Hi,

it was a known limitation, in the SCMI Clock protocol support, the lack of
dynamic allocation around per-clock rates discovery: fixed size statically
per-clock rates arrays did not scale and was increasingly a waste of memory
(see [1]).

This series aim at solving this in successive steps:

 - simplify and reduce to the minimum possible the rates data info exposed
   to the SCMI driver by scmi_clock_info
 - move away from static fixed allocation of per-clock rates arrays in
   favour of a completely dynamic runtime allocation: just allocate what
   is needed based on the effectively discovered

This is done in patches 1-6.

A further bigger optimization suggested in a past series [1] by Etienne
would be, whenever allowed by the spec, to limit upfront the number of
queries in order to simply retrieve min and max rate, that are indeed the
only rates needed by the CLK SCMI driver.

The approach proposed in [1] was open coding and duplicating some of the
functionalities already provided by SCMI iterators, though.

Patch 7-10 implement such optimization instead by:

 - reworking core SCMI iterators to support bound enumerations
 - use such new bound iterators to perform the minimum number of queries
   in order to ony retrieve min an max rate

As a final result now the rates enumeration triggered by the CLK SCMI
driver, while still allocating for all the existent rates, miminize the
number of SCMI CLK_DESCRIBE_RATE messages needed to obtain min and max.

Finally, patch 11 introduces a new clock protocol operation to be able to
trigger anytime on demand a full enumeration and obtain the full list of
rates when needed, not only min/max: this latter method is really only used
currently by some dowstream SCMI Test driver of mine.

Based on v7.0-rc1. 

Tested on JUNO and an emulated environment.

Any feeback welcome.

Thanks,
Cristian

[1]: https://lore.kernel.org/arm-scmi/aZsX-oplR6fiLBBN@pluto/T/#t
[2]: https://lore.kernel.org/20241203173908.3148794-2-etienne.carriere@foss.st.com
---

Cristian Marussi (11):
  firmware: arm_scmi: Add clock determine_rate operation
  clk: scmi: Use new determine_rate clock operation
  firmware: arm_scmi: Simplify clock rates exposed interface
  clk: scmi: Use new simplified per-clock rate properties
  firmware: arm_scmi: Drop unused clock rate interfaces
  firmware: arm_scmi: Make clock rates allocation dynamic
  firmware: arm_scmi: Harden clock parents discovery
  firmware: arm_scmi: Refactor iterators internal allocation
  firmware: arm_scmi: Add bound iterators support
  firmware: arm_scmi: Use bound iterators to minimize discovered rates
  firmware: arm_scmi: Introduce all_rates_get clock operation

 drivers/clk/clk-scmi.c                |  48 +----
 drivers/firmware/arm_scmi/clock.c     | 296 ++++++++++++++++++++------
 drivers/firmware/arm_scmi/driver.c    |  73 +++++--
 drivers/firmware/arm_scmi/protocols.h |  13 +-
 include/linux/scmi_protocol.h         |  29 +--
 5 files changed, 314 insertions(+), 145 deletions(-)

-- 
2.53.0

Hi Cristian,

Thanks for your series!

On Fri, 27 Feb 2026 at 16:33, Cristian Marussi <cristian.marussi@arm.com> wrote:
> it was a known limitation, in the SCMI Clock protocol support, the lack of
> dynamic allocation around per-clock rates discovery: fixed size statically
> per-clock rates arrays did not scale and was increasingly a waste of memory
> (see [1]).
>
> This series aim at solving this in successive steps:
>
>  - simplify and reduce to the minimum possible the rates data info exposed
>    to the SCMI driver by scmi_clock_info
>  - move away from static fixed allocation of per-clock rates arrays in
>    favour of a completely dynamic runtime allocation: just allocate what
>    is needed based on the effectively discovered
>
> This is done in patches 1-6.
>
> A further bigger optimization suggested in a past series [1] by Etienne

s/[1]/[2]/

> would be, whenever allowed by the spec, to limit upfront the number of
> queries in order to simply retrieve min and max rate, that are indeed the
> only rates needed by the CLK SCMI driver.
>
> The approach proposed in [1] was open coding and duplicating some of the

What does [1] refer to?

> functionalities already provided by SCMI iterators, though.
>
> Patch 7-10 implement such optimization instead by:
>
>  - reworking core SCMI iterators to support bound enumerations
>  - use such new bound iterators to perform the minimum number of queries
>    in order to ony retrieve min an max rate
>
> As a final result now the rates enumeration triggered by the CLK SCMI
> driver, while still allocating for all the existent rates, miminize the
> number of SCMI CLK_DESCRIBE_RATE messages needed to obtain min and max.
>
> Finally, patch 11 introduces a new clock protocol operation to be able to
> trigger anytime on demand a full enumeration and obtain the full list of
> rates when needed, not only min/max: this latter method is really only used
> currently by some dowstream SCMI Test driver of mine.
>
> Based on v7.0-rc1.
>
> Tested on JUNO and an emulated environment.

Thank you, this removes the need for increasing SCMI_MAX_NUM_RATES on
R-Car X5H, while decreasing memory usage.
Tested-by: Geert Uytterhoeven <geert+renesas@glider.be>

> [1]: https://lore.kernel.org/arm-scmi/aZsX-oplR6fiLBBN@pluto/T/#t
> [2]: https://lore.kernel.org/20241203173908.3148794-2-etienne.carriere@foss.st.com

Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds

On Mon, Mar 02, 2026 at 02:25:00PM +0100, Geert Uytterhoeven wrote:
> Hi Cristian,
> 
> Thanks for your series!
> 
> On Fri, 27 Feb 2026 at 16:33, Cristian Marussi <cristian.marussi@arm.com> wrote:
> > it was a known limitation, in the SCMI Clock protocol support, the lack of
> > dynamic allocation around per-clock rates discovery: fixed size statically
> > per-clock rates arrays did not scale and was increasingly a waste of memory
> > (see [1]).
> >
> > This series aim at solving this in successive steps:
> >
> >  - simplify and reduce to the minimum possible the rates data info exposed
> >    to the SCMI driver by scmi_clock_info
> >  - move away from static fixed allocation of per-clock rates arrays in
> >    favour of a completely dynamic runtime allocation: just allocate what
> >    is needed based on the effectively discovered
> >
> > This is done in patches 1-6.
> >
> > A further bigger optimization suggested in a past series [1] by Etienne
> 
> s/[1]/[2]/
> 
> > would be, whenever allowed by the spec, to limit upfront the number of
> > queries in order to simply retrieve min and max rate, that are indeed the
> > only rates needed by the CLK SCMI driver.
> >
> > The approach proposed in [1] was open coding and duplicating some of the
> 
> What does [1] refer to?

I messed up the refs..of course...it was just a reference to your thread
where the number of rates where staticallty raised to 64.

> 
> > functionalities already provided by SCMI iterators, though.
> >
> > Patch 7-10 implement such optimization instead by:
> >
> >  - reworking core SCMI iterators to support bound enumerations
> >  - use such new bound iterators to perform the minimum number of queries
> >    in order to ony retrieve min an max rate
> >
> > As a final result now the rates enumeration triggered by the CLK SCMI
> > driver, while still allocating for all the existent rates, miminize the
> > number of SCMI CLK_DESCRIBE_RATE messages needed to obtain min and max.
> >
> > Finally, patch 11 introduces a new clock protocol operation to be able to
> > trigger anytime on demand a full enumeration and obtain the full list of
> > rates when needed, not only min/max: this latter method is really only used
> > currently by some dowstream SCMI Test driver of mine.
> >
> > Based on v7.0-rc1.
> >
> > Tested on JUNO and an emulated environment.
> 
> Thank you, this removes the need for increasing SCMI_MAX_NUM_RATES on
> R-Car X5H, while decreasing memory usage.
> Tested-by: Geert Uytterhoeven <geert+renesas@glider.be>
> 

Thanks for testing in the real world !
Cristian