[PATCH/RFC 0/6] PSCI: Fix non-PMIC wake-up if SYSTEM_SUSPEND cuts power

[Sudeep Holla]

On 21/02/17 17:34, Geert Uytterhoeven wrote:
> Hi Sudeep,
> 
> On Tue, Feb 21, 2017 at 5:45 PM, Sudeep Holla <sudeep.holla at arm.com> wrote:
>> On 21/02/17 16:21, Geert Uytterhoeven wrote:
>>> On Tue, Feb 21, 2017 at 11:38 AM, Sudeep Holla <sudeep.holla at arm.com> wrote:
>>>> On 20/02/17 20:33, Geert Uytterhoeven wrote:
>>>>> This patch series adds support for using non-PMIC wake-up sources on the
>>>>> Renesas R-Car Gen3 (H3 or M3-W) Salvator-X development boards.
>>>>>
>>>>> Nothing in the PSCI specification requires the SoC to remain powered and
>>>>> to support wake-up sources when suspended using SYSTEM_SUSPEND.
>>>>> If the firmware implements the PSCI SYSTEM_SUSPEND operation by cutting
>>>>> power to the SoC, the only possibly wake-up sources are thus the ones
>>>>> connected to the PMIC.
>>>>
>>>> OK, but I don't see any issue with that. That's exactly how it works on
>>>
>>> How do you use other wake-up sources, like wake on LAN, UART or GPIO?
>>
>> From wakeup source configuration/management perspective, s2ram and
>> s2idle are exactly same.
> 
> From the point of view of Linux, that's indeed the case. Linux knows about e.g.
> interrupt controllers to keep awake if they're needed for one of the configured
> wake-up sources.
> 

Correct.

> PSCI does not know about the wake-up sources configured under Linux.
>

And the PSCI for a particular platform has to respect those
configurations if they powered down the interrupt controller. They need
to read those wakeups and configure the wakeup controller if required.

Also note we can't mark any device as wakeup source in the device tree
in there's dedicated wakeup controller to enable them and main interrupt
controller is powered down.

>>>> ARM Juno platform. The SoC is powered down.
>>>
>>> Good to hear this is not limited to Renesas platforms, so there's a common
>>> problem to solve.
>>
>> No, there's no problem to solve. Firmware should enter deepest sleep
>> state in the system with SYSTEM_SUSPEND from which it can wakeup of course.
> 
> While SYSTEM_SUSPEND can wake up (e.g. from PMIC), it may not support all
> wake-up sources configured from Linux. There's no API to communicate that
> information (from Linux to PSCI), or to communicate that limitation (from
> PSCI to Linux).
> 

It's platform specific. The SYSTEM_SUSPEND implementation will be aware
of the platform and it's wakeup configuration. As mention above, it
needs to take some action if required.

[...]

>>>>> Linux uses a different suspend method when other wake-up sources (e.g.
>>>>> wake on LAN, UART or GPIO) are enabled.  Hence the user no longer has to
>>>>> manually restrict "mem" suspend to "s2idle" or "shallow" states using:
>>>>
>>>> Have you explored suspend-to-idle instead ? It looks like thats exactly
>>>> what you are doing in this patch set. You also get low latency for free
>>>> as it just enters the deepest idle state on all CPUs instead of
>>>> hotplugging out all the secondaries.
>>>
>>> Yes, cfr. "s2idle" above.
>>> The user can specify to use "s2idle" manually:
>>>
>>>     $ echo s2idle > /sys/power/mem_sleep # or "shallow"
>>
>> This looks like custom file for me.
> 
> /sys/power/mem_sleep was added in v4.10-rc1, to choose which state to use
> for s2ram, cfr. Documentation/power/states.txt.
>

OK, that's new to me. I need to read and understand that then. Sorry for
the noise. I was unaware of that.

>> IIUC, the standard sysfs file for the system PM state is:
>>         /sys/power/state
>> 1. s2ram:
>>         $ echo mem > /sys/power/state
> 
> As of v4.10-rc1, this will use either "s2idle", "shallow", or "deep" mode,
> depending on availability and configuration through mem_sleep.
> ("deep" maps to PSCI SYSTEM_SUSPEND).
> 
>> 2. s2idle
>>         $ echo freeze > /sys/power/state
> 
> Correct.
> 
>>> However, how to handle this automatically, e.g. by a distro?
>>
>> As above
> 
> I meant the "mem" one, which should not pick "deep" mode if it cannot wake-up
> from that state using the configured wake-up sources.
> 
>>> On most other platforms, userspace can just do e.g.
>>>
>>>     ethtool -s eth0 wol g
>>
>> That should work.
>>
>>> to enable wake-on-LAN, and suspend to the deepest supported state using:
>>>
>>>     echo mem > /sys/power/state
>>
>> This will work only if PSCI SYSTEM_SUPEND is implemented. If the SoC
>> can't wakeup if it's powered down, then it should not use that state
>> to implement SYSTEM_SUSPEND in PSCI firmware or just return the
>> SYSTEM_SUSPEND feature is not implemented in which case "freeze" is the
>> next available state to enter.
> 
> The SoC can wake-up. It's just not guaranteed that it can wake-up using
> the wakeup-source configured from Linux. Which wakeup-sources are available
> depends on the actual PSCI implementation.  It's not specified by the PSCI
> specification.
> 
>> Just botching whatever shallow state you can enter on a particular SoC
>> into standard "mem" state sounds *horrible* to me.
> 
> That's more or less what /sys/power/mem_sleep does, though.
> 

OK, I will go through that in detail.

>>> On systems where PSCI SYSTEM_SUSPEND powers down the SoC, userspace must
>>> make sure to configure to use "s2idle" (or "shallow) instead, else the
>>> configured wake-up sources won't work.
>>
>> That's perfect. I was worried that user-space is not doing that. So to
>> summarize, PSCI firmware either:
>> 1. enters a sane and resumable state in SYSTEM_SUSPEND api
> 
> In this case, it may resume using the PMIC only.
> And there's no way for userspace (or even the kernel) to find out!
> Hence my solution to:
>   - add a DT property to indicate that PSCI will power down the SoC,
>   - use "shallow" suspend if any Linux wakeup-sources have been configured
>     and the property above is present.
> 

But what you are doing is *exactly* same as "freeze" state except the
secondaries powered off instead of idling. So my NACK is mainly for that
reason. Had you done something different, then yes, it should be considered.

>> or
>> 2. just don't implement SYSTEM_SYSTEM. Use the cpuidle+s2idle framework
>>    in Linux to enter the deepest idle state.
> 
> In that case, it indeeds falls back to cpuidle/s2idle, which works fine.
> 
>> You literally need no extra work to enter this "freeze" state if the
>> CPU_SUSPEND in PSCI can enter the deepest idle state you want to enter
>> in this "s2idle" you are referring so far.
>> Just start with:
>>
>>         $ cat /sys/power/state
>>
>> and you should see "freeze" there, if not that's the first thing to
>> check provided the platform has cpuidle working.
> 
> "freeze" is always available.
> "deep" is available if PSCI supports SYSTEM_SUSPEND.
> My third patch adds "shallow", but it can be dropped (patch 4 can fall
> through to cpu_do_idle() when needed, regardless of the existence of shallow).
> 

I understand that. But tell me how is that any different from the
"freeze" state. Why do you need that new state at all. Again just
because there's a new feature added doesn't mean we need to use it for
sake of using it. If "freeze" != "shallow", then we can explore this
further. But it's not in your case and hence I see no real need for this.

-- 
Regards,
Sudeep