For the single dd inside memcg, ext4 is now working pretty well, with
least CPU overheads:

Attached are the dd progresses for btrfs, ext4 and xfs in turn, with
btrfs performing the most fluent and fast.

Also attached the "10 memcg each running 1 dd" graph for btrfs. Well
it's unacceptably bumpy because there are no any efforts to make a
round-robin walk between the works from different memcg LRU lists.

Thanks,
Fengguang