[2/2] OPP: Disallow "opp-hz" property without a corresponding clk

  This removes the special code added by the commit 2083da24eb56 ("OPP:
Allow multiple clocks for a device"), to make it work for Qcom SoC.

In qcom-cpufreq-hw driver, we want to skip clk configuration that
happens via dev_pm_opp_set_opp(), but still want the OPP core to parse
the "opp-hz" property so we can use the freq based OPP helpers.

The DT for Qcom SoCs is fixed now and contain a valid clk entry, and we
no longer need the special handling of the same in the OPP core.

Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
---
 drivers/opp/core.c    | 14 --------------
 drivers/opp/debugfs.c |  2 +-
 2 files changed, 1 insertion(+), 15 deletions(-)
  

On Mon, Nov 21, 2022 at 12:27:48PM +0530, Viresh Kumar wrote:
> This removes the special code added by the commit 2083da24eb56 ("OPP:
> Allow multiple clocks for a device"), to make it work for Qcom SoC.
> 
> In qcom-cpufreq-hw driver, we want to skip clk configuration that
> happens via dev_pm_opp_set_opp(), but still want the OPP core to parse
> the "opp-hz" property so we can use the freq based OPP helpers.
> 
> The DT for Qcom SoCs is fixed now and contain a valid clk entry, and we
> no longer need the special handling of the same in the OPP core.

Didn't this affect also sc8280xp? Perhaps you can hold off with applying
this one for a bit until the needed devicetree changes are in linux-next
for all the affected platforms.

(It looks like Mani's series only updated sm8450 and I guess Bjorn
hasn't picked up that one up yet either.)

Johan
  

On 21-11-22, 08:22, Johan Hovold wrote:
> On Mon, Nov 21, 2022 at 12:27:48PM +0530, Viresh Kumar wrote:
> > This removes the special code added by the commit 2083da24eb56 ("OPP:
> > Allow multiple clocks for a device"), to make it work for Qcom SoC.
> > 
> > In qcom-cpufreq-hw driver, we want to skip clk configuration that
> > happens via dev_pm_opp_set_opp(), but still want the OPP core to parse
> > the "opp-hz" property so we can use the freq based OPP helpers.
> > 
> > The DT for Qcom SoCs is fixed now and contain a valid clk entry, and we
> > no longer need the special handling of the same in the OPP core.
> 
> Didn't this affect also sc8280xp?

I assumed Mani fixed all affected Qcom SoCs :(

> Perhaps you can hold off with applying
> this one for a bit until the needed devicetree changes are in linux-next
> for all the affected platforms.

Sure.

> (It looks like Mani's series only updated sm8450 and I guess Bjorn
> hasn't picked up that one up yet either.)

I applied that series today to my cpufreq/arm/linux-next, since it had
cpufreq updates too and these patches are rebased on top of them.
  

On Mon, Nov 21, 2022 at 08:22:01AM +0100, Johan Hovold wrote:
> On Mon, Nov 21, 2022 at 12:27:48PM +0530, Viresh Kumar wrote:
> > This removes the special code added by the commit 2083da24eb56 ("OPP:
> > Allow multiple clocks for a device"), to make it work for Qcom SoC.
> > 
> > In qcom-cpufreq-hw driver, we want to skip clk configuration that
> > happens via dev_pm_opp_set_opp(), but still want the OPP core to parse
> > the "opp-hz" property so we can use the freq based OPP helpers.
> > 
> > The DT for Qcom SoCs is fixed now and contain a valid clk entry, and we
> > no longer need the special handling of the same in the OPP core.
> 
> Didn't this affect also sc8280xp? Perhaps you can hold off with applying
> this one for a bit until the needed devicetree changes are in linux-next
> for all the affected platforms.
> 
> (It looks like Mani's series only updated sm8450 and I guess Bjorn
> hasn't picked up that one up yet either.)
> 

That's right. I have proposed to do the similar change to other SoCs as well
once the series was completely merged. I thought of doing so for 6.3.

Btw, there seems to be one more candidate:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm64/boot/dts/qcom/sm8250.dtsi#n2537

Looks like newer SoCs that has the GMU within the GPU block doesn't have clock
property. This is because, GMU is the one supplying clocks to the GPU unlike the
old SoCs where the clocks used to come from GCC itself.

But we do have a GMU devicetree node, so it should be a matter of adding the
clock provider support as done for cpufreq and represent it in devicetree.

I'll ping Rob Clark and see how to get it done.

Thanks,
Mani

> Johan
  

On Mon, Nov 21, 2022 at 01:08:17PM +0530, Viresh Kumar wrote:
> On 21-11-22, 08:22, Johan Hovold wrote:
> > On Mon, Nov 21, 2022 at 12:27:48PM +0530, Viresh Kumar wrote:
> > > This removes the special code added by the commit 2083da24eb56 ("OPP:
> > > Allow multiple clocks for a device"), to make it work for Qcom SoC.
> > > 
> > > In qcom-cpufreq-hw driver, we want to skip clk configuration that
> > > happens via dev_pm_opp_set_opp(), but still want the OPP core to parse
> > > the "opp-hz" property so we can use the freq based OPP helpers.
> > > 
> > > The DT for Qcom SoCs is fixed now and contain a valid clk entry, and we
> > > no longer need the special handling of the same in the OPP core.
> > 
> > Didn't this affect also sc8280xp?
> 
> I assumed Mani fixed all affected Qcom SoCs :(

I think he may have been waiting for his suggested solution to be
accepted before updating all the affected dtsi:s.
 
> > Perhaps you can hold off with applying
> > this one for a bit until the needed devicetree changes are in linux-next
> > for all the affected platforms.
> 
> Sure.

Thanks.

> > (It looks like Mani's series only updated sm8450 and I guess Bjorn
> > hasn't picked up that one up yet either.)
> 
> I applied that series today to my cpufreq/arm/linux-next, since it had
> cpufreq updates too and these patches are rebased on top of them.

Ok, I was expected the devicetree update to through Bjorn's tree as I
imagine there may be conflicts otherwise.

What was the intention here, Mani?

Johan
  

On Mon, Nov 21, 2022 at 01:08:17PM +0530, Viresh Kumar wrote:
> On 21-11-22, 08:22, Johan Hovold wrote:
> > On Mon, Nov 21, 2022 at 12:27:48PM +0530, Viresh Kumar wrote:
> > > This removes the special code added by the commit 2083da24eb56 ("OPP:
> > > Allow multiple clocks for a device"), to make it work for Qcom SoC.
> > > 
> > > In qcom-cpufreq-hw driver, we want to skip clk configuration that
> > > happens via dev_pm_opp_set_opp(), but still want the OPP core to parse
> > > the "opp-hz" property so we can use the freq based OPP helpers.
> > > 
> > > The DT for Qcom SoCs is fixed now and contain a valid clk entry, and we
> > > no longer need the special handling of the same in the OPP core.
> > 
> > Didn't this affect also sc8280xp?
> 
> I assumed Mani fixed all affected Qcom SoCs :(
> 
> > Perhaps you can hold off with applying
> > this one for a bit until the needed devicetree changes are in linux-next
> > for all the affected platforms.
> 
> Sure.
> 
> > (It looks like Mani's series only updated sm8450 and I guess Bjorn
> > hasn't picked up that one up yet either.)
> 
> I applied that series today to my cpufreq/arm/linux-next, since it had
> cpufreq updates too and these patches are rebased on top of them.
> 

Ah, I thought you applied only cpufreq patches. DTS and bindings patches are
supposed to go through Bjorn's tree. Can you please drop them?

Thanks,
Mani

> -- 
> viresh
  

On 21-11-22, 08:42, Johan Hovold wrote:
> Ok, I was expected the devicetree update to through Bjorn's tree as I
> imagine there may be conflicts otherwise.

Dropped the DT patch now.

Mani, I believe the first patch in this series is still required.
Without that, when the cpufreq driver calls dev_pm_opp_set_opp(), the
OPP core may end up calling clk_set_rate(), which shouldn't be done in
case of Qcom SoCs.

I am not sure though, how it will work with sc8280xp. Can you please
check ?
  

On Mon, Nov 21, 2022 at 02:00:36PM +0530, Viresh Kumar wrote:
> On 21-11-22, 08:42, Johan Hovold wrote:
> > Ok, I was expected the devicetree update to through Bjorn's tree as I
> > imagine there may be conflicts otherwise.
> 
> Dropped the DT patch now.
> 
> Mani, I believe the first patch in this series is still required.
> Without that, when the cpufreq driver calls dev_pm_opp_set_opp(), the
> OPP core may end up calling clk_set_rate(), which shouldn't be done in
> case of Qcom SoCs.
> 

If there is no .set_rate() callback implemented by the clock provider, it won't
hurt, right?

Thanks,
Mani

> I am not sure though, how it will work with sc8280xp. Can you please
> check ?
> 
> -- 
> viresh
  

On 22-11-22, 18:56, Manivannan Sadhasivam wrote:
> If there is no .set_rate() callback implemented by the clock provider, it won't
> hurt, right?

It shouldn't, I guess. Well, in that case, is the first patch even
required ? Maybe we should keep it, this makes clear that we won't
even call set_rate(), irrespective of the face that it is implemented
or not.

Also, the clk provider may not be part of this file later on, for
other SoC versions, and it is better in that case too.
  

On Thu, Nov 24, 2022 at 09:53:04AM +0530, Viresh Kumar wrote:
> On 22-11-22, 18:56, Manivannan Sadhasivam wrote:
> > If there is no .set_rate() callback implemented by the clock provider, it won't
> > hurt, right?
> 
> It shouldn't, I guess. Well, in that case, is the first patch even
> required ? Maybe we should keep it, this makes clear that we won't
> even call set_rate(), irrespective of the face that it is implemented
> or not.
> 

I don't think that detail is required to be made explicit. If someone cares,
they can easlily find out by glancing through the OPP code.

So IMO, we don't need patch 1/2.

> Also, the clk provider may not be part of this file later on, for
> other SoC versions, and it is better in that case too.
> 

We cannot predict what the HW guys will come up with ;) But as said above, I
don't think it is necessary to to make it explicit.

Thanks,
Mani

> -- 
> viresh
  

On 21-11-22, 13:09, Manivannan Sadhasivam wrote:
> That's right. I have proposed to do the similar change to other SoCs as well
> once the series was completely merged. I thought of doing so for 6.3.
> 
> Btw, there seems to be one more candidate:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm64/boot/dts/qcom/sm8250.dtsi#n2537
> 
> Looks like newer SoCs that has the GMU within the GPU block doesn't have clock
> property. This is because, GMU is the one supplying clocks to the GPU unlike the
> old SoCs where the clocks used to come from GCC itself.
> 
> But we do have a GMU devicetree node, so it should be a matter of adding the
> clock provider support as done for cpufreq and represent it in devicetree.
> 
> I'll ping Rob Clark and see how to get it done.

Any update on this Mani ? I want to get the hack removed if possible.
  

+ Rob Clark

On Wed, Jan 25, 2023 at 09:51:45AM +0530, Viresh Kumar wrote:
> On 21-11-22, 13:09, Manivannan Sadhasivam wrote:
> > That's right. I have proposed to do the similar change to other SoCs as well
> > once the series was completely merged. I thought of doing so for 6.3.
> > 
> > Btw, there seems to be one more candidate:
> > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/arm64/boot/dts/qcom/sm8250.dtsi#n2537
> > 
> > Looks like newer SoCs that has the GMU within the GPU block doesn't have clock
> > property. This is because, GMU is the one supplying clocks to the GPU unlike the
> > old SoCs where the clocks used to come from GCC itself.
> > 
> > But we do have a GMU devicetree node, so it should be a matter of adding the
> > clock provider support as done for cpufreq and represent it in devicetree.
> > 
> > I'll ping Rob Clark and see how to get it done.
> 
> Any update on this Mani ? I want to get the hack removed if possible.
> 

Hi Viresh,

Sorry for the delay. I've submitted the dts changes [1] to handle the CPU clocks
for the rest of the Qcom SoCs.

For the Qcom GPUs, I've CCed Rob Clark who is the maintainer.

Rob, here is the background on the issue that is being discussed in this
thread:

Viresh submitted a series [2] back in July to improve the OPP framework, but
that ended up breaking cpufreq on multiple Qcom SoCs. After investigation, it
was found that the series was expecting the clocks supplied to the OPP end
devices like CPUs/GPUs to be modeled in DT. But on Qcom platforms even though
the clocks for these nodes are supplied by a separate entity, like CPUFreq
(EPSS/OSM) for CPUs and GMU for GPUs, there was no clock property present in
the respective nodes. And these nodes are using OPP table to switch frequencies
dynamically.

While the series was merged with a hack that still allows the OPP nodes without
clock property in DT, we came to an agreement that the clock hierarchy should
be modeled properly.

So I submitted a series [3] that added clock provider support to cpufreq driver
and sourced the clock from cpufreq node to CPU nodes in DT.

Likewise, it should be handled for the adreno GPUs whose clock is managed by
GMU on newer SoCs. Can you take a look at this?

Thanks,
Mani

[1] https://lore.kernel.org/linux-arm-msm/20230215070400.5901-1-manivannan.sadhasivam@linaro.org/
[2] https://lore.kernel.org/lkml/cover.1657003420.git.viresh.kumar@linaro.org/
[3] https://lore.kernel.org/linux-arm-msm/20221117053145.10409-1-manivannan.sadhasivam@linaro.org/

> -- 
> viresh
  

On 16-02-23, 12:17, Manivannan Sadhasivam wrote:
> Sorry for the delay. I've submitted the dts changes [1] to handle the CPU clocks
> for the rest of the Qcom SoCs.
> 
> For the Qcom GPUs, I've CCed Rob Clark who is the maintainer.
> 
> Rob, here is the background on the issue that is being discussed in this
> thread:
> 
> Viresh submitted a series [2] back in July to improve the OPP framework, but
> that ended up breaking cpufreq on multiple Qcom SoCs. After investigation, it
> was found that the series was expecting the clocks supplied to the OPP end
> devices like CPUs/GPUs to be modeled in DT. But on Qcom platforms even though
> the clocks for these nodes are supplied by a separate entity, like CPUFreq
> (EPSS/OSM) for CPUs and GMU for GPUs, there was no clock property present in
> the respective nodes. And these nodes are using OPP table to switch frequencies
> dynamically.
> 
> While the series was merged with a hack that still allows the OPP nodes without
> clock property in DT, we came to an agreement that the clock hierarchy should
> be modeled properly.
> 
> So I submitted a series [3] that added clock provider support to cpufreq driver
> and sourced the clock from cpufreq node to CPU nodes in DT.
> 
> Likewise, it should be handled for the adreno GPUs whose clock is managed by
> GMU on newer SoCs. Can you take a look at this?

Any update on this ?
  

On 11-10-23, 11:18, Viresh Kumar wrote:
> On 16-02-23, 12:17, Manivannan Sadhasivam wrote:
> > Sorry for the delay. I've submitted the dts changes [1] to handle the CPU clocks
> > for the rest of the Qcom SoCs.
> > 
> > For the Qcom GPUs, I've CCed Rob Clark who is the maintainer.
> > 
> > Rob, here is the background on the issue that is being discussed in this
> > thread:
> > 
> > Viresh submitted a series [2] back in July to improve the OPP framework, but
> > that ended up breaking cpufreq on multiple Qcom SoCs. After investigation, it
> > was found that the series was expecting the clocks supplied to the OPP end
> > devices like CPUs/GPUs to be modeled in DT. But on Qcom platforms even though
> > the clocks for these nodes are supplied by a separate entity, like CPUFreq
> > (EPSS/OSM) for CPUs and GMU for GPUs, there was no clock property present in
> > the respective nodes. And these nodes are using OPP table to switch frequencies
> > dynamically.
> > 
> > While the series was merged with a hack that still allows the OPP nodes without
> > clock property in DT, we came to an agreement that the clock hierarchy should
> > be modeled properly.
> > 
> > So I submitted a series [3] that added clock provider support to cpufreq driver
> > and sourced the clock from cpufreq node to CPU nodes in DT.
> > 
> > Likewise, it should be handled for the adreno GPUs whose clock is managed by
> > GMU on newer SoCs. Can you take a look at this?
> 
> Any update on this ?

Mani,

Ping.
  

+ Dmitry

On Wed, Nov 15, 2023 at 12:02:01PM +0530, Viresh Kumar wrote:
> On 11-10-23, 11:18, Viresh Kumar wrote:
> > On 16-02-23, 12:17, Manivannan Sadhasivam wrote:
> > > Sorry for the delay. I've submitted the dts changes [1] to handle the CPU clocks
> > > for the rest of the Qcom SoCs.
> > > 
> > > For the Qcom GPUs, I've CCed Rob Clark who is the maintainer.
> > > 
> > > Rob, here is the background on the issue that is being discussed in this
> > > thread:
> > > 
> > > Viresh submitted a series [2] back in July to improve the OPP framework, but
> > > that ended up breaking cpufreq on multiple Qcom SoCs. After investigation, it
> > > was found that the series was expecting the clocks supplied to the OPP end
> > > devices like CPUs/GPUs to be modeled in DT. But on Qcom platforms even though
> > > the clocks for these nodes are supplied by a separate entity, like CPUFreq
> > > (EPSS/OSM) for CPUs and GMU for GPUs, there was no clock property present in
> > > the respective nodes. And these nodes are using OPP table to switch frequencies
> > > dynamically.
> > > 
> > > While the series was merged with a hack that still allows the OPP nodes without
> > > clock property in DT, we came to an agreement that the clock hierarchy should
> > > be modeled properly.
> > > 
> > > So I submitted a series [3] that added clock provider support to cpufreq driver
> > > and sourced the clock from cpufreq node to CPU nodes in DT.
> > > 
> > > Likewise, it should be handled for the adreno GPUs whose clock is managed by
> > > GMU on newer SoCs. Can you take a look at this?
> > 
> > Any update on this ?
> 
> Mani,
> 
> Ping.
> 

Dmitry, can you please look into this? Please read my above reply to Rob
to get the background.

- Mani

> -- 
> viresh
  

On Wed, 15 Nov 2023 at 09:55, Manivannan Sadhasivam
<manivannan.sadhasivam@linaro.org> wrote:
>
> + Dmitry
>
> On Wed, Nov 15, 2023 at 12:02:01PM +0530, Viresh Kumar wrote:
> > On 11-10-23, 11:18, Viresh Kumar wrote:
> > > On 16-02-23, 12:17, Manivannan Sadhasivam wrote:
> > > > Sorry for the delay. I've submitted the dts changes [1] to handle the CPU clocks
> > > > for the rest of the Qcom SoCs.
> > > >
> > > > For the Qcom GPUs, I've CCed Rob Clark who is the maintainer.
> > > >
> > > > Rob, here is the background on the issue that is being discussed in this
> > > > thread:
> > > >
> > > > Viresh submitted a series [2] back in July to improve the OPP framework, but
> > > > that ended up breaking cpufreq on multiple Qcom SoCs. After investigation, it
> > > > was found that the series was expecting the clocks supplied to the OPP end
> > > > devices like CPUs/GPUs to be modeled in DT. But on Qcom platforms even though
> > > > the clocks for these nodes are supplied by a separate entity, like CPUFreq
> > > > (EPSS/OSM) for CPUs and GMU for GPUs, there was no clock property present in
> > > > the respective nodes. And these nodes are using OPP table to switch frequencies
> > > > dynamically.
> > > >
> > > > While the series was merged with a hack that still allows the OPP nodes without
> > > > clock property in DT, we came to an agreement that the clock hierarchy should
> > > > be modeled properly.
> > > >
> > > > So I submitted a series [3] that added clock provider support to cpufreq driver
> > > > and sourced the clock from cpufreq node to CPU nodes in DT.
> > > >
> > > > Likewise, it should be handled for the adreno GPUs whose clock is managed by
> > > > GMU on newer SoCs. Can you take a look at this?
> > >
> > > Any update on this ?
> >
> > Mani,
> >
> > Ping.
> >
>
> Dmitry, can you please look into this? Please read my above reply to Rob
> to get the background.

The issue is that we don't have an actual clock that corresponds to
the GPU frequency. Not even a read-only one.
Can we get away by manually setting config_clocks()?

Also could you please remind me, can we sleep inside the config_clks() callback?