Re: [PATCH RFC 0/6] PSCI-via-EFI to support firmware and kernel sharing EL2 for Apple Silicon

From: Ard Biesheuvel

Date: Thu Nov 19 2026 - 10:38:36 EST




On Tue, 17 Nov 2026, at 16:25, Sven Peter wrote:
> On 17.11.26 13:10, Will Deacon wrote:
>> On Sun, Nov 15, 2026 at 07:38:35PM +0100, Sven Peter wrote:
>>>
>>>
>>> On 03.09.26 15:25, Ard Biesheuvel wrote:
>>>> On Thu, 3 Sep 2026, at 14:15, Will Deacon wrote:
>>>>> On Thu, Sep 03, 2026 at 12:42:39PM +0100, Mark Rutland wrote:
>>>>>> On Thu, Sep 03, 2026 at 01:34:56PM +0200, Ard Biesheuvel wrote:
>>>>>>> On Wed, 8 Jul 2026, at 09:15, Sven Peter wrote:
>>>>>>>> This series adds a custom EFI table that points to a PSCI entry point
>>>>>>>> (plus some other stuff that has to be available before EFI runtime
>>>>>>>> services are set up) and adds support for this new conduit to the psci
>>>>>>>> code. We can't directly use the normal EFI runtime path because that one
>>>>>>>> takes a sleeping lock and we need to be able to call into PSCI from
>>>>>>>> atomic context during e.g. cpu bringup or during idle.
>>>>>>>> It also adds support for specifying the specific MAIR attributes for EFI
>>>>>>>> runtime mappings as defined in the latest UEFI spec since Apple Silicon
>>>>>>>> is rather allergic to using Device-nGnRnE vs. Device-nGnRE for its MMIO.
>>>>>>
>>>>>>>> dt-bindings: arm: psci: Add EFI conduit
>>>>>>>> arm64/efi: Add and parse custom PSCI EFI configuration table
>>>>>>>> efi: Add EFI_MEMORY_ISA_{MASK,VALID}
>>>>>>>> arm64/efi: Honor EFI_MEMORY_ISA_MASK for Device-nGnRnE vs -nGnRE
>>>>>>>> firmware/psci: Add EFI runtime conduit
>>>>>>>> arm64: dts: apple: t8103: Add PSCI and CPU idle states
>>>>>>>>
>>>>>>>
>>>>>>> I've picked up patches #3 and #4, which are useful in their own right.
>>>>>>>
>>>>>>> I'm not sure if the arm64 maintainers will want to consider this, but
>>>>>>> I think it's a reasonable compromise, as it puts the abstraction in
>>>>>>> the right place.
>>>>>>
>>>>>> Sorry for the late reply; I've been away almost all of August and I'm
>>>>>> slowly catching up on things.
>>>>>>
>>>>>> As with last time this was proposed (in abstract), I am not happy about
>>>>>> bodging an EFI conduit into PSCI, given that the manner in which state
>>>>>> is managed is completely different from SMCCC.
>>>>>>
>>>>>> If we need a mechanism for doing hotplug and/or idle without HVC/SMC,
>>>>>> that's one thing we can consider. I don't think we should pretend that
>>>>>> it is PSCI.
>>>>>
>>>>> What do you have in mind for an alternative mechanism? I understand your
>>>>> objection to this proposal, but at least it keeps the interface fairly
>>>>> high-level and avoids opening the flood gates for a bunch of SoC-specific
>>>>> idle routines. Are you thinking of extensions to the PSCI spec or
>>>>> something else?
>>>>>
>>>>
>>>> I can't speak for Mark, of course, but I could imagine EFI being used as a
>>>> conduit to provide a callable set of interfaces (with a rigorously defined
>>>> set of constraints such as the ones in this patch, i.e., reentrancy, no FP
>>>> or SIMD, not relying on ISA features that the OS needs to know about and
>>>> enable, etc).
>>>>
>>>> There is prior art here in ACPI PRM, which does something similar, and this
>>>> also relies on EFI runtime services.
>>>>
>>>> You'd still need to implement the CPU ops and a cpuidle driver afaict (no
>>>> expert here), but at least those would be coded against an interface that
>>>> the kernel itself specified, and can be made as generic as we choose to.
>>>>
>>>> Not saying this is all great, but it might be a worthwhile compromise to
>>>> consider.
>>>
>>> Finally getting back to this series...
>>>
>>>
>>> I'd definitely like to keep mapping the resident firmware using EFI runtime
>>> services because otherwise I'd have to duplicate a whole lot of tricky code
>>> to do the same thing via some other means (e.g. a device tree nodes that
>>> contains all mappings + callbacks) and I just don't see much value in that.
>>>
>>> I'm happy to just drop the PSCI part though and replace that with a simple
>>> and somewhat generic "efi-rt" (or similar) cpuidle driver and make that one
>>> have similar requirements like I added for this RFC series and expose bunch
>>> of generic callbacks (that will just happen to resemble PSCI functions very
>>> much ;)) via a custom EFI table. That would keep PSCI out of all of this but
>>> come at the cost of an additional cpuidle driver.
>>>
>>> I don't think that would open any flood gates for SoC-specific drivers since
>>> we could just point them to the generic one and tell them to use that one if
>>> they can't do PSCI for some reason.
>>>
>>>
>>> Now before I go off and prototype this: Would this approach be fine in
>>> general with you, Mark and Will?
>>
>> tbh, if the "efi-rt" thing is just PSCI with a different conduit and that
>> conduit is properly documented, then I really don't see the point in keeping
>> it separate. However. I suspect that we probably all have slightly different
>> opinions here which isn't going to help you very much :(
>
> Yeah, well, I agree but I really don't have a strong opinion or any good
> argument if either "efi-rt" vs "psci-via-efi" is better or worse. I only
> think both are better than SoC-specific cpuidle and cpu_ops drivers and
> I have no idea how else we could do this. (We could drop the kernel to
> EL1 and do "normal" PSCI in EL2 then but that would be much worse than
> either of those options imho)
>

I agree with Mark that implementing a pseudo-PSCI that can never quite
comply with everything in the PSCI spec is probably not a great idea
unless we actually need it.

> And especially with M4+ I'd like to have something that can go upstream
> because normal wfi loses most general purpose regs there and has to be
> disabled which is not great.
>
> I can also just prototype the efi-rt cpuidle+cpu_ops thing and then we
> can take a look at it as well and see if that's better than what I have
> in this series if that helps but I'd rather not spend that time if that
> has no chance of being accepted either.
>

I think a consistent approach for both secondary bringup (+teardown?) and
cpuidle that is based on the same efi-runtime layer is perfectly acceptable,
and allows us to define it as narrowly as we can, rather than having to
consider what shared code might assume once it sees enable-method=psci on
a CPU, or a PSCI conduit (SMCCC detection, mitigations, etc etc)