[PATCH 2/2] firmware: arm_scmi: power_control: Set SCMI_SYSPOWER_IDLE in pm resume

[Sudeep Holla]

On Tue, Jul 01, 2025 at 11:07:35PM +0800, Peng Fan wrote:
> Hi Sudeep,
> 
> On Tue, Jun 24, 2025 at 10:58:07PM +0800, Peng Fan wrote:
> >On Tue, Jun 24, 2025 at 11:21:52AM +0100, Sudeep Holla wrote:
> >>On Tue, Jun 24, 2025 at 01:23:10AM +0000, Peng Fan wrote:
> >>> > 
> >>> > Just to summarise my understanding here at very high level, the issue
> >>> > exists as the second notification by an agent to the Linux to suspend
> >>> > the system wakes up the system from suspend state. Since the
> >>> > interrupts are enabled before the thaw_processes() (which eventually
> >>> > continues the execution of scmi_suspend_work_func() to set the state
> >>> > to SCMI_SYSPOWER_IDLE, the scmi_userspace_notifier() is executed
> >>> > much before and ends up ignoring the request as the state is still not
> >>> > set to SCMI_SYSPOWER_IDLE. There is a race which your patch is
> >>> > addressing.
> >>> 
> >>> Thanks for writing this down, It is very correct and clear.
> >>> 
> >>
> >>While I am not against adding bus PM ops as it can be useful elsewhere,
> >>just wonder if this usecase is a good use of it. Does setting the state
> >>before the pm_suspend() call suffice. I still need to think through the
> >>possible race with that solution, but just asking you to check if that
> >
> >There is race condition if setting the state to SCMI_SYSPOWER_IDLE before
> >pm_suspend.
> >
> >The 2nd suspend notification could runs into pm_suspend again
> >before pm_suspend update system_state to SYSTEM_SUSPEND, if my understanding
> >is correct.
> >
> >Per pm_suspend->enter_state,
> >"Make sure that no one else is trying to put the system into a sleep state",
> >not sure, but I think better not let pm_suspend to handle the race condition.
> >
> >Since syspower only has one per system(linux), the other approach is to
> >use syscore, but need a global variable for state in scmi_power_control.c,
> >because syscore_suspend/resume does not have parameter.
> >
> >we need to set state back to IDLE after linux wakeup and before kernel
> >thread scheduling. I only see two interfaces to achieve:
> >PM ops or syscore ops.
> 
> Not sure you have time to give a look. I plan to post V2 later this
> week. In V2, I would still use pm ops.
> 

Sure, please go ahead and post. I didn't get time to look at it in depth.
I agree with your comment for now, will take a look at it again with your v2.

-- 
Regards,
Sudeep