Re: [PATCH v5 00/23] Introduce SCMI Telemetry support

[David Hildenbrand (Arm)]

On 7/7/26 08:31, Subrahmanya Lingappa wrote:
> Hi Cristian,

Hi!

Thanks for having a look. Christian is out this week, so let me ask some
clarifying questions and raise some points.

> 
> Thanks for posting this. I think the direction of not forcing these
> Telemetry Data Events into hwmon/IIO/perf prematurely is the right one.

Good.

> The cover letter explains the core problem well: most DE IDs are
> platform-defined, and without external metadata the kernel cannot reliably
> know whether a DE is a thermal sensor, a performance counter, a power value,
> or something entirely platform-specific.

Right.

> 
> That said, I think the current layering is still too SCMI-shaped for the
> ABI it is trying to introduce.

Just to be clear: the data is exposed through /dev/scmi/ device nodes. It is
expected to be very, very SCMI specific as it stands :)

> 
> The thing being exposed here is not only "SCMI Telemetry"; it is a more
> general kernel concept of telemetry:
> 
>   - enumerate provider-specific data/event descriptors
>   - expose capabilities, units, data format and topology/grouping metadata
>   - configure collection rate and collection mode where supported
>   - collect samples with timestamp/freshness/validity information
>   - track generation/state changes
>   - optionally expose provider-private raw backing storage for debugging

Yes, there are some generic concepts there. However, when we talk about
generalizing, I suspect that other telemetry implementations might vary quite
already even on these generic points?

In particular when it comes about enumeration (as discussed previously, some
telemetry solutions might expose way more detailed information about events) and
configurations (and, in particular, things like configuration groups).

But also efficient ways to receive data in user space.

I mean, sure, if someone uses the exact same concept but instead of calling it
SCMI, calls it XXMI, that's one thing.

But a true generic telemetry concept would have to take way more options into
account.

So defining something actual generic and extensible might be rather challenging,
no? And for that, I would expect a pretty good reason to do so.

"Make tool X easier to implement" would not be a good reason in my option. I'm
sure you have something in mind, what would that be?

> 
> SCMI is one provider of that model. It happens to transport the data through
> SCMI v4.0 concepts such as DEs, groups, SHMTIs and TDCF.

Right.

> But another firmware
> or control-plane specification could show up later with the same broad
> telemetry model and a different wire format. Call it XXMI or YYMI: the

"could" is a bit vague. Do you have any concrete examples in mind? It will be
rather challenging to figure out what's generic without other real examples.

> protocol mechanics would differ, but the kernel consumers should not have to
> learn a new in-kernel and userspace ABI for each one.

So far, the tools that consume this data are expected to be very platform specific.

IOW, not your ordinary perf/top/whatever :)

Assuming there is some XXMI telemetry in the future, I would expect tools that
could operate across multiple platforms to build upon some generic user space
library that abstracts these things and knows how to talk to the different devices.

> 
> The common ABI should therefore avoid treating the DE numeric namespace or
> the wire/storage format as the abstraction. The useful common contract is
> closer to:
> 
>   - what is being measured: component/type/instance/name;
>   - how to interpret it: unit, exponent/rate unit, data width and signedness;
>   - how it behaves: instantaneous, accumulating, average, state/counter, etc.;
>   - how it is controlled: individual event vs group-wide collection;
>   - how fresh it is: timestamp, sequence/generation and validity/error flags.
> 
> So I would strongly consider inserting a generic telemetry layer above SCMI,
> with SCMI registering as a telemetry provider rather than owning the stable
> userspace ABI directly.

Again, I think that's rather hard. And it's not immediately clear to me why this
abstraction must be had in the kernel, instead of in user space?

So it would be good if you could elaborate on that.

> 
> Roughly:
> 
>   drivers/telemetry/
>       provider registration
>       descriptor/sample/config abstractions
>       generation tracking and poll support
>       common userspace ABI

That will be a struggle to define in a future-proof way I'm afraid.

> 
>   drivers/firmware/arm_scmi/telemetry.c
>       SCMI protocol implementation
>       TDCF/SHMTI/notification parsing
>       translation into generic telemetry descriptors/samples
> 
>   optional provider-private raw/debug path
>       SCMI SHMTI mmap or dumps, if kept, behind explicit privilege/debug
>       policy rather than as the baseline telemetry ABI

That makes the SCMI SHMTI mmap interface a second-class citizen, which IIUC is
rather bad performance wise.

> 
> With that shape, future in-kernel consumers could bind to generic telemetry
> objects and capabilities instead of SCMI internals. Architected/well-known
> DEs could later grow adapters into hwmon/IIO/perf where the semantics are
> clear, while platform-specific DEs remain available through the generic
> telemetry interface plus userspace metadata.
> 
> This would also help avoid making the SCMI protocol's internal resource
> storage part of the ABI. Right now the public protocol header exposes
> structures such as scmi_telemetry_res_info, scmi_telemetry_group and
> scmi_telemetry_de, and the chardev indexes those objects directly. That feels
> too tightly coupled for something that may need more consumers later. I would
> prefer copy-out or iterator-style provider ops instead, for example:
> 
>   info_get()
>   de_count_get()
>   de_info_get()
>   group_info_get()
>   intervals_get()
>   collection_set()
>   collection_get()
>   sample_read()
>   reset()
> 
> where the backing arrays, xarrays, lazy enumeration, SHMTI mappings and TDCF
> parsing remain private to the SCMI provider.
> 
> The sample side probably needs an explicit common result contract too. A
> sample is not only an integer value: it may have optional timestamping,
> provider data width, stale/not-yet-valid state, partial collection failures,
> hardware-fault indications, and a sequence/generation value that lets
> userspace detect races against reconfiguration or shared-memory updates. Those
> details can still be backed by SCMI-specific status codes and TDCF parsing
> internally, but I think the stable ABI should expose them in provider-neutral
> terms.
> 
> I would also be careful about freezing the raw SHMTI mmap part as a baseline
> ABI. It is useful for bring-up and high-performance tooling, but it exposes

Correct me if I'm wrong: most user space tools we expect will fall into the
category of "high-performance tooling"?

Or which tools do you have in mind that have different requirements?

> firmware-owned shared memory layout and TDCF parsing rules directly to
> userspace. If this stays, I think it should be clearly separated as a
> privileged/debug/raw provider facility, with a precise mmap contract:
> 
>   - required mmap length
>   - vm_pgoff semantics
>   - page alignment
>   - VMA flags
>   - cache/coherency expectations
>   - lifetime across reset/remove/reprobe
>   - required capability, if any
> 
> Even outside the raw mmap case, telemetry can be a sensitive high-rate view
> of system behavior. The ABI should make access policy explicit: who can
> enumerate, who can read, who can change collection state, and whether
> production systems can restrict sampling granularity or intervals
> independently of what firmware advertises.
> 
> For the stable telemetry ABI, I would start smaller: descriptor enumeration,
> sample reads with validity metadata, optional group-aware configuration, and a
> generation counter/poll mechanism so userspace can detect reconfiguration or
> partial re-enumeration races.
> 
> One other point related to this layering: configuration appears to be global
> to the SCMI instance. Any process opening /dev/scmi/tlm_N can change
> collection mode, sampling interval, DE enablement and timestamp state for
> other users. That may be OK, but it should be an explicit policy decision.
> A generic telemetry layer could make the model clearer: multi-reader with a
> privileged/single writer, or fully shared global controls, with instance-level
> locking and generation updates after each visible state change.
> 
> So my high-level suggestion for v6 would be:
> 
>   1. define the generic telemetry provider/consumer model first;
>   2. make SCMI Telemetry one provider of that model;
>   3. keep SCMI/TDCF/SHMTI details below the provider boundary;
>   4. keep raw SHMTI access separate from the stable ABI;
>   5. define common sample validity/freshness/error semantics;
>   6. add generation/poll semantics to the common ABI;
>   7. make lifetime, access policy and configuration concurrency explicit.
> 
> I do not think this invalidates the protocol work in the series. Most of the
> SCMI parsing and resource discovery can still be the provider implementation.
> The main question is where the stable kernel/userspace contract should sit.
> My preference would be that it sits at "telemetry", not at "SCMI Telemetry",
> so that SCMI, and any future XXMI/YYMI-style provider, can share one consumer
> interface.

Again, I'd like to understand why you think this abstraction should be had in
the kernel, when it could also be had in user space. I'm sure there is a good
reason you have in mind.

In particular, I am curious about tools you would expect to run across platforms
that would even care about this.

Last but not least, which other telemetry implementations do you have in mind
that are on the horizon?

In particular, I'm afraid a generic abstraction in the kernel for a handful of
specialized user space tools might take a rather long time to flesh out.

Finally, I wonder if something generic is really required when talking about an
ioctl interface. I would completely agree if we would be talking about something
like resctl, that actually exposes the data to user space in an abstracted,
easy-to-use fashion -- not just an ioctl interface to talk to a device

Thanks for taking the time to write all this up. I'll have to leave all the SCMI
details to Christian!

-- 
Cheers,

David