This adds some overheads for indirect calls to the I/O thread, but
reduces the amount of global locks.  Next step will be switching to
per-request locks.