The blockdev mapping's private_lock is fairly contended.  The buffer
LRU cache fixed a lot of that, but under page replacement load,
try_to_free_buffers is still showing up.

Moving the freeing of buffer_heads outside the lock reduces contention
in there by 30%.