Re: [PATCH RFC 01/10] firmware: arm_scmi: Add SCMI QCOM Generic Extension Protocol documentation

From
Sudeep Holla <sudeep.holla@kernel.org>
Date
2026-09-08 19:38:10
Also in
arm-scmi, linux-arm-kernel, linux-arm-msm, linux-devicetree, linux-pm, lkml
Reply to
Sibi Sankar, 2026-08-19
On Thu, Aug 20, 2026 at 01:26:57AM +0530, Sibi Sankar wrote:
On 8/13/2026 12:48 PM, Sudeep Holla wrote:
[...]
quoted
The current rationale isn't entirely clear to me. Windows historically favors
ACPI over native SCMI, meaning that supporting a proprietary vendor protocol
would require non-native workarounds potentially hidden within interfaces like
PEP. I am skeptical of this architectural direction. Since Android leverages
In an ideal world maybe but transition for all SoC capabilities to ACPI is
rarely that smooth and they do use the PEP interface.
Well I agree, then we shouldn't be bringing up the topic of OS agnostic
firmware as that is not a reality on the platforms we are talking about
then ?
quoted
the Linux kernel and already manages this through vendor modules, maintaining
those as modules seems like the optimal path until we agree on the interface
that can be merged.

Given that this patch series has been under discussion for nearly two years,
please provide a summary of the feedback that has already been incorporated
to address these concerns just for sake of argument and in your defence.

To clarify, the vendor protocol space is strictly intended for
platform-specific functionalities that cannot be standardized; it should not
be used to bypass standardization for convenience. It appears no evaluation
was done to determine if or why the standard performance protocol was
insufficient. Had that assessment occurred, I would expect the proposed
interface to align much more closely with the standard definitions as I
previously mentioned.
The design might seem contrived at the moment but it was still largely
shaped by the SCMI specification. There is a literally a 6-7 year window between
SCMI landing in mainline and the first mention of the vendor protocol identifiers
being re-useable between SoC vendors [1]. This largely shaped how Qualcomm
[1] shows the date as 2024, so 6-7 years is bit of exaggeration IMO.
used vendor protocol. The first instance of vendor protocol was a straight
forward vendor protcol implementing just MEMLAT [2] but the sudden rise in
the number of protocol eating up vendor protocol space made them club
together the class of devfreq algorithms into a singular generic extension
protocol.
I fully understand that this interface is already deployed in firmware and
that changing that firmware ABI may no longer be practical. However, I
think we need to separate two things: supporting an existing firmware ABI
and accepting that ABI as the abstraction exposed through the upstream SCMI
framework.

The former is a compatibility constraint for Qualcomm platforms. The latter
is an upstream interface and maintainability decision, and the existence of
deployed firmware cannot by itself determine that decision.
"deliberated attempt to circumvent the standard SCMI protocol template"
This was specifically targeted for having performance related tunables
outside of PERF protocol. If that is not possible, then I take back that
statement.
"expectation of immediate acceptance without modification"

I gather that ^^ are the major objections to the current series landing but
sadly none of these were raised during this 2 year period and this protocol
even had a "Reviewed-by" from the only other reviewer listed [3]. We
promised to fix all the concerns raised [4] in the next major/minor version upgrade
[5].
It is common for intermediate versions to receive Reviewed-by tags, but those
do not necessarily imply final acceptance of the entire series. If there is an
instance where I provided a Reviewed-by and later withdrew it without
providing a technical justification, please point it out.

Also, with code alone, please do not expect me to review these series
effectively. It took more than a year and half for the documentation to reach
a state where the design and implementation could be reasonably understood.
Now that we have sufficient improved documentation and a clearer understanding
of the approach, a proper and detailed review has just begun.

Therefore, I do not think it is fair to claim that these issues should have
been identified during reviews over the past two years. The lack of clear and
adequate documentation made meaningful review difficult, and I have repeatedly
requested improvements to the documentation for precisely this reason.
The current version works as is on Hamoa/Purwa/Glymur/Mahua/Kaanapali
maintaining the same ABI (that should count for something) and it would only
need a major/minor version update when a new algorithm string gets added
or a new param-ids gets added. Both of these are yet to happen.
My concern with the generic extension protocol remains the multiplexing
model. SCMI protocol IDs are themselves the mechanism by which
independently defined functionality is identified, discovered and
versioned. Introducing another namespace underneath a single vendor
protocol, where an algorithm string selects effectively independent
interfaces with their own sets of param_ids, duplicates that mechanism and
makes protocol 0x80 a container for potentially unrelated functionality.
Given that we put out the pseudo code of what the MEMLAT algorithm does
and spent the past several revisions trying to explain why generic perf nor
mpam would work for us we keep getting hit with the blanket "It appears
no evaluation was done to determine if or why the standard performance
protocol was insufficient". The only recommendation I see from your side
is move MEMLAT to it's own protocol. With the documentation/code
available to you can please let us know how current perf protocol or MPAM
can be used as a standin replacement?
This is also why I suggested treating MEMLAT as a distinct vendor protocol.
I am not suggesting that MEMLAT can necessarily be replaced by the standard
Performance protocol or MPAM. Those are separate questions. If MEMLAT
semantics do not map to either of those, then that is a perfectly valid
reason for having a vendor protocol. But "this functionality requires a
vendor protocol" does not imply that the generic SET/GET/START/STOP
multiplexing protocol is the right upstream representation of it. I am
completely against such generalisation which will end up being ambiguous
sooner or later.
[1] - https://lore.kernel.org/lkml/Zag5L9j8-oCebKFm@pluto/ (local)
[2] - https://lore.kernel.org/lkml/1667451512-9655-1-git-send-email-quic_sibis@quicinc.com/#t (local)
[3] - https://lore.kernel.org/lkml/Zo14-rQ1Jaxh5Idi@pluto/ (local)
[4] - https://lore.kernel.org/lkml/Z1GfGk0yQAVQKEVL@pluto/ (local)
[5] - https://lore.kernel.org/lkml/55fd0c34-c52c-95c5-5cd0-16fd66a4baa2@quicinc.com/ (local)
quoted
The core issue here is that the proposed vendor protocol interface has been
presented as a finished product without open discussion or a willingness to
iterate based on upstream feedback. This approach bypasses standard
development processes.
It would be way easier for us to re-design this but it would mean that we
are abandoning the current users stuck with this firmware version and that's the
only reason for trying to land this in a form that is maintainable. Either
way  please do take the call to decisively NAK the series (even if it's coming 9
revisions late and at the cost of 2 years). That way we can give up on this
and go about upstreaming using the standard development process.
I also don't think the fact that this has gone through several revisions
means that architectural concerns can no longer be raised. I appreciate the
amount of work that has gone into the series and the changes made in
response to previous reviews, but neither elapsed time nor deployed
non-standard vendor firmware can make an upstream interface immutable
before it is merged. This series is not even properly versioned, carries
RFC still without version if it is no the first RFC. Anyways, there are
series that go through 15-20 versions before it get merged. See SCMI
telemetry itself at v11 already. So, I don't understand your complaint and
all I can say is welcome to the kernel development process if its still a
for you.

So I don't see this as a choice between accepting the current generic
protocol or abandoning users of existing firmware. Sorry!

If it is easier as you mentioned, yes please re-design MEMLAT and any other
algorithm as a separate vendor protocol. If it helps take this as NACK for
the current interface. Hopefully this helps you in making proper progress
on upstreaming using the standard development process.

This is not the only or first firmware ABI Qualcomm has failed to upstream
because it was all baked and couldn't be changed. So nothing new/surprise
at that end.
That said please do consider expanding your reviewer/maintainer count
in a way that gives you the capability to review series from all SoC
vendors.
I don't quite understand this. Please elaborate.

-- 
Regards,
Sudeep
Thread (20 messages) 20 messages, 5 authors, last reply 8d ago
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help