On Thu, Jan 23, 2014 at 9:47 AM, Waiman Long [off-list ref] wrote:

Thank for the info. I am less familiar with that kind of issues on other
architecture. I will add a smp_mb__after_atomic_dec() & send out a new
patch.

SInce it's the unlock path,. you need to use the "mb__*before*"
versions, since presumably you want to protect what is inside the lock
from leaking out, not some random thing after the lock from leaking
in..

              Linus