Update bcachefs sources to 864591728963 bcachefs: Dropped superblock write is no longer a fatal error

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

61 files changed, 1171 insertions, 879 deletions

diff --git a/.bcachefs_revision b/.bcachefs_revision
index 64016eff..5d86ae9c 100644
--- a/.bcachefs_revision
+++ b/.bcachefs_revision
@@ -1 +1 @@
-55a65a994ed5fba038fda00f78416faf6f308bb8
+864591728963d416c49e502bfee56a283eda31a5
diff --git a/Makefile.compiler b/Makefile.compiler
index e0842496..8c102968 100644
--- a/Makefile.compiler
+++ b/Makefile.compiler
@@ -13,7 +13,7 @@ cc-cross-prefix = $(firstword $(foreach c, $(1), \
 			$(if $(shell command -v -- $(c)gcc 2>/dev/null), $(c))))
 
 # output directory for tests below
-TMPOUT = $(if $(KBUILD_EXTMOD),$(firstword $(KBUILD_EXTMOD))/).tmp_$$$$
+TMPOUT = .tmp_$$$$
 
 # try-run
 # Usage: option = $(call try-run, $(CC)...-o "$$TMP",option-ok,otherwise)
diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
index 3cbf8c9e..2384a5e3 100644
--- a/include/linux/blk_types.h
+++ b/include/linux/blk_types.h
@@ -6,6 +6,7 @@
 #define __LINUX_BLK_TYPES_H
 
 #include <linux/atomic.h>
+#include <linux/backing-dev.h>
 #include <linux/types.h>
 #include <linux/bvec.h>
 #include <linux/kobject.h>
diff --git a/include/linux/kmemleak.h b/include/linux/kmemleak.h
index 6a3cd1bf..93a73c07 100644
--- a/include/linux/kmemleak.h
+++ b/include/linux/kmemleak.h
@@ -26,6 +26,7 @@ extern void kmemleak_free_part(const void *ptr, size_t size) __ref;
 extern void kmemleak_free_percpu(const void __percpu *ptr) __ref;
 extern void kmemleak_update_trace(const void *ptr) __ref;
 extern void kmemleak_not_leak(const void *ptr) __ref;
+extern void kmemleak_transient_leak(const void *ptr) __ref;
 extern void kmemleak_ignore(const void *ptr) __ref;
 extern void kmemleak_scan_area(const void *ptr, size_t size, gfp_t gfp) __ref;
 extern void kmemleak_no_scan(const void *ptr) __ref;
@@ -93,6 +94,9 @@ static inline void kmemleak_update_trace(const void *ptr)
 static inline void kmemleak_not_leak(const void *ptr)
 {
 }
+static inline void kmemleak_transient_leak(const void *ptr)
+{
+}
 static inline void kmemleak_ignore(const void *ptr)
 {
 }
diff --git a/include/linux/kobject.h b/include/linux/kobject.h
index c33b2126..24096a62 100644
--- a/include/linux/kobject.h
+++ b/include/linux/kobject.h
@@ -20,6 +20,7 @@
 #include <linux/bug.h>
 #include <linux/compiler.h>
 #include <linux/kernel.h>
+#include <linux/slab.h>
 #include <linux/sysfs.h>
 #include <linux/types.h>
 #include <linux/workqueue.h>
diff --git a/libbcachefs/alloc_background.c b/libbcachefs/alloc_background.c
index 7641a3b4..94e7bc88 100644
--- a/libbcachefs/alloc_background.c
+++ b/libbcachefs/alloc_background.c
@@ -933,8 +933,6 @@ int bch2_trigger_alloc(struct btree_trans *trans,
 	if ((flags & BTREE_TRIGGER_atomic) && (flags & BTREE_TRIGGER_insert)) {
 		u64 transaction_seq = trans->journal_res.seq;
 		BUG_ON(!transaction_seq);
-		BUG_ON(transaction_seq < new_a->journal_seq_nonempty);
-		BUG_ON(transaction_seq < new_a->journal_seq_empty);
 
 		if (log_fsck_err_on(transaction_seq && new_a->journal_seq_nonempty > transaction_seq,
 				    trans, alloc_key_journal_seq_in_future,
diff --git a/libbcachefs/alloc_foreground.c b/libbcachefs/alloc_foreground.c
index 57d5f14c..6df41c33 100644
--- a/libbcachefs/alloc_foreground.c
+++ b/libbcachefs/alloc_foreground.c
@@ -107,14 +107,10 @@ void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
 		return;
 	}
 
-	percpu_down_read(&c->mark_lock);
 	spin_lock(&ob->lock);
-
 	ob->valid = false;
 	ob->data_type = 0;
-
 	spin_unlock(&ob->lock);
-	percpu_up_read(&c->mark_lock);
 
 	spin_lock(&c->freelist_lock);
 	bch2_open_bucket_hash_remove(c, ob);
diff --git a/libbcachefs/backpointers.c b/libbcachefs/backpointers.c
index 72311dc7..ebeb6a5f 100644
--- a/libbcachefs/backpointers.c
+++ b/libbcachefs/backpointers.c
@@ -178,7 +178,7 @@ static int bch2_backpointer_del(struct btree_trans *trans, struct bpos pos)
 		 bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
 }
 
-static int bch2_backpointers_maybe_flush(struct btree_trans *trans,
+static inline int bch2_backpointers_maybe_flush(struct btree_trans *trans,
 					 struct bkey_s_c visiting_k,
 					 struct bkey_buf *last_flushed)
 {
@@ -201,17 +201,30 @@ static int backpointer_target_not_found(struct btree_trans *trans,
 	 * looking at may have already been deleted - failure to find what it
 	 * pointed to is not an error:
 	 */
-	ret = bch2_backpointers_maybe_flush(trans, bp.s_c, last_flushed);
+	ret = last_flushed
+		? bch2_backpointers_maybe_flush(trans, bp.s_c, last_flushed)
+		: 0;
 	if (ret)
 		return ret;
 
 	prt_printf(&buf, "backpointer doesn't match %s it points to:\n  ",
 		   bp.v->level ? "btree node" : "extent");
 	bch2_bkey_val_to_text(&buf, c, bp.s_c);
-	prt_printf(&buf, "\n  ");
 
+	prt_printf(&buf, "\n  ");
 	bch2_bkey_val_to_text(&buf, c, target_k);
 
+	struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(target_k);
+	const union bch_extent_entry *entry;
+	struct extent_ptr_decoded p;
+	bkey_for_each_ptr_decode(target_k.k, ptrs, p, entry)
+		if (p.ptr.dev == bp.k->p.inode) {
+			prt_printf(&buf, "\n  ");
+			struct bkey_i_backpointer bp2;
+			bch2_extent_ptr_to_bp(c, bp.v->btree_id, bp.v->level, target_k, p, entry, &bp2);
+			bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&bp2.k_i));
+		}
+
 	if (fsck_err(trans, backpointer_to_missing_ptr,
 		     "%s", buf.buf))
 		ret = bch2_backpointer_del(trans, bp.k->p);
@@ -491,7 +504,7 @@ check_existing_bp:
 	struct bkey_s_c_backpointer other_bp = bkey_s_c_to_backpointer(bp_k);
 
 	struct bkey_s_c other_extent =
-		bch2_backpointer_get_key(trans, other_bp, &other_extent_iter, 0, &s->last_flushed);
+		bch2_backpointer_get_key(trans, other_bp, &other_extent_iter, 0, NULL);
 	ret = bkey_err(other_extent);
 	if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
 		ret = 0;
@@ -553,11 +566,11 @@ check_existing_bp:
 	goto err;
 missing:
 	printbuf_reset(&buf);
-	prt_str(&buf, "missing backpointer ");
-	bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&bp->k_i));
-	prt_newline(&buf);
+	prt_str(&buf, "missing backpointer\n  for:  ");
 	bch2_bkey_val_to_text(&buf, c, orig_k);
-	prt_printf(&buf, "\n  got:   ");
+	prt_printf(&buf, "\n  want: ");
+	bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&bp->k_i));
+	prt_printf(&buf, "\n  got:  ");
 	bch2_bkey_val_to_text(&buf, c, bp_k);
 
 	if (fsck_err(trans, ptr_to_missing_backpointer, "%s", buf.buf))
@@ -586,12 +599,16 @@ static int check_extent_to_backpointers(struct btree_trans *trans,
 		rcu_read_lock();
 		struct bch_dev *ca = bch2_dev_rcu_noerror(c, p.ptr.dev);
 		bool check = ca && test_bit(PTR_BUCKET_NR(ca, &p.ptr), ca->bucket_backpointer_mismatches);
+		bool empty = ca && test_bit(PTR_BUCKET_NR(ca, &p.ptr), ca->bucket_backpointer_empty);
 		rcu_read_unlock();
 
-		if (check) {
+		if (check || empty) {
 			struct bkey_i_backpointer bp;
 			bch2_extent_ptr_to_bp(c, btree, level, k, p, entry, &bp);
-			int ret = check_bp_exists(trans, s, &bp, k);
+
+			int ret = check
+				? check_bp_exists(trans, s, &bp, k)
+				: bch2_bucket_backpointer_mod(trans, k, &bp, true);
 			if (ret)
 				return ret;
 		}
@@ -825,12 +842,15 @@ static enum alloc_sector_counter data_type_to_alloc_counter(enum bch_data_type t
 	}
 }
 
-static int check_bucket_backpointer_mismatch_one(struct btree_trans *trans, struct bkey_s_c alloc_k,
-						 struct bkey_buf *last_flushed)
+static int check_bucket_backpointers_to_extents(struct btree_trans *, struct bch_dev *, struct bpos);
+
+static int check_bucket_backpointer_mismatch(struct btree_trans *trans, struct bkey_s_c alloc_k,
+					     struct bkey_buf *last_flushed)
 {
-	int ret = 0;
+	struct bch_fs *c = trans->c;
 	struct bch_alloc_v4 a_convert;
 	const struct bch_alloc_v4 *a = bch2_alloc_to_v4(alloc_k, &a_convert);
+	bool need_commit = false;
 
 	if (a->data_type == BCH_DATA_sb ||
 	    a->data_type == BCH_DATA_journal ||
@@ -846,6 +866,7 @@ static int check_bucket_backpointer_mismatch_one(struct btree_trans *trans, stru
 
 	struct btree_iter iter;
 	struct bkey_s_c bp_k;
+	int ret = 0;
 	for_each_btree_key_max_norestart(trans, iter, BTREE_ID_backpointers,
 				bucket_pos_to_bp_start(ca, alloc_k.k->p),
 				bucket_pos_to_bp_end(ca, alloc_k.k->p), 0, bp_k, ret) {
@@ -854,6 +875,17 @@ static int check_bucket_backpointer_mismatch_one(struct btree_trans *trans, stru
 
 		struct bkey_s_c_backpointer bp = bkey_s_c_to_backpointer(bp_k);
 
+		if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_backpointer_bucket_gen &&
+		    (bp.v->bucket_gen != a->gen ||
+		     bp.v->pad)) {
+			ret = bch2_backpointer_del(trans, bp_k.k->p);
+			if (ret)
+				break;
+
+			need_commit = true;
+			continue;
+		}
+
 		if (bp.v->bucket_gen != a->gen)
 			continue;
 
@@ -863,30 +895,40 @@ static int check_bucket_backpointer_mismatch_one(struct btree_trans *trans, stru
 	if (ret)
 		goto err;
 
+	if (need_commit) {
+		ret = bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc);
+		if (ret)
+			goto err;
+	}
+
 	/* Cached pointers don't have backpointers: */
 
-	if (sectors[ALLOC_dirty] != a->dirty_sectors ||
+	if (sectors[ALLOC_dirty]  != a->dirty_sectors ||
 	    sectors[ALLOC_stripe] != a->stripe_sectors) {
-		ret = bch2_backpointers_maybe_flush(trans, alloc_k, last_flushed);
-		if (ret)
+		if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_backpointer_bucket_gen) {
+			ret = bch2_backpointers_maybe_flush(trans, alloc_k, last_flushed);
+			if (ret)
+				goto err;
+		}
+
+		if (sectors[ALLOC_dirty]  > a->dirty_sectors ||
+		    sectors[ALLOC_stripe] > a->stripe_sectors) {
+			ret = check_bucket_backpointers_to_extents(trans, ca, alloc_k.k->p) ?:
+				-BCH_ERR_transaction_restart_nested;
 			goto err;
+		}
 
-		__set_bit(alloc_k.k->p.offset, ca->bucket_backpointer_mismatches);
+		if (!sectors[ALLOC_dirty] &&
+		    !sectors[ALLOC_stripe])
+			__set_bit(alloc_k.k->p.offset, ca->bucket_backpointer_empty);
+		else
+			__set_bit(alloc_k.k->p.offset, ca->bucket_backpointer_mismatches);
 	}
 err:
 	bch2_dev_put(ca);
 	return ret;
 }
 
-static int check_bucket_backpointer_mismatches(struct btree_trans *trans,
-					       struct bkey_buf *last_flushed)
-{
-	return for_each_btree_key(trans, iter, BTREE_ID_alloc,
-				     POS_MIN, BTREE_ITER_prefetch, k, ({
-		check_bucket_backpointer_mismatch_one(trans, k, last_flushed);
-	}));
-}
-
 static bool backpointer_node_has_missing(struct bch_fs *c, struct bkey_s_c k)
 {
 	switch (k.k->type) {
@@ -896,6 +938,9 @@ static bool backpointer_node_has_missing(struct bch_fs *c, struct bkey_s_c k)
 		rcu_read_lock();
 		struct bpos pos = bkey_s_c_to_btree_ptr_v2(k).v->min_key;
 		while (pos.inode <= k.k->p.inode) {
+			if (pos.inode >= c->sb.nr_devices)
+				break;
+
 			struct bch_dev *ca = bch2_dev_rcu_noerror(c, pos.inode);
 			if (!ca)
 				goto next;
@@ -941,7 +986,7 @@ err:
 }
 
 static int bch2_pin_backpointer_nodes_with_missing(struct btree_trans *trans,
-					struct bpos start, struct bpos *end)
+						   struct bpos start, struct bpos *end)
 {
 	struct bch_fs *c = trans->c;
 	int ret = 0;
@@ -1022,7 +1067,11 @@ int bch2_check_extents_to_backpointers(struct bch_fs *c)
 		ca->bucket_backpointer_mismatches = kvcalloc(BITS_TO_LONGS(ca->mi.nbuckets),
 							     sizeof(unsigned long),
 							     GFP_KERNEL);
-		if (!ca->bucket_backpointer_mismatches) {
+		ca->bucket_backpointer_empty = kvcalloc(BITS_TO_LONGS(ca->mi.nbuckets),
+							sizeof(unsigned long),
+							GFP_KERNEL);
+		if (!ca->bucket_backpointer_mismatches ||
+		    !ca->bucket_backpointer_empty) {
 			bch2_dev_put(ca);
 			ret = -BCH_ERR_ENOMEM_backpointer_mismatches_bitmap;
 			goto err_free_bitmaps;
@@ -1035,21 +1084,25 @@ int bch2_check_extents_to_backpointers(struct bch_fs *c)
 	bch2_bkey_buf_init(&s.last_flushed);
 	bkey_init(&s.last_flushed.k->k);
 
-	ret = check_bucket_backpointer_mismatches(trans, &s.last_flushed);
+	ret = for_each_btree_key(trans, iter, BTREE_ID_alloc,
+				 POS_MIN, BTREE_ITER_prefetch, k, ({
+		check_bucket_backpointer_mismatch(trans, k, &s.last_flushed);
+	}));
 	if (ret)
 		goto err;
 
-	u64 nr_buckets = 0, nr_mismatches = 0;
+	u64 nr_buckets = 0, nr_mismatches = 0, nr_empty = 0;
 	for_each_member_device(c, ca) {
-		nr_buckets += ca->mi.nbuckets;
-		nr_mismatches += bitmap_weight(ca->bucket_backpointer_mismatches, ca->mi.nbuckets);
+		nr_buckets	+= ca->mi.nbuckets;
+		nr_mismatches	+= bitmap_weight(ca->bucket_backpointer_mismatches, ca->mi.nbuckets);
+		nr_empty	+= bitmap_weight(ca->bucket_backpointer_empty, ca->mi.nbuckets);
 	}
 
-	if (!nr_mismatches)
+	if (!nr_mismatches && !nr_empty)
 		goto err;
 
 	bch_info(c, "scanning for missing backpointers in %llu/%llu buckets",
-		 nr_mismatches, nr_buckets);
+		 nr_mismatches + nr_empty, nr_buckets);
 
 	while (1) {
 		ret = bch2_pin_backpointer_nodes_with_missing(trans, s.bp_start, &s.bp_end);
@@ -1086,6 +1139,8 @@ err:
 	bch2_btree_cache_unpin(c);
 err_free_bitmaps:
 	for_each_member_device(c, ca) {
+		kvfree(ca->bucket_backpointer_empty);
+		ca->bucket_backpointer_empty = NULL;
 		kvfree(ca->bucket_backpointer_mismatches);
 		ca->bucket_backpointer_mismatches = NULL;
 	}
@@ -1122,6 +1177,25 @@ static int check_one_backpointer(struct btree_trans *trans,
 	return ret;
 }
 
+static int check_bucket_backpointers_to_extents(struct btree_trans *trans,
+						struct bch_dev *ca, struct bpos bucket)
+{
+	u32 restart_count = trans->restart_count;
+	struct bkey_buf last_flushed;
+	bch2_bkey_buf_init(&last_flushed);
+	bkey_init(&last_flushed.k->k);
+
+	int ret = for_each_btree_key_max(trans, iter, BTREE_ID_backpointers,
+				      bucket_pos_to_bp_start(ca, bucket),
+				      bucket_pos_to_bp_end(ca, bucket),
+				      0, k,
+		check_one_backpointer(trans, BBPOS_MIN, BBPOS_MAX, k, &last_flushed)
+	);
+
+	bch2_bkey_buf_exit(&last_flushed, trans->c);
+	return ret ?: trans_was_restarted(trans, restart_count);
+}
+
 static int bch2_check_backpointers_to_extents_pass(struct btree_trans *trans,
 						   struct bbpos start,
 						   struct bbpos end)
diff --git a/libbcachefs/bcachefs.h b/libbcachefs/bcachefs.h
index d27550ef..161cf2f0 100644
--- a/libbcachefs/bcachefs.h
+++ b/libbcachefs/bcachefs.h
@@ -547,17 +547,16 @@ struct bch_dev {
 
 	/*
 	 * Buckets:
-	 * Per-bucket arrays are protected by c->mark_lock, bucket_lock and
-	 * gc_gens_lock, for device resize - holding any is sufficient for
-	 * access: Or rcu_read_lock(), but only for dev_ptr_stale():
+	 * Per-bucket arrays are protected by either rcu_read_lock or
+	 * state_lock, for device resize.
 	 */
 	GENRADIX(struct bucket)	buckets_gc;
 	struct bucket_gens __rcu *bucket_gens;
 	u8			*oldest_gen;
 	unsigned long		*buckets_nouse;
-	struct rw_semaphore	bucket_lock;
 
 	unsigned long		*bucket_backpointer_mismatches;
+	unsigned long		*bucket_backpointer_empty;
 
 	struct bch_dev_usage __percpu	*usage;
 
diff --git a/libbcachefs/bcachefs_format.h b/libbcachefs/bcachefs_format.h
index 09e53bef..06809305 100644
--- a/libbcachefs/bcachefs_format.h
+++ b/libbcachefs/bcachefs_format.h
@@ -684,7 +684,8 @@ struct bch_sb_field_ext {
 	x(disk_accounting_big_endian,	BCH_VERSION(1, 15))		\
 	x(reflink_p_may_update_opts,	BCH_VERSION(1, 16))		\
 	x(inode_depth,			BCH_VERSION(1, 17))		\
-	x(persistent_inode_cursors,	BCH_VERSION(1, 18))
+	x(persistent_inode_cursors,	BCH_VERSION(1, 18))		\
+	x(autofix_errors,		BCH_VERSION(1, 19))
 
 enum bcachefs_metadata_version {
 	bcachefs_metadata_version_min = 9,
@@ -1287,14 +1288,18 @@ LE32_BITMASK(JSET_NO_FLUSH,	struct jset, flags, 5, 6);
 /* Btree: */
 
 enum btree_id_flags {
-	BTREE_ID_EXTENTS	= BIT(0),
-	BTREE_ID_SNAPSHOTS	= BIT(1),
-	BTREE_ID_SNAPSHOT_FIELD	= BIT(2),
-	BTREE_ID_DATA		= BIT(3),
+	BTREE_IS_extents	= BIT(0),
+	BTREE_IS_snapshots	= BIT(1),
+	BTREE_IS_snapshot_field	= BIT(2),
+	BTREE_IS_data		= BIT(3),
+	BTREE_IS_write_buffer	= BIT(4),
 };
 
 #define BCH_BTREE_IDS()								\
-	x(extents,		0,	BTREE_ID_EXTENTS|BTREE_ID_SNAPSHOTS|BTREE_ID_DATA,\
+	x(extents,		0,						\
+	  BTREE_IS_extents|							\
+	  BTREE_IS_snapshots|							\
+	  BTREE_IS_data,							\
 	  BIT_ULL(KEY_TYPE_whiteout)|						\
 	  BIT_ULL(KEY_TYPE_error)|						\
 	  BIT_ULL(KEY_TYPE_cookie)|						\
@@ -1302,17 +1307,20 @@ enum btree_id_flags {
 	  BIT_ULL(KEY_TYPE_reservation)|					\
 	  BIT_ULL(KEY_TYPE_reflink_p)|						\
 	  BIT_ULL(KEY_TYPE_inline_data))					\
-	x(inodes,		1,	BTREE_ID_SNAPSHOTS,			\
+	x(inodes,		1,						\
+	  BTREE_IS_snapshots,							\
 	  BIT_ULL(KEY_TYPE_whiteout)|						\
 	  BIT_ULL(KEY_TYPE_inode)|						\
 	  BIT_ULL(KEY_TYPE_inode_v2)|						\
 	  BIT_ULL(KEY_TYPE_inode_v3)|						\
 	  BIT_ULL(KEY_TYPE_inode_generation))					\
-	x(dirents,		2,	BTREE_ID_SNAPSHOTS,			\
+	x(dirents,		2,						\
+	  BTREE_IS_snapshots,							\
 	  BIT_ULL(KEY_TYPE_whiteout)|						\
 	  BIT_ULL(KEY_TYPE_hash_whiteout)|					\
 	  BIT_ULL(KEY_TYPE_dirent))						\
-	x(xattrs,		3,	BTREE_ID_SNAPSHOTS,			\
+	x(xattrs,		3,						\
+	  BTREE_IS_snapshots,							\
 	  BIT_ULL(KEY_TYPE_whiteout)|						\
 	  BIT_ULL(KEY_TYPE_cookie)|						\
 	  BIT_ULL(KEY_TYPE_hash_whiteout)|					\
@@ -1326,7 +1334,9 @@ enum btree_id_flags {
 	  BIT_ULL(KEY_TYPE_quota))						\
 	x(stripes,		6,	0,					\
 	  BIT_ULL(KEY_TYPE_stripe))						\
-	x(reflink,		7,	BTREE_ID_EXTENTS|BTREE_ID_DATA,		\
+	x(reflink,		7,						\
+	  BTREE_IS_extents|							\
+	  BTREE_IS_data,							\
 	  BIT_ULL(KEY_TYPE_reflink_v)|						\
 	  BIT_ULL(KEY_TYPE_indirect_inline_data)|				\
 	  BIT_ULL(KEY_TYPE_error))						\
@@ -1334,29 +1344,38 @@ enum btree_id_flags {
 	  BIT_ULL(KEY_TYPE_subvolume))						\
 	x(snapshots,		9,	0,					\
 	  BIT_ULL(KEY_TYPE_snapshot))						\
-	x(lru,			10,	0,					\
+	x(lru,			10,						\
+	  BTREE_IS_write_buffer,						\
 	  BIT_ULL(KEY_TYPE_set))						\
-	x(freespace,		11,	BTREE_ID_EXTENTS,			\
+	x(freespace,		11,						\
+	  BTREE_IS_extents,							\
 	  BIT_ULL(KEY_TYPE_set))						\
 	x(need_discard,		12,	0,					\
 	  BIT_ULL(KEY_TYPE_set))						\
-	x(backpointers,		13,	0,					\
+	x(backpointers,		13,						\
+	  BTREE_IS_write_buffer,						\
 	  BIT_ULL(KEY_TYPE_backpointer))					\
 	x(bucket_gens,		14,	0,					\
 	  BIT_ULL(KEY_TYPE_bucket_gens))					\
 	x(snapshot_trees,	15,	0,					\
 	  BIT_ULL(KEY_TYPE_snapshot_tree))					\
-	x(deleted_inodes,	16,	BTREE_ID_SNAPSHOT_FIELD,		\
+	x(deleted_inodes,	16,						\
+	  BTREE_IS_snapshot_field|						\
+	  BTREE_IS_write_buffer,						\
 	  BIT_ULL(KEY_TYPE_set))						\
 	x(logged_ops,		17,	0,					\
 	  BIT_ULL(KEY_TYPE_logged_op_truncate)|					\
 	  BIT_ULL(KEY_TYPE_logged_op_finsert)|					\
 	  BIT_ULL(KEY_TYPE_inode_alloc_cursor))					\
-	x(rebalance_work,	18,	BTREE_ID_SNAPSHOT_FIELD,		\
+	x(rebalance_work,	18,						\
+	  BTREE_IS_snapshot_field|						\
+	  BTREE_IS_write_buffer,						\
 	  BIT_ULL(KEY_TYPE_set)|BIT_ULL(KEY_TYPE_cookie))			\
 	x(subvolume_children,	19,	0,					\
 	  BIT_ULL(KEY_TYPE_set))						\
-	x(accounting,		20,	BTREE_ID_SNAPSHOT_FIELD,		\
+	x(accounting,		20,						\
+	  BTREE_IS_snapshot_field|						\
+	  BTREE_IS_write_buffer,						\
 	  BIT_ULL(KEY_TYPE_accounting))						\
 
 enum btree_id {
diff --git a/libbcachefs/btree_gc.c b/libbcachefs/btree_gc.c
index 24f2f3bd..dd1d9b74 100644
--- a/libbcachefs/btree_gc.c
+++ b/libbcachefs/btree_gc.c
@@ -733,16 +733,8 @@ static int bch2_gc_btrees(struct bch_fs *c)
 			continue;
 
 		ret = bch2_gc_btree(trans, btree, true);
-
-		if (mustfix_fsck_err_on(bch2_err_matches(ret, EIO),
-					trans, btree_node_read_error,
-			       "btree node read error for %s",
-			       (printbuf_reset(&buf),
-				bch2_btree_id_to_text(&buf, btree),
-				buf.buf)))
-			ret = bch2_btree_lost_data(c, btree);
 	}
-fsck_err:
+
 	printbuf_exit(&buf);
 	bch2_trans_put(trans);
 	bch_err_fn(c, ret);
@@ -811,7 +803,6 @@ static int bch2_alloc_write_key(struct btree_trans *trans,
 	old = bch2_alloc_to_v4(k, &old_convert);
 	gc = new = *old;
 
-	percpu_down_read(&c->mark_lock);
 	__bucket_m_to_alloc(&gc, *gc_bucket(ca, iter->pos.offset));
 
 	old_gc = gc;
@@ -822,7 +813,6 @@ static int bch2_alloc_write_key(struct btree_trans *trans,
 		gc.data_type = old->data_type;
 		gc.dirty_sectors = old->dirty_sectors;
 	}
-	percpu_up_read(&c->mark_lock);
 
 	/*
 	 * gc.data_type doesn't yet include need_discard & need_gc_gen states -
@@ -840,11 +830,9 @@ static int bch2_alloc_write_key(struct btree_trans *trans,
 		 * safe w.r.t. transaction restarts, so fixup the gc_bucket so
 		 * we don't run it twice:
 		 */
-		percpu_down_read(&c->mark_lock);
 		struct bucket *gc_m = gc_bucket(ca, iter->pos.offset);
 		gc_m->data_type = gc.data_type;
 		gc_m->dirty_sectors = gc.dirty_sectors;
-		percpu_up_read(&c->mark_lock);
 	}
 
 	if (fsck_err_on(new.data_type != gc.data_type,
@@ -1088,7 +1076,6 @@ static int gc_btree_gens_key(struct btree_trans *trans,
 	if (unlikely(test_bit(BCH_FS_going_ro, &c->flags)))
 		return -EROFS;
 
-	percpu_down_read(&c->mark_lock);
 	rcu_read_lock();
 	bkey_for_each_ptr(ptrs, ptr) {
 		struct bch_dev *ca = bch2_dev_rcu(c, ptr->dev);
@@ -1097,7 +1084,6 @@ static int gc_btree_gens_key(struct btree_trans *trans,
 
 		if (dev_ptr_stale(ca, ptr) > 16) {
 			rcu_read_unlock();
-			percpu_up_read(&c->mark_lock);
 			goto update;
 		}
 	}
@@ -1112,7 +1098,6 @@ static int gc_btree_gens_key(struct btree_trans *trans,
 			*gen = ptr->gen;
 	}
 	rcu_read_unlock();
-	percpu_up_read(&c->mark_lock);
 	return 0;
 update:
 	u = bch2_bkey_make_mut(trans, iter, &k, 0);
@@ -1141,7 +1126,6 @@ static int bch2_alloc_write_oldest_gen(struct btree_trans *trans, struct bch_dev
 		return ret;
 
 	a_mut->v.oldest_gen = ca->oldest_gen[iter->pos.offset];
-	alloc_data_type_set(&a_mut->v, a_mut->v.data_type);
 
 	return bch2_trans_update(trans, iter, &a_mut->k_i, 0);
 }
@@ -1254,9 +1238,16 @@ void bch2_gc_gens_async(struct bch_fs *c)
 		bch2_write_ref_put(c, BCH_WRITE_REF_gc_gens);
 }
 
-void bch2_fs_gc_init(struct bch_fs *c)
+void bch2_fs_btree_gc_exit(struct bch_fs *c)
 {
-	seqcount_init(&c->gc_pos_lock);
+}
 
+int bch2_fs_btree_gc_init(struct bch_fs *c)
+{
+	seqcount_init(&c->gc_pos_lock);
 	INIT_WORK(&c->gc_gens_work, bch2_gc_gens_work);
+
+	init_rwsem(&c->gc_lock);
+	mutex_init(&c->gc_gens_lock);
+	return 0;
 }
diff --git a/libbcachefs/btree_gc.h b/libbcachefs/btree_gc.h
index 8a47e8bd..9693a90a 100644
--- a/libbcachefs/btree_gc.h
+++ b/libbcachefs/btree_gc.h
@@ -82,6 +82,8 @@ void bch2_gc_pos_to_text(struct printbuf *, struct gc_pos *);
 
 int bch2_gc_gens(struct bch_fs *);
 void bch2_gc_gens_async(struct bch_fs *);
-void bch2_fs_gc_init(struct bch_fs *);
+
+void bch2_fs_btree_gc_exit(struct bch_fs *);
+int bch2_fs_btree_gc_init(struct bch_fs *);
 
 #endif /* _BCACHEFS_BTREE_GC_H */
diff --git a/libbcachefs/btree_io.c b/libbcachefs/btree_io.c
index d99f8a78..e371e60e 100644
--- a/libbcachefs/btree_io.c
+++ b/libbcachefs/btree_io.c
@@ -489,8 +489,8 @@ void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
 	if (b->nsets == MAX_BSETS &&
 	    !btree_node_write_in_flight(b) &&
 	    should_compact_all(c, b)) {
-		bch2_btree_node_write(c, b, SIX_LOCK_write,
-				      BTREE_WRITE_init_next_bset);
+		bch2_btree_node_write_trans(trans, b, SIX_LOCK_write,
+					    BTREE_WRITE_init_next_bset);
 		reinit_iter = true;
 	}
 
@@ -2345,6 +2345,34 @@ void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
 	}
 }
 
+void bch2_btree_node_write_trans(struct btree_trans *trans, struct btree *b,
+				 enum six_lock_type lock_type_held,
+				 unsigned flags)
+{
+	struct bch_fs *c = trans->c;
+
+	if (lock_type_held == SIX_LOCK_intent ||
+	    (lock_type_held == SIX_LOCK_read &&
+	     six_lock_tryupgrade(&b->c.lock))) {
+		__bch2_btree_node_write(c, b, flags);
+
+		/* don't cycle lock unnecessarily: */
+		if (btree_node_just_written(b) &&
+		    six_trylock_write(&b->c.lock)) {
+			bch2_btree_post_write_cleanup(c, b);
+			__bch2_btree_node_unlock_write(trans, b);
+		}
+
+		if (lock_type_held == SIX_LOCK_read)
+			six_lock_downgrade(&b->c.lock);
+	} else {
+		__bch2_btree_node_write(c, b, flags);
+		if (lock_type_held == SIX_LOCK_write &&
+		    btree_node_just_written(b))
+			bch2_btree_post_write_cleanup(c, b);
+	}
+}
+
 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
 {
 	struct bucket_table *tbl;
diff --git a/libbcachefs/btree_io.h b/libbcachefs/btree_io.h
index 9b01ca3d..6f9e4a6d 100644
--- a/libbcachefs/btree_io.h
+++ b/libbcachefs/btree_io.h
@@ -144,11 +144,13 @@ enum btree_write_flags {
 void __bch2_btree_node_write(struct bch_fs *, struct btree *, unsigned);
 void bch2_btree_node_write(struct bch_fs *, struct btree *,
 			   enum six_lock_type, unsigned);
+void bch2_btree_node_write_trans(struct btree_trans *, struct btree *,
+				 enum six_lock_type, unsigned);
 
-static inline void btree_node_write_if_need(struct bch_fs *c, struct btree *b,
+static inline void btree_node_write_if_need(struct btree_trans *trans, struct btree *b,
 					    enum six_lock_type lock_held)
 {
-	bch2_btree_node_write(c, b, lock_held, BTREE_WRITE_ONLY_IF_NEED);
+	bch2_btree_node_write_trans(trans, b, lock_held, BTREE_WRITE_ONLY_IF_NEED);
 }
 
 bool bch2_btree_flush_all_reads(struct bch_fs *);
diff --git a/libbcachefs/btree_iter.c b/libbcachefs/btree_iter.c
index 9c54891c..c291d495 100644
--- a/libbcachefs/btree_iter.c
+++ b/libbcachefs/btree_iter.c
@@ -699,6 +699,19 @@ void bch2_trans_node_add(struct btree_trans *trans,
 	bch2_trans_revalidate_updates_in_node(trans, b);
 }
 
+void bch2_trans_node_drop(struct btree_trans *trans,
+			  struct btree *b)
+{
+	struct btree_path *path;
+	unsigned i, level = b->c.level;
+
+	trans_for_each_path(trans, path, i)
+		if (path->l[level].b == b) {
+			btree_node_unlock(trans, path, level);
+			path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
+		}
+}
+
 /*
  * A btree node has been modified in such a way as to invalidate iterators - fix
  * them:
@@ -1854,7 +1867,7 @@ struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *path, struct bkey *
 			!bkey_eq(path->pos, ck->key.pos));
 
 		*u = ck->k->k;
-		k = bkey_i_to_s_c(ck->k);
+		k = (struct bkey_s_c) { u, &ck->k->v };
 	}
 
 	return k;
@@ -2144,21 +2157,18 @@ struct bkey_s_c btree_trans_peek_slot_journal(struct btree_trans *trans,
 }
 
 static noinline
-struct bkey_s_c btree_trans_peek_journal(struct btree_trans *trans,
-					 struct btree_iter *iter,
-					 struct bkey_s_c k)
+void btree_trans_peek_journal(struct btree_trans *trans,
+			      struct btree_iter *iter,
+			      struct bkey_s_c *k)
 {
 	struct btree_path *path = btree_iter_path(trans, iter);
 	struct bkey_i *next_journal =
 		bch2_btree_journal_peek(trans, iter,
-				k.k ? k.k->p : path_l(path)->b->key.k.p);
-
+				k->k ? k->k->p : path_l(path)->b->key.k.p);
 	if (next_journal) {
 		iter->k = next_journal->k;
-		k = bkey_i_to_s_c(next_journal);
+		*k = bkey_i_to_s_c(next_journal);
 	}
-
-	return k;
 }
 
 static struct bkey_i *bch2_btree_journal_peek_prev(struct btree_trans *trans,
@@ -2175,21 +2185,19 @@ static struct bkey_i *bch2_btree_journal_peek_prev(struct btree_trans *trans,
 }
 
 static noinline
-struct bkey_s_c btree_trans_peek_prev_journal(struct btree_trans *trans,
-					 struct btree_iter *iter,
-					 struct bkey_s_c k)
+void btree_trans_peek_prev_journal(struct btree_trans *trans,
+				   struct btree_iter *iter,
+				   struct bkey_s_c *k)
 {
 	struct btree_path *path = btree_iter_path(trans, iter);
 	struct bkey_i *next_journal =
 		bch2_btree_journal_peek_prev(trans, iter,
-				k.k ? k.k->p : path_l(path)->b->key.k.p);
+				k->k ? k->k->p : path_l(path)->b->key.k.p);
 
 	if (next_journal) {
 		iter->k = next_journal->k;
-		k = bkey_i_to_s_c(next_journal);
+		*k = bkey_i_to_s_c(next_journal);
 	}
-
-	return k;
 }
 
 /*
@@ -2234,10 +2242,15 @@ struct bkey_s_c btree_trans_peek_key_cache(struct btree_iter *iter, struct bpos
 	btree_path_set_should_be_locked(trans, trans->paths + iter->key_cache_path);
 
 	k = bch2_btree_path_peek_slot(trans->paths + iter->key_cache_path, &u);
-	if (k.k && !bkey_err(k)) {
-		iter->k = u;
-		k.k = &iter->k;
-	}
+	if (!k.k)
+		return k;
+
+	if ((iter->flags & BTREE_ITER_all_snapshots) &&
+	    !bpos_eq(pos, k.k->p))
+		return bkey_s_c_null;
+
+	iter->k = u;
+	k.k = &iter->k;
 	return k;
 }
 
@@ -2260,7 +2273,7 @@ static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bp
 			/* ensure that iter->k is consistent with iter->pos: */
 			bch2_btree_iter_set_pos(iter, iter->pos);
 			k = bkey_s_c_err(ret);
-			goto out;
+			break;
 		}
 
 		struct btree_path *path = btree_iter_path(trans, iter);
@@ -2270,7 +2283,7 @@ static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bp
 			/* No btree nodes at requested level: */
 			bch2_btree_iter_set_pos(iter, SPOS_MAX);
 			k = bkey_s_c_null;
-			goto out;
+			break;
 		}
 
 		btree_path_set_should_be_locked(trans, path);
@@ -2281,15 +2294,14 @@ static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bp
 		    k.k &&
 		    (k2 = btree_trans_peek_key_cache(iter, k.k->p)).k) {
 			k = k2;
-			ret = bkey_err(k);
-			if (ret) {
+			if (bkey_err(k)) {
 				bch2_btree_iter_set_pos(iter, iter->pos);
-				goto out;
+				break;
 			}
 		}
 
 		if (unlikely(iter->flags & BTREE_ITER_with_journal))
-			k = btree_trans_peek_journal(trans, iter, k);
+			btree_trans_peek_journal(trans, iter, &k);
 
 		if (unlikely((iter->flags & BTREE_ITER_with_updates) &&
 			     trans->nr_updates))
@@ -2318,12 +2330,11 @@ static struct bkey_s_c __bch2_btree_iter_peek(struct btree_iter *iter, struct bp
 			/* End of btree: */
 			bch2_btree_iter_set_pos(iter, SPOS_MAX);
 			k = bkey_s_c_null;
-			goto out;
+			break;
 		}
 	}
-out:
-	bch2_btree_iter_verify(iter);
 
+	bch2_btree_iter_verify(iter);
 	return k;
 }
 
@@ -2424,7 +2435,8 @@ struct bkey_s_c bch2_btree_iter_peek_max(struct btree_iter *iter, struct bpos en
 				continue;
 			}
 
-			if (bkey_whiteout(k.k)) {
+			if (bkey_whiteout(k.k) &&
+			    !(iter->flags & BTREE_ITER_key_cache_fill)) {
 				search_key = bkey_successor(iter, k.k->p);
 				continue;
 			}
@@ -2547,7 +2559,7 @@ static struct bkey_s_c __bch2_btree_iter_peek_prev(struct btree_iter *iter, stru
 		}
 
 		if (unlikely(iter->flags & BTREE_ITER_with_journal))
-			k = btree_trans_peek_prev_journal(trans, iter, k);
+			btree_trans_peek_prev_journal(trans, iter, &k);
 
 		if (unlikely((iter->flags & BTREE_ITER_with_updates) &&
 			     trans->nr_updates))
@@ -2784,6 +2796,11 @@ struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
 		k = bch2_btree_path_peek_slot(trans->paths + iter->path, &iter->k);
 		if (unlikely(!k.k))
 			goto out_no_locked;
+
+		if (unlikely(k.k->type == KEY_TYPE_whiteout &&
+			     (iter->flags & BTREE_ITER_filter_snapshots) &&
+			     !(iter->flags & BTREE_ITER_key_cache_fill)))
+			iter->k.type = KEY_TYPE_deleted;
 	} else {
 		struct bpos next;
 		struct bpos end = iter->pos;
@@ -3028,7 +3045,7 @@ void bch2_trans_iter_init_outlined(struct btree_trans *trans,
 			  unsigned flags)
 {
 	bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
-			       bch2_btree_iter_flags(trans, btree_id, flags),
+			       bch2_btree_iter_flags(trans, btree_id, 0, flags),
 			       _RET_IP_);
 }
 
@@ -3044,8 +3061,11 @@ void bch2_trans_node_iter_init(struct btree_trans *trans,
 	flags |= BTREE_ITER_snapshot_field;
 	flags |= BTREE_ITER_all_snapshots;
 
+	if (!depth && btree_id_cached(trans->c, btree_id))
+		flags |= BTREE_ITER_with_key_cache;
+
 	bch2_trans_iter_init_common(trans, iter, btree_id, pos, locks_want, depth,
-			       __bch2_btree_iter_flags(trans, btree_id, flags),
+			       bch2_btree_iter_flags(trans, btree_id, depth, flags),
 			       _RET_IP_);
 
 	iter->min_depth	= depth;
diff --git a/libbcachefs/btree_iter.h b/libbcachefs/btree_iter.h
index 3477fc8c..b9538e6e 100644
--- a/libbcachefs/btree_iter.h
+++ b/libbcachefs/btree_iter.h
@@ -372,6 +372,7 @@ static inline void bch2_btree_path_downgrade(struct btree_trans *trans,
 void bch2_trans_downgrade(struct btree_trans *);
 
 void bch2_trans_node_add(struct btree_trans *trans, struct btree_path *, struct btree *);
+void bch2_trans_node_drop(struct btree_trans *trans, struct btree *);
 void bch2_trans_node_reinit_iter(struct btree_trans *, struct btree *);
 
 int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter);
@@ -446,10 +447,17 @@ static inline void bch2_btree_iter_set_snapshot(struct btree_iter *iter, u32 sna
 
 void bch2_trans_iter_exit(struct btree_trans *, struct btree_iter *);
 
-static inline unsigned __bch2_btree_iter_flags(struct btree_trans *trans,
-					       unsigned btree_id,
-					       unsigned flags)
+static inline unsigned bch2_btree_iter_flags(struct btree_trans *trans,
+					     unsigned btree_id,
+					     unsigned level,
+					     unsigned flags)
 {
+	if (level || !btree_id_cached(trans->c, btree_id)) {
+		flags &= ~BTREE_ITER_cached;
+		flags &= ~BTREE_ITER_with_key_cache;
+	} else if (!(flags & BTREE_ITER_cached))
+		flags |= BTREE_ITER_with_key_cache;
+
 	if (!(flags & (BTREE_ITER_all_snapshots|BTREE_ITER_not_extents)) &&
 	    btree_id_is_extents(btree_id))
 		flags |= BTREE_ITER_is_extents;
@@ -468,19 +476,6 @@ static inline unsigned __bch2_btree_iter_flags(struct btree_trans *trans,
 	return flags;
 }
 
-static inline unsigned bch2_btree_iter_flags(struct btree_trans *trans,
-					     unsigned btree_id,
-					     unsigned flags)
-{
-	if (!btree_id_cached(trans->c, btree_id)) {
-		flags &= ~BTREE_ITER_cached;
-		flags &= ~BTREE_ITER_with_key_cache;
-	} else if (!(flags & BTREE_ITER_cached))
-		flags |= BTREE_ITER_with_key_cache;
-
-	return __bch2_btree_iter_flags(trans, btree_id, flags);
-}
-
 static inline void bch2_trans_iter_init_common(struct btree_trans *trans,
 					  struct btree_iter *iter,
 					  unsigned btree_id, struct bpos pos,
@@ -517,7 +512,7 @@ static inline void bch2_trans_iter_init(struct btree_trans *trans,
 	if (__builtin_constant_p(btree_id) &&
 	    __builtin_constant_p(flags))
 		bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
-				bch2_btree_iter_flags(trans, btree_id, flags),
+				bch2_btree_iter_flags(trans, btree_id, 0, flags),
 				_THIS_IP_);
 	else
 		bch2_trans_iter_init_outlined(trans, iter, btree_id, pos, flags);
diff --git a/libbcachefs/btree_key_cache.c b/libbcachefs/btree_key_cache.c
index 3bd40ea0..7636a5e9 100644
--- a/libbcachefs/btree_key_cache.c
+++ b/libbcachefs/btree_key_cache.c
@@ -197,7 +197,9 @@ out:
 	return ck;
 }
 
-static int btree_key_cache_create(struct btree_trans *trans, struct btree_path *path,
+static int btree_key_cache_create(struct btree_trans *trans,
+				  struct btree_path *path,
+				  struct btree_path *ck_path,
 				  struct bkey_s_c k)
 {
 	struct bch_fs *c = trans->c;
@@ -217,7 +219,7 @@ static int btree_key_cache_create(struct btree_trans *trans, struct btree_path *
 	key_u64s = min(256U, (key_u64s * 3) / 2);
 	key_u64s = roundup_pow_of_two(key_u64s);
 
-	struct bkey_cached *ck = bkey_cached_alloc(trans, path, key_u64s);
+	struct bkey_cached *ck = bkey_cached_alloc(trans, ck_path, key_u64s);
 	int ret = PTR_ERR_OR_ZERO(ck);
 	if (ret)
 		return ret;
@@ -226,19 +228,19 @@ static int btree_key_cache_create(struct btree_trans *trans, struct btree_path *
 		ck = bkey_cached_reuse(bc);
 		if (unlikely(!ck)) {
 			bch_err(c, "error allocating memory for key cache item, btree %s",
-				bch2_btree_id_str(path->btree_id));
+				bch2_btree_id_str(ck_path->btree_id));
 			return -BCH_ERR_ENOMEM_btree_key_cache_create;
 		}
 	}
 
 	ck->c.level		= 0;
-	ck->c.btree_id		= path->btree_id;
-	ck->key.btree_id	= path->btree_id;
-	ck->key.pos		= path->pos;
+	ck->c.btree_id		= ck_path->btree_id;
+	ck->key.btree_id	= ck_path->btree_id;
+	ck->key.pos		= ck_path->pos;
 	ck->flags		= 1U << BKEY_CACHED_ACCESSED;
 
 	if (unlikely(key_u64s > ck->u64s)) {
-		mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
+		mark_btree_node_locked_noreset(ck_path, 0, BTREE_NODE_UNLOCKED);
 
 		struct bkey_i *new_k = allocate_dropping_locks(trans, ret,
 				kmalloc(key_u64s * sizeof(u64), _gfp));
@@ -258,22 +260,29 @@ static int btree_key_cache_create(struct btree_trans *trans, struct btree_path *
 
 	bkey_reassemble(ck->k, k);
 
+	ret = bch2_btree_node_lock_write(trans, path, &path_l(path)->b->c);
+	if (unlikely(ret))
+		goto err;
+
 	ret = rhashtable_lookup_insert_fast(&bc->table, &ck->hash, bch2_btree_key_cache_params);
+
+	bch2_btree_node_unlock_write(trans, path, path_l(path)->b);
+
 	if (unlikely(ret)) /* raced with another fill? */
 		goto err;
 
 	atomic_long_inc(&bc->nr_keys);
 	six_unlock_write(&ck->c.lock);
 
-	enum six_lock_type lock_want = __btree_lock_want(path, 0);
+	enum six_lock_type lock_want = __btree_lock_want(ck_path, 0);
 	if (lock_want == SIX_LOCK_read)
 		six_lock_downgrade(&ck->c.lock);
-	btree_path_cached_set(trans, path, ck, (enum btree_node_locked_type) lock_want);
-	path->uptodate = BTREE_ITER_UPTODATE;
+	btree_path_cached_set(trans, ck_path, ck, (enum btree_node_locked_type) lock_want);
+	ck_path->uptodate = BTREE_ITER_UPTODATE;
 	return 0;
 err:
 	bkey_cached_free(bc, ck);
-	mark_btree_node_locked_noreset(path, 0, BTREE_NODE_UNLOCKED);
+	mark_btree_node_locked_noreset(ck_path, 0, BTREE_NODE_UNLOCKED);
 
 	return ret;
 }
@@ -293,6 +302,7 @@ static noinline int btree_key_cache_fill(struct btree_trans *trans,
 	int ret;
 
 	bch2_trans_iter_init(trans, &iter, ck_path->btree_id, ck_path->pos,
+			     BTREE_ITER_intent|
 			     BTREE_ITER_key_cache_fill|
 			     BTREE_ITER_cached_nofill);
 	iter.flags &= ~BTREE_ITER_with_journal;
@@ -306,9 +316,19 @@ static noinline int btree_key_cache_fill(struct btree_trans *trans,
 	if (unlikely(ret))
 		goto out;
 
-	ret = btree_key_cache_create(trans, ck_path, k);
+	ret = btree_key_cache_create(trans, btree_iter_path(trans, &iter), ck_path, k);
 	if (ret)
 		goto err;
+
+	if (trace_key_cache_fill_enabled()) {
+		struct printbuf buf = PRINTBUF;
+
+		bch2_bpos_to_text(&buf, ck_path->pos);
+		prt_char(&buf, ' ');
+		bch2_bkey_val_to_text(&buf, trans->c, k);
+		trace_key_cache_fill(trans, buf.buf);
+		printbuf_exit(&buf);
+	}
 out:
 	/* We're not likely to need this iterator again: */
 	bch2_set_btree_iter_dontneed(&iter);
@@ -593,8 +613,18 @@ void bch2_btree_key_cache_drop(struct btree_trans *trans,
 	bkey_cached_free(bc, ck);
 
 	mark_btree_node_locked(trans, path, 0, BTREE_NODE_UNLOCKED);
-	btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
-	path->should_be_locked = false;
+
+	struct btree_path *path2;
+	unsigned i;
+	trans_for_each_path(trans, path2, i)
+		if (path2->l[0].b == (void *) ck) {
+			__bch2_btree_path_unlock(trans, path2);
+			path2->l[0].b = ERR_PTR(-BCH_ERR_no_btree_node_drop);
+			path2->should_be_locked = false;
+			btree_path_set_dirty(path2, BTREE_ITER_NEED_TRAVERSE);
+		}
+
+	bch2_trans_verify_locks(trans);
 }
 
 static unsigned long bch2_btree_key_cache_scan(struct shrinker *shrink,
diff --git a/libbcachefs/btree_locking.c b/libbcachefs/btree_locking.c
index d343df9f..85039314 100644
--- a/libbcachefs/btree_locking.c
+++ b/libbcachefs/btree_locking.c
@@ -818,6 +818,17 @@ void bch2_trans_unlock_long(struct btree_trans *trans)
 	bch2_trans_srcu_unlock(trans);
 }
 
+void bch2_trans_unlock_write(struct btree_trans *trans)
+{
+	struct btree_path *path;
+	unsigned i;
+
+	trans_for_each_path(trans, path, i)
+		for (unsigned l = 0; l < BTREE_MAX_DEPTH; l++)
+			if (btree_node_write_locked(path, l))
+				bch2_btree_node_unlock_write(trans, path, path->l[l].b);
+}
+
 int __bch2_trans_mutex_lock(struct btree_trans *trans,
 			    struct mutex *lock)
 {
@@ -856,6 +867,9 @@ void bch2_btree_path_verify_locks(struct btree_path *path)
 		       (want == BTREE_NODE_UNLOCKED ||
 			have != BTREE_NODE_WRITE_LOCKED) &&
 		       want != have);
+
+		BUG_ON(btree_node_locked(path, l) &&
+		       path->l[l].lock_seq != six_lock_seq(&path->l[l].b->c.lock));
 	}
 }
 
diff --git a/libbcachefs/btree_locking.h b/libbcachefs/btree_locking.h
index 7474ab6c..b54ef48e 100644
--- a/libbcachefs/btree_locking.h
+++ b/libbcachefs/btree_locking.h
@@ -16,6 +16,7 @@
 void bch2_btree_lock_init(struct btree_bkey_cached_common *, enum six_lock_init_flags);
 
 void bch2_trans_unlock_noassert(struct btree_trans *);
+void bch2_trans_unlock_write(struct btree_trans *);
 
 static inline bool is_btree_node(struct btree_path *path, unsigned l)
 {
@@ -75,13 +76,6 @@ static inline void mark_btree_node_locked_noreset(struct btree_path *path,
 	path->nodes_locked |= (type + 1) << (level << 1);
 }
 
-static inline void mark_btree_node_unlocked(struct btree_path *path,
-					    unsigned level)
-{
-	EBUG_ON(btree_node_write_locked(path, level));
-	mark_btree_node_locked_noreset(path, level, BTREE_NODE_UNLOCKED);
-}
-
 static inline void mark_btree_node_locked(struct btree_trans *trans,
 					  struct btree_path *path,
 					  unsigned level,
@@ -124,19 +118,25 @@ static void btree_trans_lock_hold_time_update(struct btree_trans *trans,
 
 /* unlock: */
 
+void bch2_btree_node_unlock_write(struct btree_trans *,
+			struct btree_path *, struct btree *);
+
 static inline void btree_node_unlock(struct btree_trans *trans,
 				     struct btree_path *path, unsigned level)
 {
 	int lock_type = btree_node_locked_type(path, level);
 
 	EBUG_ON(level >= BTREE_MAX_DEPTH);
-	EBUG_ON(lock_type == BTREE_NODE_WRITE_LOCKED);
 
 	if (lock_type != BTREE_NODE_UNLOCKED) {
+		if (unlikely(lock_type == BTREE_NODE_WRITE_LOCKED)) {
+			bch2_btree_node_unlock_write(trans, path, path->l[level].b);
+			lock_type = BTREE_NODE_INTENT_LOCKED;
+		}
 		six_unlock_type(&path->l[level].b->c.lock, lock_type);
 		btree_trans_lock_hold_time_update(trans, path, level);
+		mark_btree_node_locked_noreset(path, level, BTREE_NODE_UNLOCKED);
 	}
-	mark_btree_node_unlocked(path, level);
 }
 
 static inline int btree_path_lowest_level_locked(struct btree_path *path)
@@ -163,27 +163,31 @@ static inline void __bch2_btree_path_unlock(struct btree_trans *trans,
  * succeed:
  */
 static inline void
+__bch2_btree_node_unlock_write(struct btree_trans *trans, struct btree *b)
+{
+	if (!b->c.lock.write_lock_recurse) {
+		struct btree_path *linked;
+		unsigned i;
+
+		trans_for_each_path_with_node(trans, b, linked, i)
+			linked->l[b->c.level].lock_seq++;
+	}
+
+	six_unlock_write(&b->c.lock);
+}
+
+static inline void
 bch2_btree_node_unlock_write_inlined(struct btree_trans *trans, struct btree_path *path,
 				     struct btree *b)
 {
-	struct btree_path *linked;
-	unsigned i;
-
 	EBUG_ON(path->l[b->c.level].b != b);
 	EBUG_ON(path->l[b->c.level].lock_seq != six_lock_seq(&b->c.lock));
 	EBUG_ON(btree_node_locked_type(path, b->c.level) != SIX_LOCK_write);
 
 	mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
-
-	trans_for_each_path_with_node(trans, b, linked, i)
-		linked->l[b->c.level].lock_seq++;
-
-	six_unlock_write(&b->c.lock);
+	__bch2_btree_node_unlock_write(trans, b);
 }
 
-void bch2_btree_node_unlock_write(struct btree_trans *,
-			struct btree_path *, struct btree *);
-
 int bch2_six_check_for_deadlock(struct six_lock *lock, void *p);
 
 /* lock: */
diff --git a/libbcachefs/btree_node_scan.c b/libbcachefs/btree_node_scan.c
index a5282ff1..a7f06dee 100644
--- a/libbcachefs/btree_node_scan.c
+++ b/libbcachefs/btree_node_scan.c
@@ -152,7 +152,7 @@ static inline void found_btree_node_swap(void *_l, void *_r, void *arg)
 	swap(*l, *r);
 }
 
-const struct min_heap_callbacks found_btree_node_heap_cbs = {
+static const struct min_heap_callbacks found_btree_node_heap_cbs = {
 	.less = found_btree_node_cmp_pos_less,
 	.swp = found_btree_node_swap,
 };
diff --git a/libbcachefs/btree_trans_commit.c b/libbcachefs/btree_trans_commit.c
index 9011cc3f..6b79b672 100644
--- a/libbcachefs/btree_trans_commit.c
+++ b/libbcachefs/btree_trans_commit.c
@@ -133,7 +133,7 @@ static inline int bch2_trans_lock_write(struct btree_trans *trans)
 	return 0;
 }
 
-static inline void bch2_trans_unlock_write(struct btree_trans *trans)
+static inline void bch2_trans_unlock_updates_write(struct btree_trans *trans)
 {
 	if (likely(trans->write_locked)) {
 		trans_for_each_update(trans, i)
@@ -249,7 +249,7 @@ static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 		new |= 1 << BTREE_NODE_need_write;
 	} while (!try_cmpxchg(&b->flags, &old, new));
 
-	btree_node_write_if_need(c, b, SIX_LOCK_read);
+	btree_node_write_if_need(trans, b, SIX_LOCK_read);
 	six_unlock_read(&b->c.lock);
 
 	bch2_trans_put(trans);
@@ -384,7 +384,7 @@ btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
 	struct bkey_i *new_k;
 	int ret;
 
-	bch2_trans_unlock_write(trans);
+	bch2_trans_unlock_updates_write(trans);
 	bch2_trans_unlock(trans);
 
 	new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
@@ -479,8 +479,7 @@ static int run_one_mem_trigger(struct btree_trans *trans,
 				old, flags);
 }
 
-static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
-				 bool overwrite)
+static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i)
 {
 	verify_update_old_key(trans, i);
 
@@ -507,10 +506,10 @@ static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_
 		return bch2_key_trigger(trans, i->btree_id, i->level, old, bkey_i_to_s(i->k),
 					BTREE_TRIGGER_insert|
 					BTREE_TRIGGER_overwrite|flags) ?: 1;
-	} else if (overwrite && !i->overwrite_trigger_run) {
+	} else if (!i->overwrite_trigger_run) {
 		i->overwrite_trigger_run = true;
 		return bch2_key_trigger_old(trans, i->btree_id, i->level, old, flags) ?: 1;
-	} else if (!overwrite && !i->insert_trigger_run) {
+	} else if (!i->insert_trigger_run) {
 		i->insert_trigger_run = true;
 		return bch2_key_trigger_new(trans, i->btree_id, i->level, bkey_i_to_s(i->k), flags) ?: 1;
 	} else {
@@ -519,39 +518,45 @@ static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_
 }
 
 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
-			      unsigned btree_id_start)
+			      unsigned *btree_id_updates_start)
 {
-	for (int overwrite = 1; overwrite >= 0; --overwrite) {
-		bool trans_trigger_run;
+	bool trans_trigger_run;
 
-		/*
-		 * Running triggers will append more updates to the list of updates as
-		 * we're walking it:
-		 */
-		do {
-			trans_trigger_run = false;
-
-			for (unsigned i = btree_id_start;
-			     i < trans->nr_updates && trans->updates[i].btree_id <= btree_id;
-			     i++) {
-				if (trans->updates[i].btree_id != btree_id)
-					continue;
+	/*
+	 * Running triggers will append more updates to the list of updates as
+	 * we're walking it:
+	 */
+	do {
+		trans_trigger_run = false;
 
-				int ret = run_one_trans_trigger(trans, trans->updates + i, overwrite);
-				if (ret < 0)
-					return ret;
-				if (ret)
-					trans_trigger_run = true;
+		for (unsigned i = *btree_id_updates_start;
+		     i < trans->nr_updates && trans->updates[i].btree_id <= btree_id;
+		     i++) {
+			if (trans->updates[i].btree_id < btree_id) {
+				*btree_id_updates_start = i;
+				continue;
 			}
-		} while (trans_trigger_run);
-	}
+
+			int ret = run_one_trans_trigger(trans, trans->updates + i);
+			if (ret < 0)
+				return ret;
+			if (ret)
+				trans_trigger_run = true;
+		}
+	} while (trans_trigger_run);
+
+	trans_for_each_update(trans, i)
+		BUG_ON(!(i->flags & BTREE_TRIGGER_norun) &&
+		       i->btree_id == btree_id &&
+		       btree_node_type_has_trans_triggers(i->bkey_type) &&
+		       (!i->insert_trigger_run || !i->overwrite_trigger_run));
 
 	return 0;
 }
 
 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 {
-	unsigned btree_id = 0, btree_id_start = 0;
+	unsigned btree_id = 0, btree_id_updates_start = 0;
 	int ret = 0;
 
 	/*
@@ -565,27 +570,15 @@ static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 		if (btree_id == BTREE_ID_alloc)
 			continue;
 
-		while (btree_id_start < trans->nr_updates &&
-		       trans->updates[btree_id_start].btree_id < btree_id)
-			btree_id_start++;
-
-		ret = run_btree_triggers(trans, btree_id, btree_id_start);
+		ret = run_btree_triggers(trans, btree_id, &btree_id_updates_start);
 		if (ret)
 			return ret;
 	}
 
-	for (unsigned idx = 0; idx < trans->nr_updates; idx++) {
-		struct btree_insert_entry *i = trans->updates + idx;
-
-		if (i->btree_id > BTREE_ID_alloc)
-			break;
-		if (i->btree_id == BTREE_ID_alloc) {
-			ret = run_btree_triggers(trans, BTREE_ID_alloc, idx);
-			if (ret)
-				return ret;
-			break;
-		}
-	}
+	btree_id_updates_start = 0;
+	ret = run_btree_triggers(trans, BTREE_ID_alloc, &btree_id_updates_start);
+	if (ret)
+		return ret;
 
 #ifdef CONFIG_BCACHEFS_DEBUG
 	trans_for_each_update(trans, i)
@@ -875,7 +868,7 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags
 	if (!ret && unlikely(trans->journal_replay_not_finished))
 		bch2_drop_overwrites_from_journal(trans);
 
-	bch2_trans_unlock_write(trans);
+	bch2_trans_unlock_updates_write(trans);
 
 	if (!ret && trans->journal_pin)
 		bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
diff --git a/libbcachefs/btree_types.h b/libbcachefs/btree_types.h
index baab5288..a6f251eb 100644
--- a/libbcachefs/btree_types.h
+++ b/libbcachefs/btree_types.h
@@ -790,53 +790,64 @@ static inline bool btree_node_type_has_triggers(enum btree_node_type type)
 	return BIT_ULL(type) & BTREE_NODE_TYPE_HAS_TRIGGERS;
 }
 
-static inline bool btree_node_type_is_extents(enum btree_node_type type)
+static inline bool btree_id_is_extents(enum btree_id btree)
 {
 	const u64 mask = 0
-#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_ID_EXTENTS)) << (nr + 1))
+#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_IS_extents)) << nr)
 	BCH_BTREE_IDS()
 #undef x
 	;
 
-	return BIT_ULL(type) & mask;
+	return BIT_ULL(btree) & mask;
 }
 
-static inline bool btree_id_is_extents(enum btree_id btree)
+static inline bool btree_node_type_is_extents(enum btree_node_type type)
+{
+	return type != BKEY_TYPE_btree && btree_id_is_extents(type - 1);
+}
+
+static inline bool btree_type_has_snapshots(enum btree_id btree)
 {
-	return btree_node_type_is_extents(__btree_node_type(0, btree));
+	const u64 mask = 0
+#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_IS_snapshots)) << nr)
+	BCH_BTREE_IDS()
+#undef x
+	;
+
+	return BIT_ULL(btree) & mask;
 }
 
-static inline bool btree_type_has_snapshots(enum btree_id id)
+static inline bool btree_type_has_snapshot_field(enum btree_id btree)
 {
 	const u64 mask = 0
-#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_ID_SNAPSHOTS)) << nr)
+#define x(name, nr, flags, ...)	|((!!((flags) & (BTREE_IS_snapshot_field|BTREE_IS_snapshots))) << nr)
 	BCH_BTREE_IDS()
 #undef x
 	;
 
-	return BIT_ULL(id) & mask;
+	return BIT_ULL(btree) & mask;
 }
 
-static inline bool btree_type_has_snapshot_field(enum btree_id id)
+static inline bool btree_type_has_ptrs(enum btree_id btree)
 {
 	const u64 mask = 0
-#define x(name, nr, flags, ...)	|((!!((flags) & (BTREE_ID_SNAPSHOT_FIELD|BTREE_ID_SNAPSHOTS))) << nr)
+#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_IS_data)) << nr)
 	BCH_BTREE_IDS()
 #undef x
 	;
 
-	return BIT_ULL(id) & mask;
+	return BIT_ULL(btree) & mask;
 }
 
-static inline bool btree_type_has_ptrs(enum btree_id id)
+static inline bool btree_type_uses_write_buffer(enum btree_id btree)
 {
 	const u64 mask = 0
-#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_ID_DATA)) << nr)
+#define x(name, nr, flags, ...)	|((!!((flags) & BTREE_IS_write_buffer)) << nr)
 	BCH_BTREE_IDS()
 #undef x
 	;
 
-	return BIT_ULL(id) & mask;
+	return BIT_ULL(btree) & mask;
 }
 
 struct btree_root {
diff --git a/libbcachefs/btree_update.c b/libbcachefs/btree_update.c
index b1da6dba..13d794f2 100644
--- a/libbcachefs/btree_update.c
+++ b/libbcachefs/btree_update.c
@@ -823,10 +823,17 @@ int bch2_btree_bit_mod_buffered(struct btree_trans *trans, enum btree_id btree,
 	return bch2_trans_update_buffered(trans, btree, &k);
 }
 
-static int __bch2_trans_log_msg(struct btree_trans *trans, struct printbuf *buf, unsigned u64s)
+int bch2_trans_log_msg(struct btree_trans *trans, struct printbuf *buf)
 {
+	unsigned u64s = DIV_ROUND_UP(buf->pos, sizeof(u64));
+	prt_chars(buf, '\0', u64s * sizeof(u64) - buf->pos);
+
+	int ret = buf->allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
+	if (ret)
+		return ret;
+
 	struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, jset_u64s(u64s));
-	int ret = PTR_ERR_OR_ZERO(e);
+	ret = PTR_ERR_OR_ZERO(e);
 	if (ret)
 		return ret;
 
@@ -862,7 +869,7 @@ __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
 		c->journal.early_journal_entries.nr += jset_u64s(u64s);
 	} else {
 		ret = bch2_trans_commit_do(c, NULL, NULL, commit_flags,
-			__bch2_trans_log_msg(trans, &buf, u64s));
+			bch2_trans_log_msg(trans, &buf));
 	}
 err:
 	printbuf_exit(&buf);
diff --git a/libbcachefs/btree_update.h b/libbcachefs/btree_update.h
index 58df2019..8f22ef9a 100644
--- a/libbcachefs/btree_update.h
+++ b/libbcachefs/btree_update.h
@@ -159,6 +159,7 @@ void bch2_trans_commit_hook(struct btree_trans *,
 			    struct btree_trans_commit_hook *);
 int __bch2_trans_commit(struct btree_trans *, unsigned);
 
+int bch2_trans_log_msg(struct btree_trans *, struct printbuf *);
 __printf(2, 3) int bch2_fs_log_msg(struct bch_fs *, const char *, ...);
 __printf(2, 3) int bch2_journal_log_msg(struct bch_fs *, const char *, ...);
 
diff --git a/libbcachefs/btree_update_interior.c b/libbcachefs/btree_update_interior.c
index 7d9dab95..f4aeadbe 100644
--- a/libbcachefs/btree_update_interior.c
+++ b/libbcachefs/btree_update_interior.c
@@ -238,7 +238,6 @@ static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 				       struct btree *b)
 {
 	struct bch_fs *c = trans->c;
-	unsigned i, level = b->c.level;
 
 	bch2_btree_node_lock_write_nofail(trans, path, &b->c);
 
@@ -249,13 +248,9 @@ static void bch2_btree_node_free_inmem(struct btree_trans *trans,
 	mutex_unlock(&c->btree_cache.lock);
 
 	six_unlock_write(&b->c.lock);
-	mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
+	mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 
-	trans_for_each_path(trans, path, i)
-		if (path->l[level].b == b) {
-			btree_node_unlock(trans, path, level);
-			path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
-		}
+	bch2_trans_node_drop(trans, b);
 }
 
 static void bch2_btree_node_free_never_used(struct btree_update *as,
@@ -264,8 +259,6 @@ static void bch2_btree_node_free_never_used(struct btree_update *as,
 {
 	struct bch_fs *c = as->c;
 	struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
-	struct btree_path *path;
-	unsigned i, level = b->c.level;
 
 	BUG_ON(!list_empty(&b->write_blocked));
 	BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
@@ -287,11 +280,7 @@ static void bch2_btree_node_free_never_used(struct btree_update *as,
 
 	six_unlock_intent(&b->c.lock);
 
-	trans_for_each_path(trans, path, i)
-		if (path->l[level].b == b) {
-			btree_node_unlock(trans, path, level);
-			path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
-		}
+	bch2_trans_node_drop(trans, b);
 }
 
 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
@@ -803,7 +792,7 @@ err:
 		mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
 		six_unlock_write(&b->c.lock);
 
-		btree_node_write_if_need(c, b, SIX_LOCK_intent);
+		btree_node_write_if_need(trans, b, SIX_LOCK_intent);
 		btree_node_unlock(trans, path, b->c.level);
 		bch2_path_put(trans, path_idx, true);
 	}
@@ -824,7 +813,7 @@ err:
 		b = as->new_nodes[i];
 
 		btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
-		btree_node_write_if_need(c, b, SIX_LOCK_read);
+		btree_node_write_if_need(trans, b, SIX_LOCK_read);
 		six_unlock_read(&b->c.lock);
 	}
 
@@ -1709,14 +1698,14 @@ static int btree_split(struct btree_update *as, struct btree_trans *trans,
 
 	if (n3) {
 		bch2_btree_update_get_open_buckets(as, n3);
-		bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
+		bch2_btree_node_write_trans(trans, n3, SIX_LOCK_intent, 0);
 	}
 	if (n2) {
 		bch2_btree_update_get_open_buckets(as, n2);
-		bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
+		bch2_btree_node_write_trans(trans, n2, SIX_LOCK_intent, 0);
 	}
 	bch2_btree_update_get_open_buckets(as, n1);
-	bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
+	bch2_btree_node_write_trans(trans, n1, SIX_LOCK_intent, 0);
 
 	/*
 	 * The old node must be freed (in memory) _before_ unlocking the new
@@ -1911,7 +1900,7 @@ static void __btree_increase_depth(struct btree_update *as, struct btree_trans *
 	BUG_ON(ret);
 
 	bch2_btree_update_get_open_buckets(as, n);
-	bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
+	bch2_btree_node_write_trans(trans, n, SIX_LOCK_intent, 0);
 	bch2_trans_node_add(trans, path, n);
 	six_unlock_intent(&n->c.lock);
 
@@ -2104,7 +2093,7 @@ int __bch2_foreground_maybe_merge(struct btree_trans *trans,
 	bch2_trans_verify_paths(trans);
 
 	bch2_btree_update_get_open_buckets(as, n);
-	bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
+	bch2_btree_node_write_trans(trans, n, SIX_LOCK_intent, 0);
 
 	bch2_btree_node_free_inmem(trans, trans->paths + path, b);
 	bch2_btree_node_free_inmem(trans, trans->paths + sib_path, m);
@@ -2181,7 +2170,7 @@ int bch2_btree_node_rewrite(struct btree_trans *trans,
 	bch2_btree_interior_update_will_free_node(as, b);
 
 	bch2_btree_update_get_open_buckets(as, n);
-	bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
+	bch2_btree_node_write_trans(trans, n, SIX_LOCK_intent, 0);
 
 	bch2_btree_node_free_inmem(trans, btree_iter_path(trans, iter), b);
 
@@ -2291,7 +2280,8 @@ void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
 	bool now = false, pending = false;
 
 	spin_lock(&c->btree_node_rewrites_lock);
-	if (bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
+	if (c->curr_recovery_pass > BCH_RECOVERY_PASS_journal_replay &&
+	    bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
 		list_add(&a->list, &c->btree_node_rewrites);
 		now = true;
 	} else if (!test_bit(BCH_FS_may_go_rw, &c->flags)) {
diff --git a/libbcachefs/btree_write_buffer.c b/libbcachefs/btree_write_buffer.c
index 49ce2d1e..b56c4987 100644
--- a/libbcachefs/btree_write_buffer.c
+++ b/libbcachefs/btree_write_buffer.c
@@ -312,6 +312,8 @@ static int bch2_btree_write_buffer_flush_locked(struct btree_trans *trans)
 	darray_for_each(wb->sorted, i) {
 		struct btree_write_buffered_key *k = &wb->flushing.keys.data[i->idx];
 
+		BUG_ON(!btree_type_uses_write_buffer(k->btree));
+
 		for (struct wb_key_ref *n = i + 1; n < min(i + 4, &darray_top(wb->sorted)); n++)
 			prefetch(&wb->flushing.keys.data[n->idx]);
 
@@ -632,6 +634,14 @@ int bch2_btree_write_buffer_maybe_flush(struct btree_trans *trans,
 	bch2_bkey_buf_init(&tmp);
 
 	if (!bkey_and_val_eq(referring_k, bkey_i_to_s_c(last_flushed->k))) {
+		if (trace_write_buffer_maybe_flush_enabled()) {
+			struct printbuf buf = PRINTBUF;
+
+			bch2_bkey_val_to_text(&buf, c, referring_k);
+			trace_write_buffer_maybe_flush(trans, _RET_IP_, buf.buf);
+			printbuf_exit(&buf);
+		}
+
 		bch2_bkey_buf_reassemble(&tmp, c, referring_k);
 
 		if (bkey_is_btree_ptr(referring_k.k)) {
diff --git a/libbcachefs/buckets.c b/libbcachefs/buckets.c
index 1e43ece2..345b117a 100644
--- a/libbcachefs/buckets.c
+++ b/libbcachefs/buckets.c
@@ -262,8 +262,6 @@ int bch2_check_fix_ptrs(struct btree_trans *trans,
 	struct printbuf buf = PRINTBUF;
 	int ret = 0;
 
-	percpu_down_read(&c->mark_lock);
-
 	bkey_for_each_ptr_decode(k.k, ptrs_c, p, entry_c) {
 		ret = bch2_check_fix_ptr(trans, k, p, entry_c, &do_update);
 		if (ret)
@@ -364,7 +362,6 @@ found:
 			bch_info(c, "new key %s", buf.buf);
 		}
 
-		percpu_up_read(&c->mark_lock);
 		struct btree_iter iter;
 		bch2_trans_node_iter_init(trans, &iter, btree, new->k.p, 0, level,
 					  BTREE_ITER_intent|BTREE_ITER_all_snapshots);
@@ -373,8 +370,6 @@ found:
 					  BTREE_UPDATE_internal_snapshot_node|
 					  BTREE_TRIGGER_norun);
 		bch2_trans_iter_exit(trans, &iter);
-		percpu_down_read(&c->mark_lock);
-
 		if (ret)
 			goto err;
 
@@ -382,7 +377,6 @@ found:
 			bch2_btree_node_update_key_early(trans, btree, level - 1, k, new);
 	}
 err:
-	percpu_up_read(&c->mark_lock);
 	printbuf_exit(&buf);
 	return ret;
 }
@@ -547,7 +541,8 @@ static int __mark_pointer(struct btree_trans *trans, struct bch_dev *ca,
 			  struct bkey_s_c k,
 			  const struct extent_ptr_decoded *p,
 			  s64 sectors, enum bch_data_type ptr_data_type,
-			  struct bch_alloc_v4 *a)
+			  struct bch_alloc_v4 *a,
+			  bool insert)
 {
 	u32 *dst_sectors = p->has_ec	? &a->stripe_sectors :
 		!p->ptr.cached		? &a->dirty_sectors :
@@ -557,8 +552,8 @@ static int __mark_pointer(struct btree_trans *trans, struct bch_dev *ca,
 
 	if (ret)
 		return ret;
-
-	alloc_data_type_set(a, ptr_data_type);
+	if (insert)
+		alloc_data_type_set(a, ptr_data_type);
 	return 0;
 }
 
@@ -591,7 +586,7 @@ static int bch2_trigger_pointer(struct btree_trans *trans,
 	if (flags & BTREE_TRIGGER_transactional) {
 		struct bkey_i_alloc_v4 *a = bch2_trans_start_alloc_update(trans, bucket, 0);
 		ret = PTR_ERR_OR_ZERO(a) ?:
-			__mark_pointer(trans, ca, k, &p, *sectors, bp.v.data_type, &a->v);
+			__mark_pointer(trans, ca, k, &p, *sectors, bp.v.data_type, &a->v, insert);
 		if (ret)
 			goto err;
 
@@ -603,22 +598,19 @@ static int bch2_trigger_pointer(struct btree_trans *trans,
 	}
 
 	if (flags & BTREE_TRIGGER_gc) {
-		percpu_down_read(&c->mark_lock);
 		struct bucket *g = gc_bucket(ca, bucket.offset);
 		if (bch2_fs_inconsistent_on(!g, c, "reference to invalid bucket on device %u\n  %s",
 					    p.ptr.dev,
 					    (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 			ret = -BCH_ERR_trigger_pointer;
-			goto err_unlock;
+			goto err;
 		}
 
 		bucket_lock(g);
 		struct bch_alloc_v4 old = bucket_m_to_alloc(*g), new = old;
-		ret = __mark_pointer(trans, ca, k, &p, *sectors, bp.v.data_type, &new);
+		ret = __mark_pointer(trans, ca, k, &p, *sectors, bp.v.data_type, &new, insert);
 		alloc_to_bucket(g, new);
 		bucket_unlock(g);
-err_unlock:
-		percpu_up_read(&c->mark_lock);
 
 		if (!ret)
 			ret = bch2_alloc_key_to_dev_counters(trans, ca, &old, &new, flags);
@@ -996,11 +988,10 @@ static int bch2_mark_metadata_bucket(struct btree_trans *trans, struct bch_dev *
 	struct bch_fs *c = trans->c;
 	int ret = 0;
 
-	percpu_down_read(&c->mark_lock);
 	struct bucket *g = gc_bucket(ca, b);
 	if (bch2_fs_inconsistent_on(!g, c, "reference to invalid bucket on device %u when marking metadata type %s",
 				    ca->dev_idx, bch2_data_type_str(data_type)))
-		goto err_unlock;
+		goto err;
 
 	bucket_lock(g);
 	struct bch_alloc_v4 old = bucket_m_to_alloc(*g);
@@ -1010,26 +1001,24 @@ static int bch2_mark_metadata_bucket(struct btree_trans *trans, struct bch_dev *
 			"different types of data in same bucket: %s, %s",
 			bch2_data_type_str(g->data_type),
 			bch2_data_type_str(data_type)))
-		goto err;
+		goto err_unlock;
 
 	if (bch2_fs_inconsistent_on((u64) g->dirty_sectors + sectors > ca->mi.bucket_size, c,
 			"bucket %u:%llu gen %u data type %s sector count overflow: %u + %u > bucket size",
 			ca->dev_idx, b, g->gen,
 			bch2_data_type_str(g->data_type ?: data_type),
 			g->dirty_sectors, sectors))
-		goto err;
+		goto err_unlock;
 
 	g->data_type = data_type;
 	g->dirty_sectors += sectors;
 	struct bch_alloc_v4 new = bucket_m_to_alloc(*g);
 	bucket_unlock(g);
-	percpu_up_read(&c->mark_lock);
 	ret = bch2_alloc_key_to_dev_counters(trans, ca, &old, &new, flags);
 	return ret;
-err:
-	bucket_unlock(g);
 err_unlock:
-	percpu_up_read(&c->mark_lock);
+	bucket_unlock(g);
+err:
 	return -BCH_ERR_metadata_bucket_inconsistency;
 }
 
@@ -1269,7 +1258,7 @@ int bch2_buckets_nouse_alloc(struct bch_fs *c)
 	for_each_member_device(c, ca) {
 		BUG_ON(ca->buckets_nouse);
 
-		ca->buckets_nouse = kvmalloc(BITS_TO_LONGS(ca->mi.nbuckets) *
+		ca->buckets_nouse = bch2_kvmalloc(BITS_TO_LONGS(ca->mi.nbuckets) *
 					    sizeof(unsigned long),
 					    GFP_KERNEL|__GFP_ZERO);
 		if (!ca->buckets_nouse) {
@@ -1295,10 +1284,14 @@ int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 	bool resize = ca->bucket_gens != NULL;
 	int ret;
 
-	BUG_ON(resize && ca->buckets_nouse);
+	if (resize)
+		lockdep_assert_held(&c->state_lock);
+
+	if (resize && ca->buckets_nouse)
+		return -BCH_ERR_no_resize_with_buckets_nouse;
 
-	bucket_gens = kvmalloc(struct_size(bucket_gens, b, nbuckets),
-			       GFP_KERNEL|__GFP_ZERO);
+	bucket_gens = bch2_kvmalloc(struct_size(bucket_gens, b, nbuckets),
+				    GFP_KERNEL|__GFP_ZERO);
 	if (!bucket_gens) {
 		ret = -BCH_ERR_ENOMEM_bucket_gens;
 		goto err;
@@ -1309,11 +1302,6 @@ int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 	bucket_gens->nbuckets_minus_first =
 		bucket_gens->nbuckets - bucket_gens->first_bucket;
 
-	if (resize) {
-		down_write(&ca->bucket_lock);
-		percpu_down_write(&c->mark_lock);
-	}
-
 	old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
 
 	if (resize) {
@@ -1331,11 +1319,6 @@ int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
 
 	nbuckets = ca->mi.nbuckets;
 
-	if (resize) {
-		percpu_up_write(&c->mark_lock);
-		up_write(&ca->bucket_lock);
-	}
-
 	ret = 0;
 err:
 	if (bucket_gens)
diff --git a/libbcachefs/buckets.h b/libbcachefs/buckets.h
index 3bebc4c3..a9acdd6c 100644
--- a/libbcachefs/buckets.h
+++ b/libbcachefs/buckets.h
@@ -82,16 +82,15 @@ static inline void bucket_lock(struct bucket *b)
 
 static inline struct bucket *gc_bucket(struct bch_dev *ca, size_t b)
 {
-	return genradix_ptr(&ca->buckets_gc, b);
+	return bucket_valid(ca, b)
+		? genradix_ptr(&ca->buckets_gc, b)
+		: NULL;
 }
 
 static inline struct bucket_gens *bucket_gens(struct bch_dev *ca)
 {
 	return rcu_dereference_check(ca->bucket_gens,
-				     !ca->fs ||
-				     percpu_rwsem_is_held(&ca->fs->mark_lock) ||
-				     lockdep_is_held(&ca->fs->state_lock) ||
-				     lockdep_is_held(&ca->bucket_lock));
+				     lockdep_is_held(&ca->fs->state_lock));
 }
 
 static inline u8 *bucket_gen(struct bch_dev *ca, size_t b)
diff --git a/libbcachefs/clock.c b/libbcachefs/clock.c
index 1d6b691e..1f8e035d 100644
--- a/libbcachefs/clock.c
+++ b/libbcachefs/clock.c
@@ -14,21 +14,13 @@ static inline bool io_timer_cmp(const void *l, const void *r, void __always_unus
 	return (*_l)->expire < (*_r)->expire;
 }
 
-static inline void io_timer_swp(void *l, void *r, void __always_unused *args)
-{
-	struct io_timer **_l = (struct io_timer **)l;
-	struct io_timer **_r = (struct io_timer **)r;
-
-	swap(*_l, *_r);
-}
+static const struct min_heap_callbacks callbacks = {
+	.less = io_timer_cmp,
+	.swp = NULL,
+};
 
 void bch2_io_timer_add(struct io_clock *clock, struct io_timer *timer)
 {
-	const struct min_heap_callbacks callbacks = {
-		.less = io_timer_cmp,
-		.swp = io_timer_swp,
-	};
-
 	spin_lock(&clock->timer_lock);
 
 	if (time_after_eq64((u64) atomic64_read(&clock->now), timer->expire)) {
@@ -48,11 +40,6 @@ out:
 
 void bch2_io_timer_del(struct io_clock *clock, struct io_timer *timer)
 {
-	const struct min_heap_callbacks callbacks = {
-		.less = io_timer_cmp,
-		.swp = io_timer_swp,
-	};
-
 	spin_lock(&clock->timer_lock);
 
 	for (size_t i = 0; i < clock->timers.nr; i++)
@@ -142,10 +129,6 @@ void bch2_kthread_io_clock_wait(struct io_clock *clock,
 static struct io_timer *get_expired_timer(struct io_clock *clock, u64 now)
 {
 	struct io_timer *ret = NULL;
-	const struct min_heap_callbacks callbacks = {
-		.less = io_timer_cmp,
-		.swp = io_timer_swp,
-	};
 
 	if (clock->timers.nr &&
 	    time_after_eq64(now, clock->timers.data[0]->expire)) {
diff --git a/libbcachefs/data_update.c b/libbcachefs/data_update.c
index 31b2aeb0..58521493 100644
--- a/libbcachefs/data_update.c
+++ b/libbcachefs/data_update.c
@@ -620,7 +620,7 @@ int bch2_data_update_init(struct btree_trans *trans,
 	 * and we have to check for this because we go rw before repairing the
 	 * snapshots table - just skip it, we can move it later.
 	 */
-	if (unlikely(k.k->p.snapshot && !bch2_snapshot_equiv(c, k.k->p.snapshot)))
+	if (unlikely(k.k->p.snapshot && !bch2_snapshot_exists(c, k.k->p.snapshot)))
 		return -BCH_ERR_data_update_done;
 
 	if (!bkey_get_dev_refs(c, k))
diff --git a/libbcachefs/disk_accounting.c b/libbcachefs/disk_accounting.c
index a0e49cbe..b32e91ba 100644
--- a/libbcachefs/disk_accounting.c
+++ b/libbcachefs/disk_accounting.c
@@ -79,6 +79,8 @@ static inline void accounting_key_init(struct bkey_i *k, struct disk_accounting_
 	memcpy_u64s_small(acc->v.d, d, nr);
 }
 
+static int bch2_accounting_update_sb_one(struct bch_fs *, struct bpos);
+
 int bch2_disk_accounting_mod(struct btree_trans *trans,
 			     struct disk_accounting_pos *k,
 			     s64 *d, unsigned nr, bool gc)
@@ -96,9 +98,16 @@ int bch2_disk_accounting_mod(struct btree_trans *trans,
 
 	accounting_key_init(&k_i.k, k, d, nr);
 
-	return likely(!gc)
-		? bch2_trans_update_buffered(trans, BTREE_ID_accounting, &k_i.k)
-		: bch2_accounting_mem_add(trans, bkey_i_to_s_c_accounting(&k_i.k), true);
+	if (unlikely(gc)) {
+		int ret = bch2_accounting_mem_add(trans, bkey_i_to_s_c_accounting(&k_i.k), true);
+		if (ret == -BCH_ERR_btree_insert_need_mark_replicas)
+			ret = drop_locks_do(trans,
+				bch2_accounting_update_sb_one(trans->c, disk_accounting_pos_to_bpos(k))) ?:
+				bch2_accounting_mem_add(trans, bkey_i_to_s_c_accounting(&k_i.k), true);
+		return ret;
+	} else {
+		return bch2_trans_update_buffered(trans, BTREE_ID_accounting, &k_i.k);
+	}
 }
 
 int bch2_mod_dev_cached_sectors(struct btree_trans *trans,
@@ -471,32 +480,6 @@ int bch2_fs_accounting_read(struct bch_fs *c, darray_char *out_buf, unsigned acc
 	return ret;
 }
 
-void bch2_fs_accounting_to_text(struct printbuf *out, struct bch_fs *c)
-{
-	struct bch_accounting_mem *acc = &c->accounting;
-
-	percpu_down_read(&c->mark_lock);
-	out->atomic++;
-
-	eytzinger0_for_each(i, acc->k.nr) {
-		struct disk_accounting_pos acc_k;
-		bpos_to_disk_accounting_pos(&acc_k, acc->k.data[i].pos);
-
-		bch2_accounting_key_to_text(out, &acc_k);
-
-		u64 v[BCH_ACCOUNTING_MAX_COUNTERS];
-		bch2_accounting_mem_read_counters(acc, i, v, ARRAY_SIZE(v), false);
-
-		prt_str(out, ":");
-		for (unsigned j = 0; j < acc->k.data[i].nr_counters; j++)
-			prt_printf(out, " %llu", v[j]);
-		prt_newline(out);
-	}
-
-	--out->atomic;
-	percpu_up_read(&c->mark_lock);
-}
-
 static void bch2_accounting_free_counters(struct bch_accounting_mem *acc, bool gc)
 {
 	darray_for_each(acc->k, e) {
@@ -931,10 +914,13 @@ void bch2_verify_accounting_clean(struct bch_fs *c)
 			bpos_to_disk_accounting_pos(&acc_k, k.k->p);
 
 			if (acc_k.type >= BCH_DISK_ACCOUNTING_TYPE_NR)
-				continue;
+				break;
 
-			if (acc_k.type == BCH_DISK_ACCOUNTING_inum)
+			if (!bch2_accounting_is_mem(acc_k)) {
+				struct disk_accounting_pos next = { .type = acc_k.type + 1 };
+				bch2_btree_iter_set_pos(&iter, disk_accounting_pos_to_bpos(&next));
 				continue;
+			}
 
 			bch2_accounting_mem_read(c, k.k->p, v, nr);
 
diff --git a/libbcachefs/disk_accounting.h b/libbcachefs/disk_accounting.h
index 4a3d9ff8..5360cbb3 100644
--- a/libbcachefs/disk_accounting.h
+++ b/libbcachefs/disk_accounting.h
@@ -138,7 +138,8 @@ static inline int bch2_accounting_mem_mod_locked(struct btree_trans *trans,
 	bpos_to_disk_accounting_pos(&acc_k, a.k->p);
 	bool gc = mode == BCH_ACCOUNTING_gc;
 
-	EBUG_ON(gc && !acc->gc_running);
+	if (gc && !acc->gc_running)
+		return 0;
 
 	if (!bch2_accounting_is_mem(acc_k))
 		return 0;
@@ -255,7 +256,6 @@ static inline void bch2_accounting_trans_commit_revert(struct btree_trans *trans
 
 int bch2_fs_replicas_usage_read(struct bch_fs *, darray_char *);
 int bch2_fs_accounting_read(struct bch_fs *, darray_char *, unsigned);
-void bch2_fs_accounting_to_text(struct printbuf *, struct bch_fs *);
 
 int bch2_gc_accounting_start(struct bch_fs *);
 int bch2_gc_accounting_done(struct bch_fs *);
diff --git a/libbcachefs/ec.c b/libbcachefs/ec.c
index 6b297f90..d2a5e76e 100644
--- a/libbcachefs/ec.c
+++ b/libbcachefs/ec.c
@@ -305,13 +305,12 @@ static int mark_stripe_bucket(struct btree_trans *trans,
 	}
 
 	if (flags & BTREE_TRIGGER_gc) {
-		percpu_down_read(&c->mark_lock);
 		struct bucket *g = gc_bucket(ca, bucket.offset);
 		if (bch2_fs_inconsistent_on(!g, c, "reference to invalid bucket on device %u\n  %s",
 					    ptr->dev,
 					    (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf))) {
 			ret = -BCH_ERR_mark_stripe;
-			goto err_unlock;
+			goto err;
 		}
 
 		bucket_lock(g);
@@ -319,8 +318,7 @@ static int mark_stripe_bucket(struct btree_trans *trans,
 		ret = __mark_stripe_bucket(trans, ca, s, ptr_idx, deleting, bucket, &new, flags);
 		alloc_to_bucket(g, new);
 		bucket_unlock(g);
-err_unlock:
-		percpu_up_read(&c->mark_lock);
+
 		if (!ret)
 			ret = bch2_alloc_key_to_dev_counters(trans, ca, &old, &new, flags);
 	}
@@ -1058,6 +1056,11 @@ static inline void ec_stripes_heap_swap(void *l, void *r, void *h)
 	ec_stripes_heap_set_backpointer(_h, j);
 }
 
+static const struct min_heap_callbacks callbacks = {
+	.less = ec_stripes_heap_cmp,
+	.swp = ec_stripes_heap_swap,
+};
+
 static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 {
 	ec_stripes_heap *h = &c->ec_stripes_heap;
@@ -1070,11 +1073,6 @@ static void heap_verify_backpointer(struct bch_fs *c, size_t idx)
 void bch2_stripes_heap_del(struct bch_fs *c,
 			   struct stripe *m, size_t idx)
 {
-	const struct min_heap_callbacks callbacks = {
-		.less = ec_stripes_heap_cmp,
-		.swp = ec_stripes_heap_swap,
-	};
-
 	mutex_lock(&c->ec_stripes_heap_lock);
 	heap_verify_backpointer(c, idx);
 
@@ -1085,11 +1083,6 @@ void bch2_stripes_heap_del(struct bch_fs *c,
 void bch2_stripes_heap_insert(struct bch_fs *c,
 			      struct stripe *m, size_t idx)
 {
-	const struct min_heap_callbacks callbacks = {
-		.less = ec_stripes_heap_cmp,
-		.swp = ec_stripes_heap_swap,
-	};
-
 	mutex_lock(&c->ec_stripes_heap_lock);
 	BUG_ON(min_heap_full(&c->ec_stripes_heap));
 
@@ -1108,10 +1101,6 @@ void bch2_stripes_heap_insert(struct bch_fs *c,
 void bch2_stripes_heap_update(struct bch_fs *c,
 			      struct stripe *m, size_t idx)
 {
-	const struct min_heap_callbacks callbacks = {
-		.less = ec_stripes_heap_cmp,
-		.swp = ec_stripes_heap_swap,
-	};
 	ec_stripes_heap *h = &c->ec_stripes_heap;
 	bool do_deletes;
 	size_t i;
diff --git a/libbcachefs/errcode.h b/libbcachefs/errcode.h
index 775425df..4590cd0c 100644
--- a/libbcachefs/errcode.h
+++ b/libbcachefs/errcode.h
@@ -118,6 +118,7 @@
 	x(ENOENT,			ENOENT_dev_not_found)			\
 	x(ENOENT,			ENOENT_dev_idx_not_found)		\
 	x(ENOENT,			ENOENT_inode_no_backpointer)		\
+	x(ENOENT,			ENOENT_no_snapshot_tree_subvol)		\
 	x(ENOTEMPTY,			ENOTEMPTY_dir_not_empty)		\
 	x(ENOTEMPTY,			ENOTEMPTY_subvol_not_empty)		\
 	x(EEXIST,			EEXIST_str_hash_set)			\
@@ -196,6 +197,9 @@
 	x(EINVAL,			opt_parse_error)			\
 	x(EINVAL,			remove_with_metadata_missing_unimplemented)\
 	x(EINVAL,			remove_would_lose_data)			\
+	x(EINVAL,			no_resize_with_buckets_nouse)		\
+	x(EINVAL,			inode_unpack_error)			\
+	x(EINVAL,			varint_decode_error)			\
 	x(EROFS,			erofs_trans_commit)			\
 	x(EROFS,			erofs_no_writes)			\
 	x(EROFS,			erofs_journal_err)			\
@@ -313,6 +317,7 @@ static inline long bch2_err_class(long err)
 
 #define BLK_STS_REMOVED		((__force blk_status_t)128)
 
+#include <linux/blk_types.h>
 const char *bch2_blk_status_to_str(blk_status_t);
 
 #endif /* _BCACHFES_ERRCODE_H */
diff --git a/libbcachefs/fs-common.c b/libbcachefs/fs-common.c
index 7d279f21..2c3d46ac 100644
--- a/libbcachefs/fs-common.c
+++ b/libbcachefs/fs-common.c
@@ -574,6 +574,11 @@ static inline void prt_bytes_reversed(struct printbuf *out, const void *b, unsig
 	printbuf_nul_terminate(out);
 }
 
+static inline void prt_str_reversed(struct printbuf *out, const char *s)
+{
+	prt_bytes_reversed(out, s, strlen(s));
+}
+
 static inline void reverse_bytes(void *b, size_t n)
 {
 	char *e = b + n, *s = b;
@@ -596,17 +601,20 @@ int bch2_inum_to_path(struct btree_trans *trans, subvol_inum inum, struct printb
 		struct bch_inode_unpacked inode;
 		ret = bch2_inode_find_by_inum_trans(trans, inum, &inode);
 		if (ret)
-			goto err;
+			goto disconnected;
 
 		if (!inode.bi_dir && !inode.bi_dir_offset) {
 			ret = -BCH_ERR_ENOENT_inode_no_backpointer;
-			goto err;
+			goto disconnected;
 		}
 
+		inum.subvol	= inode.bi_parent_subvol ?: inum.subvol;
+		inum.inum	= inode.bi_dir;
+
 		u32 snapshot;
 		ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 		if (ret)
-			goto err;
+			goto disconnected;
 
 		struct btree_iter d_iter;
 		struct bkey_s_c_dirent d = bch2_bkey_get_iter_typed(trans, &d_iter,
@@ -614,23 +622,19 @@ int bch2_inum_to_path(struct btree_trans *trans, subvol_inum inum, struct printb
 				0, dirent);
 		ret = bkey_err(d.s_c);
 		if (ret)
-			goto err;
+			goto disconnected;
 
 		struct qstr dirent_name = bch2_dirent_get_name(d);
 		prt_bytes_reversed(path, dirent_name.name, dirent_name.len);
 
 		prt_char(path, '/');
 
-		if (d.v->d_type == DT_SUBVOL)
-			inum.subvol = le32_to_cpu(d.v->d_parent_subvol);
-		inum.inum = d.k->p.inode;
-
 		bch2_trans_iter_exit(trans, &d_iter);
 	}
 
 	if (orig_pos == path->pos)
 		prt_char(path, '/');
-
+out:
 	ret = path->allocation_failure ? -ENOMEM : 0;
 	if (ret)
 		goto err;
@@ -639,4 +643,10 @@ int bch2_inum_to_path(struct btree_trans *trans, subvol_inum inum, struct printb
 	return 0;
 err:
 	return ret;
+disconnected:
+	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
+		goto err;
+
+	prt_str_reversed(path, "(disconnected)");
+	goto out;
 }
diff --git a/libbcachefs/fs-io-buffered.c b/libbcachefs/fs-io-buffered.c
index ff8b8df5..ab1d5db2 100644
--- a/libbcachefs/fs-io-buffered.c
+++ b/libbcachefs/fs-io-buffered.c
@@ -625,15 +625,6 @@ do_io:
 		BUG_ON(!bio_add_folio(&w->io->op.wbio.bio, folio,
 				     sectors << 9, offset << 9));
 
-		/* Check for writing past i_size: */
-		WARN_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) >
-			  round_up(i_size, block_bytes(c)) &&
-			  !test_bit(BCH_FS_emergency_ro, &c->flags),
-			  "writing past i_size: %llu > %llu (unrounded %llu)\n",
-			  bio_end_sector(&w->io->op.wbio.bio) << 9,
-			  round_up(i_size, block_bytes(c)),
-			  i_size);
-
 		w->io->op.res.sectors += reserved_sectors;
 		w->io->op.i_sectors_delta -= dirty_sectors;
 		w->io->op.new_i_size = i_size;
diff --git a/libbcachefs/fsck.c b/libbcachefs/fsck.c
index a436c072..206fc046 100644
--- a/libbcachefs/fsck.c
+++ b/libbcachefs/fsck.c
@@ -205,6 +205,36 @@ err:
 	return ret;
 }
 
+/*
+ * Find any subvolume associated with a tree of snapshots
+ * We can't rely on master_subvol - it might have been deleted.
+ */
+static int find_snapshot_tree_subvol(struct btree_trans *trans,
+				     u32 tree_id, u32 *subvol)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret;
+
+	for_each_btree_key_norestart(trans, iter, BTREE_ID_snapshots, POS_MIN, 0, k, ret) {
+		if (k.k->type != KEY_TYPE_snapshot)
+			continue;
+
+		struct bkey_s_c_snapshot s = bkey_s_c_to_snapshot(k);
+		if (le32_to_cpu(s.v->tree) != tree_id)
+			continue;
+
+		if (s.v->subvol) {
+			*subvol = le32_to_cpu(s.v->subvol);
+			goto found;
+		}
+	}
+	ret = -BCH_ERR_ENOENT_no_snapshot_tree_subvol;
+found:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
 /* Get lost+found, create if it doesn't exist: */
 static int lookup_lostfound(struct btree_trans *trans, u32 snapshot,
 			    struct bch_inode_unpacked *lostfound,
@@ -223,19 +253,24 @@ static int lookup_lostfound(struct btree_trans *trans, u32 snapshot,
 	if (ret)
 		return ret;
 
-	subvol_inum root_inum = { .subvol = le32_to_cpu(st.master_subvol) };
+	u32 subvolid;
+	ret = find_snapshot_tree_subvol(trans,
+				bch2_snapshot_tree(c, snapshot), &subvolid);
+	bch_err_msg(c, ret, "finding subvol associated with snapshot tree %u",
+		    bch2_snapshot_tree(c, snapshot));
+	if (ret)
+		return ret;
 
 	struct bch_subvolume subvol;
-	ret = bch2_subvolume_get(trans, le32_to_cpu(st.master_subvol), false, &subvol);
-	bch_err_msg(c, ret, "looking up root subvol %u for snapshot %u",
-		    le32_to_cpu(st.master_subvol), snapshot);
+	ret = bch2_subvolume_get(trans, subvolid, false, &subvol);
+	bch_err_msg(c, ret, "looking up subvol %u for snapshot %u", subvolid, snapshot);
 	if (ret)
 		return ret;
 
 	if (!subvol.inode) {
 		struct btree_iter iter;
 		struct bkey_i_subvolume *subvol = bch2_bkey_get_mut_typed(trans, &iter,
-				BTREE_ID_subvolumes, POS(0, le32_to_cpu(st.master_subvol)),
+				BTREE_ID_subvolumes, POS(0, subvolid),
 				0, subvolume);
 		ret = PTR_ERR_OR_ZERO(subvol);
 		if (ret)
@@ -245,13 +280,16 @@ static int lookup_lostfound(struct btree_trans *trans, u32 snapshot,
 		bch2_trans_iter_exit(trans, &iter);
 	}
 
-	root_inum.inum = le64_to_cpu(subvol.inode);
+	subvol_inum root_inum = {
+		.subvol = subvolid,
+		.inum = le64_to_cpu(subvol.inode)
+	};
 
 	struct bch_inode_unpacked root_inode;
 	struct bch_hash_info root_hash_info;
 	ret = lookup_inode(trans, root_inum.inum, snapshot, &root_inode);
 	bch_err_msg(c, ret, "looking up root inode %llu for subvol %u",
-		    root_inum.inum, le32_to_cpu(st.master_subvol));
+		    root_inum.inum, subvolid);
 	if (ret)
 		return ret;
 
@@ -458,7 +496,9 @@ static int reattach_inode(struct btree_trans *trans, struct bch_inode_unpacked *
 				continue;
 
 			struct bch_inode_unpacked child_inode;
-			bch2_inode_unpack(k, &child_inode);
+			ret = bch2_inode_unpack(k, &child_inode);
+			if (ret)
+				break;
 
 			if (!inode_should_reattach(&child_inode)) {
 				ret = maybe_delete_dirent(trans,
@@ -809,9 +849,8 @@ static int add_inode(struct bch_fs *c, struct inode_walker *w,
 {
 	struct bch_inode_unpacked u;
 
-	BUG_ON(bch2_inode_unpack(inode, &u));
-
-	return darray_push(&w->inodes, ((struct inode_walker_entry) {
+	return bch2_inode_unpack(inode, &u) ?:
+		darray_push(&w->inodes, ((struct inode_walker_entry) {
 		.inode		= u,
 		.snapshot	= inode.k->p.snapshot,
 	}));
@@ -1065,7 +1104,7 @@ static int get_snapshot_root_inode(struct btree_trans *trans,
 		goto err;
 	BUG();
 found_root:
-	BUG_ON(bch2_inode_unpack(k, root));
+	ret = bch2_inode_unpack(k, root);
 err:
 	bch2_trans_iter_exit(trans, &iter);
 	return ret;
@@ -1096,7 +1135,9 @@ static int check_inode(struct btree_trans *trans,
 	if (!bkey_is_inode(k.k))
 		return 0;
 
-	BUG_ON(bch2_inode_unpack(k, &u));
+	ret = bch2_inode_unpack(k, &u);
+	if (ret)
+		goto err;
 
 	if (snapshot_root->bi_inum != u.bi_inum) {
 		ret = get_snapshot_root_inode(trans, snapshot_root, u.bi_inum);
@@ -1107,7 +1148,7 @@ static int check_inode(struct btree_trans *trans,
 	if (fsck_err_on(u.bi_hash_seed		!= snapshot_root->bi_hash_seed ||
 			INODE_STR_HASH(&u)	!= INODE_STR_HASH(snapshot_root),
 			trans, inode_snapshot_mismatch,
-			"inodes in different snapshots don't match")) {
+			"inode hash info in different snapshots don't match")) {
 		u.bi_hash_seed = snapshot_root->bi_hash_seed;
 		SET_INODE_STR_HASH(&u, INODE_STR_HASH(snapshot_root));
 		do_update = true;
@@ -1318,7 +1359,9 @@ static int find_oldest_inode_needs_reattach(struct btree_trans *trans,
 			break;
 
 		struct bch_inode_unpacked parent_inode;
-		bch2_inode_unpack(k, &parent_inode);
+		ret = bch2_inode_unpack(k, &parent_inode);
+		if (ret)
+			break;
 
 		if (!inode_should_reattach(&parent_inode))
 			break;
@@ -1341,7 +1384,9 @@ static int check_unreachable_inode(struct btree_trans *trans,
 		return 0;
 
 	struct bch_inode_unpacked inode;
-	BUG_ON(bch2_inode_unpack(k, &inode));
+	ret = bch2_inode_unpack(k, &inode);
+	if (ret)
+		return ret;
 
 	if (!inode_should_reattach(&inode))
 		return 0;
@@ -2296,7 +2341,7 @@ static int check_dirent(struct btree_trans *trans, struct btree_iter *iter,
 		*hash_info = bch2_hash_info_init(c, &i->inode);
 	dir->first_this_inode = false;
 
-	ret = bch2_str_hash_check_key(trans, s, bch2_dirent_hash_desc, hash_info, iter, k);
+	ret = bch2_str_hash_check_key(trans, s, &bch2_dirent_hash_desc, hash_info, iter, k);
 	if (ret < 0)
 		goto err;
 	if (ret) {
@@ -2410,7 +2455,7 @@ static int check_xattr(struct btree_trans *trans, struct btree_iter *iter,
 		*hash_info = bch2_hash_info_init(c, &i->inode);
 	inode->first_this_inode = false;
 
-	ret = bch2_str_hash_check_key(trans, NULL, bch2_xattr_hash_desc, hash_info, iter, k);
+	ret = bch2_str_hash_check_key(trans, NULL, &bch2_xattr_hash_desc, hash_info, iter, k);
 	bch_err_fn(c, ret);
 	return ret;
 }
@@ -2653,7 +2698,9 @@ static int check_path_loop(struct btree_trans *trans, struct bkey_s_c inode_k)
 	int ret = 0;
 
 	struct bch_inode_unpacked inode;
-	BUG_ON(bch2_inode_unpack(inode_k, &inode));
+	ret = bch2_inode_unpack(inode_k, &inode);
+	if (ret)
+		return ret;
 
 	while (!inode.bi_subvol) {
 		struct btree_iter dirent_iter;
@@ -2864,7 +2911,9 @@ static int check_nlinks_find_hardlinks(struct bch_fs *c,
 
 			/* Should never fail, checked by bch2_inode_invalid: */
 			struct bch_inode_unpacked u;
-			BUG_ON(bch2_inode_unpack(k, &u));
+			_ret3 = bch2_inode_unpack(k, &u);
+			if (_ret3)
+				break;
 
 			/*
 			 * Backpointer and directory structure checks are sufficient for
@@ -2942,7 +2991,9 @@ static int check_nlinks_update_inode(struct btree_trans *trans, struct btree_ite
 	if (!bkey_is_inode(k.k))
 		return 0;
 
-	BUG_ON(bch2_inode_unpack(k, &u));
+	ret = bch2_inode_unpack(k, &u);
+	if (ret)
+		return ret;
 
 	if (S_ISDIR(u.bi_mode))
 		return 0;
diff --git a/libbcachefs/inode.c b/libbcachefs/inode.c
index d1fc44a7..04ec0520 100644
--- a/libbcachefs/inode.c
+++ b/libbcachefs/inode.c
@@ -48,10 +48,10 @@ static int inode_decode_field(const u8 *in, const u8 *end,
 	u8 *p;
 
 	if (in >= end)
-		return -1;
+		return -BCH_ERR_inode_unpack_error;
 
 	if (!*in)
-		return -1;
+		return -BCH_ERR_inode_unpack_error;
 
 	/*
 	 * position of highest set bit indicates number of bytes:
@@ -61,7 +61,7 @@ static int inode_decode_field(const u8 *in, const u8 *end,
 	bytes	= byte_table[shift - 1];
 
 	if (in + bytes > end)
-		return -1;
+		return -BCH_ERR_inode_unpack_error;
 
 	p = (u8 *) be + 16 - bytes;
 	memcpy(p, in, bytes);
@@ -177,7 +177,7 @@ static noinline int bch2_inode_unpack_v1(struct bkey_s_c_inode inode,
 		return ret;						\
 									\
 	if (field_bits > sizeof(unpacked->_name) * 8)			\
-		return -1;						\
+		return -BCH_ERR_inode_unpack_error;			\
 									\
 	unpacked->_name = field[1];					\
 	in += ret;
@@ -218,7 +218,7 @@ static int bch2_inode_unpack_v2(struct bch_inode_unpacked *unpacked,
 									\
 	unpacked->_name = v[0];						\
 	if (v[1] || v[0] != unpacked->_name)				\
-		return -1;						\
+		return -BCH_ERR_inode_unpack_error;			\
 	fieldnr++;
 
 	BCH_INODE_FIELDS_v2()
@@ -269,7 +269,7 @@ static int bch2_inode_unpack_v3(struct bkey_s_c k,
 									\
 	unpacked->_name = v[0];						\
 	if (v[1] || v[0] != unpacked->_name)				\
-		return -1;						\
+		return -BCH_ERR_inode_unpack_error;			\
 	fieldnr++;
 
 	BCH_INODE_FIELDS_v3()
@@ -886,7 +886,7 @@ bch2_inode_alloc_cursor_get(struct btree_trans *trans, u64 cpu, u64 *min, u64 *m
 {
 	struct bch_fs *c = trans->c;
 
-	u64 cursor_idx = c->opts.shard_inode_numbers ? cpu : 0;
+	u64 cursor_idx = c->opts.inodes_32bit ? 0 : cpu + 1;
 
 	cursor_idx &= ~(~0ULL << c->opts.shard_inode_numbers_bits);
 
@@ -907,19 +907,16 @@ bch2_inode_alloc_cursor_get(struct btree_trans *trans, u64 cpu, u64 *min, u64 *m
 	if (ret)
 		goto err;
 
-	cursor->v.bits = c->opts.shard_inode_numbers_bits;
+	if (c->opts.inodes_32bit) {
+		*min = BLOCKDEV_INODE_MAX;
+		*max = INT_MAX;
+	} else {
+		cursor->v.bits = c->opts.shard_inode_numbers_bits;
 
-	unsigned bits = (c->opts.inodes_32bit ? 31 : 63);
-	if (c->opts.shard_inode_numbers) {
-		bits -= cursor->v.bits;
+		unsigned bits = 63 - c->opts.shard_inode_numbers_bits;
 
-		*min = (cpu << bits);
+		*min = max(cpu << bits, (u64) INT_MAX + 1);
 		*max = (cpu << bits) | ~(ULLONG_MAX << bits);
-
-		*min = max_t(u64, *min, BLOCKDEV_INODE_MAX);
-	} else {
-		*min = BLOCKDEV_INODE_MAX;
-		*max = ~(ULLONG_MAX << bits);
 	}
 
 	if (le64_to_cpu(cursor->v.idx)  < *min)
diff --git a/libbcachefs/inode_format.h b/libbcachefs/inode_format.h
index 1b93e189..b99a5bf1 100644
--- a/libbcachefs/inode_format.h
+++ b/libbcachefs/inode_format.h
@@ -102,7 +102,8 @@ struct bch_inode_generation {
 	x(bi_subvol,			32)	\
 	x(bi_parent_subvol,		32)	\
 	x(bi_nocow,			8)	\
-	x(bi_depth,			32)
+	x(bi_depth,			32)	\
+	x(bi_inodes_32bit,		8)
 
 /* subset of BCH_INODE_FIELDS */
 #define BCH_INODE_OPTS()			\
@@ -115,7 +116,8 @@ struct bch_inode_generation {
 	x(foreground_target,		16)	\
 	x(background_target,		16)	\
 	x(erasure_code,			16)	\
-	x(nocow,			8)
+	x(nocow,			8)	\
+	x(inodes_32bit,			8)
 
 enum inode_opt_id {
 #define x(name, ...)				\
diff --git a/libbcachefs/io_write.c b/libbcachefs/io_write.c
index 20da357e..3e71860f 100644
--- a/libbcachefs/io_write.c
+++ b/libbcachefs/io_write.c
@@ -1356,6 +1356,9 @@ err:
 	if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 		goto retry;
 
+	bch2_trans_put(trans);
+	darray_exit(&buckets);
+
 	if (ret) {
 		struct printbuf buf = PRINTBUF;
 		bch2_write_op_error(&buf, op);
@@ -1366,9 +1369,6 @@ err:
 		op->flags |= BCH_WRITE_SUBMITTED;
 	}
 
-	bch2_trans_put(trans);
-	darray_exit(&buckets);
-
 	/* fallback to cow write path? */
 	if (!(op->flags & BCH_WRITE_SUBMITTED)) {
 		closure_sync(&op->cl);
diff --git a/libbcachefs/journal_io.c b/libbcachefs/journal_io.c
index 7f2efe85..e1773ac2 100644
--- a/libbcachefs/journal_io.c
+++ b/libbcachefs/journal_io.c
@@ -1114,8 +1114,10 @@ reread:
 				       (printbuf_reset(&err),
 					prt_str(&err, "journal "),
 					bch2_csum_err_msg(&err, csum_type, j->csum, csum),
-					err.buf)))
+					err.buf))) {
 			saw_bad = true;
+			bch2_fatal_error(c);
+		}
 
 		ret = bch2_encrypt(c, JSET_CSUM_TYPE(j), journal_nonce(j),
 			     j->encrypted_start,
diff --git a/libbcachefs/move.c b/libbcachefs/move.c
index 9b066c61..c493ea62 100644
--- a/libbcachefs/move.c
+++ b/libbcachefs/move.c
@@ -414,7 +414,9 @@ static struct bch_io_opts *bch2_move_get_io_opts(struct btree_trans *trans,
 				continue;
 
 			struct bch_inode_unpacked inode;
-			BUG_ON(bch2_inode_unpack(k, &inode));
+			_ret3 = bch2_inode_unpack(k, &inode);
+			if (_ret3)
+				break;
 
 			struct snapshot_io_opts_entry e = { .snapshot = k.k->p.snapshot };
 			bch2_inode_opts_get(&e.io_opts, trans->c, &inode);
diff --git a/libbcachefs/opts.h b/libbcachefs/opts.h
index 2df5c8f7..e763d52e 100644
--- a/libbcachefs/opts.h
+++ b/libbcachefs/opts.h
@@ -222,15 +222,10 @@ enum fsck_err_opts {
 	  BCH_SB_ERASURE_CODE,		false,				\
 	  NULL,		"Enable erasure coding (DO NOT USE YET)")	\
 	x(inodes_32bit,			u8,				\
-	  OPT_FS|OPT_FORMAT,						\
+	  OPT_FS|OPT_INODE|OPT_FORMAT|OPT_MOUNT|OPT_RUNTIME,		\
 	  OPT_BOOL(),							\
 	  BCH_SB_INODE_32BIT,		true,				\
 	  NULL,		"Constrain inode numbers to 32 bits")		\
-	x(shard_inode_numbers,		u8,				\
-	  OPT_FS|OPT_FORMAT,						\
-	  OPT_BOOL(),							\
-	  BCH_SB_SHARD_INUMS,		true,				\
-	  NULL,		"Shard new inode numbers by CPU id")		\
 	x(shard_inode_numbers_bits,	u8,				\
 	  OPT_FS|OPT_FORMAT,						\
 	  OPT_UINT(0, 8),						\
diff --git a/libbcachefs/recovery.c b/libbcachefs/recovery.c
index 7849916e..98825437 100644
--- a/libbcachefs/recovery.c
+++ b/libbcachefs/recovery.c
@@ -107,6 +107,12 @@ out:
 	return ret;
 }
 
+static void kill_btree(struct bch_fs *c, enum btree_id btree)
+{
+	bch2_btree_id_root(c, btree)->alive = false;
+	bch2_shoot_down_journal_keys(c, btree, 0, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
+}
+
 /* for -o reconstruct_alloc: */
 static void bch2_reconstruct_alloc(struct bch_fs *c)
 {
@@ -157,16 +163,9 @@ static void bch2_reconstruct_alloc(struct bch_fs *c)
 	bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
 
-	bch2_shoot_down_journal_keys(c, BTREE_ID_alloc,
-				     0, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
-	bch2_shoot_down_journal_keys(c, BTREE_ID_backpointers,
-				     0, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
-	bch2_shoot_down_journal_keys(c, BTREE_ID_need_discard,
-				     0, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
-	bch2_shoot_down_journal_keys(c, BTREE_ID_freespace,
-				     0, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
-	bch2_shoot_down_journal_keys(c, BTREE_ID_bucket_gens,
-				     0, BTREE_MAX_DEPTH, POS_MIN, SPOS_MAX);
+	for (unsigned i = 0; i < btree_id_nr_alive(c); i++)
+		if (btree_id_is_alloc(i))
+			kill_btree(c, i);
 }
 
 /*
@@ -573,9 +572,6 @@ static int read_btree_roots(struct bch_fs *c)
 		if (!r->alive)
 			continue;
 
-		if (btree_id_is_alloc(i) && c->opts.reconstruct_alloc)
-			continue;
-
 		printbuf_reset(&buf);
 		bch2_btree_id_level_to_text(&buf, i, r->level);
 
@@ -785,6 +781,11 @@ int bch2_fs_recovery(struct bch_fs *c)
 
 	c->opts.recovery_passes |= bch2_recovery_passes_from_stable(le64_to_cpu(ext->recovery_passes_required[0]));
 
+	if (c->sb.version_upgrade_complete < bcachefs_metadata_version_autofix_errors) {
+		SET_BCH_SB_ERROR_ACTION(c->disk_sb.sb, BCH_ON_ERROR_fix_safe);
+		write_sb = true;
+	}
+
 	if (write_sb)
 		bch2_write_super(c);
 	mutex_unlock(&c->sb_lock);
@@ -882,15 +883,15 @@ use_clean:
 	c->journal_replay_seq_start	= last_seq;
 	c->journal_replay_seq_end	= blacklist_seq - 1;
 
-	if (c->opts.reconstruct_alloc)
-		bch2_reconstruct_alloc(c);
-
 	zero_out_btree_mem_ptr(&c->journal_keys);
 
 	ret = journal_replay_early(c, clean);
 	if (ret)
 		goto err;
 
+	if (c->opts.reconstruct_alloc)
+		bch2_reconstruct_alloc(c);
+
 	/*
 	 * After an unclean shutdown, skip then next few journal sequence
 	 * numbers as they may have been referenced by btree writes that
diff --git a/libbcachefs/recovery_passes_types.h b/libbcachefs/recovery_passes_types.h
index f967b23d..71baad41 100644
--- a/libbcachefs/recovery_passes_types.h
+++ b/libbcachefs/recovery_passes_types.h
@@ -49,7 +49,7 @@
 	x(fs_upgrade_for_subvolumes,		22, 0)					\
 	x(check_inodes,				24, PASS_FSCK)				\
 	x(check_extents,			25, PASS_FSCK)				\
-	x(check_indirect_extents,		26, PASS_FSCK)				\
+	x(check_indirect_extents,		26, PASS_ONLINE|PASS_FSCK)		\
 	x(check_dirents,			27, PASS_FSCK)				\
 	x(check_xattrs,				28, PASS_FSCK)				\
 	x(check_root,				29, PASS_ONLINE|PASS_FSCK)		\
diff --git a/libbcachefs/sb-counters_format.h b/libbcachefs/sb-counters_format.h
index 62ea4782..fdcf598f 100644
--- a/libbcachefs/sb-counters_format.h
+++ b/libbcachefs/sb-counters_format.h
@@ -2,86 +2,91 @@
 #ifndef _BCACHEFS_SB_COUNTERS_FORMAT_H
 #define _BCACHEFS_SB_COUNTERS_FORMAT_H
 
-#define BCH_PERSISTENT_COUNTERS()				\
-	x(io_read,					0)	\
-	x(io_write,					1)	\
-	x(io_move,					2)	\
-	x(bucket_invalidate,				3)	\
-	x(bucket_discard,				4)	\
-	x(bucket_alloc,					5)	\
-	x(bucket_alloc_fail,				6)	\
-	x(btree_cache_scan,				7)	\
-	x(btree_cache_reap,				8)	\
-	x(btree_cache_cannibalize,			9)	\
-	x(btree_cache_cannibalize_lock,			10)	\
-	x(btree_cache_cannibalize_lock_fail,		11)	\
-	x(btree_cache_cannibalize_unlock,		12)	\
-	x(btree_node_write,				13)	\
-	x(btree_node_read,				14)	\
-	x(btree_node_compact,				15)	\
-	x(btree_node_merge,				16)	\
-	x(btree_node_split,				17)	\
-	x(btree_node_rewrite,				18)	\
-	x(btree_node_alloc,				19)	\
-	x(btree_node_free,				20)	\
-	x(btree_node_set_root,				21)	\
-	x(btree_path_relock_fail,			22)	\
-	x(btree_path_upgrade_fail,			23)	\
-	x(btree_reserve_get_fail,			24)	\
-	x(journal_entry_full,				25)	\
-	x(journal_full,					26)	\
-	x(journal_reclaim_finish,			27)	\
-	x(journal_reclaim_start,			28)	\
-	x(journal_write,				29)	\
-	x(read_promote,					30)	\
-	x(read_bounce,					31)	\
-	x(read_split,					33)	\
-	x(read_retry,					32)	\
-	x(read_reuse_race,				34)	\
-	x(move_extent_read,				35)	\
-	x(move_extent_write,				36)	\
-	x(move_extent_finish,				37)	\
-	x(move_extent_fail,				38)	\
-	x(move_extent_start_fail,			39)	\
-	x(copygc,					40)	\
-	x(copygc_wait,					41)	\
-	x(gc_gens_end,					42)	\
-	x(gc_gens_start,				43)	\
-	x(trans_blocked_journal_reclaim,		44)	\
-	x(trans_restart_btree_node_reused,		45)	\
-	x(trans_restart_btree_node_split,		46)	\
-	x(trans_restart_fault_inject,			47)	\
-	x(trans_restart_iter_upgrade,			48)	\
-	x(trans_restart_journal_preres_get,		49)	\
-	x(trans_restart_journal_reclaim,		50)	\
-	x(trans_restart_journal_res_get,		51)	\
-	x(trans_restart_key_cache_key_realloced,	52)	\
-	x(trans_restart_key_cache_raced,		53)	\
-	x(trans_restart_mark_replicas,			54)	\
-	x(trans_restart_mem_realloced,			55)	\
-	x(trans_restart_memory_allocation_failure,	56)	\
-	x(trans_restart_relock,				57)	\
-	x(trans_restart_relock_after_fill,		58)	\
-	x(trans_restart_relock_key_cache_fill,		59)	\
-	x(trans_restart_relock_next_node,		60)	\
-	x(trans_restart_relock_parent_for_fill,		61)	\
-	x(trans_restart_relock_path,			62)	\
-	x(trans_restart_relock_path_intent,		63)	\
-	x(trans_restart_too_many_iters,			64)	\
-	x(trans_restart_traverse,			65)	\
-	x(trans_restart_upgrade,			66)	\
-	x(trans_restart_would_deadlock,			67)	\
-	x(trans_restart_would_deadlock_write,		68)	\
-	x(trans_restart_injected,			69)	\
-	x(trans_restart_key_cache_upgrade,		70)	\
-	x(trans_traverse_all,				71)	\
-	x(transaction_commit,				72)	\
-	x(write_super,					73)	\
-	x(trans_restart_would_deadlock_recursion_limit,	74)	\
-	x(trans_restart_write_buffer_flush,		75)	\
-	x(trans_restart_split_race,			76)	\
-	x(write_buffer_flush_slowpath,			77)	\
-	x(write_buffer_flush_sync,			78)
+enum counters_flags {
+	TYPE_COUNTER	= BIT(0),	/* event counters */
+	TYPE_SECTORS	= BIT(1),	/* amount counters, the unit is sectors */
+};
+
+#define BCH_PERSISTENT_COUNTERS()					\
+	x(io_read,					0,	TYPE_SECTORS)	\
+	x(io_write,					1,	TYPE_SECTORS)	\
+	x(io_move,					2,	TYPE_SECTORS)	\
+	x(bucket_invalidate,				3,	TYPE_COUNTER)	\
+	x(bucket_discard,				4,	TYPE_COUNTER)	\
+	x(bucket_alloc,					5,	TYPE_COUNTER)	\
+	x(bucket_alloc_fail,				6,	TYPE_COUNTER)	\
+	x(btree_cache_scan,				7,	TYPE_COUNTER)	\
+	x(btree_cache_reap,				8,	TYPE_COUNTER)	\
+	x(btree_cache_cannibalize,			9,	TYPE_COUNTER)	\
+	x(btree_cache_cannibalize_lock,			10,	TYPE_COUNTER)	\
+	x(btree_cache_cannibalize_lock_fail,		11,	TYPE_COUNTER)	\
+	x(btree_cache_cannibalize_unlock,		12,	TYPE_COUNTER)	\
+	x(btree_node_write,				13,	TYPE_COUNTER)	\
+	x(btree_node_read,				14,	TYPE_COUNTER)	\
+	x(btree_node_compact,				15,	TYPE_COUNTER)	\
+	x(btree_node_merge,				16,	TYPE_COUNTER)	\
+	x(btree_node_split,				17,	TYPE_COUNTER)	\
+	x(btree_node_rewrite,				18,	TYPE_COUNTER)	\
+	x(btree_node_alloc,				19,	TYPE_COUNTER)	\
+	x(btree_node_free,				20,	TYPE_COUNTER)	\
+	x(btree_node_set_root,				21,	TYPE_COUNTER)	\
+	x(btree_path_relock_fail,			22,	TYPE_COUNTER)	\
+	x(btree_path_upgrade_fail,			23,	TYPE_COUNTER)	\
+	x(btree_reserve_get_fail,			24,	TYPE_COUNTER)	\
+	x(journal_entry_full,				25,	TYPE_COUNTER)	\
+	x(journal_full,					26,	TYPE_COUNTER)	\
+	x(journal_reclaim_finish,			27,	TYPE_COUNTER)	\
+	x(journal_reclaim_start,			28,	TYPE_COUNTER)	\
+	x(journal_write,				29,	TYPE_COUNTER)	\
+	x(read_promote,					30,	TYPE_COUNTER)	\
+	x(read_bounce,					31,	TYPE_COUNTER)	\
+	x(read_split,					33,	TYPE_COUNTER)	\
+	x(read_retry,					32,	TYPE_COUNTER)	\
+	x(read_reuse_race,				34,	TYPE_COUNTER)	\
+	x(move_extent_read,				35,	TYPE_SECTORS)	\
+	x(move_extent_write,				36,	TYPE_SECTORS)	\
+	x(move_extent_finish,				37,	TYPE_SECTORS)	\
+	x(move_extent_fail,				38,	TYPE_COUNTER)	\
+	x(move_extent_start_fail,			39,	TYPE_COUNTER)	\
+	x(copygc,					40,	TYPE_COUNTER)	\
+	x(copygc_wait,					41,	TYPE_COUNTER)	\
+	x(gc_gens_end,					42,	TYPE_COUNTER)	\
+	x(gc_gens_start,				43,	TYPE_COUNTER)	\
+	x(trans_blocked_journal_reclaim,		44,	TYPE_COUNTER)	\
+	x(trans_restart_btree_node_reused,		45,	TYPE_COUNTER)	\
+	x(trans_restart_btree_node_split,		46,	TYPE_COUNTER)	\
+	x(trans_restart_fault_inject,			47,	TYPE_COUNTER)	\
+	x(trans_restart_iter_upgrade,			48,	TYPE_COUNTER)	\
+	x(trans_restart_journal_preres_get,		49,	TYPE_COUNTER)	\
+	x(trans_restart_journal_reclaim,		50,	TYPE_COUNTER)	\
+	x(trans_restart_journal_res_get,		51,	TYPE_COUNTER)	\
+	x(trans_restart_key_cache_key_realloced,	52,	TYPE_COUNTER)	\
+	x(trans_restart_key_cache_raced,		53,	TYPE_COUNTER)	\
+	x(trans_restart_mark_replicas,			54,	TYPE_COUNTER)	\
+	x(trans_restart_mem_realloced,			55,	TYPE_COUNTER)	\
+	x(trans_restart_memory_allocation_failure,	56,	TYPE_COUNTER)	\
+	x(trans_restart_relock,				57,	TYPE_COUNTER)	\
+	x(trans_restart_relock_after_fill,		58,	TYPE_COUNTER)	\
+	x(trans_restart_relock_key_cache_fill,		59,	TYPE_COUNTER)	\
+	x(trans_restart_relock_next_node,		60,	TYPE_COUNTER)	\
+	x(trans_restart_relock_parent_for_fill,		61,	TYPE_COUNTER)	\
+	x(trans_restart_relock_path,			62,	TYPE_COUNTER)	\
+	x(trans_restart_relock_path_intent,		63,	TYPE_COUNTER)	\
+	x(trans_restart_too_many_iters,			64,	TYPE_COUNTER)	\
+	x(trans_restart_traverse,			65,	TYPE_COUNTER)	\
+	x(trans_restart_upgrade,			66,	TYPE_COUNTER)	\
+	x(trans_restart_would_deadlock,			67,	TYPE_COUNTER)	\
+	x(trans_restart_would_deadlock_write,		68,	TYPE_COUNTER)	\
+	x(trans_restart_injected,			69,	TYPE_COUNTER)	\
+	x(trans_restart_key_cache_upgrade,		70,	TYPE_COUNTER)	\
+	x(trans_traverse_all,				71,	TYPE_COUNTER)	\
+	x(transaction_commit,				72,	TYPE_COUNTER)	\
+	x(write_super,					73,	TYPE_COUNTER)	\
+	x(trans_restart_would_deadlock_recursion_limit,	74,	TYPE_COUNTER)	\
+	x(trans_restart_write_buffer_flush,		75,	TYPE_COUNTER)	\
+	x(trans_restart_split_race,			76,	TYPE_COUNTER)	\
+	x(write_buffer_flush_slowpath,			77,	TYPE_COUNTER)	\
+	x(write_buffer_flush_sync,			78,	TYPE_COUNTER)
 
 enum bch_persistent_counters {
 #define x(t, n, ...) BCH_COUNTER_##t,
diff --git a/libbcachefs/sb-downgrade.c b/libbcachefs/sb-downgrade.c
index fe453e17..051214fd 100644
--- a/libbcachefs/sb-downgrade.c
+++ b/libbcachefs/sb-downgrade.c
@@ -83,7 +83,6 @@
 	  BIT_ULL(BCH_RECOVERY_PASS_check_inodes),		\
 	  BCH_FSCK_ERR_inode_has_child_snapshots_wrong)		\
 	x(backpointer_bucket_gen,				\
-	  BIT_ULL(BCH_RECOVERY_PASS_check_backpointers_to_extents)|\
 	  BIT_ULL(BCH_RECOVERY_PASS_check_extents_to_backpointers),\
 	  BCH_FSCK_ERR_backpointer_to_missing_ptr,		\
 	  BCH_FSCK_ERR_ptr_to_missing_backpointer)		\
diff --git a/libbcachefs/six.c b/libbcachefs/six.c
index 617d07e5..537bf049 100644
--- a/libbcachefs/six.c
+++ b/libbcachefs/six.c
@@ -616,8 +616,6 @@ void six_unlock_ip(struct six_lock *lock, enum six_lock_type type, unsigned long
 
 	if (type != SIX_LOCK_write)
 		six_release(&lock->dep_map, ip);
-	else
-		lock->seq++;
 
 	if (type == SIX_LOCK_intent &&
 	    lock->intent_lock_recurse) {
@@ -625,6 +623,15 @@ void six_unlock_ip(struct six_lock *lock, enum six_lock_type type, unsigned long
 		return;
 	}
 
+	if (type == SIX_LOCK_write &&
+	    lock->write_lock_recurse) {
+		--lock->write_lock_recurse;
+		return;
+	}
+
+	if (type == SIX_LOCK_write)
+		lock->seq++;
+
 	do_six_unlock_type(lock, type);
 }
 EXPORT_SYMBOL_GPL(six_unlock_ip);
@@ -735,13 +742,13 @@ void six_lock_increment(struct six_lock *lock, enum six_lock_type type)
 			atomic_add(l[type].lock_val, &lock->state);
 		}
 		break;
+	case SIX_LOCK_write:
+		lock->write_lock_recurse++;
+		fallthrough;
 	case SIX_LOCK_intent:
 		EBUG_ON(!(atomic_read(&lock->state) & SIX_LOCK_HELD_intent));
 		lock->intent_lock_recurse++;
 		break;
-	case SIX_LOCK_write:
-		BUG();
-		break;
 	}
 }
 EXPORT_SYMBOL_GPL(six_lock_increment);
diff --git a/libbcachefs/six.h b/libbcachefs/six.h
index 68d46fd7..c142e06b 100644
--- a/libbcachefs/six.h
+++ b/libbcachefs/six.h
@@ -137,6 +137,7 @@ struct six_lock {
 	atomic_t		state;
 	u32			seq;
 	unsigned		intent_lock_recurse;
+	unsigned		write_lock_recurse;
 	struct task_struct	*owner;
 	unsigned __percpu	*readers;
 	raw_spinlock_t		wait_lock;
diff --git a/libbcachefs/snapshot.c b/libbcachefs/snapshot.c
index 99f04551..cf6b3256 100644
--- a/libbcachefs/snapshot.c
+++ b/libbcachefs/snapshot.c
@@ -2,6 +2,7 @@
 
 #include "bcachefs.h"
 #include "bkey_buf.h"
+#include "btree_cache.h"
 #include "btree_key_cache.h"
 #include "btree_update.h"
 #include "buckets.h"
@@ -279,23 +280,6 @@ fsck_err:
 	return ret;
 }
 
-static void __set_is_ancestor_bitmap(struct bch_fs *c, u32 id)
-{
-	struct snapshot_t *t = snapshot_t_mut(c, id);
-	u32 parent = id;
-
-	while ((parent = bch2_snapshot_parent_early(c, parent)) &&
-	       parent - id - 1 < IS_ANCESTOR_BITMAP)
-		__set_bit(parent - id - 1, t->is_ancestor);
-}
-
-static void set_is_ancestor_bitmap(struct bch_fs *c, u32 id)
-{
-	mutex_lock(&c->snapshot_table_lock);
-	__set_is_ancestor_bitmap(c, id);
-	mutex_unlock(&c->snapshot_table_lock);
-}
-
 static int __bch2_mark_snapshot(struct btree_trans *trans,
 		       enum btree_id btree, unsigned level,
 		       struct bkey_s_c old, struct bkey_s_c new,
@@ -317,6 +301,7 @@ static int __bch2_mark_snapshot(struct btree_trans *trans,
 	if (new.k->type == KEY_TYPE_snapshot) {
 		struct bkey_s_c_snapshot s = bkey_s_c_to_snapshot(new);
 
+		t->live		= true;
 		t->parent	= le32_to_cpu(s.v->parent);
 		t->children[0]	= le32_to_cpu(s.v->children[0]);
 		t->children[1]	= le32_to_cpu(s.v->children[1]);
@@ -335,7 +320,11 @@ static int __bch2_mark_snapshot(struct btree_trans *trans,
 			t->skip[2]	= 0;
 		}
 
-		__set_is_ancestor_bitmap(c, id);
+		u32 parent = id;
+
+		while ((parent = bch2_snapshot_parent_early(c, parent)) &&
+		       parent - id - 1 < IS_ANCESTOR_BITMAP)
+			__set_bit(parent - id - 1, t->is_ancestor);
 
 		if (BCH_SNAPSHOT_DELETED(s.v)) {
 			set_bit(BCH_FS_need_delete_dead_snapshots, &c->flags);
@@ -365,70 +354,6 @@ int bch2_snapshot_lookup(struct btree_trans *trans, u32 id,
 				       BTREE_ITER_with_updates, snapshot, s);
 }
 
-static int bch2_snapshot_live(struct btree_trans *trans, u32 id)
-{
-	struct bch_snapshot v;
-	int ret;
-
-	if (!id)
-		return 0;
-
-	ret = bch2_snapshot_lookup(trans, id, &v);
-	if (bch2_err_matches(ret, ENOENT))
-		bch_err(trans->c, "snapshot node %u not found", id);
-	if (ret)
-		return ret;
-
-	return !BCH_SNAPSHOT_DELETED(&v);
-}
-
-/*
- * If @k is a snapshot with just one live child, it's part of a linear chain,
- * which we consider to be an equivalence class: and then after snapshot
- * deletion cleanup, there should only be a single key at a given position in
- * this equivalence class.
- *
- * This sets the equivalence class of @k to be the child's equivalence class, if
- * it's part of such a linear chain: this correctly sets equivalence classes on
- * startup if we run leaf to root (i.e. in natural key order).
- */
-static int bch2_snapshot_set_equiv(struct btree_trans *trans, struct bkey_s_c k)
-{
-	struct bch_fs *c = trans->c;
-	unsigned i, nr_live = 0, live_idx = 0;
-	struct bkey_s_c_snapshot snap;
-	u32 id = k.k->p.offset, child[2];
-
-	if (k.k->type != KEY_TYPE_snapshot)
-		return 0;
-
-	snap = bkey_s_c_to_snapshot(k);
-
-	child[0] = le32_to_cpu(snap.v->children[0]);
-	child[1] = le32_to_cpu(snap.v->children[1]);
-
-	for (i = 0; i < 2; i++) {
-		int ret = bch2_snapshot_live(trans, child[i]);
-
-		if (ret < 0)
-			return ret;
-
-		if (ret)
-			live_idx = i;
-		nr_live += ret;
-	}
-
-	mutex_lock(&c->snapshot_table_lock);
-
-	snapshot_t_mut(c, id)->equiv = nr_live == 1
-		? snapshot_t_mut(c, child[live_idx])->equiv
-		: id;
-
-	mutex_unlock(&c->snapshot_table_lock);
-
-	return 0;
-}
-
 /* fsck: */
 
 static u32 bch2_snapshot_child(struct bch_fs *c, u32 id, unsigned child)
@@ -570,6 +495,9 @@ static int check_snapshot_tree(struct btree_trans *trans,
 		goto err;
 	}
 
+	if (!st.v->master_subvol)
+		goto out;
+
 	ret = bch2_subvolume_get(trans, le32_to_cpu(st.v->master_subvol), false, &subvol);
 	if (ret && !bch2_err_matches(ret, ENOENT))
 		goto err;
@@ -613,6 +541,7 @@ static int check_snapshot_tree(struct btree_trans *trans,
 		u->v.master_subvol = cpu_to_le32(subvol_id);
 		st = snapshot_tree_i_to_s_c(u);
 	}
+out:
 err:
 fsck_err:
 	bch2_trans_iter_exit(trans, &snapshot_iter);
@@ -913,7 +842,7 @@ static int check_snapshot_exists(struct btree_trans *trans, u32 id)
 {
 	struct bch_fs *c = trans->c;
 
-	if (bch2_snapshot_equiv(c, id))
+	if (bch2_snapshot_exists(c, id))
 		return 0;
 
 	/* Do we need to reconstruct the snapshot_tree entry as well? */
@@ -962,8 +891,7 @@ static int check_snapshot_exists(struct btree_trans *trans, u32 id)
 
 	return  bch2_btree_insert_trans(trans, BTREE_ID_snapshots, &snapshot->k_i, 0) ?:
 		bch2_mark_snapshot(trans, BTREE_ID_snapshots, 0,
-				   bkey_s_c_null, bkey_i_to_s(&snapshot->k_i), 0) ?:
-		bch2_snapshot_set_equiv(trans, bkey_i_to_s_c(&snapshot->k_i));
+				   bkey_s_c_null, bkey_i_to_s(&snapshot->k_i), 0);
 }
 
 /* Figure out which snapshot nodes belong in the same tree: */
@@ -1061,7 +989,7 @@ int bch2_reconstruct_snapshots(struct bch_fs *c)
 		snapshot_id_list_to_text(&buf, t);
 
 		darray_for_each(*t, id) {
-			if (fsck_err_on(!bch2_snapshot_equiv(c, *id),
+			if (fsck_err_on(!bch2_snapshot_exists(c, *id),
 					trans, snapshot_node_missing,
 					"snapshot node %u from tree %s missing, recreate?", *id, buf.buf)) {
 				if (t->nr > 1) {
@@ -1094,10 +1022,12 @@ int bch2_check_key_has_snapshot(struct btree_trans *trans,
 	struct printbuf buf = PRINTBUF;
 	int ret = 0;
 
-	if (fsck_err_on(!bch2_snapshot_equiv(c, k.k->p.snapshot),
+	if (fsck_err_on(!bch2_snapshot_exists(c, k.k->p.snapshot),
 			trans, bkey_in_missing_snapshot,
 			"key in missing snapshot %s, delete?",
-			(bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
+			(bch2_btree_id_to_text(&buf, iter->btree_id),
+			 prt_char(&buf, ' '),
+			 bch2_bkey_val_to_text(&buf, c, k), buf.buf)))
 		ret = bch2_btree_delete_at(trans, iter,
 					    BTREE_UPDATE_internal_snapshot_node) ?: 1;
 fsck_err:
@@ -1111,13 +1041,11 @@ fsck_err:
 int bch2_snapshot_node_set_deleted(struct btree_trans *trans, u32 id)
 {
 	struct btree_iter iter;
-	struct bkey_i_snapshot *s;
-	int ret = 0;
-
-	s = bch2_bkey_get_mut_typed(trans, &iter,
+	struct bkey_i_snapshot *s =
+		bch2_bkey_get_mut_typed(trans, &iter,
 				    BTREE_ID_snapshots, POS(0, id),
 				    0, snapshot);
-	ret = PTR_ERR_OR_ZERO(s);
+	int ret = PTR_ERR_OR_ZERO(s);
 	if (unlikely(ret)) {
 		bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT),
 					trans->c, "missing snapshot %u", id);
@@ -1305,10 +1233,6 @@ static int create_snapids(struct btree_trans *trans, u32 parent, u32 tree,
 			goto err;
 
 		new_snapids[i]	= iter.pos.offset;
-
-		mutex_lock(&c->snapshot_table_lock);
-		snapshot_t_mut(c, new_snapids[i])->equiv = new_snapids[i];
-		mutex_unlock(&c->snapshot_table_lock);
 	}
 err:
 	bch2_trans_iter_exit(trans, &iter);
@@ -1414,129 +1338,153 @@ int bch2_snapshot_node_create(struct btree_trans *trans, u32 parent,
  * that key to snapshot leaf nodes, where we can mutate it
  */
 
-static int delete_dead_snapshots_process_key(struct btree_trans *trans,
-			       struct btree_iter *iter,
-			       struct bkey_s_c k,
-			       snapshot_id_list *deleted,
-			       snapshot_id_list *equiv_seen,
-			       struct bpos *last_pos)
+struct snapshot_interior_delete {
+	u32	id;
+	u32	live_child;
+};
+typedef DARRAY(struct snapshot_interior_delete) interior_delete_list;
+
+static inline u32 interior_delete_has_id(interior_delete_list *l, u32 id)
 {
-	int ret = bch2_check_key_has_snapshot(trans, iter, k);
-	if (ret)
-		return ret < 0 ? ret : 0;
+	darray_for_each(*l, i)
+		if (i->id == id)
+			return i->live_child;
+	return 0;
+}
 
-	struct bch_fs *c = trans->c;
-	u32 equiv = bch2_snapshot_equiv(c, k.k->p.snapshot);
-	if (!equiv) /* key for invalid snapshot node, but we chose not to delete */
+static unsigned __live_child(struct snapshot_table *t, u32 id,
+			     snapshot_id_list *delete_leaves,
+			     interior_delete_list *delete_interior)
+{
+	struct snapshot_t *s = __snapshot_t(t, id);
+	if (!s)
 		return 0;
 
-	if (!bkey_eq(k.k->p, *last_pos))
-		equiv_seen->nr = 0;
+	for (unsigned i = 0; i < ARRAY_SIZE(s->children); i++)
+		if (s->children[i] &&
+		    !snapshot_list_has_id(delete_leaves, s->children[i]) &&
+		    !interior_delete_has_id(delete_interior, s->children[i]))
+			return s->children[i];
 
-	if (snapshot_list_has_id(deleted, k.k->p.snapshot))
-		return bch2_btree_delete_at(trans, iter,
-					    BTREE_UPDATE_internal_snapshot_node);
+	for (unsigned i = 0; i < ARRAY_SIZE(s->children); i++) {
+		u32 live_child = s->children[i]
+			? __live_child(t, s->children[i], delete_leaves, delete_interior)
+			: 0;
+		if (live_child)
+			return live_child;
+	}
 
-	if (!bpos_eq(*last_pos, k.k->p) &&
-	    snapshot_list_has_id(equiv_seen, equiv))
-		return bch2_btree_delete_at(trans, iter,
-					    BTREE_UPDATE_internal_snapshot_node);
+	return 0;
+}
 
-	*last_pos = k.k->p;
+static unsigned live_child(struct bch_fs *c, u32 id,
+			   snapshot_id_list *delete_leaves,
+			   interior_delete_list *delete_interior)
+{
+	rcu_read_lock();
+	u32 ret = __live_child(rcu_dereference(c->snapshots), id,
+			       delete_leaves, delete_interior);
+	rcu_read_unlock();
+	return ret;
+}
 
-	ret = snapshot_list_add_nodup(c, equiv_seen, equiv);
-	if (ret)
-		return ret;
+static int delete_dead_snapshots_process_key(struct btree_trans *trans,
+					     struct btree_iter *iter,
+					     struct bkey_s_c k,
+					     snapshot_id_list *delete_leaves,
+					     interior_delete_list *delete_interior)
+{
+	if (snapshot_list_has_id(delete_leaves, k.k->p.snapshot))
+		return bch2_btree_delete_at(trans, iter,
+					    BTREE_UPDATE_internal_snapshot_node);
 
-	/*
-	 * When we have a linear chain of snapshot nodes, we consider
-	 * those to form an equivalence class: we're going to collapse
-	 * them all down to a single node, and keep the leaf-most node -
-	 * which has the same id as the equivalence class id.
-	 *
-	 * If there are multiple keys in different snapshots at the same
-	 * position, we're only going to keep the one in the newest
-	 * snapshot (we delete the others above) - the rest have been
-	 * overwritten and are redundant, and for the key we're going to keep we
-	 * need to move it to the equivalance class ID if it's not there
-	 * already.
-	 */
-	if (equiv != k.k->p.snapshot) {
+	u32 live_child = interior_delete_has_id(delete_interior, k.k->p.snapshot);
+	if (live_child) {
 		struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, k);
 		int ret = PTR_ERR_OR_ZERO(new);
 		if (ret)
 			return ret;
 
-		new->k.p.snapshot = equiv;
-
-		struct btree_iter new_iter;
-		bch2_trans_iter_init(trans, &new_iter, iter->btree_id, new->k.p,
-				     BTREE_ITER_all_snapshots|
-				     BTREE_ITER_cached|
-				     BTREE_ITER_intent);
+		new->k.p.snapshot = live_child;
 
-		ret =   bch2_btree_iter_traverse(&new_iter) ?:
-			bch2_trans_update(trans, &new_iter, new,
-					BTREE_UPDATE_internal_snapshot_node) ?:
-			bch2_btree_delete_at(trans, iter,
-					BTREE_UPDATE_internal_snapshot_node);
-		bch2_trans_iter_exit(trans, &new_iter);
+		struct btree_iter dst_iter;
+		struct bkey_s_c dst_k = bch2_bkey_get_iter(trans, &dst_iter,
+							   iter->btree_id, new->k.p,
+							   BTREE_ITER_all_snapshots|
+							   BTREE_ITER_intent);
+		ret = bkey_err(dst_k);
 		if (ret)
 			return ret;
+
+		ret =   (bkey_deleted(dst_k.k)
+			 ? bch2_trans_update(trans, &dst_iter, new,
+					     BTREE_UPDATE_internal_snapshot_node)
+			 : 0) ?:
+			bch2_btree_delete_at(trans, iter,
+					     BTREE_UPDATE_internal_snapshot_node);
+		bch2_trans_iter_exit(trans, &dst_iter);
+		return ret;
 	}
 
 	return 0;
 }
 
-static int bch2_snapshot_needs_delete(struct btree_trans *trans, struct bkey_s_c k)
+/*
+ * For a given snapshot, if it doesn't have a subvolume that points to it, and
+ * it doesn't have child snapshot nodes - it's now redundant and we can mark it
+ * as deleted.
+ */
+static int check_should_delete_snapshot(struct btree_trans *trans, struct bkey_s_c k,
+					snapshot_id_list *delete_leaves,
+					interior_delete_list *delete_interior)
 {
-	struct bkey_s_c_snapshot snap;
-	u32 children[2];
-	int ret;
-
 	if (k.k->type != KEY_TYPE_snapshot)
 		return 0;
 
-	snap = bkey_s_c_to_snapshot(k);
-	if (BCH_SNAPSHOT_DELETED(snap.v) ||
-	    BCH_SNAPSHOT_SUBVOL(snap.v))
+	struct bch_fs *c = trans->c;
+	struct bkey_s_c_snapshot s = bkey_s_c_to_snapshot(k);
+	unsigned live_children = 0;
+
+	if (BCH_SNAPSHOT_SUBVOL(s.v))
 		return 0;
 
-	children[0] = le32_to_cpu(snap.v->children[0]);
-	children[1] = le32_to_cpu(snap.v->children[1]);
+	for (unsigned i = 0; i < 2; i++) {
+		u32 child = le32_to_cpu(s.v->children[i]);
 
-	ret   = bch2_snapshot_live(trans, children[0]) ?:
-		bch2_snapshot_live(trans, children[1]);
-	if (ret < 0)
-		return ret;
-	return !ret;
-}
+		live_children += child &&
+			!snapshot_list_has_id(delete_leaves, child);
+	}
 
-/*
- * For a given snapshot, if it doesn't have a subvolume that points to it, and
- * it doesn't have child snapshot nodes - it's now redundant and we can mark it
- * as deleted.
- */
-static int bch2_delete_redundant_snapshot(struct btree_trans *trans, struct bkey_s_c k)
-{
-	int ret = bch2_snapshot_needs_delete(trans, k);
+	if (live_children == 0) {
+		return snapshot_list_add(c, delete_leaves, s.k->p.offset);
+	} else if (live_children == 1) {
+		struct snapshot_interior_delete d = {
+			.id		= s.k->p.offset,
+			.live_child	= live_child(c, s.k->p.offset, delete_leaves, delete_interior),
+		};
+
+		if (!d.live_child) {
+			bch_err(c, "error finding live child of snapshot %u", d.id);
+			return -EINVAL;
+		}
 
-	return ret <= 0
-		? ret
-		: bch2_snapshot_node_set_deleted(trans, k.k->p.offset);
+		return darray_push(delete_interior, d);
+	} else {
+		return 0;
+	}
 }
 
 static inline u32 bch2_snapshot_nth_parent_skip(struct bch_fs *c, u32 id, u32 n,
-						snapshot_id_list *skip)
+						interior_delete_list *skip)
 {
 	rcu_read_lock();
-	while (snapshot_list_has_id(skip, id))
+	while (interior_delete_has_id(skip, id))
 		id = __bch2_snapshot_parent(c, id);
 
 	while (n--) {
 		do {
 			id = __bch2_snapshot_parent(c, id);
-		} while (snapshot_list_has_id(skip, id));
+		} while (interior_delete_has_id(skip, id));
 	}
 	rcu_read_unlock();
 
@@ -1545,7 +1493,7 @@ static inline u32 bch2_snapshot_nth_parent_skip(struct bch_fs *c, u32 id, u32 n,
 
 static int bch2_fix_child_of_deleted_snapshot(struct btree_trans *trans,
 					      struct btree_iter *iter, struct bkey_s_c k,
-					      snapshot_id_list *deleted)
+					      interior_delete_list *deleted)
 {
 	struct bch_fs *c = trans->c;
 	u32 nr_deleted_ancestors = 0;
@@ -1555,7 +1503,7 @@ static int bch2_fix_child_of_deleted_snapshot(struct btree_trans *trans,
 	if (k.k->type != KEY_TYPE_snapshot)
 		return 0;
 
-	if (snapshot_list_has_id(deleted, k.k->p.offset))
+	if (interior_delete_has_id(deleted, k.k->p.offset))
 		return 0;
 
 	s = bch2_bkey_make_mut_noupdate_typed(trans, k, snapshot);
@@ -1564,7 +1512,7 @@ static int bch2_fix_child_of_deleted_snapshot(struct btree_trans *trans,
 		return ret;
 
 	darray_for_each(*deleted, i)
-		nr_deleted_ancestors += bch2_snapshot_is_ancestor(c, s->k.p.offset, *i);
+		nr_deleted_ancestors += bch2_snapshot_is_ancestor(c, s->k.p.offset, i->id);
 
 	if (!nr_deleted_ancestors)
 		return 0;
@@ -1582,7 +1530,7 @@ static int bch2_fix_child_of_deleted_snapshot(struct btree_trans *trans,
 		for (unsigned j = 0; j < ARRAY_SIZE(s->v.skip); j++) {
 			u32 id = le32_to_cpu(s->v.skip[j]);
 
-			if (snapshot_list_has_id(deleted, id)) {
+			if (interior_delete_has_id(deleted, id)) {
 				id = bch2_snapshot_nth_parent_skip(c,
 							parent,
 							depth > 1
@@ -1601,51 +1549,44 @@ static int bch2_fix_child_of_deleted_snapshot(struct btree_trans *trans,
 
 int bch2_delete_dead_snapshots(struct bch_fs *c)
 {
-	struct btree_trans *trans;
-	snapshot_id_list deleted = { 0 };
-	snapshot_id_list deleted_interior = { 0 };
-	int ret = 0;
-
 	if (!test_and_clear_bit(BCH_FS_need_delete_dead_snapshots, &c->flags))
 		return 0;
 
-	trans = bch2_trans_get(c);
+	struct btree_trans *trans = bch2_trans_get(c);
+	snapshot_id_list delete_leaves = {};
+	interior_delete_list delete_interior = {};
+	int ret = 0;
 
 	/*
 	 * For every snapshot node: If we have no live children and it's not
 	 * pointed to by a subvolume, delete it:
 	 */
-	ret = for_each_btree_key_commit(trans, iter, BTREE_ID_snapshots,
-			POS_MIN, 0, k,
-			NULL, NULL, 0,
-		bch2_delete_redundant_snapshot(trans, k));
-	bch_err_msg(c, ret, "deleting redundant snapshots");
+	ret = for_each_btree_key(trans, iter, BTREE_ID_snapshots, POS_MIN, 0, k,
+		check_should_delete_snapshot(trans, k, &delete_leaves, &delete_interior));
+	bch_err_msg(c, ret, "walking snapshots");
 	if (ret)
 		goto err;
 
-	ret = for_each_btree_key(trans, iter, BTREE_ID_snapshots,
-				 POS_MIN, 0, k,
-		bch2_snapshot_set_equiv(trans, k));
-	bch_err_msg(c, ret, "in bch2_snapshots_set_equiv");
-	if (ret)
+	if (!delete_leaves.nr && !delete_interior.nr)
 		goto err;
 
-	ret = for_each_btree_key(trans, iter, BTREE_ID_snapshots,
-				 POS_MIN, 0, k, ({
-		if (k.k->type != KEY_TYPE_snapshot)
-			continue;
+	{
+		struct printbuf buf = PRINTBUF;
+		prt_printf(&buf, "deleting leaves");
+		darray_for_each(delete_leaves, i)
+			prt_printf(&buf, " %u", *i);
 
-		BCH_SNAPSHOT_DELETED(bkey_s_c_to_snapshot(k).v)
-			? snapshot_list_add(c, &deleted, k.k->p.offset)
-			: 0;
-	}));
-	bch_err_msg(c, ret, "walking snapshots");
-	if (ret)
-		goto err;
+		prt_printf(&buf, " interior");
+		darray_for_each(delete_interior, i)
+			prt_printf(&buf, " %u->%u", i->id, i->live_child);
+
+		ret = commit_do(trans, NULL, NULL, 0, bch2_trans_log_msg(trans, &buf));
+		printbuf_exit(&buf);
+		if (ret)
+			goto err;
+	}
 
 	for (unsigned btree = 0; btree < BTREE_ID_NR; btree++) {
-		struct bpos last_pos = POS_MIN;
-		snapshot_id_list equiv_seen = { 0 };
 		struct disk_reservation res = { 0 };
 
 		if (!btree_type_has_snapshots(btree))
@@ -1655,33 +1596,24 @@ int bch2_delete_dead_snapshots(struct bch_fs *c)
 				btree, POS_MIN,
 				BTREE_ITER_prefetch|BTREE_ITER_all_snapshots, k,
 				&res, NULL, BCH_TRANS_COMMIT_no_enospc,
-			delete_dead_snapshots_process_key(trans, &iter, k, &deleted,
-							  &equiv_seen, &last_pos));
+			delete_dead_snapshots_process_key(trans, &iter, k,
+							  &delete_leaves,
+							  &delete_interior));
 
 		bch2_disk_reservation_put(c, &res);
-		darray_exit(&equiv_seen);
 
 		bch_err_msg(c, ret, "deleting keys from dying snapshots");
 		if (ret)
 			goto err;
 	}
 
-	bch2_trans_unlock(trans);
-	down_write(&c->snapshot_create_lock);
-
-	ret = for_each_btree_key(trans, iter, BTREE_ID_snapshots,
-				 POS_MIN, 0, k, ({
-		u32 snapshot = k.k->p.offset;
-		u32 equiv = bch2_snapshot_equiv(c, snapshot);
-
-		equiv != snapshot
-			? snapshot_list_add(c, &deleted_interior, snapshot)
-			: 0;
-	}));
-
-	bch_err_msg(c, ret, "walking snapshots");
-	if (ret)
-		goto err_create_lock;
+	darray_for_each(delete_leaves, i) {
+		ret = commit_do(trans, NULL, NULL, 0,
+			bch2_snapshot_node_delete(trans, *i));
+		bch_err_msg(c, ret, "deleting snapshot %u", *i);
+		if (ret)
+			goto err;
+	}
 
 	/*
 	 * Fixing children of deleted snapshots can't be done completely
@@ -1691,30 +1623,20 @@ int bch2_delete_dead_snapshots(struct bch_fs *c)
 	ret = for_each_btree_key_commit(trans, iter, BTREE_ID_snapshots, POS_MIN,
 				  BTREE_ITER_intent, k,
 				  NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
-		bch2_fix_child_of_deleted_snapshot(trans, &iter, k, &deleted_interior));
+		bch2_fix_child_of_deleted_snapshot(trans, &iter, k, &delete_interior));
 	if (ret)
-		goto err_create_lock;
-
-	darray_for_each(deleted, i) {
-		ret = commit_do(trans, NULL, NULL, 0,
-			bch2_snapshot_node_delete(trans, *i));
-		bch_err_msg(c, ret, "deleting snapshot %u", *i);
-		if (ret)
-			goto err_create_lock;
-	}
+		goto err;
 
-	darray_for_each(deleted_interior, i) {
+	darray_for_each(delete_interior, i) {
 		ret = commit_do(trans, NULL, NULL, 0,
-			bch2_snapshot_node_delete(trans, *i));
-		bch_err_msg(c, ret, "deleting snapshot %u", *i);
+			bch2_snapshot_node_delete(trans, i->id));
+		bch_err_msg(c, ret, "deleting snapshot %u", i->id);
 		if (ret)
-			goto err_create_lock;
+			goto err;
 	}
-err_create_lock:
-	up_write(&c->snapshot_create_lock);
 err:
-	darray_exit(&deleted_interior);
-	darray_exit(&deleted);
+	darray_exit(&delete_interior);
+	darray_exit(&delete_leaves);
 	bch2_trans_put(trans);
 	bch_err_fn(c, ret);
 	return ret;
@@ -1767,37 +1689,36 @@ int __bch2_key_has_snapshot_overwrites(struct btree_trans *trans,
 	return ret;
 }
 
-static int bch2_check_snapshot_needs_deletion(struct btree_trans *trans, struct bkey_s_c k)
+static bool interior_snapshot_needs_delete(struct bkey_s_c_snapshot snap)
 {
-	struct bch_fs *c = trans->c;
-	struct bkey_s_c_snapshot snap;
-	int ret = 0;
+	/* If there's one child, it's redundant and keys will be moved to the child */
+	return !!snap.v->children[0] + !!snap.v->children[1] == 1;
+}
 
+static int bch2_check_snapshot_needs_deletion(struct btree_trans *trans, struct bkey_s_c k)
+{
 	if (k.k->type != KEY_TYPE_snapshot)
 		return 0;
 
-	snap = bkey_s_c_to_snapshot(k);
+	struct bkey_s_c_snapshot snap = bkey_s_c_to_snapshot(k);
 	if (BCH_SNAPSHOT_DELETED(snap.v) ||
-	    bch2_snapshot_equiv(c, k.k->p.offset) != k.k->p.offset ||
-	    (ret = bch2_snapshot_needs_delete(trans, k)) > 0) {
-		set_bit(BCH_FS_need_delete_dead_snapshots, &c->flags);
-		return 0;
-	}
+	    interior_snapshot_needs_delete(snap))
+		set_bit(BCH_FS_need_delete_dead_snapshots, &trans->c->flags);
 
-	return ret;
+	return 0;
 }
 
 int bch2_snapshots_read(struct bch_fs *c)
 {
+	/*
+	 * Initializing the is_ancestor bitmaps requires ancestors to already be
+	 * initialized - so mark in reverse:
+	 */
 	int ret = bch2_trans_run(c,
-		for_each_btree_key(trans, iter, BTREE_ID_snapshots,
-				   POS_MIN, 0, k,
+		for_each_btree_key_reverse(trans, iter, BTREE_ID_snapshots,
+				   POS_MAX, 0, k,
 			__bch2_mark_snapshot(trans, BTREE_ID_snapshots, 0, bkey_s_c_null, k, 0) ?:
-			bch2_snapshot_set_equiv(trans, k) ?:
-			bch2_check_snapshot_needs_deletion(trans, k)) ?:
-		for_each_btree_key(trans, iter, BTREE_ID_snapshots,
-				   POS_MIN, 0, k,
-			   (set_is_ancestor_bitmap(c, k.k->p.offset), 0)));
+			bch2_check_snapshot_needs_deletion(trans, k)));
 	bch_err_fn(c, ret);
 
 	/*
diff --git a/libbcachefs/snapshot.h b/libbcachefs/snapshot.h
index ae23d45f..00373cf3 100644
--- a/libbcachefs/snapshot.h
+++ b/libbcachefs/snapshot.h
@@ -119,19 +119,19 @@ static inline u32 bch2_snapshot_root(struct bch_fs *c, u32 id)
 	return id;
 }
 
-static inline u32 __bch2_snapshot_equiv(struct bch_fs *c, u32 id)
+static inline bool __bch2_snapshot_exists(struct bch_fs *c, u32 id)
 {
 	const struct snapshot_t *s = snapshot_t(c, id);
-	return s ? s->equiv : 0;
+	return s ? s->live : 0;
 }
 
-static inline u32 bch2_snapshot_equiv(struct bch_fs *c, u32 id)
+static inline bool bch2_snapshot_exists(struct bch_fs *c, u32 id)
 {
 	rcu_read_lock();
-	id = __bch2_snapshot_equiv(c, id);
+	bool ret = __bch2_snapshot_exists(c, id);
 	rcu_read_unlock();
 
-	return id;
+	return ret;
 }
 
 static inline int bch2_snapshot_is_internal_node(struct bch_fs *c, u32 id)
diff --git a/libbcachefs/str_hash.c b/libbcachefs/str_hash.c
index c3276a7e..f5977c5c 100644
--- a/libbcachefs/str_hash.c
+++ b/libbcachefs/str_hash.c
@@ -101,38 +101,108 @@ static int hash_pick_winner(struct btree_trans *trans,
 	}
 }
 
-int bch2_str_hash_check_key(struct btree_trans *trans,
-			    struct snapshots_seen *s,
-			    const struct bch_hash_desc desc,
-			    struct bch_hash_info *hash_info,
-			    struct btree_iter *k_iter, struct bkey_s_c hash_k)
+static int repair_inode_hash_info(struct btree_trans *trans,
+				  struct bch_inode_unpacked *snapshot_root)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret = 0;
+
+	for_each_btree_key_reverse_norestart(trans, iter, BTREE_ID_inodes,
+					     SPOS(0, snapshot_root->bi_inum, snapshot_root->bi_snapshot - 1),
+					     BTREE_ITER_all_snapshots, k, ret) {
+		if (k.k->p.offset != snapshot_root->bi_inum)
+			break;
+		if (!bkey_is_inode(k.k))
+			continue;
+
+		struct bch_inode_unpacked inode;
+		ret = bch2_inode_unpack(k, &inode);
+		if (ret)
+			break;
+
+		if (fsck_err_on(inode.bi_hash_seed	!= snapshot_root->bi_hash_seed ||
+				INODE_STR_HASH(&inode)	!= INODE_STR_HASH(snapshot_root),
+				trans, inode_snapshot_mismatch,
+				"inode hash info in different snapshots don't match")) {
+			inode.bi_hash_seed = snapshot_root->bi_hash_seed;
+			SET_INODE_STR_HASH(&inode, INODE_STR_HASH(snapshot_root));
+			ret = __bch2_fsck_write_inode(trans, &inode) ?:
+				bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc) ?:
+				-BCH_ERR_transaction_restart_nested;
+			break;
+		}
+	}
+fsck_err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+/*
+ * All versions of the same inode in different snapshots must have the same hash
+ * seed/type: verify that the hash info we're using matches the root
+ */
+static int check_inode_hash_info_matches_root(struct btree_trans *trans, u64 inum,
+					      struct bch_hash_info *hash_info)
 {
 	struct bch_fs *c = trans->c;
-	struct btree_iter iter = { NULL };
-	struct printbuf buf = PRINTBUF;
+	struct btree_iter iter;
 	struct bkey_s_c k;
-	u64 hash;
 	int ret = 0;
 
-	if (hash_k.k->type != desc.key_type)
-		return 0;
+	for_each_btree_key_reverse_norestart(trans, iter, BTREE_ID_inodes, SPOS(0, inum, U32_MAX),
+					     BTREE_ITER_all_snapshots, k, ret) {
+		if (k.k->p.offset != inum)
+			break;
+		if (bkey_is_inode(k.k))
+			goto found;
+	}
+	bch_err(c, "%s(): inum %llu not found", __func__, inum);
+	ret = -BCH_ERR_fsck_repair_unimplemented;
+	goto err;
+found:;
+	struct bch_inode_unpacked inode;
+	ret = bch2_inode_unpack(k, &inode);
+	if (ret)
+		goto err;
 
-	hash = desc.hash_bkey(hash_info, hash_k);
+	struct bch_hash_info hash2 = bch2_hash_info_init(c, &inode);
+	if (memcmp(hash_info, &hash2, sizeof(hash2))) {
+		ret = repair_inode_hash_info(trans, &inode);
+		if (!ret) {
+			bch_err(c, "inode hash info mismatch with root, but mismatch not found");
+			ret = -BCH_ERR_fsck_repair_unimplemented;
+		}
+	}
+err:
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
 
-	if (likely(hash == hash_k.k->p.offset))
-		return 0;
+int __bch2_str_hash_check_key(struct btree_trans *trans,
+			      struct snapshots_seen *s,
+			      const struct bch_hash_desc *desc,
+			      struct bch_hash_info *hash_info,
+			      struct btree_iter *k_iter, struct bkey_s_c hash_k)
+{
+	struct bch_fs *c = trans->c;
+	struct btree_iter iter = { NULL };
+	struct printbuf buf = PRINTBUF;
+	struct bkey_s_c k;
+	int ret = 0;
 
+	u64 hash = desc->hash_bkey(hash_info, hash_k);
 	if (hash_k.k->p.offset < hash)
 		goto bad_hash;
 
-	for_each_btree_key_norestart(trans, iter, desc.btree_id,
+	for_each_btree_key_norestart(trans, iter, desc->btree_id,
 				     SPOS(hash_k.k->p.inode, hash, hash_k.k->p.snapshot),
 				     BTREE_ITER_slots, k, ret) {
 		if (bkey_eq(k.k->p, hash_k.k->p))
 			break;
 
-		if (k.k->type == desc.key_type &&
-		    !desc.cmp_bkey(k, hash_k))
+		if (k.k->type == desc->key_type &&
+		    !desc->cmp_bkey(k, hash_k))
 			goto duplicate_entries;
 
 		if (bkey_deleted(k.k)) {
@@ -145,16 +215,23 @@ out:
 	printbuf_exit(&buf);
 	return ret;
 bad_hash:
+	/*
+	 * Before doing any repair, check hash_info itself:
+	 */
+	ret = check_inode_hash_info_matches_root(trans, hash_k.k->p.inode, hash_info);
+	if (ret)
+		goto out;
+
 	if (fsck_err(trans, hash_table_key_wrong_offset,
 		     "hash table key at wrong offset: btree %s inode %llu offset %llu, hashed to %llu\n  %s",
-		     bch2_btree_id_str(desc.btree_id), hash_k.k->p.inode, hash_k.k->p.offset, hash,
+		     bch2_btree_id_str(desc->btree_id), hash_k.k->p.inode, hash_k.k->p.offset, hash,
 		     (printbuf_reset(&buf),
 		      bch2_bkey_val_to_text(&buf, c, hash_k), buf.buf))) {
 		struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, hash_k);
 		if (IS_ERR(new))
 			return PTR_ERR(new);
 
-		k = bch2_hash_set_or_get_in_snapshot(trans, &iter, desc, hash_info,
+		k = bch2_hash_set_or_get_in_snapshot(trans, &iter, *desc, hash_info,
 				       (subvol_inum) { 0, hash_k.k->p.inode },
 				       hash_k.k->p.snapshot, new,
 				       STR_HASH_must_create|
@@ -166,9 +243,9 @@ bad_hash:
 		if (k.k)
 			goto duplicate_entries;
 
-		ret =   bch2_hash_delete_at(trans, desc, hash_info, k_iter,
+		ret =   bch2_hash_delete_at(trans, *desc, hash_info, k_iter,
 					    BTREE_UPDATE_internal_snapshot_node) ?:
-			bch2_fsck_update_backpointers(trans, s, desc, hash_info, new) ?:
+			bch2_fsck_update_backpointers(trans, s, *desc, hash_info, new) ?:
 			bch2_trans_commit(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc) ?:
 			-BCH_ERR_transaction_restart_nested;
 		goto out;
@@ -176,7 +253,7 @@ bad_hash:
 fsck_err:
 	goto out;
 duplicate_entries:
-	ret = hash_pick_winner(trans, desc, hash_info, hash_k, k);
+	ret = hash_pick_winner(trans, *desc, hash_info, hash_k, k);
 	if (ret < 0)
 		goto out;
 
@@ -192,14 +269,14 @@ duplicate_entries:
 
 	switch (ret) {
 	case 0:
-		ret = bch2_hash_delete_at(trans, desc, hash_info, k_iter, 0);
+		ret = bch2_hash_delete_at(trans, *desc, hash_info, k_iter, 0);
 		break;
 	case 1:
-		ret = bch2_hash_delete_at(trans, desc, hash_info, &iter, 0);
+		ret = bch2_hash_delete_at(trans, *desc, hash_info, &iter, 0);
 		break;
 	case 2:
-		ret = fsck_rename_dirent(trans, s, desc, hash_info, bkey_s_c_to_dirent(hash_k)) ?:
-			bch2_hash_delete_at(trans, desc, hash_info, k_iter, 0);
+		ret = fsck_rename_dirent(trans, s, *desc, hash_info, bkey_s_c_to_dirent(hash_k)) ?:
+			bch2_hash_delete_at(trans, *desc, hash_info, k_iter, 0);
 		goto out;
 	}
 
diff --git a/libbcachefs/str_hash.h b/libbcachefs/str_hash.h
index 0c20f3af..55a4ac7b 100644
--- a/libbcachefs/str_hash.h
+++ b/libbcachefs/str_hash.h
@@ -394,10 +394,25 @@ int bch2_hash_delete(struct btree_trans *trans,
 }
 
 struct snapshots_seen;
-int bch2_str_hash_check_key(struct btree_trans *,
-			    struct snapshots_seen *,
-			    const struct bch_hash_desc,
-			    struct bch_hash_info *,
-			    struct btree_iter *, struct bkey_s_c);
+int __bch2_str_hash_check_key(struct btree_trans *,
+			      struct snapshots_seen *,
+			      const struct bch_hash_desc *,
+			      struct bch_hash_info *,
+			      struct btree_iter *, struct bkey_s_c);
+
+static inline int bch2_str_hash_check_key(struct btree_trans *trans,
+			    struct snapshots_seen *s,
+			    const struct bch_hash_desc *desc,
+			    struct bch_hash_info *hash_info,
+			    struct btree_iter *k_iter, struct bkey_s_c hash_k)
+{
+	if (hash_k.k->type != desc->key_type)
+		return 0;
+
+	if (likely(desc->hash_bkey(hash_info, hash_k) == hash_k.k->p.offset))
+		return 0;
+
+	return __bch2_str_hash_check_key(trans, s, desc, hash_info, k_iter, hash_k);
+}
 
 #endif /* _BCACHEFS_STR_HASH_H */
diff --git a/libbcachefs/subvolume.c b/libbcachefs/subvolume.c
index 0e756e35..e3d04752 100644
--- a/libbcachefs/subvolume.c
+++ b/libbcachefs/subvolume.c
@@ -409,26 +409,56 @@ static int bch2_subvolumes_reparent(struct btree_trans *trans, u32 subvolid_to_d
  */
 static int __bch2_subvolume_delete(struct btree_trans *trans, u32 subvolid)
 {
-	struct btree_iter iter;
-	struct bkey_s_c_subvolume subvol;
-	u32 snapid;
-	int ret = 0;
+	struct btree_iter subvol_iter = {}, snapshot_iter = {}, snapshot_tree_iter = {};
 
-	subvol = bch2_bkey_get_iter_typed(trans, &iter,
+	struct bkey_s_c_subvolume subvol =
+		bch2_bkey_get_iter_typed(trans, &subvol_iter,
 				BTREE_ID_subvolumes, POS(0, subvolid),
 				BTREE_ITER_cached|BTREE_ITER_intent,
 				subvolume);
-	ret = bkey_err(subvol);
+	int ret = bkey_err(subvol);
 	bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), trans->c,
 				"missing subvolume %u", subvolid);
 	if (ret)
-		return ret;
+		goto err;
 
-	snapid = le32_to_cpu(subvol.v->snapshot);
+	u32 snapid = le32_to_cpu(subvol.v->snapshot);
+
+	struct bkey_s_c_snapshot snapshot =
+		bch2_bkey_get_iter_typed(trans, &snapshot_iter,
+				BTREE_ID_snapshots, POS(0, snapid),
+				0, snapshot);
+	ret = bkey_err(subvol);
+	bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), trans->c,
+				"missing snapshot %u", snapid);
+	if (ret)
+		goto err;
+
+	u32 treeid = le32_to_cpu(snapshot.v->tree);
 
-	ret =   bch2_btree_delete_at(trans, &iter, 0) ?:
+	struct bkey_s_c_snapshot_tree snapshot_tree =
+		bch2_bkey_get_iter_typed(trans, &snapshot_tree_iter,
+				BTREE_ID_snapshot_trees, POS(0, treeid),
+				0, snapshot_tree);
+
+	if (le32_to_cpu(snapshot_tree.v->master_subvol) == subvolid) {
+		struct bkey_i_snapshot_tree *snapshot_tree_mut =
+			bch2_bkey_make_mut_typed(trans, &snapshot_tree_iter,
+						 &snapshot_tree.s_c,
+						 0, snapshot_tree);
+		ret = PTR_ERR_OR_ZERO(snapshot_tree_mut);
+		if (ret)
+			goto err;
+
+		snapshot_tree_mut->v.master_subvol = 0;
+	}
+
+	ret =   bch2_btree_delete_at(trans, &subvol_iter, 0) ?:
 		bch2_snapshot_node_set_deleted(trans, snapid);
-	bch2_trans_iter_exit(trans, &iter);
+err:
+	bch2_trans_iter_exit(trans, &snapshot_tree_iter);
+	bch2_trans_iter_exit(trans, &snapshot_iter);
+	bch2_trans_iter_exit(trans, &subvol_iter);
 	return ret;
 }
 
diff --git a/libbcachefs/subvolume_types.h b/libbcachefs/subvolume_types.h
index f2ec4277..1549d6da 100644
--- a/libbcachefs/subvolume_types.h
+++ b/libbcachefs/subvolume_types.h
@@ -9,13 +9,13 @@ typedef DARRAY(u32) snapshot_id_list;
 #define IS_ANCESTOR_BITMAP	128
 
 struct snapshot_t {
+	bool			live;
 	u32			parent;
 	u32			skip[3];
 	u32			depth;
 	u32			children[2];
 	u32			subvol; /* Nonzero only if a subvolume points to this node: */
 	u32			tree;
-	u32			equiv;
 	unsigned long		is_ancestor[BITS_TO_LONGS(IS_ANCESTOR_BITMAP)];
 };
 
diff --git a/libbcachefs/super-io.c b/libbcachefs/super-io.c
index dbc09e30..8037ccba 100644
--- a/libbcachefs/super-io.c
+++ b/libbcachefs/super-io.c
@@ -1084,9 +1084,16 @@ int bch2_write_super(struct bch_fs *c)
 				": Superblock write was silently dropped! (seq %llu expected %llu)",
 				le64_to_cpu(ca->sb_read_scratch->seq),
 				ca->disk_sb.seq);
-			bch2_fs_fatal_error(c, "%s", buf.buf);
+
+			if (c->opts.errors != BCH_ON_ERROR_continue &&
+			    c->opts.errors != BCH_ON_ERROR_fix_safe) {
+				ret = -BCH_ERR_erofs_sb_err;
+				bch2_fs_fatal_error(c, "%s", buf.buf);
+			} else {
+				bch_err(c, "%s", buf.buf);
+			}
+
 			printbuf_exit(&buf);
-			ret = -BCH_ERR_erofs_sb_err;
 		}
 
 		if (le64_to_cpu(ca->sb_read_scratch->seq) > ca->disk_sb.seq) {
diff --git a/libbcachefs/super.c b/libbcachefs/super.c
index bd6aecea..d97ea7bd 100644
--- a/libbcachefs/super.c
+++ b/libbcachefs/super.c
@@ -563,6 +563,7 @@ 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_fs_btree_gc_exit(c);
 	bch2_journal_keys_put_initial(c);
 	bch2_find_btree_nodes_exit(&c->found_btree_nodes);
 	BUG_ON(atomic_read(&c->journal_keys.ref));
@@ -770,13 +771,9 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	spin_lock_init(&c->recovery_pass_lock);
 	sema_init(&c->online_fsck_mutex, 1);
 
-	init_rwsem(&c->gc_lock);
-	mutex_init(&c->gc_gens_lock);
-
 	for (i = 0; i < BCH_TIME_STAT_NR; i++)
 		bch2_time_stats_init(&c->times[i]);
 
-	bch2_fs_gc_init(c);
 	bch2_fs_copygc_init(c);
 	bch2_fs_btree_key_cache_init_early(&c->btree_key_cache);
 	bch2_fs_btree_iter_init_early(c);
@@ -911,6 +908,7 @@ static struct bch_fs *bch2_fs_alloc(struct bch_sb *sb, struct bch_opts opts)
 	    bch2_fs_btree_cache_init(c) ?:
 	    bch2_fs_btree_key_cache_init(&c->btree_key_cache) ?:
 	    bch2_fs_btree_interior_update_init(c) ?:
+	    bch2_fs_btree_gc_init(c) ?:
 	    bch2_fs_buckets_waiting_for_journal_init(c) ?:
 	    bch2_fs_btree_write_buffer_init(c) ?:
 	    bch2_fs_subvolumes_init(c) ?:
@@ -1306,8 +1304,6 @@ static struct bch_dev *__bch2_dev_alloc(struct bch_fs *c,
 	init_completion(&ca->ref_completion);
 	init_completion(&ca->io_ref_completion);
 
-	init_rwsem(&ca->bucket_lock);
-
 	INIT_WORK(&ca->io_error_work, bch2_io_error_work);
 
 	bch2_time_stats_quantiles_init(&ca->io_latency[READ]);
diff --git a/libbcachefs/sysfs.c b/libbcachefs/sysfs.c
index 97733c76..a7eb1f51 100644
--- a/libbcachefs/sysfs.c
+++ b/libbcachefs/sysfs.c
@@ -203,7 +203,6 @@ read_attribute(disk_groups);
 
 read_attribute(has_data);
 read_attribute(alloc_debug);
-read_attribute(accounting);
 read_attribute(usage_base);
 
 #define x(t, n, ...) read_attribute(t);
@@ -397,9 +396,6 @@ SHOW(bch2_fs)
 	if (attr == &sysfs_alloc_debug)
 		bch2_fs_alloc_debug_to_text(out, c);
 
-	if (attr == &sysfs_accounting)
-		bch2_fs_accounting_to_text(out, c);
-
 	if (attr == &sysfs_usage_base)
 		bch2_fs_usage_base_to_text(out, c);
 
@@ -509,15 +505,22 @@ SHOW(bch2_fs_counters)
 
 	printbuf_tabstop_push(out, 32);
 
-	#define x(t, ...) \
+	#define x(t, n, f, ...) \
 		if (attr == &sysfs_##t) {					\
 			counter             = percpu_u64_get(&c->counters[BCH_COUNTER_##t]);\
 			counter_since_mount = counter - c->counters_on_mount[BCH_COUNTER_##t];\
+			if (f & TYPE_SECTORS) {					\
+				counter <<= 9;					\
+				counter_since_mount <<= 9;			\
+			}							\
+										\
 			prt_printf(out, "since mount:\t");			\
+			(f & TYPE_COUNTER) ? prt_u64(out, counter_since_mount) :\
 			prt_human_readable_u64(out, counter_since_mount);	\
 			prt_newline(out);					\
 										\
 			prt_printf(out, "since filesystem creation:\t");	\
+			(f & TYPE_COUNTER) ? prt_u64(out, counter) :		\
 			prt_human_readable_u64(out, counter);			\
 			prt_newline(out);					\
 		}
@@ -595,7 +598,6 @@ struct attribute *bch2_fs_internal_files[] = {
 
 	&sysfs_disk_groups,
 	&sysfs_alloc_debug,
-	&sysfs_accounting,
 	&sysfs_usage_base,
 	NULL
 };
diff --git a/libbcachefs/trace.h b/libbcachefs/trace.h
index 7baf66be..9d40b7d4 100644
--- a/libbcachefs/trace.h
+++ b/libbcachefs/trace.h
@@ -1338,6 +1338,12 @@ TRACE_EVENT(trans_restart_key_cache_key_realloced,
 		  __entry->new_u64s)
 );
 
+DEFINE_EVENT(transaction_event,	trans_restart_write_buffer_flush,
+	TP_PROTO(struct btree_trans *trans,
+		 unsigned long caller_ip),
+	TP_ARGS(trans, caller_ip)
+);
+
 TRACE_EVENT(path_downgrade,
 	TP_PROTO(struct btree_trans *trans,
 		 unsigned long caller_ip,
@@ -1374,10 +1380,21 @@ TRACE_EVENT(path_downgrade,
 		  __entry->pos_snapshot)
 );
 
-DEFINE_EVENT(transaction_event,	trans_restart_write_buffer_flush,
-	TP_PROTO(struct btree_trans *trans,
-		 unsigned long caller_ip),
-	TP_ARGS(trans, caller_ip)
+TRACE_EVENT(key_cache_fill,
+	TP_PROTO(struct btree_trans *trans, const char *key),
+	TP_ARGS(trans, key),
+
+	TP_STRUCT__entry(
+		__array(char,		trans_fn, 32	)
+		__string(key,		key			)
+	),
+
+	TP_fast_assign(
+		strscpy(__entry->trans_fn, trans->fn, sizeof(__entry->trans_fn));
+		__assign_str(key);
+	),
+
+	TP_printk("%s %s", __entry->trans_fn, __get_str(key))
 );
 
 TRACE_EVENT(write_buffer_flush,
@@ -1436,6 +1453,24 @@ TRACE_EVENT(write_buffer_flush_slowpath,
 	TP_printk("%zu/%zu", __entry->slowpath, __entry->total)
 );
 
+TRACE_EVENT(write_buffer_maybe_flush,
+	TP_PROTO(struct btree_trans *trans, unsigned long caller_ip, const char *key),
+	TP_ARGS(trans, caller_ip, key),
+
+	TP_STRUCT__entry(
+		__array(char,			trans_fn, 32	)
+		__field(unsigned long,		caller_ip	)
+		__string(key,			key		)
+	),
+
+	TP_fast_assign(
+		strscpy(__entry->trans_fn, trans->fn, sizeof(__entry->trans_fn));
+		__assign_str(key);
+	),
+
+	TP_printk("%s %pS %s", __entry->trans_fn, (void *) __entry->caller_ip, __get_str(key))
+);
+
 DEFINE_EVENT(fs_str, rebalance_extent,
 	TP_PROTO(struct bch_fs *c, const char *str),
 	TP_ARGS(c, str)
diff --git a/libbcachefs/util.h b/libbcachefs/util.h
index c292b9ce..1a172011 100644
--- a/libbcachefs/util.h
+++ b/libbcachefs/util.h
@@ -55,6 +55,16 @@ static inline size_t buf_pages(void *p, size_t len)
 			    PAGE_SIZE);
 }
 
+static inline void *bch2_kvmalloc(size_t n, gfp_t flags)
+{
+	void *p = unlikely(n >= INT_MAX)
+		? vmalloc(n)
+		: kvmalloc(n, flags & ~__GFP_ZERO);
+	if (p && (flags & __GFP_ZERO))
+		memset(p, 0, n);
+	return p;
+}
+
 #define init_heap(heap, _size, gfp)					\
 ({									\
 	(heap)->nr = 0;						\
diff --git a/libbcachefs/varint.c b/libbcachefs/varint.c
index 6a78553d..6620ecae 100644
--- a/libbcachefs/varint.c
+++ b/libbcachefs/varint.c
@@ -9,6 +9,7 @@
 #include <valgrind/memcheck.h>
 #endif
 
+#include "errcode.h"
 #include "varint.h"
 
 /**
@@ -53,7 +54,7 @@ int bch2_varint_decode(const u8 *in, const u8 *end, u64 *out)
 	u64 v;
 
 	if (unlikely(in + bytes > end))
-		return -1;
+		return -BCH_ERR_varint_decode_error;
 
 	if (likely(bytes < 9)) {
 		__le64 v_le = 0;
@@ -115,7 +116,7 @@ int bch2_varint_decode_fast(const u8 *in, const u8 *end, u64 *out)
 	unsigned bytes = ffz(*in) + 1;
 
 	if (unlikely(in + bytes > end))
-		return -1;
+		return -BCH_ERR_varint_decode_error;
 
 	if (likely(bytes < 9)) {
 		v >>= bytes;