[PATCH v4 10/13] firmware: arm_sdei: Add support for CPU and system power states

[will.deacon@arm.com (Will Deacon)]

On Tue, Oct 24, 2017 at 06:34:19PM +0100, James Morse wrote:
> Hi Will,
> 
> On 18/10/17 18:17, Will Deacon wrote:
> > On Tue, Oct 17, 2017 at 06:44:29PM +0100, James Morse wrote:
> >> When a CPU enters an idle lower-power state or is powering off, we
> >> need to mask SDE events so that no events can be delivered while we
> >> are messing with the MMU as the registered entry points won't be valid.
> >>
> >> If the system reboots, we want to unregister all events and mask the CPUs.
> >> For kexec this allows us to hand a clean slate to the next kernel
> >> instead of relying on it to call sdei_{private,system}_data_reset().
> >>
> >> For hibernate we unregister all events and re-register them on restore,
> >> in case we restored with the SDE code loaded at a different address.
> >> (e.g. KASLR).
> >>
> >> Add all the notifiers necessary to do this. We only support shared events
> >> so all events are left registered and enabled over CPU hotplug.
> 
> >>  static void sdei_smccc_smc(unsigned long function_id,
> >>  			   unsigned long arg0, unsigned long arg1,
> >>  			   unsigned long arg2, unsigned long arg3,
> >> @@ -544,9 +742,36 @@ static int sdei_probe(struct platform_device *pdev)
> >>  		return 0;
> >>  	}
> >>  
> >> +	err = cpuhp_setup_state_nocalls(CPUHP_AP_ARM_SDEI_STARTING, "SDEI",
> >> +				&sdei_cpuhp_up, &sdei_cpuhp_down);
> >> +	if (err) {
> >> +		pr_warn("Failed to register CPU hotplug notifier...\n");
> >> +		return err;
> >> +	}
> > 
> > What prevents CPU hotplug events coming in here?
> 
> Nothing, but what would trigger them? This is after the arch code has brought up
> secondaries, but before user space is running.

But as we recently discovered (ae2e972dae3c), userspace can be running
really early due to an initrd and I don't think that necessarily rules out
hotplug operations.

> No events are registered by this point, so all that could go wrong is a
> freshly-arrived CPU is unmasked, then this probe method fails and assumes it
> left cores masked. If you think its possible I can post a patch to bolt it down
> some more.

If the race is benign, could you avoid using the _nocalls registration
function and drop the subsequent IPI?

Will