[PATCH] ARM: OMAP4: cpuidle: fix: call cpu_cluster_pm_exit conditionally

[Santosh Shilimkar]

On Thursday 29 August 2013 01:15 PM, Kevin Hilman wrote:
> +Santosh
> 
> "Strashko, Grygorii" <grygorii.strashko at ti.com> writes:
> 
>> Hi Vladimir, Kevin
>>
>> On 08/27/2013 06:54 PM, Kevin Hilman wrote:
>>> Vladimir Murzin <murzin.v at gmail.com> writes:
>>>
>>>> We call cpu_cluster_pm_enter for dev->cpu == 0 only, but
>>>> cpu_cluster_pm_exit called without that check.
>>>>
>>>> Because of that unhandled page fault may happen:
>>>>
>>
>> [...]
>>
>>>>
>>>> It is supposed that sar_base is initialized in irq_save_context, which
>>>> is called on CPU_CLUSTER_PM_ENTER notification. If this notification
>>>> has been missed and CPU_CLUSTER_PM_EXIT is received sar_base is NULL.
>>>>
>>>> Fix it by calling CPU_CLUSTER_PM_{ENTER,EXIT} under the same condition.
>>
>> Could you check, if revert of the following patch will solve the issue, pls?
>> commit e7457253494fff660a72bc0cedeee97491ccd173
>> "ARM: OMAP4+: CPUidle: Deprecate use of omap4_mpuss_read_prev_context_state()"
>>
>>>>
>>>> Signed-off-by: Vladimir Murzin <murzin.v at gmail.com>
>>>
>>> Good catch.
>>
>> Yes, but It seems, that CPUIdle logic is unclear for OAMP4 .
>> The above issue may happen if CPU1 enter/exit LP while CPU0:
>> - not enter at all (somewhere inside "coupled" core);
>> - still entering LP (somewhere before call to omap4_enter_lowpower());
>>
>> The question is - Should first CPUx, who exited from LP(C3) state,
>> restore Cluster context, or it should be done by CPU0 only?
>> (on OMAP4 CPUs may return from C3 async).
> 
> Well, they're not *supposed* to return async on OMAP4.  IIUC, only CPU0
> wakes up and then it's CPU0s job to wake up CPU1. However, the crash
> reported here certainly suggests that CPU1 exiting before CPU0, so
> one of the possibilities you suggest above is probably happening (I
> suspect the latter.)
> 
> It looks like we might still need to check the actual hardware state
> there to avoid those potential cases.
> 
The subject patch is good enough to avoid the double notifier call chain
even though its not harmful its UN-necessary.

And then the sar_base check should be in place as well to avoid the
reported issue.

Regards,
Santosh