An alternative might be to re-engineer the solution to provide a low-resume-latency request interface which the scheduler can use instead of a timer hack to keep it alive - the existing QoS interface isn't suitable because it's global and wakes all the CPUs up to get the new latency requirement in effect. That would not strictly be a big.LITTLE feature, but as its the only client it would likely need to be carried around in the same patch set.

There is already a suitable state disable API in cpuidle, but determining which states to disable still requires reading the exit latency which is only available in this driver information.

In summary, I don't like this fix although I accept that it works and it'll continue to work for TC2 and any other platform where the idle driver cannot be changed at runtime.

What do you think?