Update bcachefs sources to 1b14994029 bcachefs: Fragmentation LRU

29 files changed, 474 insertions, 344 deletions

diff --git a/.bcachefs_revision b/.bcachefs_revision
index bfcbcf58..7b2b6b45 100644
--- a/.bcachefs_revision
+++ b/.bcachefs_revision
@@ -1 +1 @@
-8dbfede1d9e6483c682956c7c8a4900a65f98dde
+1b149940290c0ef39070b4afaadab84a65bba034
diff --git a/include/trace/events/bcachefs.h b/include/trace/events/bcachefs.h
index 10e51bb3..b9dca1d2 100644
--- a/include/trace/events/bcachefs.h
+++ b/include/trace/events/bcachefs.h
@@ -723,8 +723,8 @@ TRACE_EVENT(move_data,
 TRACE_EVENT(evacuate_bucket,
 	TP_PROTO(struct bch_fs *c, struct bpos *bucket,
 		 unsigned sectors, unsigned bucket_size,
-		 int ret),
-	TP_ARGS(c, bucket, sectors, bucket_size, ret),
+		 u64 fragmentation, int ret),
+	TP_ARGS(c, bucket, sectors, bucket_size, fragmentation, ret),
 
 	TP_STRUCT__entry(
 		__field(dev_t,		dev		)
@@ -732,6 +732,7 @@ TRACE_EVENT(evacuate_bucket,
 		__field(u64,		bucket		)
 		__field(u32,		sectors		)
 		__field(u32,		bucket_size	)
+		__field(u64,		fragmentation	)
 		__field(int,		ret		)
 	),
 
@@ -741,14 +742,15 @@ TRACE_EVENT(evacuate_bucket,
 		__entry->bucket			= bucket->offset;
 		__entry->sectors		= sectors;
 		__entry->bucket_size		= bucket_size;
+		__entry->fragmentation		= fragmentation;
 		__entry->ret			= ret;
 	),
 
-	TP_printk("%d,%d %llu:%llu sectors %u/%u ret %i",
+	TP_printk("%d,%d %llu:%llu sectors %u/%u fragmentation %llu ret %i",
 		  MAJOR(__entry->dev), MINOR(__entry->dev),
 		  __entry->member, __entry->bucket,
 		  __entry->sectors, __entry->bucket_size,
-		  __entry->ret)
+		  __entry->fragmentation, __entry->ret)
 );
 
 TRACE_EVENT(copygc,
diff --git a/libbcachefs/alloc_background.c b/libbcachefs/alloc_background.c
index 2db44365..755faa34 100644
--- a/libbcachefs/alloc_background.c
+++ b/libbcachefs/alloc_background.c
@@ -9,6 +9,7 @@
 #include "btree_update.h"
 #include "btree_update_interior.h"
 #include "btree_gc.h"
+#include "btree_write_buffer.h"
 #include "buckets.h"
 #include "buckets_waiting_for_journal.h"
 #include "clock.h"
@@ -414,6 +415,8 @@ void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c
 	prt_newline(out);
 	prt_printf(out, "io_time[WRITE]    %llu",	a->io_time[WRITE]);
 	prt_newline(out);
+	prt_printf(out, "fragmentation     %llu",	a->fragmentation_lru);
+	prt_newline(out);
 	prt_printf(out, "bp_start          %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a));
 	prt_newline(out);
 
@@ -909,8 +912,8 @@ int bch2_trans_mark_alloc(struct btree_trans *trans,
 	    !new_a->io_time[READ])
 		new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 
-	old_lru = alloc_lru_idx(*old_a);
-	new_lru = alloc_lru_idx(*new_a);
+	old_lru = alloc_lru_idx_read(*old_a);
+	new_lru = alloc_lru_idx_read(*new_a);
 
 	if (old_lru != new_lru) {
 		ret = bch2_lru_change(trans, new->k.p.inode,
@@ -920,6 +923,18 @@ int bch2_trans_mark_alloc(struct btree_trans *trans,
 			return ret;
 	}
 
+	new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
+					bch_dev_bkey_exists(c, new->k.p.inode));
+
+	if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
+		ret = bch2_lru_change(trans,
+				BCH_LRU_FRAGMENTATION_START,
+				bucket_to_u64(new->k.p),
+				old_a->fragmentation_lru, new_a->fragmentation_lru);
+		if (ret)
+			return ret;
+	}
+
 	if (old_a->gen != new_a->gen) {
 		ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
 		if (ret)
@@ -1775,15 +1790,11 @@ static int invalidate_one_bucket(struct btree_trans *trans,
 	if (ret)
 		goto out;
 
-	if (lru_pos_time(lru_iter->pos) != alloc_lru_idx(a->v)) {
-		prt_str(&buf, "alloc key does not point back to lru entry when invalidating bucket:");
-		goto err;
-	}
+	/* We expect harmless races here due to the btree write buffer: */
+	if (lru_pos_time(lru_iter->pos) != alloc_lru_idx_read(a->v))
+		goto out;
 
-	if (a->v.data_type != BCH_DATA_cached) {
-		prt_str(&buf, "lru entry points to non cached bucket:");
-		goto err;
-	}
+	BUG_ON(a->v.data_type != BCH_DATA_cached);
 
 	if (!a->v.cached_sectors)
 		bch_err(c, "invalidating empty bucket, confused");
@@ -1845,6 +1856,10 @@ static void bch2_do_invalidates_work(struct work_struct *work)
 
 	bch2_trans_init(&trans, c, 0, 0);
 
+	ret = bch2_btree_write_buffer_flush(&trans);
+	if (ret)
+		goto err;
+
 	for_each_member_device(ca, c, i) {
 		s64 nr_to_invalidate =
 			should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
@@ -1860,7 +1875,7 @@ static void bch2_do_invalidates_work(struct work_struct *work)
 			break;
 		}
 	}
-
+err:
 	bch2_trans_exit(&trans);
 	bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
 }
diff --git a/libbcachefs/alloc_background.h b/libbcachefs/alloc_background.h
index b3c2f1e0..96ac8f39 100644
--- a/libbcachefs/alloc_background.h
+++ b/libbcachefs/alloc_background.h
@@ -64,11 +64,24 @@ static inline enum bch_data_type alloc_data_type(struct bch_alloc_v4 a,
 				 a.stripe, a, data_type);
 }
 
-static inline u64 alloc_lru_idx(struct bch_alloc_v4 a)
+static inline u64 alloc_lru_idx_read(struct bch_alloc_v4 a)
 {
 	return a.data_type == BCH_DATA_cached ? a.io_time[READ] : 0;
 }
 
+static inline u64 alloc_lru_idx_fragmentation(struct bch_alloc_v4 a,
+					      struct bch_dev *ca)
+{
+	if (a.data_type != BCH_DATA_btree &&
+	    a.data_type != BCH_DATA_user)
+		return 0;
+
+	if (a.dirty_sectors >= ca->mi.bucket_size)
+		return 0;
+
+	return div_u64((u64) a.dirty_sectors * (1ULL << 31), ca->mi.bucket_size);
+}
+
 static inline u64 alloc_freespace_genbits(struct bch_alloc_v4 a)
 {
 	return ((u64) alloc_gc_gen(a) >> 4) << 56;
diff --git a/libbcachefs/backpointers.c b/libbcachefs/backpointers.c
index c269fc73..e001f419 100644
--- a/libbcachefs/backpointers.c
+++ b/libbcachefs/backpointers.c
@@ -738,7 +738,7 @@ static size_t btree_nodes_fit_in_ram(struct bch_fs *c)
 
 	si_meminfo(&i);
 	mem_bytes = i.totalram * i.mem_unit;
-	return (mem_bytes >> 1) / btree_bytes(c);
+	return div_u64(mem_bytes >> 1, btree_bytes(c));
 }
 
 int bch2_get_btree_in_memory_pos(struct btree_trans *trans,
diff --git a/libbcachefs/bcachefs.h b/libbcachefs/bcachefs.h
index c8620127..b10e39d8 100644
--- a/libbcachefs/bcachefs.h
+++ b/libbcachefs/bcachefs.h
@@ -927,7 +927,6 @@ struct bch_fs {
 
 	/* COPYGC */
 	struct task_struct	*copygc_thread;
-	copygc_heap		copygc_heap;
 	struct write_point	copygc_write_point;
 	s64			copygc_wait;
 	bool			copygc_running;
diff --git a/libbcachefs/bcachefs_format.h b/libbcachefs/bcachefs_format.h
index 7574eccd..fdd0050c 100644
--- a/libbcachefs/bcachefs_format.h
+++ b/libbcachefs/bcachefs_format.h
@@ -988,6 +988,7 @@ struct bch_alloc_v4 {
 	__u64			io_time[2];
 	__u32			stripe;
 	__u32			nr_external_backpointers;
+	__u64			fragmentation_lru;
 } __packed __aligned(8);
 
 #define BCH_ALLOC_V4_U64s_V0	6
@@ -1559,7 +1560,8 @@ struct bch_sb_field_journal_seq_blacklist {
 	x(inode_v3,			23)		\
 	x(unwritten_extents,		24)		\
 	x(bucket_gens,			25)		\
-	x(lru_v2,			26)
+	x(lru_v2,			26)		\
+	x(fragmentation_lru,		27)
 
 enum bcachefs_metadata_version {
 	bcachefs_metadata_version_min = 9,
diff --git a/libbcachefs/btree_iter.c b/libbcachefs/btree_iter.c
index 3977bb1f..dc2b2a08 100644
--- a/libbcachefs/btree_iter.c
+++ b/libbcachefs/btree_iter.c
@@ -1174,17 +1174,10 @@ int bch2_btree_path_traverse_one(struct btree_trans *trans,
 
 	path->uptodate = BTREE_ITER_UPTODATE;
 out:
-	if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted) {
-		struct printbuf buf = PRINTBUF;
-
-		prt_printf(&buf, "ret %s (%i) trans->restarted %s (%i)\n",
-			   bch2_err_str(ret), ret,
-			   bch2_err_str(trans->restarted), trans->restarted);
-#ifdef CONFIG_BCACHEFS_DEBUG
-		bch2_prt_backtrace(&buf, &trans->last_restarted);
-#endif
-		panic("%s", buf.buf);
-	}
+	if (bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted)
+		panic("ret %s (%i) trans->restarted %s (%i)\n",
+		      bch2_err_str(ret), ret,
+		      bch2_err_str(trans->restarted), trans->restarted);
 	bch2_btree_path_verify(trans, path);
 	return ret;
 }
@@ -1367,14 +1360,14 @@ void bch2_trans_restart_error(struct btree_trans *trans, u32 restart_count)
 {
 	panic("trans->restart_count %u, should be %u, last restarted by %pS\n",
 	      trans->restart_count, restart_count,
-	      (void *) trans->last_begin_ip);
+	      (void *) trans->last_restarted_ip);
 }
 
 void bch2_trans_in_restart_error(struct btree_trans *trans)
 {
 	panic("in transaction restart: %s, last restarted by %pS\n",
 	      bch2_err_str(trans->restarted),
-	      (void *) trans->last_begin_ip);
+	      (void *) trans->last_restarted_ip);
 }
 
 noinline __cold
@@ -2872,7 +2865,7 @@ u32 bch2_trans_begin(struct btree_trans *trans)
 	if (unlikely(time_after(jiffies, trans->srcu_lock_time + msecs_to_jiffies(10))))
 		bch2_trans_reset_srcu_lock(trans);
 
-	trans->last_begin_ip = _RET_IP_;
+	trans->last_restarted_ip = _RET_IP_;
 	if (trans->restarted) {
 		bch2_btree_path_traverse_all(trans);
 		trans->notrace_relock_fail = false;
@@ -3053,10 +3046,6 @@ void bch2_trans_exit(struct btree_trans *trans)
 	if (trans->paths)
 		mempool_free(trans->paths, &c->btree_paths_pool);
 
-#ifdef CONFIG_BCACHEFS_DEBUG
-	darray_exit(&trans->last_restarted);
-#endif
-
 	trans->mem	= (void *) 0x1;
 	trans->paths	= (void *) 0x1;
 }
diff --git a/libbcachefs/btree_iter.h b/libbcachefs/btree_iter.h
index bbbbe52b..0ede02c3 100644
--- a/libbcachefs/btree_iter.h
+++ b/libbcachefs/btree_iter.h
@@ -251,10 +251,6 @@ static inline int btree_trans_restart_nounlock(struct btree_trans *trans, int er
 	BUG_ON(err <= 0);
 	BUG_ON(!bch2_err_matches(err, BCH_ERR_transaction_restart));
 
-#ifdef CONFIG_BCACHEFS_DEBUG
-	bch2_save_backtrace(&trans->last_restarted, current);
-#endif
-
 	trans->restarted = err;
 	return -err;
 }
diff --git a/libbcachefs/btree_types.h b/libbcachefs/btree_types.h
index 97ff267c..ad73cd2e 100644
--- a/libbcachefs/btree_types.h
+++ b/libbcachefs/btree_types.h
@@ -442,10 +442,7 @@ struct btree_trans {
 	bool			notrace_relock_fail:1;
 	enum bch_errcode	restarted:16;
 	u32			restart_count;
-#ifdef CONFIG_BCACHEFS_DEBUG
-	bch_stacktrace		last_restarted;
-#endif
-	unsigned long		last_begin_ip;
+	unsigned long		last_restarted_ip;
 	unsigned long		srcu_lock_time;
 
 	/*
diff --git a/libbcachefs/btree_update.h b/libbcachefs/btree_update.h
index 771e4b23..ee1d1593 100644
--- a/libbcachefs/btree_update.h
+++ b/libbcachefs/btree_update.h
@@ -58,6 +58,9 @@ int bch2_btree_delete_extent_at(struct btree_trans *, struct btree_iter *,
 				unsigned, unsigned);
 int bch2_btree_delete_at(struct btree_trans *, struct btree_iter *, unsigned);
 
+int bch2_btree_insert_nonextent(struct btree_trans *, enum btree_id,
+				struct bkey_i *, enum btree_update_flags);
+
 int __bch2_btree_insert(struct btree_trans *, enum btree_id, struct bkey_i *,
 			enum btree_update_flags);
 int bch2_btree_insert(struct bch_fs *, enum btree_id, struct bkey_i *,
diff --git a/libbcachefs/btree_update_leaf.c b/libbcachefs/btree_update_leaf.c
index 0195b13d..20ad7989 100644
--- a/libbcachefs/btree_update_leaf.c
+++ b/libbcachefs/btree_update_leaf.c
@@ -56,9 +56,10 @@ static void verify_update_old_key(struct btree_trans *trans, struct btree_insert
 			k = bkey_i_to_s_c(j_k);
 	}
 
-	i->old_k.needs_whiteout = k.k->needs_whiteout;
+	u = *k.k;
+	u.needs_whiteout = i->old_k.needs_whiteout;
 
-	BUG_ON(memcmp(&i->old_k, k.k, sizeof(struct bkey)));
+	BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
 	BUG_ON(i->old_v != k.v);
 #endif
 }
@@ -1306,12 +1307,45 @@ static noinline int extent_back_merge(struct btree_trans *trans,
 	return 0;
 }
 
+/*
+ * When deleting, check if we need to emit a whiteout (because we're overwriting
+ * something in an ancestor snapshot)
+ */
+static int need_whiteout_for_snapshot(struct btree_trans *trans,
+				      enum btree_id btree_id, struct bpos pos)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	u32 snapshot = pos.snapshot;
+	int ret;
+
+	if (!bch2_snapshot_parent(trans->c, pos.snapshot))
+		return 0;
+
+	pos.snapshot++;
+
+	for_each_btree_key_norestart(trans, iter, btree_id, pos,
+			   BTREE_ITER_ALL_SNAPSHOTS|
+			   BTREE_ITER_NOPRESERVE, k, ret) {
+		if (!bkey_eq(k.k->p, pos))
+			break;
+
+		if (bch2_snapshot_is_ancestor(trans->c, snapshot,
+					      k.k->p.snapshot)) {
+			ret = !bkey_whiteout(k.k);
+			break;
+		}
+	}
+	bch2_trans_iter_exit(trans, &iter);
+
+	return ret;
+}
 int bch2_trans_update_extent(struct btree_trans *trans,
 			     struct btree_iter *orig_iter,
 			     struct bkey_i *insert,
 			     enum btree_update_flags flags)
 {
-	struct btree_iter iter, update_iter;
+	struct btree_iter iter;
 	struct bpos start = bkey_start_pos(&insert->k);
 	struct bkey_i *update;
 	struct bkey_s_c k;
@@ -1359,16 +1393,8 @@ int bch2_trans_update_extent(struct btree_trans *trans,
 
 			bch2_cut_back(start, update);
 
-			bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
-					     BTREE_ITER_NOT_EXTENTS|
-					     BTREE_ITER_ALL_SNAPSHOTS|
-					     BTREE_ITER_INTENT);
-			ret   = bch2_btree_iter_traverse(&update_iter) ?:
-				bch2_trans_update(trans, &update_iter, update,
-						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
-						  flags);
-			bch2_trans_iter_exit(trans, &update_iter);
-
+			ret = bch2_btree_insert_nonextent(trans, btree_id, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 			if (ret)
 				goto err;
 		}
@@ -1382,15 +1408,8 @@ int bch2_trans_update_extent(struct btree_trans *trans,
 			bch2_cut_front(start, update);
 			bch2_cut_back(insert->k.p, update);
 
-			bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
-					     BTREE_ITER_NOT_EXTENTS|
-					     BTREE_ITER_ALL_SNAPSHOTS|
-					     BTREE_ITER_INTENT);
-			ret   = bch2_btree_iter_traverse(&update_iter) ?:
-				bch2_trans_update(trans, &update_iter, update,
-						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
-						  flags);
-			bch2_trans_iter_exit(trans, &update_iter);
+			ret = bch2_btree_insert_nonextent(trans, btree_id, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 			if (ret)
 				goto err;
 		}
@@ -1402,21 +1421,15 @@ int bch2_trans_update_extent(struct btree_trans *trans,
 
 			bkey_init(&update->k);
 			update->k.p = k.k->p;
+			update->k.p.snapshot = insert->k.p.snapshot;
 
-			if (insert->k.p.snapshot != k.k->p.snapshot) {
-				update->k.p.snapshot = insert->k.p.snapshot;
+			if (insert->k.p.snapshot != k.k->p.snapshot ||
+			    (btree_type_has_snapshots(btree_id) &&
+			     need_whiteout_for_snapshot(trans, btree_id, update->k.p)))
 				update->k.type = KEY_TYPE_whiteout;
-			}
-
-			bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
-					     BTREE_ITER_NOT_EXTENTS|
-					     BTREE_ITER_INTENT);
-			ret   = bch2_btree_iter_traverse(&update_iter) ?:
-				bch2_trans_update(trans, &update_iter, update,
-						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
-						  flags);
-			bch2_trans_iter_exit(trans, &update_iter);
 
+			ret = bch2_btree_insert_nonextent(trans, btree_id, update,
+						  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 			if (ret)
 				goto err;
 		}
@@ -1468,40 +1481,6 @@ err:
 	return ret;
 }
 
-/*
- * When deleting, check if we need to emit a whiteout (because we're overwriting
- * something in an ancestor snapshot)
- */
-static int need_whiteout_for_snapshot(struct btree_trans *trans,
-				      enum btree_id btree_id, struct bpos pos)
-{
-	struct btree_iter iter;
-	struct bkey_s_c k;
-	u32 snapshot = pos.snapshot;
-	int ret;
-
-	if (!bch2_snapshot_parent(trans->c, pos.snapshot))
-		return 0;
-
-	pos.snapshot++;
-
-	for_each_btree_key_norestart(trans, iter, btree_id, pos,
-			   BTREE_ITER_ALL_SNAPSHOTS|
-			   BTREE_ITER_NOPRESERVE, k, ret) {
-		if (!bkey_eq(k.k->p, pos))
-			break;
-
-		if (bch2_snapshot_is_ancestor(trans->c, snapshot,
-					      k.k->p.snapshot)) {
-			ret = !bkey_whiteout(k.k);
-			break;
-		}
-	}
-	bch2_trans_iter_exit(trans, &iter);
-
-	return ret;
-}
-
 static int __must_check
 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
 				struct bkey_i *k, enum btree_update_flags flags,
@@ -1747,8 +1726,23 @@ void bch2_trans_commit_hook(struct btree_trans *trans,
 	trans->hooks = h;
 }
 
-int __bch2_btree_insert(struct btree_trans *trans,
-			enum btree_id id,
+int bch2_btree_insert_nonextent(struct btree_trans *trans,
+				enum btree_id btree, struct bkey_i *k,
+				enum btree_update_flags flags)
+{
+	struct btree_iter iter;
+	int ret;
+
+	bch2_trans_iter_init(trans, &iter, btree, k->k.p,
+			     BTREE_ITER_NOT_EXTENTS|
+			     BTREE_ITER_INTENT);
+	ret   = bch2_btree_iter_traverse(&iter) ?:
+		bch2_trans_update(trans, &iter, k, flags);
+	bch2_trans_iter_exit(trans, &iter);
+	return ret;
+}
+
+int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id,
 			struct bkey_i *k, enum btree_update_flags flags)
 {
 	struct btree_iter iter;
diff --git a/libbcachefs/btree_write_buffer.c b/libbcachefs/btree_write_buffer.c
index 05b755a0..6285532e 100644
--- a/libbcachefs/btree_write_buffer.c
+++ b/libbcachefs/btree_write_buffer.c
@@ -88,6 +88,8 @@ static union btree_write_buffer_state btree_write_buffer_switch(struct btree_wri
 	while (old.idx == 0 ? wb->state.ref0 : wb->state.ref1)
 		cpu_relax();
 
+	smp_mb();
+
 	return old;
 }
 
diff --git a/libbcachefs/buckets_types.h b/libbcachefs/buckets_types.h
index 1dbba7d9..2a9dab90 100644
--- a/libbcachefs/buckets_types.h
+++ b/libbcachefs/buckets_types.h
@@ -89,15 +89,4 @@ struct disk_reservation {
 	unsigned		nr_replicas;
 };
 
-struct copygc_heap_entry {
-	u8			dev;
-	u8			gen;
-	u8			replicas;
-	u32			fragmentation;
-	u32			sectors;
-	u64			bucket;
-};
-
-typedef HEAP(struct copygc_heap_entry) copygc_heap;
-
 #endif /* _BUCKETS_TYPES_H */
diff --git a/libbcachefs/data_update.c b/libbcachefs/data_update.c
index 199e894e..de0575f6 100644
--- a/libbcachefs/data_update.c
+++ b/libbcachefs/data_update.c
@@ -22,9 +22,10 @@ static int insert_snapshot_whiteouts(struct btree_trans *trans,
 				     struct bpos new_pos)
 {
 	struct bch_fs *c = trans->c;
-	struct btree_iter iter, update_iter;
-	struct bkey_s_c k;
+	struct btree_iter iter, iter2;
+	struct bkey_s_c k, k2;
 	snapshot_id_list s;
+	struct bkey_i *update;
 	int ret;
 
 	if (!btree_type_has_snapshots(id))
@@ -32,10 +33,7 @@ static int insert_snapshot_whiteouts(struct btree_trans *trans,
 
 	darray_init(&s);
 
-	if (bkey_eq(old_pos, new_pos))
-		return 0;
-
-	if (!snapshot_t(c, old_pos.snapshot)->children[0])
+	if (!bch2_snapshot_has_children(c, old_pos.snapshot))
 		return 0;
 
 	bch2_trans_iter_init(trans, &iter, id, old_pos,
@@ -47,33 +45,39 @@ static int insert_snapshot_whiteouts(struct btree_trans *trans,
 		if (ret)
 			break;
 
+		if (!k.k)
+			break;
+
 		if (!bkey_eq(old_pos, k.k->p))
 			break;
 
-		if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, old_pos.snapshot)) {
-			struct bkey_i *update;
+		if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, old_pos.snapshot) &&
+		    !snapshot_list_has_ancestor(c, &s, k.k->p.snapshot)) {
+			struct bpos whiteout_pos = new_pos;
 
-			if (snapshot_list_has_ancestor(c, &s, k.k->p.snapshot))
-				continue;
+			whiteout_pos.snapshot = k.k->p.snapshot;
 
-			update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
+			bch2_trans_iter_init(trans, &iter2, id, whiteout_pos,
+					     BTREE_ITER_NOT_EXTENTS|
+					     BTREE_ITER_INTENT);
+			k2 = bch2_btree_iter_peek_slot(&iter2);
+			ret = bkey_err(k2);
 
-			ret = PTR_ERR_OR_ZERO(update);
-			if (ret)
-				break;
+			if (!ret && k2.k->type == KEY_TYPE_deleted) {
+				update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
+				ret = PTR_ERR_OR_ZERO(update);
+				if (ret)
+					break;
 
-			bkey_init(&update->k);
-			update->k.p = new_pos;
-			update->k.p.snapshot = k.k->p.snapshot;
+				bkey_init(&update->k);
+				update->k.p		= whiteout_pos;
+				update->k.type		= KEY_TYPE_whiteout;
+
+				ret = bch2_trans_update(trans, &iter2, update,
+							BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
+			}
+			bch2_trans_iter_exit(trans, &iter2);
 
-			bch2_trans_iter_init(trans, &update_iter, id, update->k.p,
-					     BTREE_ITER_NOT_EXTENTS|
-					     BTREE_ITER_ALL_SNAPSHOTS|
-					     BTREE_ITER_INTENT);
-			ret   = bch2_btree_iter_traverse(&update_iter) ?:
-				bch2_trans_update(trans, &update_iter, update,
-					  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
-			bch2_trans_iter_exit(trans, &update_iter);
 			if (ret)
 				break;
 
@@ -229,9 +233,21 @@ static int __bch2_data_update_index_update(struct btree_trans *trans,
 
 		next_pos = insert->k.p;
 
-		ret   = insert_snapshot_whiteouts(trans, m->btree_id,
-						  k.k->p, insert->k.p) ?:
-			bch2_trans_update(trans, &iter, insert,
+		if (!bkey_eq(bkey_start_pos(&insert->k), bkey_start_pos(k.k))) {
+			ret = insert_snapshot_whiteouts(trans, m->btree_id, k.k->p,
+							bkey_start_pos(&insert->k));
+			if (ret)
+				goto err;
+		}
+
+		if (!bkey_eq(insert->k.p, k.k->p)) {
+			ret = insert_snapshot_whiteouts(trans, m->btree_id,
+							k.k->p, insert->k.p);
+			if (ret)
+				goto err;
+		}
+
+		ret   = bch2_trans_update(trans, &iter, insert,
 				BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 			bch2_trans_commit(trans, &op->res,
 				NULL,
diff --git a/libbcachefs/error.c b/libbcachefs/error.c
index c2882c59..1dae649f 100644
--- a/libbcachefs/error.c
+++ b/libbcachefs/error.c
@@ -98,7 +98,6 @@ static struct fsck_err_state *fsck_err_get(struct bch_fs *c, const char *fmt)
 
 	INIT_LIST_HEAD(&s->list);
 	s->fmt = fmt;
-	s->buf = PRINTBUF;
 	list_add(&s->list, &c->fsck_errors);
 	return s;
 }
@@ -111,9 +110,23 @@ int bch2_fsck_err(struct bch_fs *c, unsigned flags, const char *fmt, ...)
 	struct printbuf buf = PRINTBUF, *out = &buf;
 	int ret = -BCH_ERR_fsck_ignore;
 
+	va_start(args, fmt);
+	prt_vprintf(out, fmt, args);
+	va_end(args);
+
 	mutex_lock(&c->fsck_error_lock);
 	s = fsck_err_get(c, fmt);
 	if (s) {
+		if (s->last_msg && !strcmp(buf.buf, s->last_msg)) {
+			ret = s->ret;
+			mutex_unlock(&c->fsck_error_lock);
+			printbuf_exit(&buf);
+			return ret;
+		}
+
+		kfree(s->last_msg);
+		s->last_msg = kstrdup(buf.buf, GFP_KERNEL);
+
 		if (c->opts.ratelimit_errors &&
 		    !(flags & FSCK_NO_RATELIMIT) &&
 		    s->nr >= FSCK_ERR_RATELIMIT_NR) {
@@ -123,8 +136,6 @@ int bch2_fsck_err(struct bch_fs *c, unsigned flags, const char *fmt, ...)
 				print = false;
 		}
 
-		printbuf_reset(&s->buf);
-		out = &s->buf;
 		s->nr++;
 	}
 
@@ -133,10 +144,6 @@ int bch2_fsck_err(struct bch_fs *c, unsigned flags, const char *fmt, ...)
 		prt_printf(out, bch2_log_msg(c, ""));
 #endif
 
-	va_start(args, fmt);
-	prt_vprintf(out, fmt, args);
-	va_end(args);
-
 	if (test_bit(BCH_FS_FSCK_DONE, &c->flags)) {
 		if (c->opts.errors != BCH_ON_ERROR_continue ||
 		    !(flags & (FSCK_CAN_FIX|FSCK_CAN_IGNORE))) {
@@ -190,6 +197,9 @@ int bch2_fsck_err(struct bch_fs *c, unsigned flags, const char *fmt, ...)
 	else if (suppressing)
 		bch_err(c, "Ratelimiting new instances of previous error");
 
+	if (s)
+		s->ret = ret;
+
 	mutex_unlock(&c->fsck_error_lock);
 
 	printbuf_exit(&buf);
@@ -214,11 +224,11 @@ void bch2_flush_fsck_errs(struct bch_fs *c)
 	mutex_lock(&c->fsck_error_lock);
 
 	list_for_each_entry_safe(s, n, &c->fsck_errors, list) {
-		if (s->ratelimited)
-			bch_err(c, "Saw %llu errors like:\n    %s", s->nr, s->buf.buf);
+		if (s->ratelimited && s->last_msg)
+			bch_err(c, "Saw %llu errors like:\n    %s", s->nr, s->last_msg);
 
 		list_del(&s->list);
-		printbuf_exit(&s->buf);
+		kfree(s->last_msg);
 		kfree(s);
 	}
 
diff --git a/libbcachefs/error.h b/libbcachefs/error.h
index 9991879d..91c7e4ee 100644
--- a/libbcachefs/error.h
+++ b/libbcachefs/error.h
@@ -103,7 +103,8 @@ struct fsck_err_state {
 	const char		*fmt;
 	u64			nr;
 	bool			ratelimited;
-	struct printbuf		buf;
+	int			ret;
+	char			*last_msg;
 };
 
 #define FSCK_CAN_FIX		(1 << 0)
diff --git a/libbcachefs/fsck.c b/libbcachefs/fsck.c
index 52bb00b5..f2768a74 100644
--- a/libbcachefs/fsck.c
+++ b/libbcachefs/fsck.c
@@ -605,6 +605,17 @@ static int ref_visible(struct bch_fs *c, struct snapshots_seen *s,
 		: bch2_snapshot_is_ancestor(c, src, dst);
 }
 
+static int ref_visible2(struct bch_fs *c,
+		       u32 src, struct snapshots_seen *src_seen,
+		       u32 dst, struct snapshots_seen *dst_seen)
+{
+	if (dst > src) {
+		swap(dst, src);
+		swap(dst_seen, src_seen);
+	}
+	return key_visible_in_snapshot(c, src_seen, dst, src);
+}
+
 #define for_each_visible_inode(_c, _s, _w, _snapshot, _i)				\
 	for (_i = (_w)->inodes.data; _i < (_w)->inodes.data + (_w)->inodes.nr &&	\
 	     (_i)->snapshot <= (_snapshot); _i++)					\
@@ -1158,10 +1169,102 @@ fsck_err:
 	return ret;
 }
 
+struct extent_end {
+	u32			snapshot;
+	u64			offset;
+	struct snapshots_seen	seen;
+};
+
+typedef DARRAY(struct extent_end) extent_ends;
+
+static int check_overlapping_extents(struct btree_trans *trans,
+			      struct snapshots_seen *seen,
+			      extent_ends *extent_ends,
+			      struct bkey_s_c k,
+			      struct btree_iter *iter)
+{
+	struct bch_fs *c = trans->c;
+	struct extent_end *i;
+	struct printbuf buf = PRINTBUF;
+	int ret = 0;
+
+	darray_for_each(*extent_ends, i) {
+		/* duplicate, due to transaction restart: */
+		if (i->offset	== k.k->p.offset &&
+		    i->snapshot == k.k->p.snapshot)
+			continue;
+
+		if (!ref_visible2(c,
+				  k.k->p.snapshot, seen,
+				  i->snapshot, &i->seen))
+			continue;
+
+		if (fsck_err_on(i->offset > bkey_start_offset(k.k), c,
+				"overlapping extents: extent in snapshot %u ends at %llu overlaps with\n%s",
+				i->snapshot,
+				i->offset,
+				(printbuf_reset(&buf),
+				 bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
+			struct bkey_i *update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+			if ((ret = PTR_ERR_OR_ZERO(update)))
+				goto err;
+			bkey_reassemble(update, k);
+			ret = bch2_trans_update_extent(trans, iter, update, 0);
+			if (!ret)
+				goto err;
+		}
+	}
+err:
+fsck_err:
+	printbuf_exit(&buf);
+	return ret;
+}
+
+static int extent_ends_at(extent_ends *extent_ends,
+			  struct snapshots_seen *seen,
+			  struct bkey_s_c k)
+{
+	struct extent_end *i, n = (struct extent_end) {
+		.snapshot	= k.k->p.snapshot,
+		.offset		= k.k->p.offset,
+		.seen		= *seen,
+	};
+
+	n.seen.ids.data = kmemdup(seen->ids.data,
+			      sizeof(seen->ids.data[0]) * seen->ids.size,
+			      GFP_KERNEL);
+	if (!n.seen.ids.data)
+		return -ENOMEM;
+
+	darray_for_each(*extent_ends, i) {
+		if (i->snapshot == k.k->p.snapshot) {
+			snapshots_seen_exit(&i->seen);
+			*i = n;
+			return 0;
+		}
+
+		if (i->snapshot >= k.k->p.snapshot)
+			break;
+	}
+
+	return darray_insert_item(extent_ends, i - extent_ends->data, n);
+}
+
+static void extent_ends_reset(extent_ends *extent_ends)
+{
+	struct extent_end *i;
+
+	darray_for_each(*extent_ends, i)
+		snapshots_seen_exit(&i->seen);
+
+	extent_ends->nr = 0;
+}
+
 static int check_extent(struct btree_trans *trans, struct btree_iter *iter,
 			struct bkey_s_c k,
 			struct inode_walker *inode,
-			struct snapshots_seen *s)
+			struct snapshots_seen *s,
+			extent_ends *extent_ends)
 {
 	struct bch_fs *c = trans->c;
 	struct inode_walker_entry *i;
@@ -1189,24 +1292,20 @@ static int check_extent(struct btree_trans *trans, struct btree_iter *iter,
 		ret = check_i_sectors(trans, inode);
 		if (ret)
 			goto err;
+
+		extent_ends_reset(extent_ends);
 	}
 
 	BUG_ON(!iter->path->should_be_locked);
-#if 0
-	if (bkey_gt(prev.k->k.p, bkey_start_pos(k.k))) {
-		char buf1[200];
-		char buf2[200];
 
-		bch2_bkey_val_to_text(&PBUF(buf1), c, bkey_i_to_s_c(prev.k));
-		bch2_bkey_val_to_text(&PBUF(buf2), c, k);
+	ret = check_overlapping_extents(trans, s, extent_ends, k, iter);
+	if (ret)
+		goto err;
+
+	ret = extent_ends_at(extent_ends, s, k);
+	if (ret)
+		goto err;
 
-		if (fsck_err(c, "overlapping extents:\n%s\n%s", buf1, buf2)) {
-			ret = fix_overlapping_extent(trans, k, prev.k->k.p)
-				?: -BCH_ERR_transaction_restart_nested;
-			goto out;
-		}
-	}
-#endif
 	ret = __walk_inode(trans, inode, equiv);
 	if (ret < 0)
 		goto err;
@@ -1304,13 +1403,9 @@ static int check_extents(struct bch_fs *c)
 	struct btree_trans trans;
 	struct btree_iter iter;
 	struct bkey_s_c k;
+	extent_ends extent_ends = { 0 };
 	int ret = 0;
 
-#if 0
-	struct bkey_buf prev;
-	bch2_bkey_buf_init(&prev);
-	prev.k->k = KEY(0, 0, 0);
-#endif
 	snapshots_seen_init(&s);
 	bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
 
@@ -1321,10 +1416,10 @@ static int check_extents(struct bch_fs *c)
 			BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS, k,
 			NULL, NULL,
 			BTREE_INSERT_LAZY_RW|BTREE_INSERT_NOFAIL,
-		check_extent(&trans, &iter, k, &w, &s));
-#if 0
-	bch2_bkey_buf_exit(&prev, c);
-#endif
+		check_extent(&trans, &iter, k, &w, &s, &extent_ends));
+
+	extent_ends_reset(&extent_ends);
+	darray_exit(&extent_ends);
 	inode_walker_exit(&w);
 	bch2_trans_exit(&trans);
 	snapshots_seen_exit(&s);
diff --git a/libbcachefs/lru.c b/libbcachefs/lru.c
index 07fb41ca..9eec12a9 100644
--- a/libbcachefs/lru.c
+++ b/libbcachefs/lru.c
@@ -49,7 +49,6 @@ void bch2_lru_pos_to_text(struct printbuf *out, struct bpos lru)
 static int __bch2_lru_set(struct btree_trans *trans, u16 lru_id,
 			u64 dev_bucket, u64 time, unsigned key_type)
 {
-	struct btree_iter iter;
 	struct bkey_i *k;
 	int ret = 0;
 
@@ -69,13 +68,7 @@ static int __bch2_lru_set(struct btree_trans *trans, u16 lru_id,
 	EBUG_ON(lru_pos_time(k->k.p) != time);
 	EBUG_ON(k->k.p.offset != dev_bucket);
 
-	bch2_trans_iter_init(trans, &iter, BTREE_ID_lru,
-			     k->k.p, BTREE_ITER_INTENT);
-
-	ret = bch2_btree_iter_traverse(&iter) ?:
-		bch2_trans_update(trans, &iter, k, 0);
-	bch2_trans_iter_exit(trans, &iter);
-	return ret;
+	return bch2_trans_update_buffered(trans, BTREE_ID_lru, k);
 }
 
 int bch2_lru_del(struct btree_trans *trans, u16 lru_id, u64 dev_bucket, u64 time)
@@ -99,6 +92,13 @@ int bch2_lru_change(struct btree_trans *trans,
 		bch2_lru_set(trans, lru_id, dev_bucket, new_time);
 }
 
+static const char * const bch2_lru_types[] = {
+#define x(n) #n,
+	BCH_LRU_TYPES()
+#undef x
+	NULL
+};
+
 static int bch2_check_lru_key(struct btree_trans *trans,
 			      struct btree_iter *lru_iter,
 			      struct bkey_s_c lru_k)
@@ -110,7 +110,9 @@ static int bch2_check_lru_key(struct btree_trans *trans,
 	const struct bch_alloc_v4 *a;
 	struct printbuf buf1 = PRINTBUF;
 	struct printbuf buf2 = PRINTBUF;
+	enum bch_lru_type type = lru_type(lru_k);
 	struct bpos alloc_pos = u64_to_bucket(lru_k.k->p.offset);
+	u64 idx;
 	int ret;
 
 	if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_pos), c,
@@ -126,11 +128,21 @@ static int bch2_check_lru_key(struct btree_trans *trans,
 
 	a = bch2_alloc_to_v4(k, &a_convert);
 
+	switch (type) {
+	case BCH_LRU_read:
+		idx = alloc_lru_idx_read(*a);
+		break;
+	case BCH_LRU_fragmentation:
+		idx = a->fragmentation_lru;
+		break;
+	}
+
 	if (fsck_err_on(lru_k.k->type != KEY_TYPE_set ||
-			a->data_type != BCH_DATA_cached ||
-			a->io_time[READ] != lru_pos_time(lru_k.k->p), c,
-			"incorrect lru entry (time %llu) %s\n"
+			lru_pos_time(lru_k.k->p) != idx, c,
+			"incorrect lru entry: lru %s time %llu\n"
+			"  %s\n"
 			"  for %s",
+			bch2_lru_types[type],
 			lru_pos_time(lru_k.k->p),
 			(bch2_bkey_val_to_text(&buf1, c, lru_k), buf1.buf),
 			(bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) {
diff --git a/libbcachefs/lru.h b/libbcachefs/lru.h
index b8d9848c..78a60769 100644
--- a/libbcachefs/lru.h
+++ b/libbcachefs/lru.h
@@ -22,6 +22,27 @@ static inline u64 lru_pos_time(struct bpos pos)
 	return pos.inode & ~(~0ULL << LRU_TIME_BITS);
 }
 
+#define BCH_LRU_TYPES()		\
+	x(read)			\
+	x(fragmentation)
+
+enum bch_lru_type {
+#define x(n) BCH_LRU_##n,
+	BCH_LRU_TYPES()
+#undef x
+};
+
+#define BCH_LRU_FRAGMENTATION_START	((1U << 16) - 1)
+
+static inline enum bch_lru_type lru_type(struct bkey_s_c l)
+{
+	u16 lru_id = l.k->p.inode >> 48;
+
+	if (lru_id == BCH_LRU_FRAGMENTATION_START)
+		return BCH_LRU_fragmentation;
+	return BCH_LRU_read;
+}
+
 int bch2_lru_invalid(const struct bch_fs *, struct bkey_s_c, unsigned, struct printbuf *);
 void bch2_lru_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 
diff --git a/libbcachefs/move.c b/libbcachefs/move.c
index e7eb55bd..7dac9264 100644
--- a/libbcachefs/move.c
+++ b/libbcachefs/move.c
@@ -653,13 +653,13 @@ failed_to_evacuate:
 	printbuf_exit(&buf);
 }
 
-int __bch2_evacuate_bucket(struct moving_context *ctxt,
+int __bch2_evacuate_bucket(struct btree_trans *trans,
+			   struct moving_context *ctxt,
 			   struct bpos bucket, int gen,
 			   struct data_update_opts _data_opts)
 {
 	struct bch_fs *c = ctxt->c;
 	struct bch_io_opts io_opts = bch2_opts_to_inode_opts(c->opts);
-	struct btree_trans trans;
 	struct btree_iter iter;
 	struct bkey_buf sk;
 	struct bch_backpointer bp;
@@ -668,17 +668,17 @@ int __bch2_evacuate_bucket(struct moving_context *ctxt,
 	struct bkey_s_c k;
 	struct data_update_opts data_opts;
 	unsigned dirty_sectors, bucket_size;
+	u64 fragmentation;
 	u64 bp_offset = 0, cur_inum = U64_MAX;
 	int ret = 0;
 
 	bch2_bkey_buf_init(&sk);
-	bch2_trans_init(&trans, c, 0, 0);
 
-	bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
+	bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 			     bucket, BTREE_ITER_CACHED);
-	ret = lockrestart_do(&trans,
+	ret = lockrestart_do(trans,
 			bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
-	bch2_trans_iter_exit(&trans, &iter);
+	bch2_trans_iter_exit(trans, &iter);
 
 	if (ret) {
 		bch_err(c, "%s: error looking up alloc key: %s", __func__, bch2_err_str(ret));
@@ -688,17 +688,18 @@ int __bch2_evacuate_bucket(struct moving_context *ctxt,
 	a = bch2_alloc_to_v4(k, &a_convert);
 	dirty_sectors = a->dirty_sectors;
 	bucket_size = bch_dev_bkey_exists(c, bucket.inode)->mi.bucket_size;
+	fragmentation = a->fragmentation_lru;
 
-	ret = bch2_btree_write_buffer_flush(&trans);
+	ret = bch2_btree_write_buffer_flush(trans);
 	if (ret) {
 		bch_err(c, "%s: error flushing btree write buffer: %s", __func__, bch2_err_str(ret));
 		goto err;
 	}
 
-	while (!(ret = move_ratelimit(&trans, ctxt))) {
-		bch2_trans_begin(&trans);
+	while (!(ret = move_ratelimit(trans, ctxt))) {
+		bch2_trans_begin(trans);
 
-		ret = bch2_get_next_backpointer(&trans, bucket, gen,
+		ret = bch2_get_next_backpointer(trans, bucket, gen,
 						&bp_offset, &bp,
 						BTREE_ITER_CACHED);
 		if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
@@ -713,7 +714,7 @@ int __bch2_evacuate_bucket(struct moving_context *ctxt,
 			struct bkey_s_c k;
 			unsigned i = 0;
 
-			k = bch2_backpointer_get_key(&trans, &iter,
+			k = bch2_backpointer_get_key(trans, &iter,
 						bucket, bp_offset, bp);
 			ret = bkey_err(k);
 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
@@ -726,9 +727,9 @@ int __bch2_evacuate_bucket(struct moving_context *ctxt,
 			bch2_bkey_buf_reassemble(&sk, c, k);
 			k = bkey_i_to_s_c(sk.k);
 
-			ret = move_get_io_opts(&trans, &io_opts, k, &cur_inum);
+			ret = move_get_io_opts(trans, &io_opts, k, &cur_inum);
 			if (ret) {
-				bch2_trans_iter_exit(&trans, &iter);
+				bch2_trans_iter_exit(trans, &iter);
 				continue;
 			}
 
@@ -742,15 +743,15 @@ int __bch2_evacuate_bucket(struct moving_context *ctxt,
 				i++;
 			}
 
-			ret = bch2_move_extent(&trans, &iter, ctxt, io_opts,
+			ret = bch2_move_extent(trans, &iter, ctxt, io_opts,
 					       bp.btree_id, k, data_opts);
-			bch2_trans_iter_exit(&trans, &iter);
+			bch2_trans_iter_exit(trans, &iter);
 
 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 				continue;
 			if (ret == -ENOMEM) {
 				/* memory allocation failure, wait for some IO to finish */
-				bch2_move_ctxt_wait_for_io(ctxt, &trans);
+				bch2_move_ctxt_wait_for_io(ctxt, trans);
 				continue;
 			}
 			if (ret)
@@ -762,7 +763,7 @@ int __bch2_evacuate_bucket(struct moving_context *ctxt,
 		} else {
 			struct btree *b;
 
-			b = bch2_backpointer_get_node(&trans, &iter,
+			b = bch2_backpointer_get_node(trans, &iter,
 						bucket, bp_offset, bp);
 			ret = PTR_ERR_OR_ZERO(b);
 			if (ret == -BCH_ERR_backpointer_to_overwritten_btree_node)
@@ -774,8 +775,8 @@ int __bch2_evacuate_bucket(struct moving_context *ctxt,
 			if (!b)
 				goto next;
 
-			ret = bch2_btree_node_rewrite(&trans, &iter, b, 0);
-			bch2_trans_iter_exit(&trans, &iter);
+			ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
+			bch2_trans_iter_exit(trans, &iter);
 
 			if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 				continue;
@@ -792,17 +793,16 @@ next:
 		bp_offset++;
 	}
 
-	trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, ret);
+	trace_evacuate_bucket(c, &bucket, dirty_sectors, bucket_size, fragmentation, ret);
 
 	if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) && gen >= 0) {
-		bch2_trans_unlock(&trans);
+		bch2_trans_unlock(trans);
 		move_ctxt_wait_event(ctxt, NULL, list_empty(&ctxt->reads));
 		closure_sync(&ctxt->cl);
 		if (!ctxt->write_error)
-			verify_bucket_evacuated(&trans, bucket, gen);
+			verify_bucket_evacuated(trans, bucket, gen);
 	}
 err:
-	bch2_trans_exit(&trans);
 	bch2_bkey_buf_exit(&sk, c);
 	return ret;
 }
@@ -815,12 +815,15 @@ int bch2_evacuate_bucket(struct bch_fs *c,
 			 struct write_point_specifier wp,
 			 bool wait_on_copygc)
 {
+	struct btree_trans trans;
 	struct moving_context ctxt;
 	int ret;
 
+	bch2_trans_init(&trans, c, 0, 0);
 	bch2_moving_ctxt_init(&ctxt, c, rate, stats, wp, wait_on_copygc);
-	ret = __bch2_evacuate_bucket(&ctxt, bucket, gen, data_opts);
+	ret = __bch2_evacuate_bucket(&trans, &ctxt, bucket, gen, data_opts);
 	bch2_moving_ctxt_exit(&ctxt);
+	bch2_trans_exit(&trans);
 
 	return ret;
 }
diff --git a/libbcachefs/move.h b/libbcachefs/move.h
index aef61380..c5a7c0ad 100644
--- a/libbcachefs/move.h
+++ b/libbcachefs/move.h
@@ -66,7 +66,8 @@ int bch2_move_data(struct bch_fs *,
 		   bool,
 		   move_pred_fn, void *);
 
-int __bch2_evacuate_bucket(struct moving_context *,
+int __bch2_evacuate_bucket(struct btree_trans *,
+			   struct moving_context *,
 			   struct bpos, int,
 			   struct data_update_opts);
 int bch2_evacuate_bucket(struct bch_fs *, struct bpos, int,
diff --git a/libbcachefs/movinggc.c b/libbcachefs/movinggc.c
index f0ab65ff..1a646433 100644
--- a/libbcachefs/movinggc.c
+++ b/libbcachefs/movinggc.c
@@ -10,6 +10,7 @@
 #include "alloc_foreground.h"
 #include "btree_iter.h"
 #include "btree_update.h"
+#include "btree_write_buffer.h"
 #include "buckets.h"
 #include "clock.h"
 #include "disk_groups.h"
@@ -19,6 +20,7 @@
 #include "eytzinger.h"
 #include "io.h"
 #include "keylist.h"
+#include "lru.h"
 #include "move.h"
 #include "movinggc.h"
 #include "super-io.h"
@@ -31,138 +33,105 @@
 #include <linux/sort.h>
 #include <linux/wait.h>
 
-static inline int fragmentation_cmp(copygc_heap *heap,
-				   struct copygc_heap_entry l,
-				   struct copygc_heap_entry r)
+static int bch2_bucket_is_movable(struct btree_trans *trans,
+				  struct bpos bucket, u64 time, u8 *gen)
 {
-	return cmp_int(l.fragmentation, r.fragmentation);
-}
-
-static int find_buckets_to_copygc(struct bch_fs *c)
-{
-	copygc_heap *h = &c->copygc_heap;
-	struct btree_trans trans;
 	struct btree_iter iter;
 	struct bkey_s_c k;
+	struct bch_alloc_v4 _a;
+	const struct bch_alloc_v4 *a;
 	int ret;
 
-	bch2_trans_init(&trans, c, 0, 0);
+	if (bch2_bucket_is_open(trans->c, bucket.inode, bucket.offset))
+		return 0;
 
-	/*
-	 * Find buckets with lowest sector counts, skipping completely
-	 * empty buckets, by building a maxheap sorted by sector count,
-	 * and repeatedly replacing the maximum element until all
-	 * buckets have been visited.
-	 */
-	h->used = 0;
-
-	for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
-			   BTREE_ITER_PREFETCH, k, ret) {
-		struct bch_dev *ca = bch_dev_bkey_exists(c, iter.pos.inode);
-		struct copygc_heap_entry e;
-		struct bch_alloc_v4 a_convert;
-		const struct bch_alloc_v4 *a;
-
-		a = bch2_alloc_to_v4(k, &a_convert);
-
-		if ((a->data_type != BCH_DATA_btree &&
-		     a->data_type != BCH_DATA_user) ||
-		    a->dirty_sectors >= ca->mi.bucket_size ||
-		    bch2_bucket_is_open(c, iter.pos.inode, iter.pos.offset))
-			continue;
+	bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, bucket, 0);
+	k = bch2_btree_iter_peek_slot(&iter);
+	ret = bkey_err(k);
+	bch2_trans_iter_exit(trans, &iter);
+
+	if (ret)
+		return ret;
 
-		e = (struct copygc_heap_entry) {
-			.dev		= iter.pos.inode,
-			.gen		= a->gen,
-			.replicas	= 1 + a->stripe_redundancy,
-			.fragmentation	= div_u64((u64) a->dirty_sectors * (1ULL << 31),
-						  ca->mi.bucket_size),
-			.sectors	= a->dirty_sectors,
-			.bucket		= iter.pos.offset,
-		};
-		heap_add_or_replace(h, e, -fragmentation_cmp, NULL);
+	a = bch2_alloc_to_v4(k, &_a);
+	*gen = a->gen;
+	ret = (a->data_type == BCH_DATA_btree ||
+	       a->data_type == BCH_DATA_user) &&
+		a->fragmentation_lru &&
+		a->fragmentation_lru <= time;
 
+	if (ret) {
+		struct printbuf buf = PRINTBUF;
+
+		bch2_bkey_val_to_text(&buf, trans->c, k);
+		pr_debug("%s", buf.buf);
+		printbuf_exit(&buf);
 	}
-	bch2_trans_iter_exit(&trans, &iter);
 
-	bch2_trans_exit(&trans);
 	return ret;
 }
 
+static int bch2_copygc_next_bucket(struct btree_trans *trans,
+				   struct bpos *bucket, u8 *gen, struct bpos *pos)
+{
+	struct btree_iter iter;
+	struct bkey_s_c k;
+	int ret;
+
+	ret = for_each_btree_key2_upto(trans, iter, BTREE_ID_lru,
+				  bpos_max(*pos, lru_pos(BCH_LRU_FRAGMENTATION_START, 0, 0)),
+				  lru_pos(BCH_LRU_FRAGMENTATION_START, U64_MAX, LRU_TIME_MAX),
+				  0, k, ({
+		*bucket = u64_to_bucket(k.k->p.offset);
+
+		bch2_bucket_is_movable(trans, *bucket, lru_pos_time(k.k->p), gen);
+	}));
+
+	*pos = iter.pos;
+	if (ret < 0)
+		return ret;
+	return ret ? 0 : -ENOENT;
+}
+
 static int bch2_copygc(struct bch_fs *c)
 {
-	copygc_heap *h = &c->copygc_heap;
-	struct copygc_heap_entry e;
 	struct bch_move_stats move_stats;
-	struct bch_dev *ca;
-	unsigned dev_idx;
-	size_t heap_size = 0;
+	struct btree_trans trans;
 	struct moving_context ctxt;
 	struct data_update_opts data_opts = {
 		.btree_insert_flags = BTREE_INSERT_USE_RESERVE|JOURNAL_WATERMARK_copygc,
 	};
+	struct bpos bucket;
+	struct bpos pos;
+	u8 gen = 0;
+	unsigned nr_evacuated;
 	int ret = 0;
 
 	bch2_move_stats_init(&move_stats, "copygc");
-
-	for_each_rw_member(ca, c, dev_idx)
-		heap_size += ca->mi.nbuckets >> 7;
-
-	if (h->size < heap_size) {
-		free_heap(&c->copygc_heap);
-		if (!init_heap(&c->copygc_heap, heap_size, GFP_KERNEL)) {
-			bch_err(c, "error allocating copygc heap");
-			return 0;
-		}
-	}
-
-	ret = find_buckets_to_copygc(c);
-	if (ret) {
-		bch2_fs_fatal_error(c, "error walking buckets to copygc!");
-		return ret;
-	}
-
-	if (!h->used) {
-		s64 wait = S64_MAX, dev_wait;
-		u64 dev_min_wait_fragmented = 0;
-		u64 dev_min_wait_allowed = 0;
-		int dev_min_wait = -1;
-
-		for_each_rw_member(ca, c, dev_idx) {
-			struct bch_dev_usage usage = bch2_dev_usage_read(ca);
-			s64 allowed = ((__dev_buckets_available(ca, usage, RESERVE_none) *
-					       ca->mi.bucket_size) >> 1);
-			s64 fragmented = usage.d[BCH_DATA_user].fragmented;
-
-			dev_wait = max(0LL, allowed - fragmented);
-
-			if (dev_min_wait < 0 || dev_wait < wait) {
-				dev_min_wait = dev_idx;
-				dev_min_wait_fragmented = fragmented;
-				dev_min_wait_allowed	= allowed;
-			}
-		}
-
-		bch_err_ratelimited(c, "copygc requested to run but found no buckets to move! dev %u fragmented %llu allowed %llu",
-				    dev_min_wait, dev_min_wait_fragmented, dev_min_wait_allowed);
-		return 0;
-	}
-
-	heap_resort(h, fragmentation_cmp, NULL);
-
 	bch2_moving_ctxt_init(&ctxt, c, NULL, &move_stats,
 			      writepoint_ptr(&c->copygc_write_point),
 			      false);
+	bch2_trans_init(&trans, c, 0, 0);
+
+	ret = bch2_btree_write_buffer_flush(&trans);
+	BUG_ON(ret);
 
-	/* not correct w.r.t. device removal */
-	while (h->used && !ret) {
-		BUG_ON(!heap_pop(h, e, -fragmentation_cmp, NULL));
-		ret = __bch2_evacuate_bucket(&ctxt, POS(e.dev, e.bucket), e.gen,
-					     data_opts);
+	for (nr_evacuated = 0, pos = POS_MIN;
+	     nr_evacuated < 32 && !ret;
+	     nr_evacuated++, pos = bpos_nosnap_successor(pos)) {
+		ret = bch2_copygc_next_bucket(&trans, &bucket, &gen, &pos) ?:
+			__bch2_evacuate_bucket(&trans, &ctxt, bucket, gen, data_opts);
+		if (bkey_eq(pos, POS_MAX))
+			break;
 	}
 
+	bch2_trans_exit(&trans);
 	bch2_moving_ctxt_exit(&ctxt);
 
+	/* no entries in LRU btree found, or got to end: */
+	if (ret == -ENOENT)
+		ret = 0;
+
 	if (ret < 0 && !bch2_err_matches(ret, EROFS))
 		bch_err(c, "error from bch2_move_data() in copygc: %s", bch2_err_str(ret));
 
diff --git a/libbcachefs/recovery.c b/libbcachefs/recovery.c
index 178f0642..1976d5fa 100644
--- a/libbcachefs/recovery.c
+++ b/libbcachefs/recovery.c
@@ -1105,6 +1105,9 @@ int bch2_fs_recovery(struct bch_fs *c)
 			c->opts.version_upgrade	= true;
 			c->opts.fsck		= true;
 			c->opts.fix_errors	= FSCK_OPT_YES;
+		} else if (c->sb.version < bcachefs_metadata_version_fragmentation_lru) {
+			bch_info(c, "version prior to backpointers, upgrade required");
+			c->opts.version_upgrade	= true;
 		}
 	}
 
diff --git a/libbcachefs/subvolume.c b/libbcachefs/subvolume.c
index 1805c854..ba281104 100644
--- a/libbcachefs/subvolume.c
+++ b/libbcachefs/subvolume.c
@@ -512,8 +512,7 @@ int bch2_snapshot_node_create(struct btree_trans *trans, u32 parent,
 		n->v.pad	= 0;
 		SET_BCH_SNAPSHOT_SUBVOL(&n->v, true);
 
-		ret   = bch2_trans_update(trans, &iter, &n->k_i, 0) ?:
-			bch2_mark_snapshot(trans, bkey_s_c_null, bkey_i_to_s_c(&n->k_i), 0);
+		ret   = bch2_trans_update(trans, &iter, &n->k_i, 0);
 		if (ret)
 			goto err;
 
diff --git a/libbcachefs/subvolume.h b/libbcachefs/subvolume.h
index b6740eab..7c488c3d 100644
--- a/libbcachefs/subvolume.h
+++ b/libbcachefs/subvolume.h
@@ -8,15 +8,15 @@
 void bch2_snapshot_to_text(struct printbuf *, struct bch_fs *, struct bkey_s_c);
 int bch2_snapshot_invalid(const struct bch_fs *, struct bkey_s_c,
 			  unsigned, struct printbuf *);
+int bch2_mark_snapshot(struct btree_trans *, struct bkey_s_c,
+		       struct bkey_s_c, unsigned);
 
 #define bch2_bkey_ops_snapshot ((struct bkey_ops) {		\
 	.key_invalid	= bch2_snapshot_invalid,		\
 	.val_to_text	= bch2_snapshot_to_text,		\
+	.atomic_trigger	= bch2_mark_snapshot,			\
 })
 
-int bch2_mark_snapshot(struct btree_trans *, struct bkey_s_c,
-		       struct bkey_s_c, unsigned);
-
 static inline struct snapshot_t *snapshot_t(struct bch_fs *c, u32 id)
 {
 	return genradix_ptr(&c->snapshots, U32_MAX - id);
@@ -68,6 +68,13 @@ static inline bool bch2_snapshot_is_ancestor(struct bch_fs *c, u32 id, u32 ances
 	return id == ancestor;
 }
 
+static inline bool bch2_snapshot_has_children(struct bch_fs *c, u32 id)
+{
+	struct snapshot_t *t = snapshot_t(c, id);
+
+	return (t->children[0]|t->children[1]) != 0;
+}
+
 static inline bool snapshot_list_has_id(snapshot_id_list *s, u32 id)
 {
 	u32 *i;
diff --git a/libbcachefs/super.c b/libbcachefs/super.c
index 8bed1185..3b7cf9e8 100644
--- a/libbcachefs/super.c
+++ b/libbcachefs/super.c
@@ -488,7 +488,6 @@ static void __bch2_fs_free(struct bch_fs *c)
 	kfree(rcu_dereference_protected(c->disk_groups, 1));
 	kfree(c->journal_seq_blacklist_table);
 	kfree(c->unused_inode_hints);
-	free_heap(&c->copygc_heap);
 
 	if (c->io_complete_wq)
 		destroy_workqueue(c->io_complete_wq);
diff --git a/libbcachefs/util.c b/libbcachefs/util.c
index a9790954..bf5ffb47 100644
--- a/libbcachefs/util.c
+++ b/libbcachefs/util.c
@@ -433,8 +433,8 @@ static const struct time_unit {
 	{ "us",		NSEC_PER_USEC	 },
 	{ "ms",		NSEC_PER_MSEC	 },
 	{ "s",		NSEC_PER_SEC	 },
-	{ "m",          NSEC_PER_SEC * 60},
-	{ "h",          NSEC_PER_SEC * 3600},
+	{ "m",          (u64) NSEC_PER_SEC * 60},
+	{ "h",          (u64) NSEC_PER_SEC * 3600},
 	{ "eon",        U64_MAX          },
 };
 
diff --git a/linux/six.c b/linux/six.c
index 41337a7f..5a6eadc0 100644
--- a/linux/six.c
+++ b/linux/six.c
@@ -833,19 +833,12 @@ struct six_lock_count six_lock_counts(struct six_lock *lock)
 {
 	struct six_lock_count ret;
 
-	ret.n[SIX_LOCK_read]	= 0;
+	ret.n[SIX_LOCK_read]	= !lock->readers
+		? lock->state.read_lock
+		: pcpu_read_count(lock);
 	ret.n[SIX_LOCK_intent]	= lock->state.intent_lock + lock->intent_lock_recurse;
 	ret.n[SIX_LOCK_write]	= lock->state.seq & 1;
 
-	if (!lock->readers)
-		ret.n[SIX_LOCK_read] += lock->state.read_lock;
-	else {
-		int cpu;
-
-		for_each_possible_cpu(cpu)
-			ret.n[SIX_LOCK_read] += *per_cpu_ptr(lock->readers, cpu);
-	}
-
 	return ret;
 }
 EXPORT_SYMBOL_GPL(six_lock_counts);