On Wed, Mar 06 2019 at 15:12 -0700, Stephen Boyd wrote:
Quoting Lina Iyer (2019-03-04 09:14:50)
On Fri, Mar 01 2019 at 10:58 -0700, Stephen Boyd wrote:
>Quoting Lina Iyer (2019-02-27 14:29:13)
>> Hi Stephen,
>> On Tue, Feb 26 2019 at 17:49 -0700, Stephen Boyd wrote:
>Ok, can you explain why it's even a problem for the TCSes to be active
>during suspend? I would hope that for suspend/resume, if this is
>actually a problem, the RPMh driver itself can block suspend with a
>driver suspend callback that checks for idleness.
The RSC can transmit TCS executed from Linux and when all the CPUs have
powered down, could execute a firmware in the RSC to deliver the sleep
state requests. The firmware cannot run when there are active requests
being processed. To ensure that case, we bail out of sleep or suspend,
when the last CPU is powering down, if there are active requests.
Ok, do we actually bail out or just pick a shallower idle state that
wouldn't trigger the firmware to run something that may conflict with
the active requests (i.e. some light CPU sleep mode)? The commit text
seems to imply we block certain idle states.
We bail out of idle and let cpuidle determine the state again. We don't
go into a shallower state.
>But I suspect that in
>the system wide suspend/resume case, any callers that could make TCS
>requests are child devices of the RPMh controller and therefore they
>would already be suspended if they didn't have anything pending they're
>waiting for a response on or they would be blocking suspend themselves
>if they're waiting for the response. So why are we even checking the
>TCSes in system suspend path at all? Assume that callers know what
>they're doing and will block suspend if they care?
In suspend, they probably would do what you mention above. All CPUs
might conincidentally be idle at the same idle, when a request is being
>Following that same logic, is this more of an API that is planned for
>use by CPU idle? Where the case is much more of a runtime PM design.
>Even then, I don't get it. A device that's runtime active and making
>RPMh requests might need to block some forms of CPU idle states because
>a request hasn't been processed yet that may change the decision for
>certain deep idle states?
A process waiting on a RPMH request, may let the CPU go to sleep and
therefore this is a possibility.
Ok thanks for the info. Can these details be included in the commit text
so we don't lose sight of the bigger picture? And can this patch series
be combined with a larger cpuidle/suspend patch series so we don't have
to review this in isolation? I don't understand the need to add more
APIs that aren't used yet.