[GIT PULL] ARM64: juno: SCPI and dependent drivers for v4.3

[Sudeep Holla]

Hi Olof,

On Fri, Aug 21, 2015 at 5:17 PM, Olof Johansson <olof at lixom.net> wrote:
> Hi,
>
> On Thu, Aug 13, 2015 at 03:53:04PM +0100, Sudeep Holla wrote:
>>
>>
>> On 13/08/15 15:24, Catalin Marinas wrote:
>> >On Thu, Aug 13, 2015 at 02:40:26PM +0200, Olof Johansson wrote:
>> >>On Wed, Aug 05, 2015 at 11:22:51AM +0100, Sudeep Holla wrote:
>> >>>Here is the first pull request for SCPI, clock and cpufreq support on
>> >>>ARM64 Juno development platform for v4.3
>> >>
>> >>I'd like to hear from Catalin and Will on why ARM is doing this. Catalin
>> >>and Will are pushing back very hard at vendors who don't use PSCI for
>> >>CPU power management, and then they have coworkers that are looking to
>> >>merge this directly into the kernel.
>> >
>> >Actually, the arm64 CPU power management story is pretty much handled by
>> >Lorenzo and Sudeep in my team, so I trust their judgement ;). But I
>> >agree with you that a better explanation on the reasons behind this is
>> >needed, otherwise we risk confusing messages. I'll try to summarise it
>> >below but Sudeep can reply with more details.
>> >
>>
>> (I will use the acronomys directly as Catalin has already explained the
>> difference)
>>
>> Thanks Catalin for the detailed explanation. I agree that this pull
>> request has not much details, but the patch series makes reference only
>> to the clocks and features provided by the SCP(System Control Processor)
>> to be used by Linux(non-secure) and nothing about CPU power states that
>> are handled by PSCI. This SCPI is completely independent and orthogonal
>> to PSCI.
>
> Yes, I am aware that PSCI and SCPI are two confusingly similar acronyms that
> represent separate things. My concern though is that we're letting in a poorly
> architected solution into the kernel with SCPI. PSCI is already choosing to
> abstract this outside of the kernel. Is there a reason SCPI isn't using the
> same approach?
>

I am not sure we can compare the CPU power states with the CPU DVFS.
In case of power states, we need to ensure all the exception levels have
saved the context and hence PSCI abstraction makes sense for the kernel.
However the CPU DVFS is mostly driven by the kernel based on the load,
so we talk directly to the power controller from the kernel.

In way SCPI is abstracting by hiding all the regulators and PLLs you need
to handle. So I may be missing to understand when you say SCPI also use
PSCI approach.

>> >PSCI only covers CPU boot/cpuidle and platform suspend. It does not have
>> >any provisions for DVFS and I'm not sure this could be easily
>> >standardised (or whether we could find enough supporters to get it
>> >adopted).
>> >
>>
>> Though we are not campaigning it as aggressively as PSCI.
>
> Which, given the resistance to pick up PSCI, this is unlikely to be used by
> anyone...
>

Agreed even I am not sure how widely will that be accepted, but we need to
support the systems which have SCPI right ? Like Juno for now.

>> >SCPI is a (standard) interface to a coprocessor handling clocks etc. to
>> >be used with cpufreq. This patchset adds support for cpufreq/dvs which
>> >is not handled by PSCI. So, just to clarify the acronyms:
>> >
>> >SCPI - System Control and Power Interface - it's a protocol for
>> >communicating with a System Control Processor (e.g. Cortex-M3) via
>> >mailboxes. It is used in Linux for DVFS but this is *not* a secure/EL3
>> >firmware interface (a.k.a. SMC calls). It indeed a firmware interface
>> >but the firmware runs on a separate coprocessor and not on the CPU itself.
>> >
>> >PSCI - Power State Coordination Interface - it is a secure firmware
>> >interface for coordinating the CPU power states (idle, not DVFS). The
>> >firmware may use an SCPI communication channel (e.g. deeper sleep states
>> >where clock changes are needed).
>> >
>>
>> I will add some more information to (hopefully) clarify the doubts.
>>
>> SCP(System Control Processor) offers control and management of the
>> core/cluster power states, various power domain DVFS including the
>> core/cluster, certain system clocks configuration, thermal sensors and
>> many others.
>>
>> Though the core/cluster power states are also handled this SCP,
>> they are accessible only through secure channels and PSCI firmware
>> implementations on the platform rely on that internally. Non-secure
>> access/requests are not possible.
>>
>> Just to summarize, this SCP helps to offload all the power management
>> work to a dedicated low power processor which can be always running.
>> On Juno, its Cortex-M while on AMD Seattle it's Cortex-A5.
>
> I understand what the SCP is used for. What I don't understand is why you're
> adding all this complexity in the kernel for it.
>

I am not sure how we can avoid that though. Any thoughts ?

> Also, is there only one known implementation of this today on Juno? Are there
> others coming soon? Single-user frameworks tend to be hard to get right, and
> it's usually much better to wait until it's clear it will fit more than just
> the original target.
>

No argument on that, the specification is not reviewed as much as PSCI but we do
have version, so can be fixed/enhanced in future. It's no different
w.r.t PSCI in that
respect IMO.

Yes currently only on Juno, but future ARM Ltd. platforms have plans to reuse
the same.

> Anyway, the git server you provided the link to in the original email
> (linux-arm.org) seems to be offline so I couldn't fetch a fresh copy of
> the branch to look at.
>

Extremely sorry for that(not much under my control :(), I will get it
fixed ASAP.
I am current away attending LPC.

> It's pretty clear this is not 4.3 material at this point, so I'll take
> a look at it later.
>

Fair enough, I understand that its too late for v4.3

Regards,
Sudeep