Re: [PATCH v15 10/35] firmware: arm_scmi: Add Telemetry resources enumeration
From: Cristian Marussi
Date: Wed Nov 18 2026 - 13:43:26 EST
On Wed, Nov 18, 2026 at 09:26:23AM -0800, Jonathan Cameron wrote:
> On Wed, 18 Nov 2026 09:09:33 +0000
> Cristian Marussi <cristian.marussi@xxxxxxx> wrote:
>
> > On Mon, Nov 16, 2026 at 01:36:38PM -0800, Jonathan Cameron wrote:
> > > On Thu, 12 Nov 2026 19:56:10 +0000
> > > Cristian Marussi <cristian.marussi@xxxxxxx> wrote:
> > >
> >
> > Hi,
> >
> > > > On Thu, Nov 12, 2026 at 10:01:45AM -0800, Jonathan Cameron wrote:
> > > > > On Mon, 9 Nov 2026 09:11:56 +0000
> > > > > Cristian Marussi <cristian.marussi@xxxxxxx> wrote:
> > > > >
> > > > > > Add basic Telemetry allocation and initialization routines, introducing and
> > > > > > wiring-up the logic needed to enumerate all the SCMI Telemetry resources
> > > > > > optionally using a lazy on-demand scheme: this approach factors in the
> > > > > > flexibility needed to minimize the amount of initialization work based on
> > > > > > the effective access pattern used by the upper layer ABI.
> > > > > >
> > > > > > Signed-off-by: Cristian Marussi <cristian.marussi@xxxxxxx>
> > > > >
> > > > > Hi Cristian,
> > > > >
> > > > > This stretches across the patch set but first elements related to it are
> > > > > in this patch. How do we handle larger values of de_data_sz?
> > > > >
> > > >
> > > > We dont. This was somethig that was on my todo list initially, but then
> > > > I realized that there was NO FW-side plan to actually implement anything
> > > > larger than 64bit.
> > >
> > > Understood, some set of firmware that you have isn't doing this. I don't
> > > necessarily mind not supporting it initially but like the other features
> > > it is very useful to have an implementation to make sure it doesn't break
> > > the ABI. I don't want to go through a 'move to version 2 ABI if you want
> > > this feature' path.
> > >
> >
> > Agreed that it would be better if avoided.
> >
> > > Let's just say I hit this for a reason. Whether that reason makes it into
> > > a production firmware is an open question at the moment.
> > >
> > > FWIW I have both QEMU based emulation (I'm working through the process to be
> > > able to share that) and the changes to fit in via what I think is the
> > > cleanest ABI change. Kernel / userspace change below.
> > > I modified this from my original proposal to take into account the
> > > desire to add it later (even if to me it smells like papering over a bug
> > > rather than adding a feature!). See below for what that required.
> > > Not only shmti is tested. The LLM found bugs in it's own fast channel
> > > code, so take that with a pinch of salt!
> > >
> >
> > I skimmed through your patch and it is mostly in line with what I was
> > imagining (not even protyped though), i.e. in a nutshell
> >
> > - an ABI feature bit to add MUTIWORD support
> > - a flag bit to activate this feature on all _READ IOCTLs at will
> > - a repurpose of the pad field in scmi_tlm_sample
> >
> > What I find awkward is the usage of pad as an index: that will expose
> > a lot of duplicate data since all DEs multi-words are now scattered
> > through a number of sample with the same ID and TS, and a bit cumbersome
> > to handle in userspace to reconstruct the multi-word value.
> >
> > My non-tested-nor-prototyped idea was instead to repurpose the pad field
> > as a length, so as to pack all the multi-words payload as the trailing
> > part of a flexible array member (an extrnal array really being an ABI)
> > and use a single structure to read a complete DE.
> >
> > In this case the problem would be to maintain backward compatibility
> > with the current scmi_tlm_de_sample.
>
> Indeed, that's what drove the just using an index approach.
>
> >
> > At first I was thinking that we could simply introduce a new alternative
> > scmi_tlm_de_multi_sample structure that is used in place of the current
> > scmi_tlm_de_sample ONLY when feataure bits is available AND requested: so
> > userspace knows that when sets flag.MULTIWORD it has to use the
> > multi-word structure in input where pad is a length and there are multipe vals.
> >
> > While this would work for all the bulk ops since they take a __u64 pointer
> > to a payload structure really (so it does NOT really matter what they are
> > carrying as long as it is consistent with the advertised feature bits and
> > request flag), it CANNOT work for DE_READ where scmi_tlm_de_sample is
> > used directly...so a change in the current, not-merged, ABI would be
> > needed anyway....
> >
> > Then...I thought that maybe we dont even need this...and this part could be
> > judged horrific maybe, so feel free to stop reading if your eyes starts
> > bleeding :P
> >
> > Currently we have
> >
> > struct scmi_tlm_de_sample {
> > __u32 id;
> > __u32 pad;
> > __u64 tstamp;
> > __u64 val;
> > };
> >
> > with pad being unused and the whole structure that can be used alone
> > or packed into a container struct.
> >
> > WHAT IF... when MULTIWORD feature is enabled AND requested we repurpose
> > the meaning NOT ONLY of <pad> but also of <val> and use this latter NOT
> > as a payload for a single value returned but as a pointer to an external
> > array of words counted by pad ?
> >
> > The structure size remains the same, while its interpretation vary based
> > on the requested flag when such feature is available.
> >
> > It may be a bit cumbersome still on userspace for bulk requests since
> > scmi_tlm_batch will contain a pointer to an array of possibly varied
> > sized items, so item_sz represents the common basic scmi_tlm_de_sample
> > size which in turn carries the sizing inside the repurposed <pad>
> > field inside for each single sample item
>
> Are they variable sized, or indirections? Today I think
> scmi_tlm_data_read() simply points to an array of scmi_tlm_de_sample.
> In this scheme it continues to do that but some of those then have
> pointers on to additional storage. That's a more complex beast
> to manage.
>
> >
> > ...so you can anyway check the provided user-input sizes it is just
> > scattered all over and as a consequence also the user provided data
> > would have to be a bit fragmented since
> >
> > - batch points to an array of scmi_tlm_de_sample sized by items_sz
> > - each sample in turn points to its array of values sized by <pad>
> I read this as implying they all do this including the small ones.
>
> > - you can really mix both kind of DEs in the same request: some
> > can carry the single word value embedded in val, multiwords can
> > point to teh <pad> sized external array.
>
> But nope. It is a grab bag of different types that happen to be
> the same length.
>
> >
> > ...mmm sort of start liking this hoorr...or I am missing something ?
> >
> > ... O_o ... it would need a tad of documentation of course ...
>
> Quite a bit! And probably a helper library.
>
> >
> > If you have not fainted in despair, I (and some LLM) could prototype
> > something like this in one of the next series ... certainly not in V16.
> >
> > ... thoughts ? insults ? :P
>
> It's an option. More compact than the index one, though overheads of
> allocations might undo that unless they tend to be many line DEs
> and does avoid duplication of the metadata and timestamp (if there) which
> are guaranteed to match in every line. Give we are trending towards
> basically having this breaking ABI change (be it with an opt in)
> it maybe comes down to invasiveness of changes in userspace that would
> be needed to accommodate it.
>
> Probably a case for a prototype and see what it looks like.
>
I will protype this later on in the coming weeks/series and see how much
painful it is to support this schema with my test tool as a starter...
> Anyhow, I guess now we have two options on the table we can conclude
> it is 'probably' going to be a safe ABI extension to make in a follow
> up series. Not ideal, but I'd rather this didn't substantially slow
> down the progress to getting to full feature set.
>
Yes it does not seem anyway something that breaks backward compatibility
in one way or another by using ABIT features bits and flags.
...but I need to prototype something for my own peace of mind :D
Thanks,
Cristian