Re: [PATCH 2/2] firmware/psci: Set pm_set_resume/suspend_via_firmware() on qcom

[Konrad Dybcio]

On 28.12.2023 12:50, Sudeep Holla wrote:
> On Thu, Dec 28, 2023 at 12:47:51PM +0100, Konrad Dybcio wrote:
>> On 28.12.2023 11:28, Sudeep Holla wrote:
>>> On Wed, Dec 27, 2023 at 11:15:31PM +0100, Konrad Dybcio wrote:
>>>> Most Qualcomm platforms implementing PSCI (ab)use CPU_SUSPEND for
>>>> entering various stages of suspend, across the SoC. These range from a
>>>> simple WFI to a full-fledged power collapse of the entire chip
>>>> (mostly, anyway).
>>>>
>>>> Some device drivers are curious to know whether "the firmware" (which is
>>>> often assumed to be ACPI) takes care of suspending or resuming the
>>>> platform. Set the flag that reports this behavior on the aforementioned
>>>> chips.
>>>>
>>>> Some newer Qualcomm chips ship with firmware that actually advertises
>>>> PSCI SYSTEM_SUSPEND, so the compatible list should only grow slightly.
>>>>
>>>
>>> NACK, just use suspend-to-idle if SYSTEM_SUSPEND is not advertised. It is
>>> designed for such platforms especially on x86/ACPI which don't advertise
>>> Sx states. I see no reason why that doesn't work on ARM platforms as well.
>>
>> Not sure if I got the message through well, but the bottom line is, on
>> Qualcomm platforms the "idle" states aren't actually just "idle" (read:
>> they're not like S0ix). All but the most shallow ones shut down quite a
>> chunk of the entire SoC, with the lowest ones being essentially S3 with
>> power being cut off from the entire chip, except for the memory rail.
>>
> 
> No I understood that and S2I is exactly what you need.
> Have you checked if S2I already works as intended on these platforms ?
Yes, simple CPU idling works OOTB and the SoC power collapse only works
given the developer doesn't cut corners when bringing up the platform
(read: works on some of the ones we support, trying hard to expand that
group!)

> What extra do you achieve with this hack by advertising fake S2R ?
Uh.. unless I misunderstood something, I'm not advertising s2ram..
Quite the opposite, I'm making sure only s2idle is allowed.

Admittedly, my mistake, I managed to misname the function which I
didn't spot before wrapping this patch up for sending..

> S2I will have less latency compared to S2R and the mem_sleep will be
> automatically set to S2I if S2R is not supported, so no userspace impact
> as well.
Yes, everything that differs them is abstracted in TZ or through the
power management coprocessor.

> 
> So please let us know what this change provide extra over S2I ? Until
> then my NACK still stands.
The main point here is to advertise pm_resume/suspend_via_firmware,
which represents that Qualcomm does a lot of not-very-obvious power
wire plumbing within their PSCI impl.

Konrad