[PATCH] ARM: tegra: cpuidle: use CPUIDLE_FLAG_TIMER_STOP flag

[daniel.lezcano@linaro.org (Daniel Lezcano)]

On 07/22/2013 05:15 AM, Joseph Lo wrote:
> On Fri, 2013-07-19 at 18:52 +0800, Daniel Lezcano wrote:
>> On 07/19/2013 09:14 AM, Joseph Lo wrote:
>>> On Thu, 2013-07-18 at 20:41 +0800, Daniel Lezcano wrote:
>>>> On 07/18/2013 01:08 PM, Joseph Lo wrote:
>>>>> On Thu, 2013-07-18 at 04:31 +0800, Stephen Warren wrote:
>>>>>> On 07/17/2013 04:15 AM, Joseph Lo wrote:
>>>>>>> On Wed, 2013-07-17 at 03:51 +0800, Stephen Warren wrote:
>>>>>>>> On 07/16/2013 05:17 AM, Joseph Lo wrote:
>>>>>>>>> On Tue, 2013-07-16 at 02:04 +0800, Stephen Warren wrote:
>>>>>>>>>> On 06/25/2013 03:23 AM, Joseph Lo wrote:
>>>>>>>>>>> Use the CPUIDLE_FLAG_TIMER_STOP and let the cpuidle framework
>>>>>>>>>>> to handle the CLOCK_EVT_NOTIFY_BROADCAST_ENTER/EXIT when entering
>>>>>>>>>>> this state.
>>>>>> ... [ discussion of issues with Joesph's patches applied]
>>>>>>>
>>>>>>> OK. I did more stress tests last night and today. I found it cause by
>>>>>>> the patch "ARM: tegra: cpuidle: use CPUIDLE_FLAG_TIMER_STOP flag" and
>>>>>>> only impact the Tegra20 platform. The hot plug regression seems due to
>>>>>>> this patch. After dropping this patch on top of v3.11-rc1, the Tegra20
>>>>>>> can back to normal.
>>>>>>>
>>>>>>> And the hop plug and suspend stress test can pass on Tegra30/114 too.
>>>>>>>
>>>>>>> Can the other two patch series for Tegra114 to support CPU idle power
>>>>>>> down mode and system suspend still moving forward, not be blocked by
>>>>>>> this patch?
>>>>>>>
>>>>>>> Looks the CPUIDLE_FLAG_TIMER_STOP flag still cause some other issue for
>>>>>>> hot plug on Tegra20, I will continue to check this. You can just drop
>>>>>>> this patch.
>>>>>>
>>>>>> OK, if I drop that patch, then everything on Tegra20 and Tegra30 seems
>>>>>> fine again.
>>>>>>
>>>>>> However, I've found some new and exciting issue on Tegra114!
>>>>>>
>>>>>> With unmodified v3.11-rc1, I can do the following without issue:
>>>>>>
>>>>>> * Unplug/replug CPUs, so that I had all combinations of CPU 1, 2, 3
>>>>>> plugged/unpplugged (with CPU 0 always plugged).
>>>>>>
>>>>>> * Unplug/replug CPUs, so that I had all combinations of CPU 0, 1, 2, 3
>>>>>> plugged/unpplugged (with the obvious exception of never having all CPUs
>>>>>> unplugged).
>>>>>>
>>>>>> However, if I try this with your Tegra114 cpuidle and suspend patches
>>>>>> applied, I see the following issues:
>>>>>>
>>>>>> 1) If I boot, unplug CPU 0, then replug CPU 0, the system immediately
>>>>>> hard-hangs.
>>>>>>
>>>>> Sorry, I didn't apply the hotplug stress test on CPU0 too much. Because
>>>>> all of our use cases and stress tests are focused on secondary CPUs
>>>>> only.
>>>>>
>>>>> After doing some tests, here is my summary. This issue happens after we
>>>>> support CPU idle power-down mode and relates to PMC or flow controller I
>>>>> believe.
>>>>>
>>>>> i) on top of v3.11-rc1 (only support WFI in CPU idle)
>>>>> When hot unplug CPU0, the PMC can power gate and put it into reset
>>>>> state. This is what I expect and also true on all the other secondary
>>>>> CPUs. The flow controller can maintain the CPU power state machine as
>>>>> what we want.
>>>>>
>>>>> ii) on top of v3.11-rc1 + CPU idle power down mode support
>>>>> a) I saw most of the time the CPU0,1,2,3 were in power down and reset
>>>>> status. That means the idle power down mode works fine.
>>>>>
>>>>> b) Testing with the CPU hotplug stress test with the secondary CPUs (not
>>>>> include CPU0), the result is good too.
>>>>>
>>>>> c) Testing hot plug on CPU0 with CPUIDLE_FLAG_TIMER_STOP apply or not
>>>>> apply (Note 1), the result shows not good to me. The CPU0 have already
>>>>> gone into WFI and the flow controller is set as WAITFOREVENT mode. But
>>>>> the PMC always can't power gate CPU0 and sometimes can put it into
>>>>> reset, but sometimes can't. That's why you can see it hanging after
>>>>> unplug CPU0 sometimes.
>>>>
>>>> Are sure coupled idle state support hotplug and especially the cpu0
>>>> hotplug ?
>>>
>>> Tegra114 didn't use coupled idle framework.
>>
>> Ok, so the problem occurs with the CPUIDLE_FLAG_TIMER_STOP flag only on
>> tegra114, right ?
>>
>> Sorry, I am a bit lost :)
>>
> Here are the issues that happen after apply CPUIDLE_FLAG_TIMER_STOP.
> 1) Tegra114/30
> The warning message at kernel/time/tick-broadcast.c:667
> tick_broadcast_oneshot_control could be triggered when doing CPU hot
> plug stress test.

With the fix for tick-broadcast.c [1] ?

> 2) Tegra20
> The system is easy to stick or become lag.
> The CPU hot plug is easy to cause system stick too.
> 
> The fix I suggested in another mail looks can fix all the issues above.
> I verified it again today on 3 different Tegra SoC platforms.

Not sure your patch fixes the problem.

I am wondering if there isn't a underlaying problem which surface with
the flag.

Thanks !
  -- Daniel

[1]
https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=ea8deb8dfa6b0e8d1b3d1051585706739b46656c

-- 
 <http://www.linaro.org/> Linaro.org ? Open source software for ARM SoCs

Follow Linaro:  <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog