff_t pos_out,
-			 loff_t *req_count, unsigned int remap_flags)
-{
-	struct inode *inode_in = file_in->f_mapping->host;
-	struct inode *inode_out = file_out->f_mapping->host;
-	uint64_t count = *req_count;
-	uint64_t bcount;
-	loff_t size_in, size_out;
-	loff_t bs = inode_out->i_sb->s_blocksize;
-	int ret;
-
-	/* The start of both ranges must be aligned to an fs block. */
-	if (!IS_ALIGNED(pos_in, bs) || !IS_ALIGNED(pos_out, bs))
-		return -EINVAL;
-
-	/* Ensure offsets don't wrap. */
-	if (pos_in + count < pos_in || pos_out + count < pos_out)
-		return -EINVAL;
-
-	size_in = i_size_read(inode_in);
-	size_out = i_size_read(inode_out);
-
-	/* Dedupe requires both ranges to be within EOF. */
-	if ((remap_flags & REMAP_FILE_DEDUP) &&
-	    (pos_in >= size_in || pos_in + count > size_in ||
-	     pos_out >= size_out || pos_out + count > size_out))
-		return -EINVAL;
-
-	/* Ensure the infile range is within the infile. */
-	if (pos_in >= size_in)
-		return -EINVAL;
-	count = min(count, size_in - (uint64_t)pos_in);
-
-	ret = generic_access_check_limits(file_in, pos_in, &count);
-	if (ret)
-		return ret;
-
-	ret = generic_write_check_limits(file_out, pos_out, &count);
-	if (ret)
-		return ret;
-
-	/*
-	 * If the user wanted us to link to the infile's EOF, round up to the
-	 * next block boundary for this check.
-	 *
-	 * Otherwise, make sure the count is also block-aligned, having
-	 * already confirmed the starting offsets' block alignment.
-	 */
-	if (pos_in + count == size_in) {
-		bcount = ALIGN(size_in, bs) - pos_in;
-	} else {
-		if (!IS_ALIGNED(count, bs))
-			count = ALIGN_DOWN(count, bs);
-		bcount = count;
-	}
-
-	/* Don't allow overlapped cloning within the same file. */
-	if (inode_in == inode_out &&
-	    pos_out + bcount > pos_in &&
-	    pos_out < pos_in + bcount)
-		return -EINVAL;
-
-	/*
-	 * We shortened the request but the caller can't deal with that, so
-	 * bounce the request back to userspace.
-	 */
-	if (*req_count != count && !(remap_flags & REMAP_FILE_CAN_SHORTEN))
-		return -EINVAL;
-
-	*req_count = count;
-	return 0;
-}
-
-static int generic_remap_check_len(struct inode *inode_in,
-				   struct inode *inode_out,
-				   loff_t pos_out,
-				   loff_t *len,
-				   unsigned int remap_flags)
-{
-	u64 blkmask = i_blocksize(inode_in) - 1;
-	loff_t new_len = *len;
-
-	if ((*len & blkmask) == 0)
-		return 0;
-
-	if ((remap_flags & REMAP_FILE_DEDUP) ||
-	    pos_out + *len < i_size_read(inode_out))
-		new_len &= ~blkmask;
-
-	if (new_len == *len)
-		return 0;
-
-	if (remap_flags & REMAP_FILE_CAN_SHORTEN) {
-		*len = new_len;
-		return 0;
-	}
-
-	return (remap_flags & REMAP_FILE_DEDUP) ? -EBADE : -EINVAL;
-}
-
-static int generic_remap_file_range_prep(struct file *file_in, loff_t pos_in,
-				  struct file *file_out, loff_t pos_out,
-				  loff_t *len, unsigned int remap_flags)
-{
-	struct inode *inode_in = file_inode(file_in);
-	struct inode *inode_out = file_inode(file_out);
-	bool same_inode = (inode_in == inode_out);
-	int ret;
-
-	/* Don't touch certain kinds of inodes */
-	if (IS_IMMUTABLE(inode_out))
-		return -EPERM;
-
-	if (IS_SWAPFILE(inode_in) || IS_SWAPFILE(inode_out))
-		return -ETXTBSY;
-
-	/* Don't reflink dirs, pipes, sockets... */
-	if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode))
-		return -EISDIR;
-	if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode))
-		return -EINVAL;
-
-	/* Zero length dedupe exits immediately; reflink goes to EOF. */
-	if (*len == 0) {
-		loff_t isize = i_size_read(inode_in);
-
-		if ((remap_flags & REMAP_FILE_DEDUP) || pos_in == isize)
-			return 0;
-		if (pos_in > isize)
-			return -EINVAL;
-		*len = isize - pos_in;
-		if (*len == 0)
-			return 0;
-	}
-
-	/* Check that we don't violate system file offset limits. */
-	ret = generic_remap_checks(file_in, pos_in, file_out, pos_out, len,
-			remap_flags);
-	if (ret)
-		return ret;
-
-	/* Wait for the completion of any pending IOs on both files */
-	inode_dio_wait(inode_in);
-	if (!same_inode)
-		inode_dio_wait(inode_out);
-
-	ret = filemap_write_and_wait_range(inode_in->i_mapping,
-			pos_in, pos_in + *len - 1);
-	if (ret)
-		return ret;
-
-	ret = filemap_write_and_wait_range(inode_out->i_mapping,
-			pos_out, pos_out + *len - 1);
-	if (ret)
-		return ret;
-
-	/*
-	 * Check that the extents are the same.
-	 */
-	if (remap_flags & REMAP_FILE_DEDUP) {
-		bool		is_same = false;
-
-		ret = vfs_dedupe_file_range_compare(inode_in, pos_in,
-				inode_out, pos_out, *len, &is_same);
-		if (ret)
-			return ret;
-		if (!is_same)
-			return -EBADE;
-	}
-
-	ret = generic_remap_check_len(inode_in, inode_out, pos_out, len,
-			remap_flags);
-	if (ret)
-		return ret;
-
-	/* If can't alter the file contents, we're done. */
-	if (!(remap_flags & REMAP_FILE_DEDUP)) {
-		/* Update the timestamps, since we can alter file contents. */
-		if (!(file_out->f_mode & FMODE_NOCMTIME)) {
-			ret = file_update_time(file_out);
-			if (ret)
-				return ret;
-		}
-
-		/*
-		 * Clear the security bits if the process is not being run by
-		 * root.  This keeps people from modifying setuid and setgid
-		 * binaries.
-		 */
-		ret = file_remove_privs(file_out);
-		if (ret)
-			return ret;
-	}
-
-	return 0;
-}
-
 loff_t bch2_remap_file_range(struct file *file_src, loff_t pos_src,
 			     struct file *file_dst, loff_t pos_dst,
 			     loff_t len, unsigned remap_flags)
@@ -3244,7 +3014,7 @@ static loff_t page_hole_offset(struct address_space *mapping, loff_t offset)
 	loff_t ret = -1;
 
 	page = find_lock_entry(mapping, index);
-	if (!page || radix_tree_exception(page))
+	if (!page || xa_is_value(page))
 		return offset;
 
 	pg_offset = __page_hole_offset(page, offset & (PAGE_SIZE - 1));
diff --git a/fs/bcachefs/fs-io.h b/fs/bcachefs/fs-io.h
index 1b593ea707d5..7063556d289b 100644
--- a/fs/bcachefs/fs-io.h
+++ b/fs/bcachefs/fs-io.h
@@ -35,10 +35,6 @@ int bch2_fsync(struct file *, loff_t, loff_t, int);
 int bch2_truncate(struct bch_inode_info *, struct iattr *);
 long bch2_fallocate_dispatch(struct file *, int, loff_t, loff_t);
 
-#define REMAP_FILE_ADVISORY		(0)
-#define REMAP_FILE_DEDUP		(1 << 0)
-#define REMAP_FILE_CAN_SHORTEN		(1 << 1)
-
 loff_t bch2_remap_file_range(struct file *, loff_t, struct file *,
 			     loff_t, loff_t, unsigned);
 
diff --git a/fs/bcachefs/fs.c b/fs/bcachefs/fs.c
index 13b8bbcdb694..1c89a1b2c2d0 100644
--- a/fs/bcachefs/fs.c
+++ b/fs/bcachefs/fs.c
@@ -142,8 +142,6 @@ retry:
 				  &inode->ei_journal_seq,
 				  BTREE_INSERT_NOUNLOCK|
 				  BTREE_INSERT_NOFAIL);
-	if (ret == -EINTR)
-		goto retry;
 
 	/*
 	 * the btree node lock protects inode->ei_inode, not ei_update_lock;
@@ -152,6 +150,11 @@ retry:
 	if (!ret)
 		bch2_inode_update_after_write(c, inode, &inode_u, fields);
 
+	bch2_trans_iter_put(&trans, iter);
+
+	if (ret == -EINTR)
+		goto retry;
+
 	bch2_trans_exit(&trans);
 	return ret < 0 ? ret : 0;
 }
@@ -963,15 +966,6 @@ static int bch2_vfs_readdir(struct file *file, struct dir_context *ctx)
 	return bch2_readdir(c, inode->v.i_ino, ctx);
 }
 
-static int bch2_clone_file_range(struct file *file_src, loff_t pos_src,
-				 struct file *file_dst, loff_t pos_dst,
-				 u64 len)
-{
-	return bch2_remap_file_range(file_src, pos_src,
-				     file_dst, pos_dst,
-				     len, 0);
-}
-
 static const struct file_operations bch_file_operations = {
 	.llseek		= bch2_llseek,
 	.read_iter	= bch2_read_iter,
@@ -989,7 +983,7 @@ static const struct file_operations bch_file_operations = {
 #ifdef CONFIG_COMPAT
 	.compat_ioctl	= bch2_compat_fs_ioctl,
 #endif
-	.clone_file_range = bch2_clone_file_range,
+	.remap_file_range = bch2_remap_file_range,
 };
 
 static const struct inode_operations bch_file_inode_operations = {
@@ -1520,7 +1514,7 @@ static struct dentry *bch2_mount(struct file_system_type *fs_type,
 
 	sb->s_bdi->congested_fn		= bch2_congested;
 	sb->s_bdi->congested_data	= c;
-	sb->s_bdi->ra_pages		= VM_MAX_READAHEAD * 1024 / PAGE_SIZE;
+	sb->s_bdi->ra_pages		= VM_READAHEAD_PAGES;
 
 	for_each_online_member(ca, c, i) {
 		struct block_device *bdev = ca->disk_sb.bdev;
diff --git a/fs/bcachefs/fsck.c b/fs/bcachefs/fsck.c
index 9ef532d875e8..3ab621c62c43 100644
--- a/fs/bcachefs/fsck.c
+++ b/fs/bcachefs/fsck.c
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: GPL-2.0
 
 #include "bcachefs.h"
+#include "bkey_on_stack.h"
 #include "btree_update.h"
 #include "dirent.h"
 #include "error.h"
@@ -81,7 +82,6 @@ static int remove_dirent(struct btree_trans *trans,
 	return __bch2_trans_do(trans, NULL, NULL,
 			       BTREE_INSERT_NOFAIL|
 			       BTREE_INSERT_LAZY_RW,
-			       TRANS_RESET_MEM,
 			       __remove_dirent(trans, dirent));
 }
 
@@ -182,8 +182,6 @@ static int hash_redo_key(const struct bch_hash_desc desc,
 	struct bkey_i delete;
 	struct bkey_i *tmp;
 
-	bch2_trans_reset(trans, TRANS_RESET_MEM);
-
 	tmp = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 	if (IS_ERR(tmp))
 		return PTR_ERR(tmp);
@@ -194,11 +192,8 @@ static int hash_redo_key(const struct bch_hash_desc desc,
 	delete.k.p = k_iter->pos;
 	bch2_trans_update(trans, k_iter, &delete, 0);
 
-	return  bch2_hash_set(trans, desc, &h->info, k_iter->pos.inode,
-			      tmp, BCH_HASH_SET_MUST_CREATE) ?:
-		bch2_trans_commit(trans, NULL, NULL,
-				  BTREE_INSERT_NOFAIL|
-				  BTREE_INSERT_LAZY_RW);
+	return bch2_hash_set(trans, desc, &h->info, k_iter->pos.inode,
+			     tmp, BCH_HASH_SET_MUST_CREATE);
 }
 
 static int fsck_hash_delete_at(struct btree_trans *trans,
@@ -320,10 +315,9 @@ static int hash_check_key(struct btree_trans *trans,
 			desc.btree_id, k.k->p.offset,
 			hashed, h->chain->pos.offset,
 			(bch2_bkey_val_to_text(&PBUF(buf), c, k), buf))) {
-		do {
-			ret = hash_redo_key(desc, trans, h, k_iter, k, hashed);
-		} while (ret == -EINTR);
-
+		ret = __bch2_trans_do(trans, NULL, NULL,
+				      BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
+			hash_redo_key(desc, trans, h, k_iter, k, hashed));
 		if (ret) {
 			bch_err(c, "hash_redo_key err %i", ret);
 			return ret;
@@ -387,7 +381,6 @@ static int check_dirent_hash(struct btree_trans *trans, struct hash_check *h,
 		ret = __bch2_trans_do(trans, NULL, NULL,
 				      BTREE_INSERT_NOFAIL|
 				      BTREE_INSERT_LAZY_RW,
-				      TRANS_RESET_MEM,
 			(bch2_trans_update(trans, iter, &d->k_i, 0), 0));
 		if (ret)
 			goto err;
@@ -410,11 +403,10 @@ err_redo:
 		     k->k->p.offset, hash, h->chain->pos.offset,
 		     (bch2_bkey_val_to_text(&PBUF(buf), c,
 					    *k), buf))) {
-		do {
-			ret = hash_redo_key(bch2_dirent_hash_desc, trans,
-					    h, iter, *k, hash);
-		} while (ret == -EINTR);
-
+		ret = __bch2_trans_do(trans, NULL, NULL,
+				      BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
+			hash_redo_key(bch2_dirent_hash_desc, trans,
+				      h, iter, *k, hash));
 		if (ret)
 			bch_err(c, "hash_redo_key err %i", ret);
 		else
@@ -431,6 +423,42 @@ static int bch2_inode_truncate(struct bch_fs *c, u64 inode_nr, u64 new_size)
 			POS(inode_nr + 1, 0), NULL);
 }
 
+static int bch2_fix_overlapping_extent(struct btree_trans *trans,
+				       struct btree_iter *iter,
+				       struct bkey_s_c k, struct bpos cut_at)
+{
+	struct btree_iter *u_iter;
+	struct bkey_i *u;
+	int ret;
+
+	u = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+	ret = PTR_ERR_OR_ZERO(u);
+	if (ret)
+		return ret;
+
+	bkey_reassemble(u, k);
+	bch2_cut_front(cut_at, u);
+
+	u_iter = bch2_trans_copy_iter(trans, iter);
+	ret = PTR_ERR_OR_ZERO(u_iter);
+	if (ret)
+		return ret;
+
+	/*
+	 * We don't want to go through the
+	 * extent_handle_overwrites path:
+	 */
+	__bch2_btree_iter_set_pos(u_iter, u->k.p, false);
+
+	/*
+	 * XXX: this is going to leave disk space
+	 * accounting slightly wrong
+	 */
+	ret = bch2_trans_update(trans, u_iter, u, 0);
+	bch2_trans_iter_put(trans, u_iter);
+	return ret;
+}
+
 /*
  * Walk extents: verify that extents have a corresponding S_ISREG inode, and
  * that i_size an i_sectors are consistent
@@ -442,17 +470,40 @@ static int check_extents(struct bch_fs *c)
 	struct btree_trans trans;
 	struct btree_iter *iter;
 	struct bkey_s_c k;
+	struct bkey_on_stack prev;
 	u64 i_sectors;
 	int ret = 0;
 
+	bkey_on_stack_init(&prev);
+	prev.k->k = KEY(0, 0, 0);
 	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
 
 	bch_verbose(c, "checking extents");
 
 	iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS,
-				   POS(BCACHEFS_ROOT_INO, 0), 0);
+				   POS(BCACHEFS_ROOT_INO, 0),
+				   BTREE_ITER_INTENT);
 retry:
 	for_each_btree_key_continue(iter, 0, k, ret) {
+		if (bkey_cmp(prev.k->k.p, bkey_start_pos(k.k)) > 0) {
+			char buf1[200];
+			char buf2[200];
+
+			bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(prev.k));
+			bch2_bkey_val_to_text(&PBUF(buf2), c, k);
+
+			if (fsck_err(c, "overlapping extents:\n%s\n%s", buf1, buf2)) {
+				ret = __bch2_trans_do(&trans, NULL, NULL,
+						      BTREE_INSERT_NOFAIL|
+						      BTREE_INSERT_LAZY_RW,
+						bch2_fix_overlapping_extent(&trans,
+								iter, k, prev.k->k.p));
+				if (ret)
+					goto err;
+			}
+		}
+		bkey_on_stack_reassemble(&prev, c, k);
+
 		ret = walk_inode(&trans, &w, k.k->p.inode);
 		if (ret)
 			break;
@@ -477,7 +528,8 @@ retry:
 			!(w.inode.bi_flags & BCH_INODE_I_SECTORS_DIRTY) &&
 			w.inode.bi_sectors !=
 			(i_sectors = bch2_count_inode_sectors(&trans, w.cur_inum)),
-			c, "i_sectors wrong: got %llu, should be %llu",
+			c, "inode %llu has incorrect i_sectors: got %llu, should be %llu",
+			w.inode.bi_inum,
 			w.inode.bi_sectors, i_sectors)) {
 			struct bkey_inode_buf p;
 
@@ -519,6 +571,7 @@ err:
 fsck_err:
 	if (ret == -EINTR)
 		goto retry;
+	bkey_on_stack_exit(&prev, c);
 	return bch2_trans_exit(&trans) ?: ret;
 }
 
@@ -660,7 +713,6 @@ retry:
 			ret = __bch2_trans_do(&trans, NULL, NULL,
 					      BTREE_INSERT_NOFAIL|
 					      BTREE_INSERT_LAZY_RW,
-					      TRANS_RESET_MEM,
 				(bch2_trans_update(&trans, iter, &n->k_i, 0), 0));
 			kfree(n);
 			if (ret)
@@ -986,12 +1038,12 @@ retry:
 		if (!ret)
 			continue;
 
-		if (fsck_err_on(!inode_bitmap_test(&dirs_done, k.k->p.inode), c,
+		if (fsck_err_on(!inode_bitmap_test(&dirs_done, k.k->p.offset), c,
 				"unreachable directory found (inum %llu)",
-				k.k->p.inode)) {
+				k.k->p.offset)) {
 			bch2_trans_unlock(&trans);
 
-			ret = reattach_inode(c, lostfound_inode, k.k->p.inode);
+			ret = reattach_inode(c, lostfound_inode, k.k->p.offset);
 			if (ret) {
 				goto err;
 			}
@@ -1275,7 +1327,6 @@ static int check_inode(struct btree_trans *trans,
 		ret = __bch2_trans_do(trans, NULL, NULL,
 				      BTREE_INSERT_NOFAIL|
 				      BTREE_INSERT_LAZY_RW,
-				      TRANS_RESET_MEM,
 			(bch2_trans_update(trans, iter, &p.inode.k_i, 0), 0));
 		if (ret)
 			bch_err(c, "error in fsck: error %i "
@@ -1302,18 +1353,18 @@ static int bch2_gc_walk_inodes(struct bch_fs *c,
 	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
 
 	iter = bch2_trans_get_iter(&trans, BTREE_ID_INODES,
-				   POS(range_start, 0), 0);
+				   POS(0, range_start), 0);
 	nlinks_iter = genradix_iter_init(links, 0);
 
 	while ((k = bch2_btree_iter_peek(iter)).k &&
 	       !(ret2 = bkey_err(k))) {
 peek_nlinks:	link = genradix_iter_peek(&nlinks_iter, links);
 
-		if (!link && (!k.k || iter->pos.inode >= range_end))
+		if (!link && (!k.k || iter->pos.offset >= range_end))
 			break;
 
 		nlinks_pos = range_start + nlinks_iter.pos;
-		if (iter->pos.inode > nlinks_pos) {
+		if (iter->pos.offset > nlinks_pos) {
 			/* Should have been caught by dirents pass: */
 			need_fsck_err_on(link && link->count, c,
 				"missing inode %llu (nlink %u)",
@@ -1322,7 +1373,7 @@ peek_nlinks:	link = genradix_iter_peek(&nlinks_iter, links);
 			goto peek_nlinks;
 		}
 
-		if (iter->pos.inode < nlinks_pos || !link)
+		if (iter->pos.offset < nlinks_pos || !link)
 			link = &zero_links;
 
 		if (k.k && k.k->type == KEY_TYPE_inode) {
@@ -1338,7 +1389,7 @@ peek_nlinks:	link = genradix_iter_peek(&nlinks_iter, links);
 				nlinks_pos, link->count);
 		}
 
-		if (nlinks_pos == iter->pos.inode)
+		if (nlinks_pos == iter->pos.offset)
 			genradix_iter_advance(&nlinks_iter, links);
 
 		bch2_btree_iter_next(iter);
diff --git a/fs/bcachefs/inode.c b/fs/bcachefs/inode.c
index e811b98d0f03..7d20f082ad45 100644
--- a/fs/bcachefs/inode.c
+++ b/fs/bcachefs/inode.c
@@ -98,7 +98,7 @@ void bch2_inode_pack(struct bkey_inode_buf *packed,
 	unsigned bytes;
 
 	bkey_inode_init(&packed->inode.k_i);
-	packed->inode.k.p.inode		= inode->bi_inum;
+	packed->inode.k.p.offset	= inode->bi_inum;
 	packed->inode.v.bi_hash_seed	= inode->bi_hash_seed;
 	packed->inode.v.bi_flags	= cpu_to_le32(inode->bi_flags);
 	packed->inode.v.bi_mode		= cpu_to_le16(inode->bi_mode);
@@ -149,7 +149,7 @@ int bch2_inode_unpack(struct bkey_s_c_inode inode,
 	unsigned fieldnr = 0, field_bits;
 	int ret;
 
-	unpacked->bi_inum	= inode.k->p.inode;
+	unpacked->bi_inum	= inode.k->p.offset;
 	unpacked->bi_hash_seed	= inode.v->bi_hash_seed;
 	unpacked->bi_flags	= le32_to_cpu(inode.v->bi_flags);
 	unpacked->bi_mode	= le16_to_cpu(inode.v->bi_mode);
@@ -188,7 +188,7 @@ struct btree_iter *bch2_inode_peek(struct btree_trans *trans,
 	struct bkey_s_c k;
 	int ret;
 
-	iter = bch2_trans_get_iter(trans, BTREE_ID_INODES, POS(inum, 0),
+	iter = bch2_trans_get_iter(trans, BTREE_ID_INODES, POS(0, inum),
 				   BTREE_ITER_SLOTS|flags);
 	if (IS_ERR(iter))
 		return iter;
@@ -232,13 +232,13 @@ const char *bch2_inode_invalid(const struct bch_fs *c, struct bkey_s_c k)
 		struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 		struct bch_inode_unpacked unpacked;
 
-	if (k.k->p.offset)
-		return "nonzero offset";
+	if (k.k->p.inode)
+		return "nonzero k.p.inode";
 
 	if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 		return "incorrect value size";
 
-	if (k.k->p.inode < BLOCKDEV_INODE_MAX)
+	if (k.k->p.offset < BLOCKDEV_INODE_MAX)
 		return "fs inode in blockdev range";
 
 	if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
@@ -280,8 +280,8 @@ void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c,
 const char *bch2_inode_generation_invalid(const struct bch_fs *c,
 					  struct bkey_s_c k)
 {
-	if (k.k->p.offset)
-		return "nonzero offset";
+	if (k.k->p.inode)
+		return "nonzero k.p.inode";
 
 	if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation))
 		return "incorrect value size";
@@ -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(0, start),
+			   BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
+		if (bkey_cmp(iter->pos, POS(0, max)) > 0)
+			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.offset;
+	inode_u->bi_inum	= k.k->p.offset;
+	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)
@@ -454,7 +443,7 @@ int bch2_inode_rm(struct bch_fs *c, u64 inode_nr)
 
 	bch2_trans_init(&trans, c, 0, 0);
 
-	iter = bch2_trans_get_iter(&trans, BTREE_ID_INODES, POS(inode_nr, 0),
+	iter = bch2_trans_get_iter(&trans, BTREE_ID_INODES, POS(0, inode_nr),
 				   BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 	do {
 		struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
@@ -486,10 +475,10 @@ int bch2_inode_rm(struct bch_fs *c, u64 inode_nr)
 
 		if (!bi_generation) {
 			bkey_init(&delete.k);
-			delete.k.p.inode = inode_nr;
+			delete.k.p.offset = inode_nr;
 		} else {
 			bkey_inode_generation_init(&delete.k_i);
-			delete.k.p.inode = inode_nr;
+			delete.k.p.offset = inode_nr;
 			delete.v.bi_generation = cpu_to_le32(bi_generation);
 		}
 
@@ -511,21 +500,20 @@ int bch2_inode_find_by_inum_trans(struct btree_trans *trans, u64 inode_nr,
 	int ret;
 
 	iter = bch2_trans_get_iter(trans, BTREE_ID_INODES,
-			POS(inode_nr, 0), BTREE_ITER_SLOTS);
+			POS(0, inode_nr), BTREE_ITER_SLOTS);
 	if (IS_ERR(iter))
 		return PTR_ERR(iter);
 
 	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/fs/bcachefs/io.c b/fs/bcachefs/io.c
index f2a2c45a02ad..19059702428a 100644
--- a/fs/bcachefs/io.c
+++ b/fs/bcachefs/io.c
@@ -124,10 +124,10 @@ void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw)
 
 void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
 {
+	struct bvec_iter_all iter;
 	struct bio_vec *bv;
-	unsigned i;
 
-	bio_for_each_segment_all(bv, bio, i)
+	bio_for_each_segment_all(bv, bio, iter)
 		if (bv->bv_page != ZERO_PAGE(0))
 			mempool_free(bv->bv_page, &c->bio_bounce_pages);
 	bio->bi_vcnt = 0;
@@ -325,7 +325,7 @@ int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
 			bch2_disk_reservation_init(c, 0);
 		struct bkey_i delete;
 
-		bch2_trans_reset(trans, TRANS_RESET_MEM);
+		bch2_trans_begin(trans);
 
 		ret = bkey_err(k);
 		if (ret)
@@ -399,7 +399,7 @@ int bch2_write_index_default(struct bch_write_op *op)
 				   BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 
 	do {
-		bch2_trans_reset(&trans, TRANS_RESET_MEM);
+		bch2_trans_begin(&trans);
 
 		k = bch2_keylist_front(keys);
 
@@ -546,9 +546,14 @@ static void __bch2_write_index(struct bch_write_op *op)
 	 * particularly want to plumb io_opts all the way through the btree
 	 * update stack right now
 	 */
-	for_each_keylist_key(keys, k)
+	for_each_keylist_key(keys, k) {
 		bch2_rebalance_add_key(c, bkey_i_to_s_c(k), &op->opts);
 
+		if (bch2_bkey_is_incompressible(bkey_i_to_s_c(k)))
+			bch2_check_set_feature(op->c, BCH_FEATURE_incompressible);
+
+	}
+
 	if (!bch2_keylist_empty(keys)) {
 		u64 sectors_start = keylist_sectors(keys);
 		int ret = op->index_update_fn(op);
@@ -784,8 +789,9 @@ static enum prep_encoded_ret {
 	/* Can we just write the entire extent as is? */
 	if (op->crc.uncompressed_size == op->crc.live_size &&
 	    op->crc.compressed_size <= wp->sectors_free &&
-	    op->crc.compression_type == op->compression_type) {
-		if (!op->crc.compression_type &&
+	    (op->crc.compression_type == op->compression_type ||
+	     op->incompressible)) {
+		if (!crc_is_compressed(op->crc) &&
 		    op->csum_type != op->crc.csum_type &&
 		    bch2_write_rechecksum(c, op, op->csum_type))
 			return PREP_ENCODED_CHECKSUM_ERR;
@@ -797,7 +803,7 @@ static enum prep_encoded_ret {
 	 * If the data is compressed and we couldn't write the entire extent as
 	 * is, we have to decompress it:
 	 */
-	if (op->crc.compression_type) {
+	if (crc_is_compressed(op->crc)) {
 		struct bch_csum csum;
 
 		if (bch2_write_decrypt(op))
@@ -864,6 +870,7 @@ static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
 		ret = -EIO;
 		goto err;
 	case PREP_ENCODED_CHECKSUM_ERR:
+		BUG();
 		goto csum_err;
 	case PREP_ENCODED_DO_WRITE:
 		/* XXX look for bug here */
@@ -908,11 +915,13 @@ static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
 		       bch2_csum_type_is_encryption(op->crc.csum_type));
 		BUG_ON(op->compression_type && !bounce);
 
-		crc.compression_type = op->compression_type
-			?  bch2_bio_compress(c, dst, &dst_len, src, &src_len,
-					     op->compression_type)
+		crc.compression_type = op->incompressible
+			? BCH_COMPRESSION_TYPE_incompressible
+			: op->compression_type
+			? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
+					    op->compression_type)
 			: 0;
-		if (!crc.compression_type) {
+		if (!crc_is_compressed(crc)) {
 			dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
 			dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
 
@@ -933,7 +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;
@@ -941,7 +950,7 @@ static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
 		}
 
 		if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
-		    !crc.compression_type &&
+		    !crc_is_compressed(crc) &&
 		    bch2_csum_type_is_encryption(op->crc.csum_type) ==
 		    bch2_csum_type_is_encryption(op->csum_type)) {
 			/*
@@ -1060,6 +1069,12 @@ again:
 					BKEY_EXTENT_U64s_MAX))
 			goto flush_io;
 
+		if ((op->flags & BCH_WRITE_FROM_INTERNAL) &&
+		    percpu_ref_is_dying(&c->writes)) {
+			ret = -EROFS;
+			goto err;
+		}
+
 		wp = bch2_alloc_sectors_start(c,
 			op->target,
 			op->opts.erasure_code,
@@ -1212,7 +1227,8 @@ void bch2_write(struct closure *cl)
 
 	if (c->opts.nochanges ||
 	    !percpu_ref_tryget(&c->writes)) {
-		__bcache_io_error(c, "read only");
+		if (!(op->flags & BCH_WRITE_FROM_INTERNAL))
+			__bcache_io_error(c, "read only");
 		op->error = -EROFS;
 		goto err;
 	}
@@ -1338,6 +1354,7 @@ static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
 
 static struct promote_op *__promote_alloc(struct bch_fs *c,
 					  enum btree_id btree_id,
+					  struct bkey_s_c k,
 					  struct bpos pos,
 					  struct extent_ptr_decoded *pick,
 					  struct bch_io_opts opts,
@@ -1394,8 +1411,7 @@ static struct promote_op *__promote_alloc(struct bch_fs *c,
 			(struct data_opts) {
 				.target = opts.promote_target
 			},
-			btree_id,
-			bkey_s_c_null);
+			btree_id, k);
 	BUG_ON(ret);
 
 	return op;
@@ -1437,7 +1453,7 @@ static struct promote_op *promote_alloc(struct bch_fs *c,
 				  k.k->type == KEY_TYPE_reflink_v
 				  ? BTREE_ID_REFLINK
 				  : BTREE_ID_EXTENTS,
-				  pos, pick, opts, sectors, rbio);
+				  k, pos, pick, opts, sectors, rbio);
 	if (!promote)
 		return NULL;
 
@@ -1690,33 +1706,39 @@ static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
 	}
 }
 
-static void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
+static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
+				   struct bch_read_bio *rbio)
 {
 	struct bch_fs *c = rbio->c;
-	struct btree_trans trans;
-	struct btree_iter *iter;
-	struct bkey_s_c k;
-	struct bkey_on_stack new;
-	struct bch_extent_crc_unpacked new_crc;
 	u64 data_offset = rbio->pos.offset - rbio->pick.crc.offset;
-	int ret;
-
-	if (rbio->pick.crc.compression_type)
-		return;
+	struct bch_extent_crc_unpacked new_crc;
+	struct btree_iter *iter = NULL;
+	struct bkey_i *new;
+	struct bkey_s_c k;
+	int ret = 0;
 
-	bkey_on_stack_init(&new);
-	bch2_trans_init(&trans, c, 0, 0);
-retry:
-	bch2_trans_begin(&trans);
+	if (crc_is_compressed(rbio->pick.crc))
+		return 0;
 
-	iter = bch2_trans_get_iter(&trans, BTREE_ID_EXTENTS, rbio->pos,
+	iter = bch2_trans_get_iter(trans, BTREE_ID_EXTENTS, rbio->pos,
 				   BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
+	if ((ret = PTR_ERR_OR_ZERO(iter)))
+		goto out;
+
 	k = bch2_btree_iter_peek_slot(iter);
-	if (IS_ERR_OR_NULL(k.k))
+	if ((ret = bkey_err(k)))
 		goto out;
 
-	bkey_on_stack_reassemble(&new, c, k);
-	k = bkey_i_to_s_c(new.k);
+	/*
+	 * going to be temporarily appending another checksum entry:
+	 */
+	new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
+				 BKEY_EXTENT_U64s_MAX * 8);
+	if ((ret = PTR_ERR_OR_ZERO(new)))
+		goto out;
+
+	bkey_reassemble(new, k);
+	k = bkey_i_to_s_c(new);
 
 	if (bversion_cmp(k.k->version, rbio->version) ||
 	    !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
@@ -1732,21 +1754,23 @@ retry:
 			bkey_start_offset(k.k) - data_offset, k.k->size,
 			rbio->pick.crc.csum_type)) {
 		bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
+		ret = 0;
 		goto out;
 	}
 
-	if (!bch2_bkey_narrow_crcs(new.k, new_crc))
+	if (!bch2_bkey_narrow_crcs(new, new_crc))
 		goto out;
 
-	bch2_trans_update(&trans, iter, new.k, 0);
-	ret = bch2_trans_commit(&trans, NULL, NULL,
-				BTREE_INSERT_NOFAIL|
-				BTREE_INSERT_NOWAIT);
-	if (ret == -EINTR)
-		goto retry;
+	bch2_trans_update(trans, iter, new, 0);
 out:
-	bch2_trans_exit(&trans);
-	bkey_on_stack_exit(&new, c);
+	bch2_trans_iter_put(trans, iter);
+	return ret;
+}
+
+static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
+{
+	bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL,
+		      __bch2_rbio_narrow_crcs(&trans, rbio));
 }
 
 /* Inner part that may run in process context */
@@ -1786,7 +1810,7 @@ static void __bch2_read_endio(struct work_struct *work)
 	crc.offset     += rbio->offset_into_extent;
 	crc.live_size	= bvec_iter_sectors(rbio->bvec_iter);
 
-	if (crc.compression_type != BCH_COMPRESSION_TYPE_none) {
+	if (crc_is_compressed(crc)) {
 		bch2_encrypt_bio(c, crc.csum_type, nonce, src);
 		if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
 			goto decompression_err;
@@ -1883,7 +1907,7 @@ static void bch2_read_endio(struct bio *bio)
 	}
 
 	if (rbio->narrow_crcs ||
-	    rbio->pick.crc.compression_type ||
+	    crc_is_compressed(rbio->pick.crc) ||
 	    bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
 		context = RBIO_CONTEXT_UNBOUND,	wq = system_unbound_wq;
 	else if (rbio->pick.crc.csum_type)
@@ -1994,7 +2018,7 @@ int __bch2_read_extent(struct bch_fs *c, struct bch_read_bio *orig,
 
 	EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
 
-	if (pick.crc.compression_type != BCH_COMPRESSION_TYPE_none ||
+	if (crc_is_compressed(pick.crc) ||
 	    (pick.crc.csum_type != BCH_CSUM_NONE &&
 	     (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
 	      (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
@@ -2009,7 +2033,7 @@ int __bch2_read_extent(struct bch_fs *c, struct bch_read_bio *orig,
 					&rbio, &bounce, &read_full);
 
 	if (!read_full) {
-		EBUG_ON(pick.crc.compression_type);
+		EBUG_ON(crc_is_compressed(pick.crc));
 		EBUG_ON(pick.crc.csum_type &&
 			(bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
 			 bvec_iter_sectors(iter) != pick.crc.live_size ||
diff --git a/fs/bcachefs/io.h b/fs/bcachefs/io.h
index 45c950942d78..e45dcf9635ae 100644
--- a/fs/bcachefs/io.h
+++ b/fs/bcachefs/io.h
@@ -31,10 +31,11 @@ enum bch_write_flags {
 	BCH_WRITE_ONLY_SPECIFIED_DEVS	= (1 << 6),
 	BCH_WRITE_NOPUT_RESERVATION	= (1 << 7),
 	BCH_WRITE_WROTE_DATA_INLINE	= (1 << 8),
+	BCH_WRITE_FROM_INTERNAL		= (1 << 9),
 
 	/* Internal: */
-	BCH_WRITE_JOURNAL_SEQ_PTR	= (1 << 9),
-	BCH_WRITE_SKIP_CLOSURE_PUT	= (1 << 10),
+	BCH_WRITE_JOURNAL_SEQ_PTR	= (1 << 10),
+	BCH_WRITE_SKIP_CLOSURE_PUT	= (1 << 11),
 };
 
 static inline u64 *op_journal_seq(struct bch_write_op *op)
@@ -78,6 +79,7 @@ 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/fs/bcachefs/io_types.h b/fs/bcachefs/io_types.h
index c37b7d7401e9..684e4c9a5d98 100644
--- a/fs/bcachefs/io_types.h
+++ b/fs/bcachefs/io_types.h
@@ -104,7 +104,8 @@ 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;
 	u16			target;
diff --git a/fs/bcachefs/journal.c b/fs/bcachefs/journal.c
index 9f03a479c9a2..0a4538b3dc60 100644
--- a/fs/bcachefs/journal.c
+++ b/fs/bcachefs/journal.c
@@ -376,7 +376,8 @@ unlock:
 		goto retry;
 
 	if (ret == -ENOSPC) {
-		BUG_ON(!can_discard && (flags & JOURNAL_RES_GET_RESERVED));
+		WARN_ONCE(!can_discard && (flags & JOURNAL_RES_GET_RESERVED),
+			  "JOURNAL_RES_GET_RESERVED set but journal full");
 
 		/*
 		 * Journal is full - can't rely on reclaim from work item due to
diff --git a/fs/bcachefs/journal_io.c b/fs/bcachefs/journal_io.c
index 7112a25d0600..39bb2154cce1 100644
--- a/fs/bcachefs/journal_io.c
+++ b/fs/bcachefs/journal_io.c
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: GPL-2.0
 #include "bcachefs.h"
 #include "alloc_foreground.h"
+#include "btree_io.h"
 #include "buckets.h"
 #include "checksum.h"
 #include "error.h"
@@ -138,7 +139,8 @@ static void journal_entry_null_range(void *start, void *end)
 
 static int journal_validate_key(struct bch_fs *c, struct jset *jset,
 				struct jset_entry *entry,
-				struct bkey_i *k, enum btree_node_type key_type,
+				unsigned level, enum btree_id btree_id,
+				struct bkey_i *k,
 				const char *type, int write)
 {
 	void *next = vstruct_next(entry);
@@ -171,14 +173,13 @@ 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 (!write)
+		bch2_bkey_compat(level, btree_id, version,
+			    JSET_BIG_ENDIAN(jset), write,
+			    NULL, bkey_to_packed(k));
 
-	if (!write &&
-	    version < bcachefs_metadata_version_bkey_renumber)
-		bch2_bkey_renumber(key_type, bkey_to_packed(k), write);
-
-	invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(k), key_type);
+	invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(k),
+				    __btree_node_type(level, btree_id));
 	if (invalid) {
 		char buf[160];
 
@@ -192,9 +193,10 @@ static int journal_validate_key(struct bch_fs *c, struct jset *jset,
 		return 0;
 	}
 
-	if (write &&
-	    version < bcachefs_metadata_version_bkey_renumber)
-		bch2_bkey_renumber(key_type, bkey_to_packed(k), write);
+	if (write)
+		bch2_bkey_compat(level, btree_id, version,
+			    JSET_BIG_ENDIAN(jset), write,
+			    NULL, bkey_to_packed(k));
 fsck_err:
 	return ret;
 }
@@ -207,10 +209,10 @@ static int journal_entry_validate_btree_keys(struct bch_fs *c,
 	struct bkey_i *k;
 
 	vstruct_for_each(entry, k) {
-		int ret = journal_validate_key(c, jset, entry, k,
-				__btree_node_type(entry->level,
-						  entry->btree_id),
-				"key", write);
+		int ret = journal_validate_key(c, jset, entry,
+					       entry->level,
+					       entry->btree_id,
+					       k, "key", write);
 		if (ret)
 			return ret;
 	}
@@ -240,7 +242,7 @@ static int journal_entry_validate_btree_root(struct bch_fs *c,
 		return 0;
 	}
 
-	return journal_validate_key(c, jset, entry, k, BKEY_TYPE_BTREE,
+	return journal_validate_key(c, jset, entry, 1, entry->btree_id, k,
 				    "btree root", write);
 fsck_err:
 	return ret;
@@ -1016,8 +1018,7 @@ void bch2_journal_write(struct closure *cl)
 	if (bch2_csum_type_is_encryption(JSET_CSUM_TYPE(jset)))
 		validate_before_checksum = true;
 
-	if (le32_to_cpu(jset->version) <
-	    bcachefs_metadata_version_bkey_renumber)
+	if (le32_to_cpu(jset->version) < bcachefs_metadata_version_max)
 		validate_before_checksum = true;
 
 	if (validate_before_checksum &&
@@ -1041,9 +1042,16 @@ void bch2_journal_write(struct closure *cl)
 	bytes = vstruct_bytes(jset);
 	memset((void *) jset + bytes, 0, (sectors << 9) - bytes);
 
+retry_alloc:
 	spin_lock(&j->lock);
 	ret = journal_write_alloc(j, w, sectors);
 
+	if (ret && j->can_discard) {
+		spin_unlock(&j->lock);
+		bch2_journal_do_discards(j);
+		goto retry_alloc;
+	}
+
 	/*
 	 * write is allocated, no longer need to account for it in
 	 * bch2_journal_space_available():
diff --git a/fs/bcachefs/journal_reclaim.c b/fs/bcachefs/journal_reclaim.c
index 695b2c8ba03b..db3afd908474 100644
--- a/fs/bcachefs/journal_reclaim.c
+++ b/fs/bcachefs/journal_reclaim.c
@@ -290,38 +290,6 @@ void bch2_journal_pin_put(struct journal *j, u64 seq)
 	}
 }
 
-static inline void __journal_pin_add(struct journal *j,
-				     u64 seq,
-				     struct journal_entry_pin *pin,
-				     journal_pin_flush_fn flush_fn)
-{
-	struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
-
-	BUG_ON(journal_pin_active(pin));
-	BUG_ON(!atomic_read(&pin_list->count));
-
-	atomic_inc(&pin_list->count);
-	pin->seq	= seq;
-	pin->flush	= flush_fn;
-
-	list_add(&pin->list, flush_fn ? &pin_list->list : &pin_list->flushed);
-
-	/*
-	 * If the journal is currently full,  we might want to call flush_fn
-	 * immediately:
-	 */
-	journal_wake(j);
-}
-
-void bch2_journal_pin_add(struct journal *j, u64 seq,
-			  struct journal_entry_pin *pin,
-			  journal_pin_flush_fn flush_fn)
-{
-	spin_lock(&j->lock);
-	__journal_pin_add(j, seq, pin, flush_fn);
-	spin_unlock(&j->lock);
-}
-
 static inline void __journal_pin_drop(struct journal *j,
 				      struct journal_entry_pin *pin)
 {
@@ -354,42 +322,46 @@ void bch2_journal_pin_drop(struct journal *j,
 	spin_unlock(&j->lock);
 }
 
-void bch2_journal_pin_update(struct journal *j, u64 seq,
-			     struct journal_entry_pin *pin,
-			     journal_pin_flush_fn flush_fn)
+void __bch2_journal_pin_add(struct journal *j, u64 seq,
+			    struct journal_entry_pin *pin,
+			    journal_pin_flush_fn flush_fn)
 {
+	struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
+
 	spin_lock(&j->lock);
 
-	if (pin->seq != seq) {
-		__journal_pin_drop(j, pin);
-		__journal_pin_add(j, seq, pin, flush_fn);
-	} else {
-		struct journal_entry_pin_list *pin_list =
-			journal_seq_pin(j, seq);
+	__journal_pin_drop(j, pin);
+
+	BUG_ON(!atomic_read(&pin_list->count));
 
-		list_move(&pin->list, &pin_list->list);
-	}
+	atomic_inc(&pin_list->count);
+	pin->seq	= seq;
+	pin->flush	= flush_fn;
+
+	list_add(&pin->list, flush_fn ? &pin_list->list : &pin_list->flushed);
 
 	spin_unlock(&j->lock);
+
+	/*
+	 * If the journal is currently full,  we might want to call flush_fn
+	 * immediately:
+	 */
+	journal_wake(j);
 }
 
-void bch2_journal_pin_add_if_older(struct journal *j,
-				  struct journal_entry_pin *src_pin,
-				  struct journal_entry_pin *pin,
-				  journal_pin_flush_fn flush_fn)
+void bch2_journal_pin_copy(struct journal *j,
+			   struct journal_entry_pin *dst,
+			   struct journal_entry_pin *src,
+			   journal_pin_flush_fn flush_fn)
 {
-	spin_lock(&j->lock);
-
-	if (journal_pin_active(src_pin) &&
-	    (!journal_pin_active(pin) ||
-	     src_pin->seq < pin->seq)) {
-		__journal_pin_drop(j, pin);
-		__journal_pin_add(j, src_pin->seq, pin, flush_fn);
-	}
-
-	spin_unlock(&j->lock);
+	if (journal_pin_active(src) &&
+	    (!journal_pin_active(dst) || src->seq < dst->seq))
+		__bch2_journal_pin_add(j, src->seq, dst, flush_fn);
 }
 
+/**
+ * bch2_journal_pin_flush: ensure journal pin callback is no longer running
+ */
 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
 {
 	BUG_ON(journal_pin_active(pin));
diff --git a/fs/bcachefs/journal_reclaim.h b/fs/bcachefs/journal_reclaim.h
index 9bf982a17797..883a0a5680af 100644
--- a/fs/bcachefs/journal_reclaim.h
+++ b/fs/bcachefs/journal_reclaim.h
@@ -29,16 +29,24 @@ journal_seq_pin(struct journal *j, u64 seq)
 }
 
 void bch2_journal_pin_put(struct journal *, u64);
-
-void bch2_journal_pin_add(struct journal *, u64, struct journal_entry_pin *,
-			  journal_pin_flush_fn);
-void bch2_journal_pin_update(struct journal *, u64, struct journal_entry_pin *,
-			     journal_pin_flush_fn);
 void bch2_journal_pin_drop(struct journal *, struct journal_entry_pin *);
-void bch2_journal_pin_add_if_older(struct journal *,
-				  struct journal_entry_pin *,
-				  struct journal_entry_pin *,
-				  journal_pin_flush_fn);
+
+void __bch2_journal_pin_add(struct journal *, u64, struct journal_entry_pin *,
+			    journal_pin_flush_fn);
+
+static inline void bch2_journal_pin_add(struct journal *j, u64 seq,
+					struct journal_entry_pin *pin,
+					journal_pin_flush_fn flush_fn)
+{
+	if (unlikely(!journal_pin_active(pin)))
+		__bch2_journal_pin_add(j, seq, pin, flush_fn);
+}
+
+void bch2_journal_pin_copy(struct journal *,
+			   struct journal_entry_pin *,
+			   struct journal_entry_pin *,
+			   journal_pin_flush_fn);
+
 void bch2_journal_pin_flush(struct journal *, struct journal_entry_pin *);
 
 void bch2_journal_do_discards(struct journal *);
diff --git a/fs/bcachefs/migrate.c b/fs/bcachefs/migrate.c
index 1ef62a189e33..e26fa1608f39 100644
--- a/fs/bcachefs/migrate.c
+++ b/fs/bcachefs/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/fs/bcachefs/move.c b/fs/bcachefs/move.c
index 20885b605b50..4afda95f4017 100644
--- a/fs/bcachefs/move.c
+++ b/fs/bcachefs/move.c
@@ -215,6 +215,9 @@ int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 			    enum btree_id btree_id,
 			    struct bkey_s_c k)
 {
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
+	struct extent_ptr_decoded p;
 	int ret;
 
 	m->btree_id	= btree_id;
@@ -223,9 +226,14 @@ int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 	m->nr_ptrs_reserved = 0;
 
 	bch2_write_op_init(&m->op, c, io_opts);
-	m->op.compression_type =
-		bch2_compression_opt_to_type[io_opts.background_compression ?:
-					     io_opts.compression];
+
+	if (!bch2_bkey_is_incompressible(k))
+		m->op.compression_type =
+			bch2_compression_opt_to_type[io_opts.background_compression ?:
+						     io_opts.compression];
+	else
+		m->op.incompressible = true;
+
 	m->op.target	= data_opts.target,
 	m->op.write_point = wp;
 
@@ -235,7 +243,8 @@ int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 	m->op.flags |= BCH_WRITE_ONLY_SPECIFIED_DEVS|
 		BCH_WRITE_PAGES_STABLE|
 		BCH_WRITE_PAGES_OWNED|
-		BCH_WRITE_DATA_ENCODED;
+		BCH_WRITE_DATA_ENCODED|
+		BCH_WRITE_FROM_INTERNAL;
 
 	m->op.nr_replicas	= 1;
 	m->op.nr_replicas_required = 1;
@@ -265,14 +274,11 @@ int bch2_migrate_write_init(struct bch_fs *c, struct migrate_write *m,
 		break;
 	}
 	case DATA_REWRITE: {
-		struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-		const union bch_extent_entry *entry;
-		struct extent_ptr_decoded p;
 		unsigned compressed_sectors = 0;
 
 		bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
 			if (!p.ptr.cached &&
-			    p.crc.compression_type != BCH_COMPRESSION_TYPE_none &&
+			    crc_is_compressed(p.crc) &&
 			    bch2_dev_in_target(c, p.ptr.dev, data_opts.target))
 				compressed_sectors += p.crc.compressed_size;
 
@@ -300,12 +306,12 @@ static void move_free(struct closure *cl)
 {
 	struct moving_io *io = container_of(cl, struct moving_io, cl);
 	struct moving_context *ctxt = io->write.ctxt;
+	struct bvec_iter_all iter;
 	struct bio_vec *bv;
-	unsigned i;
 
 	bch2_disk_reservation_put(io->write.op.c, &io->write.op.res);
 
-	bio_for_each_segment_all(bv, &io->write.op.wbio.bio, i)
+	bio_for_each_segment_all(bv, &io->write.op.wbio.bio, iter)
 		if (bv->bv_page)
 			__free_page(bv->bv_page);
 
diff --git a/fs/bcachefs/opts.h b/fs/bcachefs/opts.h
index 1c05effa71e6..ba4903352343 100644
--- a/fs/bcachefs/opts.h
+++ b/fs/bcachefs/opts.h
@@ -255,6 +255,11 @@ enum opt_type {
 	  OPT_BOOL(),							\
 	  NO_SB_OPT,			false,				\
 	  NULL,		"Don't replay the journal")			\
+	x(keep_journal,			u8,				\
+	  OPT_MOUNT,							\
+	  OPT_BOOL(),							\
+	  NO_SB_OPT,			false,				\
+	  NULL,		"Don't free journal entries/keys after startup")\
 	x(noexcl,			u8,				\
 	  OPT_MOUNT,							\
 	  OPT_BOOL(),							\
diff --git a/fs/bcachefs/rebalance.c b/fs/bcachefs/rebalance.c
index 612385e9d4e4..ab1934325948 100644
--- a/fs/bcachefs/rebalance.c
+++ b/fs/bcachefs/rebalance.c
@@ -17,50 +17,52 @@
 #include <linux/sched/cputime.h>
 #include <trace/events/bcachefs.h>
 
-static inline bool rebalance_ptr_pred(struct bch_fs *c,
-				      struct extent_ptr_decoded p,
-				      struct bch_io_opts *io_opts)
+/*
+ * Check if an extent should be moved:
+ * returns -1 if it should not be moved, or
+ * device of pointer that should be moved, if known, or INT_MAX if unknown
+ */
+static int __bch2_rebalance_pred(struct bch_fs *c,
+				 struct bkey_s_c k,
+				 struct bch_io_opts *io_opts)
 {
-	if (io_opts->background_target &&
-	    !bch2_dev_in_target(c, p.ptr.dev, io_opts->background_target) &&
-	    !p.ptr.cached)
-		return true;
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
+	const union bch_extent_entry *entry;
+	struct extent_ptr_decoded p;
 
 	if (io_opts->background_compression &&
-	    p.crc.compression_type !=
-	    bch2_compression_opt_to_type[io_opts->background_compression])
-		return true;
-
-	return false;
+	    !bch2_bkey_is_incompressible(k))
+		bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
+			if (!p.ptr.cached &&
+			    p.crc.compression_type !=
+			    bch2_compression_opt_to_type[io_opts->background_compression])
+				return p.ptr.dev;
+
+	if (io_opts->background_target)
+		bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
+			if (!p.ptr.cached &&
+			    !bch2_dev_in_target(c, p.ptr.dev, io_opts->background_target))
+				return p.ptr.dev;
+
+	return -1;
 }
 
 void bch2_rebalance_add_key(struct bch_fs *c,
 			    struct bkey_s_c k,
 			    struct bch_io_opts *io_opts)
 {
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const union bch_extent_entry *entry;
-	struct extent_ptr_decoded p;
+	atomic64_t *counter;
+	int dev;
 
-	if (!io_opts->background_target &&
-	    !io_opts->background_compression)
+	dev = __bch2_rebalance_pred(c, k, io_opts);
+	if (dev < 0)
 		return;
 
-	bkey_for_each_ptr_decode(k.k, ptrs, p, entry)
-		if (rebalance_ptr_pred(c, p, io_opts)) {
-			struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
+	counter = dev < INT_MAX
+		? &bch_dev_bkey_exists(c, dev)->rebalance_work
+		: &c->rebalance.work_unknown_dev;
 
-			if (atomic64_add_return(p.crc.compressed_size,
-						&ca->rebalance_work) ==
-			    p.crc.compressed_size)
-				rebalance_wakeup(c);
-		}
-}
-
-void bch2_rebalance_add_work(struct bch_fs *c, u64 sectors)
-{
-	if (atomic64_add_return(sectors, &c->rebalance.work_unknown_dev) ==
-	    sectors)
+	if (atomic64_add_return(k.k->size, counter) == k.k->size)
 		rebalance_wakeup(c);
 }
 
@@ -69,26 +71,20 @@ static enum data_cmd rebalance_pred(struct bch_fs *c, void *arg,
 				    struct bch_io_opts *io_opts,
 				    struct data_opts *data_opts)
 {
-	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
-	const union bch_extent_entry *entry;
-	struct extent_ptr_decoded p;
-	unsigned nr_replicas = 0;
-
-	bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
-		nr_replicas += !p.ptr.cached;
-
-		if (rebalance_ptr_pred(c, p, io_opts))
-			goto found;
+	if (__bch2_rebalance_pred(c, k, io_opts) >= 0) {
+		data_opts->target		= io_opts->background_target;
+		data_opts->btree_insert_flags	= 0;
+		return DATA_ADD_REPLICAS;
+	} else {
+		return DATA_SKIP;
 	}
+}
 
-	if (nr_replicas < io_opts->data_replicas)
-		goto found;
-
-	return DATA_SKIP;
-found:
-	data_opts->target		= io_opts->background_target;
-	data_opts->btree_insert_flags	= 0;
-	return DATA_ADD_REPLICAS;
+void bch2_rebalance_add_work(struct bch_fs *c, u64 sectors)
+{
+	if (atomic64_add_return(sectors, &c->rebalance.work_unknown_dev) ==
+	    sectors)
+		rebalance_wakeup(c);
 }
 
 struct rebalance_work {
diff --git a/fs/bcachefs/recovery.c b/fs/bcachefs/recovery.c
index 8ecd4abc8eeb..a4d0eec2ea3e 100644
--- a/fs/bcachefs/recovery.c
+++ b/fs/bcachefs/recovery.c
@@ -27,43 +27,173 @@
 
 /* iterate over keys read from the journal: */
 
-struct journal_iter bch2_journal_iter_init(struct journal_keys *keys,
-					   enum btree_id id)
+static struct journal_key *journal_key_search(struct journal_keys *journal_keys,
+					      enum btree_id id, unsigned level,
+					      struct bpos pos)
 {
-	return (struct journal_iter) {
-		.keys		= keys,
-		.k		= keys->d,
-		.btree_id	= id,
-	};
+	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) ?:
+		     cmp_int(level,	journal_keys->d[m].level) ?:
+		     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) ?:
+		cmp_int(level,	journal_keys->d[l].level) ?:
+		bkey_cmp(pos,	journal_keys->d[l].k->k.p)) > 0);
+
+	BUG_ON(l &&
+	       (cmp_int(id,	journal_keys->d[l - 1].btree_id) ?:
+		cmp_int(level,	journal_keys->d[l - 1].level) ?:
+		bkey_cmp(pos,	journal_keys->d[l - 1].k->k.p)) <= 0);
+
+	return l < journal_keys->nr ? journal_keys->d + l : NULL;
+}
+
+static struct bkey_i *bch2_journal_iter_peek(struct journal_iter *iter)
+{
+	if (iter->k &&
+	    iter->k < iter->keys->d + iter->keys->nr &&
+	    iter->k->btree_id	== iter->btree_id &&
+	    iter->k->level	== iter->level)
+		return iter->k->k;
+
+	iter->k = NULL;
+	return NULL;
 }
 
-struct bkey_s_c bch2_journal_iter_peek(struct journal_iter *iter)
+static void bch2_journal_iter_advance(struct journal_iter *iter)
 {
+	if (iter->k)
+		iter->k++;
+}
+
+static void bch2_journal_iter_init(struct journal_iter *iter,
+				   struct journal_keys *journal_keys,
+				   enum btree_id id, unsigned level,
+				   struct bpos pos)
+{
+	iter->btree_id	= id;
+	iter->level	= level;
+	iter->keys	= journal_keys;
+	iter->k		= journal_key_search(journal_keys, id, level, pos);
+}
+
+static struct bkey_s_c bch2_journal_iter_peek_btree(struct btree_and_journal_iter *iter)
+{
+	return iter->btree
+		? bch2_btree_iter_peek(iter->btree)
+		: bch2_btree_node_iter_peek_unpack(&iter->node_iter,
+						   iter->b, &iter->unpacked);
+}
+
+static void bch2_journal_iter_advance_btree(struct btree_and_journal_iter *iter)
+{
+	if (iter->btree)
+		bch2_btree_iter_next(iter->btree);
+	else
+		bch2_btree_node_iter_advance(&iter->node_iter, iter->b);
+}
+
+void bch2_btree_and_journal_iter_advance(struct btree_and_journal_iter *iter)
+{
+	switch (iter->last) {
+	case none:
+		break;
+	case btree:
+		bch2_journal_iter_advance_btree(iter);
+		break;
+	case journal:
+		bch2_journal_iter_advance(&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) {
-		if (iter->k == iter->keys->d + iter->keys->nr)
+		struct bkey_s_c btree_k		=
+			bch2_journal_iter_peek_btree(iter);
+		struct bkey_s_c journal_k	=
+			bkey_i_to_s_c(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_journal_iter_advance_btree(iter);
+
+			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;
+		}
 
-		if (iter->k->btree_id == iter->btree_id)
-			return bkey_i_to_s_c(iter->k->k);
+		ret = iter->last == journal ? journal_k : btree_k;
 
-		iter->k++;
+		if (iter->b &&
+		    bkey_cmp(ret.k->p, iter->b->data->max_key) > 0) {
+			iter->journal.k = NULL;
+			iter->last = none;
+			return bkey_s_c_null;
+		}
+
+		if (!bkey_deleted(ret.k))
+			break;
+
+		bch2_btree_and_journal_iter_advance(iter);
 	}
 
-	return bkey_s_c_null;
+	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);
+}
+
+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)
+{
+	memset(iter, 0, sizeof(*iter));
+
+	iter->btree = bch2_trans_get_iter(trans, id, pos, 0);
+	bch2_journal_iter_init(&iter->journal, journal_keys, id, 0, pos);
 }
 
-struct bkey_s_c bch2_journal_iter_next(struct journal_iter *iter)
+void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *iter,
+						struct journal_keys *journal_keys,
+						struct btree *b)
 {
-	if (iter->k == iter->keys->d + iter->keys->nr)
-		return bkey_s_c_null;
+	memset(iter, 0, sizeof(*iter));
 
-	iter->k++;
-	return bch2_journal_iter_peek(iter);
+	iter->b = b;
+	bch2_btree_node_iter_init_from_start(&iter->node_iter, iter->b);
+	bch2_journal_iter_init(&iter->journal, journal_keys,
+			       b->btree_id, b->level, b->data->min_key);
 }
 
 /* sort and dedup all keys in the journal: */
 
-static void journal_entries_free(struct list_head *list)
+void bch2_journal_entries_free(struct list_head *list)
 {
 
 	while (!list_empty(list)) {
@@ -75,13 +205,17 @@ static void journal_entries_free(struct list_head *list)
 	}
 }
 
+/*
+ * When keys compare equal, oldest compares first:
+ */
 static int journal_sort_key_cmp(const void *_l, const void *_r)
 {
 	const struct journal_key *l = _l;
 	const struct journal_key *r = _r;
 
-	return cmp_int(l->btree_id, r->btree_id) ?:
-		bkey_cmp(l->pos, r->pos) ?:
+	return  cmp_int(l->btree_id,	r->btree_id) ?:
+		cmp_int(l->level,	r->level) ?:
+		bkey_cmp(l->k->k.p, r->k->k.p) ?:
 		cmp_int(l->journal_seq, r->journal_seq) ?:
 		cmp_int(l->journal_offset, r->journal_offset);
 }
@@ -91,27 +225,14 @@ static int journal_sort_seq_cmp(const void *_l, const void *_r)
 	const struct journal_key *l = _l;
 	const struct journal_key *r = _r;
 
-	return cmp_int(l->journal_seq, r->journal_seq) ?:
-		cmp_int(l->btree_id, r->btree_id) ?:
-		bkey_cmp(l->pos, r->pos);
-}
-
-static void journal_keys_sift(struct journal_keys *keys, struct journal_key *i)
-{
-	while (i + 1 < keys->d + keys->nr &&
-	       journal_sort_key_cmp(i, i + 1) > 0) {
-		swap(i[0], i[1]);
-		i++;
-	}
+	return  cmp_int(r->level,	l->level) ?:
+		cmp_int(l->journal_seq, r->journal_seq) ?:
+		cmp_int(l->btree_id,	r->btree_id) ?:
+		bkey_cmp(l->k->k.p,	r->k->k.p);
 }
 
-static void journal_keys_free(struct journal_keys *keys)
+void bch2_journal_keys_free(struct journal_keys *keys)
 {
-	struct journal_key *i;
-
-	for_each_journal_key(*keys, i)
-		if (i->allocated)
-			kfree(i->k);
 	kvfree(keys->d);
 	keys->d = NULL;
 	keys->nr = 0;
@@ -122,15 +243,15 @@ static struct journal_keys journal_keys_sort(struct list_head *journal_entries)
 	struct journal_replay *p;
 	struct jset_entry *entry;
 	struct bkey_i *k, *_n;
-	struct journal_keys keys = { NULL }, keys_deduped = { NULL };
-	struct journal_key *i;
+	struct journal_keys keys = { NULL };
+	struct journal_key *src, *dst;
 	size_t nr_keys = 0;
 
 	list_for_each_entry(p, journal_entries, list)
 		for_each_jset_key(k, _n, entry, &p->j)
 			nr_keys++;
 
-	keys.journal_seq_base = keys_deduped.journal_seq_base =
+	keys.journal_seq_base =
 		le64_to_cpu(list_first_entry(journal_entries,
 					     struct journal_replay,
 					     list)->j.seq);
@@ -139,91 +260,33 @@ static struct journal_keys journal_keys_sort(struct list_head *journal_entries)
 	if (!keys.d)
 		goto err;
 
-	keys_deduped.d = kvmalloc(sizeof(keys.d[0]) * nr_keys * 2, GFP_KERNEL);
-	if (!keys_deduped.d)
-		goto err;
-
 	list_for_each_entry(p, journal_entries, list)
 		for_each_jset_key(k, _n, entry, &p->j)
 			keys.d[keys.nr++] = (struct journal_key) {
 				.btree_id	= entry->btree_id,
-				.pos		= bkey_start_pos(&k->k),
+				.level		= entry->level,
 				.k		= k,
 				.journal_seq	= le64_to_cpu(p->j.seq) -
 					keys.journal_seq_base,
 				.journal_offset	= k->_data - p->j._data,
 			};
 
-	sort(keys.d, nr_keys, sizeof(keys.d[0]), journal_sort_key_cmp, NULL);
-
-	i = keys.d;
-	while (i < keys.d + keys.nr) {
-		if (i + 1 < keys.d + keys.nr &&
-		    i[0].btree_id == i[1].btree_id &&
-		    !bkey_cmp(i[0].pos, i[1].pos)) {
-			if (bkey_cmp(i[0].k->k.p, i[1].k->k.p) <= 0) {
-				i++;
-			} else {
-				bch2_cut_front(i[1].k->k.p, i[0].k);
-				i[0].pos = i[1].k->k.p;
-				journal_keys_sift(&keys, i);
-			}
-			continue;
-		}
+	sort(keys.d, keys.nr, sizeof(keys.d[0]), journal_sort_key_cmp, NULL);
 
-		if (i + 1 < keys.d + keys.nr &&
-		    i[0].btree_id == i[1].btree_id &&
-		    bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)) > 0) {
-			if ((cmp_int(i[0].journal_seq, i[1].journal_seq) ?:
-			     cmp_int(i[0].journal_offset, i[1].journal_offset)) < 0) {
-				if (bkey_cmp(i[0].k->k.p, i[1].k->k.p) <= 0) {
-					bch2_cut_back(bkey_start_pos(&i[1].k->k), i[0].k);
-				} else {
-					struct bkey_i *split =
-						kmalloc(bkey_bytes(i[0].k), GFP_KERNEL);
-
-					if (!split)
-						goto err;
-
-					bkey_copy(split, i[0].k);
-					bch2_cut_back(bkey_start_pos(&i[1].k->k), split);
-					keys_deduped.d[keys_deduped.nr++] = (struct journal_key) {
-						.btree_id	= i[0].btree_id,
-						.allocated	= true,
-						.pos		= bkey_start_pos(&split->k),
-						.k		= split,
-						.journal_seq	= i[0].journal_seq,
-						.journal_offset	= i[0].journal_offset,
-					};
-
-					bch2_cut_front(i[1].k->k.p, i[0].k);
-					i[0].pos = i[1].k->k.p;
-					journal_keys_sift(&keys, i);
-					continue;
-				}
-			} else {
-				if (bkey_cmp(i[0].k->k.p, i[1].k->k.p) >= 0) {
-					i[1] = i[0];
-					i++;
-					continue;
-				} else {
-					bch2_cut_front(i[0].k->k.p, i[1].k);
-					i[1].pos = i[0].k->k.p;
-					journal_keys_sift(&keys, i + 1);
-					continue;
-				}
-			}
-		}
+	src = dst = keys.d;
+	while (src < keys.d + keys.nr) {
+		while (src + 1 < keys.d + keys.nr &&
+		       src[0].btree_id	== src[1].btree_id &&
+		       src[0].level	== src[1].level &&
+		       !bkey_cmp(src[0].k->k.p, src[1].k->k.p))
+			src++;
 
-		keys_deduped.d[keys_deduped.nr++] = *i++;
+		*dst++ = *src++;
 	}
 
-	kvfree(keys.d);
-	return keys_deduped;
+	keys.nr = dst - keys.d;
 err:
-	journal_keys_free(&keys_deduped);
-	kvfree(keys.d);
-	return (struct journal_keys) { NULL };
+	return keys;
 }
 
 /* journal replay: */
@@ -274,11 +337,6 @@ retry:
 
 		atomic_end = bpos_min(k->k.p, iter->l[0].b->key.k.p);
 
-		split_iter = bch2_trans_copy_iter(&trans, iter);
-		ret = PTR_ERR_OR_ZERO(split_iter);
-		if (ret)
-			goto err;
-
 		split = bch2_trans_kmalloc(&trans, bkey_bytes(&k->k));
 		ret = PTR_ERR_OR_ZERO(split);
 		if (ret)
@@ -297,12 +355,25 @@ retry:
 		}
 
 		bkey_copy(split, k);
-		bch2_cut_front(split_iter->pos, split);
+		bch2_cut_front(iter->pos, split);
 		bch2_cut_back(atomic_end, split);
 
+		split_iter = bch2_trans_copy_iter(&trans, iter);
+		ret = PTR_ERR_OR_ZERO(split_iter);
+		if (ret)
+			goto err;
+
+		/*
+		 * It's important that we don't go through the
+		 * extent_handle_overwrites() and extent_update_to_keys() path
+		 * here: journal replay is supposed to treat extents like
+		 * regular keys
+		 */
+		__bch2_btree_iter_set_pos(split_iter, split->k.p, false);
 		bch2_trans_update(&trans, split_iter, split, !remark
 				  ? BTREE_TRIGGER_NORUN
 				  : BTREE_TRIGGER_NOOVERWRITES);
+
 		bch2_btree_iter_set_pos(iter, split->k.p);
 	} while (bkey_cmp(iter->pos, k->k.p) < 0);
 
@@ -328,27 +399,40 @@ err:
 }
 
 static int __bch2_journal_replay_key(struct btree_trans *trans,
-				     enum btree_id id, struct bkey_i *k)
+				     enum btree_id id, unsigned level,
+				     struct bkey_i *k)
 {
 	struct btree_iter *iter;
+	int ret;
 
-	iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
-				   BTREE_ITER_INTENT);
+	iter = bch2_trans_get_node_iter(trans, id, k->k.p,
+					BTREE_MAX_DEPTH, level,
+					BTREE_ITER_INTENT);
 	if (IS_ERR(iter))
 		return PTR_ERR(iter);
 
-	bch2_trans_update(trans, iter, k, BTREE_TRIGGER_NORUN);
-	return 0;
+	/*
+	 * iter->flags & BTREE_ITER_IS_EXTENTS triggers the update path to run
+	 * extent_handle_overwrites() and extent_update_to_keys() - but we don't
+	 * want that here, journal replay is supposed to treat extents like
+	 * regular keys:
+	 */
+	__bch2_btree_iter_set_pos(iter, k->k.p, false);
+
+	ret   = bch2_btree_iter_traverse(iter) ?:
+		bch2_trans_update(trans, iter, k, BTREE_TRIGGER_NORUN);
+	bch2_trans_iter_put(trans, iter);
+	return ret;
 }
 
 static int bch2_journal_replay_key(struct bch_fs *c, enum btree_id id,
-				   struct bkey_i *k)
+				   unsigned level, struct bkey_i *k)
 {
 	return bch2_trans_do(c, NULL, NULL,
 			     BTREE_INSERT_NOFAIL|
 			     BTREE_INSERT_LAZY_RW|
 			     BTREE_INSERT_JOURNAL_REPLAY,
-			     __bch2_journal_replay_key(&trans, id, k));
+			     __bch2_journal_replay_key(&trans, id, level, k));
 }
 
 static int bch2_journal_replay(struct bch_fs *c,
@@ -360,15 +444,21 @@ static int bch2_journal_replay(struct bch_fs *c,
 
 	sort(keys.d, keys.nr, sizeof(keys.d[0]), journal_sort_seq_cmp, NULL);
 
+	if (keys.nr)
+		replay_now_at(j, keys.journal_seq_base);
+
 	for_each_journal_key(keys, i) {
-		replay_now_at(j, keys.journal_seq_base + i->journal_seq);
+		if (!i->level)
+			replay_now_at(j, keys.journal_seq_base + i->journal_seq);
 
+		if (i->level)
+			ret = bch2_journal_replay_key(c, i->btree_id, i->level, i->k);
 		if (i->btree_id == BTREE_ID_ALLOC)
 			ret = bch2_alloc_replay_key(c, i->k);
-		else if (btree_node_type_is_extents(i->btree_id))
+		else if (i->k->k.size)
 			ret = bch2_extent_replay_key(c, i->btree_id, i->k);
 		else
-			ret = bch2_journal_replay_key(c, i->btree_id, i->k);
+			ret = bch2_journal_replay_key(c, i->btree_id, i->level, i->k);
 
 		if (ret) {
 			bch_err(c, "journal replay: error %d while replaying key",
@@ -707,8 +797,6 @@ int bch2_fs_recovery(struct bch_fs *c)
 	const char *err = "cannot allocate memory";
 	struct bch_sb_field_clean *clean = NULL;
 	u64 journal_seq;
-	LIST_HEAD(journal_entries);
-	struct journal_keys journal_keys = { NULL };
 	bool wrote = false, write_sb = false;
 	int ret;
 
@@ -727,33 +815,33 @@ int bch2_fs_recovery(struct bch_fs *c)
 		set_bit(BCH_FS_REBUILD_REPLICAS, &c->flags);
 	}
 
-	if (!c->sb.clean || c->opts.fsck) {
+	if (!c->sb.clean || c->opts.fsck || c->opts.keep_journal) {
 		struct jset *j;
 
-		ret = bch2_journal_read(c, &journal_entries);
+		ret = bch2_journal_read(c, &c->journal_entries);
 		if (ret)
 			goto err;
 
-		if (mustfix_fsck_err_on(c->sb.clean && !journal_empty(&journal_entries), c,
+		if (mustfix_fsck_err_on(c->sb.clean && !journal_empty(&c->journal_entries), c,
 				"filesystem marked clean but journal not empty")) {
 			c->sb.compat &= ~(1ULL << BCH_COMPAT_FEAT_ALLOC_INFO);
 			SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
 			c->sb.clean = false;
 		}
 
-		if (!c->sb.clean && list_empty(&journal_entries)) {
+		if (!c->sb.clean && list_empty(&c->journal_entries)) {
 			bch_err(c, "no journal entries found");
 			ret = BCH_FSCK_REPAIR_IMPOSSIBLE;
 			goto err;
 		}
 
-		journal_keys = journal_keys_sort(&journal_entries);
-		if (!journal_keys.d) {
+		c->journal_keys = journal_keys_sort(&c->journal_entries);
+		if (!c->journal_keys.d) {
 			ret = -ENOMEM;
 			goto err;
 		}
 
-		j = &list_last_entry(&journal_entries,
+		j = &list_last_entry(&c->journal_entries,
 				     struct journal_replay, list)->j;
 
 		ret = verify_superblock_clean(c, &clean, j);
@@ -765,7 +853,14 @@ int bch2_fs_recovery(struct bch_fs *c)
 		journal_seq = le64_to_cpu(clean->journal_seq) + 1;
 	}
 
-	ret = journal_replay_early(c, clean, &journal_entries);
+	if (!c->sb.clean &&
+	    !(c->sb.features & (1ULL << BCH_FEATURE_extents_above_btree_updates))) {
+		bch_err(c, "filesystem needs recovery from older version; run fsck from older bcachefs-tools to fix");
+		ret = -EINVAL;
+		goto err;
+	}
+
+	ret = journal_replay_early(c, clean, &c->journal_entries);
 	if (ret)
 		goto err;
 
@@ -783,15 +878,15 @@ int bch2_fs_recovery(struct bch_fs *c)
 
 	ret = bch2_blacklist_table_initialize(c);
 
-	if (!list_empty(&journal_entries)) {
+	if (!list_empty(&c->journal_entries)) {
 		ret = verify_journal_entries_not_blacklisted_or_missing(c,
-							&journal_entries);
+							&c->journal_entries);
 		if (ret)
 			goto err;
 	}
 
 	ret = bch2_fs_journal_start(&c->journal, journal_seq,
-				    &journal_entries);
+				    &c->journal_entries);
 	if (ret)
 		goto err;
 
@@ -801,14 +896,14 @@ int bch2_fs_recovery(struct bch_fs *c)
 
 	bch_verbose(c, "starting alloc read");
 	err = "error reading allocation information";
-	ret = bch2_alloc_read(c, &journal_keys);
+	ret = bch2_alloc_read(c, &c->journal_keys);
 	if (ret)
 		goto err;
 	bch_verbose(c, "alloc read done");
 
 	bch_verbose(c, "starting stripes_read");
 	err = "error reading stripes";
-	ret = bch2_stripes_read(c, &journal_keys);
+	ret = bch2_stripes_read(c, &c->journal_keys);
 	if (ret)
 		goto err;
 	bch_verbose(c, "stripes_read done");
@@ -824,7 +919,7 @@ int bch2_fs_recovery(struct bch_fs *c)
 		 */
 		bch_info(c, "starting metadata mark and sweep");
 		err = "error in mark and sweep";
-		ret = bch2_gc(c, NULL, true, true);
+		ret = bch2_gc(c, &c->journal_keys, true, true);
 		if (ret)
 			goto err;
 		bch_verbose(c, "mark and sweep done");
@@ -835,7 +930,7 @@ int bch2_fs_recovery(struct bch_fs *c)
 	    test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags)) {
 		bch_info(c, "starting mark and sweep");
 		err = "error in mark and sweep";
-		ret = bch2_gc(c, &journal_keys, true, false);
+		ret = bch2_gc(c, &c->journal_keys, true, false);
 		if (ret)
 			goto err;
 		bch_verbose(c, "mark and sweep done");
@@ -856,7 +951,7 @@ int bch2_fs_recovery(struct bch_fs *c)
 
 	bch_verbose(c, "starting journal replay");
 	err = "journal replay failed";
-	ret = bch2_journal_replay(c, journal_keys);
+	ret = bch2_journal_replay(c, c->journal_keys);
 	if (ret)
 		goto err;
 	bch_verbose(c, "journal replay done");
@@ -922,8 +1017,7 @@ int bch2_fs_recovery(struct bch_fs *c)
 			c->disk_sb.sb->version_min =
 				le16_to_cpu(bcachefs_metadata_version_min);
 		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] |= BCH_SB_FEATURES_ALL;
 		write_sb = true;
 	}
 
@@ -953,8 +1047,10 @@ fsck_err:
 	set_bit(BCH_FS_FSCK_DONE, &c->flags);
 	bch2_flush_fsck_errs(c);
 
-	journal_keys_free(&journal_keys);
-	journal_entries_free(&journal_entries);
+	if (!c->opts.keep_journal) {
+		bch2_journal_keys_free(&c->journal_keys);
+		bch2_journal_entries_free(&c->journal_entries);
+	}
 	kfree(clean);
 	if (ret)
 		bch_err(c, "Error in recovery: %s (%i)", err, ret);
@@ -1042,8 +1138,7 @@ int bch2_fs_initialize(struct bch_fs *c)
 	c->disk_sb.sb->version = c->disk_sb.sb->version_min =
 		le16_to_cpu(bcachefs_metadata_version_current);
 	c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_atomic_nlink;
-	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] |= BCH_SB_FEATURES_ALL;
 
 	SET_BCH_SB_INITIALIZED(c->disk_sb.sb, true);
 	SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
diff --git a/fs/bcachefs/recovery.h b/fs/bcachefs/recovery.h
index 479ea46f8dcb..19f2f172a26b 100644
--- a/fs/bcachefs/recovery.h
+++ b/fs/bcachefs/recovery.h
@@ -2,32 +2,50 @@
 #ifndef _BCACHEFS_RECOVERY_H
 #define _BCACHEFS_RECOVERY_H
 
-struct journal_keys {
-	struct journal_key {
-		enum btree_id	btree_id:8;
-		unsigned	allocated:1;
-		struct bpos	pos;
-		struct bkey_i	*k;
-		u32		journal_seq;
-		u32		journal_offset;
-	}			*d;
-	size_t			nr;
-	u64			journal_seq_base;
-};
-
 #define for_each_journal_key(keys, i)				\
 	for (i = (keys).d; i < (keys).d + (keys).nr; (i)++)
 
 struct journal_iter {
+	enum btree_id		btree_id;
+	unsigned		level;
 	struct journal_keys	*keys;
 	struct journal_key	*k;
-	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 *);
+/*
+ * Iterate over keys in the btree, with keys from the journal overlaid on top:
+ */
+
+struct btree_and_journal_iter {
+	struct btree_iter	*btree;
+
+	struct btree		*b;
+	struct btree_node_iter	node_iter;
+	struct bkey		unpacked;
+
+	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 *);
+
+void bch2_btree_and_journal_iter_init(struct btree_and_journal_iter *,
+				      struct btree_trans *,
+				      struct journal_keys *,
+				      enum btree_id, struct bpos);
+void bch2_btree_and_journal_iter_init_node_iter(struct btree_and_journal_iter *,
+						struct journal_keys *,
+						struct btree *);
+
+void bch2_journal_keys_free(struct journal_keys *);
+void bch2_journal_entries_free(struct list_head *);
 
 int bch2_fs_recovery(struct bch_fs *);
 int bch2_fs_initialize(struct bch_fs *);
diff --git a/fs/bcachefs/reflink.c b/fs/bcachefs/reflink.c
index 3b8c74ca3725..2f223be74926 100644
--- a/fs/bcachefs/reflink.c
+++ b/fs/bcachefs/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;
 }
@@ -185,7 +184,7 @@ s64 bch2_remap_range(struct bch_fs *c,
 				       BTREE_ITER_INTENT);
 
 	while (1) {
-		bch2_trans_reset(&trans, TRANS_RESET_MEM);
+		bch2_trans_begin(&trans);
 
 		trans.mem_top = 0;
 
diff --git a/fs/bcachefs/reflink.h b/fs/bcachefs/reflink.h
index ac23b855858c..5445c1cf0797 100644
--- a/fs/bcachefs/reflink.h
+++ b/fs/bcachefs/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/fs/bcachefs/replicas.c b/fs/bcachefs/replicas.c
index 366888b1b36d..be4908575f72 100644
--- a/fs/bcachefs/replicas.c
+++ b/fs/bcachefs/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/fs/bcachefs/str_hash.h b/fs/bcachefs/str_hash.h
index f2779159a6b8..dea9b7252b88 100644
--- a/fs/bcachefs/str_hash.h
+++ b/fs/bcachefs/str_hash.h
@@ -163,6 +163,7 @@ bch2_hash_lookup(struct btree_trans *trans,
 			break;
 		}
 	}
+	bch2_trans_iter_put(trans, iter);
 
 	return ERR_PTR(ret ?: -ENOENT);
 }
@@ -187,6 +188,9 @@ bch2_hash_hole(struct btree_trans *trans,
 			return iter;
 	}
 
+	iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
+	bch2_trans_iter_put(trans, iter);
+
 	return ERR_PTR(ret ?: -ENOSPC);
 }
 
@@ -262,10 +266,8 @@ 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 +321,16 @@ 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 */
diff --git a/fs/bcachefs/super-io.c b/fs/bcachefs/super-io.c
index 43927853210a..6596764c8421 100644
--- a/fs/bcachefs/super-io.c
+++ b/fs/bcachefs/super-io.c
@@ -956,6 +956,9 @@ int bch2_fs_mark_dirty(struct bch_fs *c)
 	mutex_lock(&c->sb_lock);
 	SET_BCH_SB_CLEAN(c->disk_sb.sb, false);
 	c->disk_sb.sb->compat[0] &= ~(1ULL << BCH_COMPAT_FEAT_ALLOC_METADATA);
+	c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_new_extent_overwrite;
+	c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_extents_above_btree_updates;
+	c->disk_sb.sb->features[0] |= 1ULL << BCH_FEATURE_btree_updates_journalled;
 	ret = bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
@@ -1086,6 +1089,8 @@ void bch2_fs_mark_clean(struct bch_fs *c)
 
 	c->disk_sb.sb->compat[0] |= 1ULL << BCH_COMPAT_FEAT_ALLOC_INFO;
 	c->disk_sb.sb->compat[0] |= 1ULL << BCH_COMPAT_FEAT_ALLOC_METADATA;
+	c->disk_sb.sb->features[0] &= ~(1ULL << BCH_FEATURE_extents_above_btree_updates);
+	c->disk_sb.sb->features[0] &= ~(1ULL << BCH_FEATURE_btree_updates_journalled);
 
 	u64s = sizeof(*sb_clean) / sizeof(u64) + c->journal.entry_u64s_reserved;
 
diff --git a/fs/bcachefs/super.c b/fs/bcachefs/super.c
index 38920fff4500..d2c275ce79ab 100644
--- a/fs/bcachefs/super.c
+++ b/fs/bcachefs/super.c
@@ -500,6 +500,8 @@ static void bch2_fs_free(struct bch_fs *c)
 	bch2_io_clock_exit(&c->io_clock[WRITE]);
 	bch2_io_clock_exit(&c->io_clock[READ]);
 	bch2_fs_compress_exit(c);
+	bch2_journal_keys_free(&c->journal_keys);
+	bch2_journal_entries_free(&c->journal_entries);
 	percpu_free_rwsem(&c->mark_lock);
 	kfree(c->usage_scratch);
 	free_percpu(c->usage[1]);
@@ -549,6 +551,10 @@ void bch2_fs_stop(struct bch_fs *c)
 
 	cancel_work_sync(&c->journal_seq_blacklist_gc_work);
 
+	mutex_lock(&c->state_lock);
+	bch2_fs_read_only(c);
+	mutex_unlock(&c->state_lock);
+
 	for_each_member_device(ca, c, i)
 		if (ca->kobj.state_in_sysfs &&
 		    ca->disk_sb.bdev)
@@ -572,10 +578,6 @@ void bch2_fs_stop(struct bch_fs *c)
 	closure_sync(&c->cl);
 	closure_debug_destroy(&c->cl);
 
-	mutex_lock(&c->state_lock);
-	bch2_fs_read_only(c);
-	mutex_unlock(&c->state_lock);
-
 	/* btree prefetch might have kicked off reads in the background: */
 	bch2_btree_flush_all_reads(c);
 
@@ -674,6 +676,7 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	INIT_LIST_HEAD(&c->list);
 
 	INIT_LIST_HEAD(&c->btree_interior_update_list);
+	INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
 	mutex_init(&c->btree_reserve_cache_lock);
 	mutex_init(&c->btree_interior_update_lock);
 
@@ -688,6 +691,8 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	INIT_WORK(&c->journal_seq_blacklist_gc_work,
 		  bch2_blacklist_entries_gc);
 
+	INIT_LIST_HEAD(&c->journal_entries);
+
 	INIT_LIST_HEAD(&c->fsck_errors);
 	mutex_init(&c->fsck_error_lock);
 
diff --git a/fs/bcachefs/sysfs.c b/fs/bcachefs/sysfs.c
index 602def1ee95a..d78ffcc0e8a4 100644
--- a/fs/bcachefs/sysfs.c
+++ b/fs/bcachefs/sysfs.c
@@ -276,7 +276,7 @@ static ssize_t bch2_compression_stats(struct bch_fs *c, char *buf)
 			struct extent_ptr_decoded p;
 
 			extent_for_each_ptr_decode(e, p, entry) {
-				if (p.crc.compression_type == BCH_COMPRESSION_TYPE_none) {
+				if (!crc_is_compressed(p.crc)) {
 					nr_uncompressed_extents++;
 					uncompressed_sectors += e.k->size;
 				} else {
diff --git a/fs/bcachefs/tests.c b/fs/bcachefs/tests.c
index 8f9b0cca17da..4dcace650416 100644
--- a/fs/bcachefs/tests.c
+++ b/fs/bcachefs/tests.c
@@ -18,7 +18,7 @@ static void delete_test_keys(struct bch_fs *c)
 				      NULL);
 	BUG_ON(ret);
 
-	ret = bch2_btree_delete_range(c, BTREE_ID_DIRENTS,
+	ret = bch2_btree_delete_range(c, BTREE_ID_XATTRS,
 				      POS(0, 0), POS(0, U64_MAX),
 				      NULL);
 	BUG_ON(ret);
@@ -37,14 +37,14 @@ static void test_delete(struct bch_fs *c, u64 nr)
 
 	bch2_trans_init(&trans, c, 0, 0);
 
-	iter = bch2_trans_get_iter(&trans, BTREE_ID_DIRENTS, k.k.p,
+	iter = bch2_trans_get_iter(&trans, BTREE_ID_XATTRS, k.k.p,
 				   BTREE_ITER_INTENT);
 
 	ret = bch2_btree_iter_traverse(iter);
 	BUG_ON(ret);
 
-	bch2_trans_update(&trans, iter, &k.k_i, 0);
-	ret = bch2_trans_commit(&trans, NULL, NULL, 0);
+	ret = __bch2_trans_do(&trans, NULL, NULL, 0,
+		bch2_trans_update(&trans, iter, &k.k_i, 0));
 	BUG_ON(ret);
 
 	pr_info("deleting once");
@@ -69,14 +69,14 @@ static void test_delete_written(struct bch_fs *c, u64 nr)
 
 	bch2_trans_init(&trans, c, 0, 0);
 
-	iter = bch2_trans_get_iter(&trans, BTREE_ID_DIRENTS, k.k.p,
+	iter = bch2_trans_get_iter(&trans, BTREE_ID_XATTRS, k.k.p,
 				   BTREE_ITER_INTENT);
 
 	ret = bch2_btree_iter_traverse(iter);
 	BUG_ON(ret);
 
-	bch2_trans_update(&trans, iter, &k.k_i, 0);
-	ret = bch2_trans_commit(&trans, NULL, NULL, 0);
+	ret = __bch2_trans_do(&trans, NULL, NULL, 0,
+		bch2_trans_update(&trans, iter, &k.k_i, 0));
 	BUG_ON(ret);
 
 	bch2_journal_flush_all_pins(&c->journal);
@@ -107,7 +107,7 @@ static void test_iterate(struct bch_fs *c, u64 nr)
 		bkey_cookie_init(&k.k_i);
 		k.k.p.offset = i;
 
-		ret = bch2_btree_insert(c, BTREE_ID_DIRENTS, &k.k_i,
+		ret = bch2_btree_insert(c, BTREE_ID_XATTRS, &k.k_i,
 					NULL, NULL, 0);
 		BUG_ON(ret);
 	}
@@ -116,9 +116,13 @@ static void test_iterate(struct bch_fs *c, u64 nr)
 
 	i = 0;
 
-	for_each_btree_key(&trans, iter, BTREE_ID_DIRENTS,
-			   POS_MIN, 0, k, ret)
+	for_each_btree_key(&trans, iter, BTREE_ID_XATTRS,
+			   POS_MIN, 0, k, ret) {
+		if (k.k->p.inode)
+			break;
+
 		BUG_ON(k.k->p.offset != i++);
+	}
 
 	BUG_ON(i != nr);
 
@@ -202,7 +206,7 @@ static void test_iterate_slots(struct bch_fs *c, u64 nr)
 		bkey_cookie_init(&k.k_i);
 		k.k.p.offset = i * 2;
 
-		ret = bch2_btree_insert(c, BTREE_ID_DIRENTS, &k.k_i,
+		ret = bch2_btree_insert(c, BTREE_ID_XATTRS, &k.k_i,
 					NULL, NULL, 0);
 		BUG_ON(ret);
 	}
@@ -211,8 +215,11 @@ static void test_iterate_slots(struct bch_fs *c, u64 nr)
 
 	i = 0;
 
-	for_each_btree_key(&trans, iter, BTREE_ID_DIRENTS, POS_MIN,
+	for_each_btree_key(&trans, iter, BTREE_ID_XATTRS, POS_MIN,
 			   0, k, ret) {
+		if (k.k->p.inode)
+			break;
+
 		BUG_ON(k.k->p.offset != i);
 		i += 2;
 	}
@@ -224,11 +231,12 @@ static void test_iterate_slots(struct bch_fs *c, u64 nr)
 
 	i = 0;
 
-	for_each_btree_key(&trans, iter, BTREE_ID_DIRENTS, POS_MIN,
+	for_each_btree_key(&trans, iter, BTREE_ID_XATTRS, POS_MIN,
 			   BTREE_ITER_SLOTS, k, ret) {
+		BUG_ON(k.k->p.offset != i);
 		BUG_ON(bkey_deleted(k.k) != (i & 1));
-		BUG_ON(k.k->p.offset != i++);
 
+		i++;
 		if (i == nr * 2)
 			break;
 	}
@@ -307,7 +315,7 @@ static void test_peek_end(struct bch_fs *c, u64 nr)
 
 	bch2_trans_init(&trans, c, 0, 0);
 
-	iter = bch2_trans_get_iter(&trans, BTREE_ID_DIRENTS, POS_MIN, 0);
+	iter = bch2_trans_get_iter(&trans, BTREE_ID_XATTRS, POS_MIN, 0);
 
 	k = bch2_btree_iter_peek(iter);
 	BUG_ON(k.k);
@@ -409,18 +417,24 @@ static u64 test_rand(void)
 
 static void rand_insert(struct bch_fs *c, u64 nr)
 {
+	struct btree_trans trans;
 	struct bkey_i_cookie k;
 	int ret;
 	u64 i;
 
+	bch2_trans_init(&trans, c, 0, 0);
+
 	for (i = 0; i < nr; i++) {
 		bkey_cookie_init(&k.k_i);
 		k.k.p.offset = test_rand();
 
-		ret = bch2_btree_insert(c, BTREE_ID_DIRENTS, &k.k_i,
-					NULL, NULL, 0);
+		ret = __bch2_trans_do(&trans, NULL, NULL, 0,
+			__bch2_btree_insert(&trans, BTREE_ID_XATTRS, &k.k_i));
+
 		BUG_ON(ret);
 	}
+
+	bch2_trans_exit(&trans);
 }
 
 static void rand_lookup(struct bch_fs *c, u64 nr)
@@ -433,7 +447,7 @@ static void rand_lookup(struct bch_fs *c, u64 nr)
 	bch2_trans_init(&trans, c, 0, 0);
 
 	for (i = 0; i < nr; i++) {
-		iter = bch2_trans_get_iter(&trans, BTREE_ID_DIRENTS,
+		iter = bch2_trans_get_iter(&trans, BTREE_ID_XATTRS,
 					   POS(0, test_rand()), 0);
 
 		k = bch2_btree_iter_peek(iter);
@@ -454,7 +468,7 @@ static void rand_mixed(struct bch_fs *c, u64 nr)
 	bch2_trans_init(&trans, c, 0, 0);
 
 	for (i = 0; i < nr; i++) {
-		iter = bch2_trans_get_iter(&trans, BTREE_ID_DIRENTS,
+		iter = bch2_trans_get_iter(&trans, BTREE_ID_XATTRS,
 					   POS(0, test_rand()), 0);
 
 		k = bch2_btree_iter_peek(iter);
@@ -465,8 +479,9 @@ static void rand_mixed(struct bch_fs *c, u64 nr)
 			bkey_cookie_init(&k.k_i);
 			k.k.p = iter->pos;
 
-			bch2_trans_update(&trans, iter, &k.k_i, 0);
-			ret = bch2_trans_commit(&trans, NULL, NULL, 0);
+			ret = __bch2_trans_do(&trans, NULL, NULL, 0,
+				bch2_trans_update(&trans, iter, &k.k_i, 0));
+
 			BUG_ON(ret);
 		}
 
@@ -476,20 +491,50 @@ static void rand_mixed(struct bch_fs *c, u64 nr)
 	bch2_trans_exit(&trans);
 }
 
+static int __do_delete(struct btree_trans *trans, struct bpos pos)
+{
+	struct btree_iter *iter;
+	struct bkey_i delete;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	iter = bch2_trans_get_iter(trans, BTREE_ID_XATTRS, pos,
+				   BTREE_ITER_INTENT);
+	ret = PTR_ERR_OR_ZERO(iter);
+	if (ret)
+		goto err;
+
+	k = bch2_btree_iter_peek(iter);
+	ret = bkey_err(k);
+	if (ret)
+		goto err;
+
+	bkey_init(&delete.k);
+	delete.k.p = k.k->p;
+
+	bch2_trans_update(trans, iter, &delete, 0);
+err:
+	bch2_trans_iter_put(trans, iter);
+	return ret;
+}
+
 static void rand_delete(struct bch_fs *c, u64 nr)
 {
-	struct bkey_i k;
+	struct btree_trans trans;
 	int ret;
 	u64 i;
 
+	bch2_trans_init(&trans, c, 0, 0);
+
 	for (i = 0; i < nr; i++) {
-		bkey_init(&k.k);
-		k.k.p.offset = test_rand();
+		struct bpos pos = POS(0, test_rand());
 
-		ret = bch2_btree_insert(c, BTREE_ID_DIRENTS, &k,
-					NULL, NULL, 0);
+		ret = __bch2_trans_do(&trans, NULL, NULL, 0,
+			__do_delete(&trans, pos));
 		BUG_ON(ret);
 	}
+
+	bch2_trans_exit(&trans);
 }
 
 static void seq_insert(struct bch_fs *c, u64 nr)
@@ -505,12 +550,13 @@ static void seq_insert(struct bch_fs *c, u64 nr)
 
 	bch2_trans_init(&trans, c, 0, 0);
 
-	for_each_btree_key(&trans, iter, BTREE_ID_DIRENTS, POS_MIN,
+	for_each_btree_key(&trans, iter, BTREE_ID_XATTRS, POS_MIN,
 			   BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
 		insert.k.p = iter->pos;
 
-		bch2_trans_update(&trans, iter, &insert.k_i, 0);
-		ret = bch2_trans_commit(&trans, NULL, NULL, 0);
+		ret = __bch2_trans_do(&trans, NULL, NULL, 0,
+			bch2_trans_update(&trans, iter, &insert.k_i, 0));
+
 		BUG_ON(ret);
 
 		if (++i == nr)
@@ -528,7 +574,7 @@ static void seq_lookup(struct bch_fs *c, u64 nr)
 
 	bch2_trans_init(&trans, c, 0, 0);
 
-	for_each_btree_key(&trans, iter, BTREE_ID_DIRENTS, POS_MIN, 0, k, ret)
+	for_each_btree_key(&trans, iter, BTREE_ID_XATTRS, POS_MIN, 0, k, ret)
 		;
 	bch2_trans_exit(&trans);
 }
@@ -542,14 +588,15 @@ static void seq_overwrite(struct bch_fs *c, u64 nr)
 
 	bch2_trans_init(&trans, c, 0, 0);
 
-	for_each_btree_key(&trans, iter, BTREE_ID_DIRENTS, POS_MIN,
+	for_each_btree_key(&trans, iter, BTREE_ID_XATTRS, POS_MIN,
 			   BTREE_ITER_INTENT, k, ret) {
 		struct bkey_i_cookie u;
 
 		bkey_reassemble(&u.k_i, k);
 
-		bch2_trans_update(&trans, iter, &u.k_i, 0);
-		ret = bch2_trans_commit(&trans, NULL, NULL, 0);
+		ret = __bch2_trans_do(&trans, NULL, NULL, 0,
+			bch2_trans_update(&trans, iter, &u.k_i, 0));
+
 		BUG_ON(ret);
 	}
 	bch2_trans_exit(&trans);
@@ -559,7 +606,7 @@ static void seq_delete(struct bch_fs *c, u64 nr)
 {
 	int ret;
 
-	ret = bch2_btree_delete_range(c, BTREE_ID_DIRENTS,
+	ret = bch2_btree_delete_range(c, BTREE_ID_XATTRS,
 				      POS(0, 0), POS(0, U64_MAX),
 				      NULL);
 	BUG_ON(ret);
diff --git a/fs/bcachefs/util.h b/fs/bcachefs/util.h
index 2b19a0038045..0128daba5970 100644
--- a/fs/bcachefs/util.h
+++ b/fs/bcachefs/util.h
@@ -664,35 +664,6 @@ static inline void memset_u64s_tail(void *s, int c, unsigned bytes)
 	memset(s + bytes, c, rem);
 }
 
-static inline struct bio_vec next_contig_bvec(struct bio *bio,
-					      struct bvec_iter *iter)
-{
-	struct bio_vec bv = bio_iter_iovec(bio, *iter);
-
-	bio_advance_iter(bio, iter, bv.bv_len);
-#ifndef CONFIG_HIGHMEM
-	while (iter->bi_size) {
-		struct bio_vec next = bio_iter_iovec(bio, *iter);
-
-		if (page_address(bv.bv_page) + bv.bv_offset + bv.bv_len !=
-		    page_address(next.bv_page) + next.bv_offset)
-			break;
-
-		bv.bv_len += next.bv_len;
-		bio_advance_iter(bio, iter, next.bv_len);
-	}
-#endif
-	return bv;
-}
-
-#define __bio_for_each_contig_segment(bv, bio, iter, start)		\
-	for (iter = (start);						\
-	     (iter).bi_size &&						\
-		((bv = next_contig_bvec((bio), &(iter))), 1);)
-
-#define bio_for_each_contig_segment(bv, bio, iter)			\
-	__bio_for_each_contig_segment(bv, bio, iter, (bio)->bi_iter)
-
 void sort_cmp_size(void *base, size_t num, size_t size,
 	  int (*cmp_func)(const void *, const void *, size_t),
 	  void (*swap_func)(void *, void *, size_t));