| Message ID | 20200821140026.19643-1-ansuelsmth@gmail.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Add Krait Cache Scaling support | expand |
+Vincent/Saravana/Sibi On 21-08-20, 16:00, Ansuel Smith wrote: > This adds Krait Cache scaling support using the cpufreq notifier. > I have some doubt about where this should be actually placed (clk or cpufreq)? > Also the original idea was to create a dedicated cpufreq driver (like it's done in > the codeaurora qcom repo) by copying the cpufreq-dt driver and adding the cache > scaling logic but i still don't know what is better. Have a very similar driver or > add a dedicated driver only for the cache using the cpufreq notifier and do the > scale on every freq transition. > Thanks to everyone who will review or answer these questions. Saravana was doing something with devfreq to solve such issues if I wasn't mistaken. Sibi ?
On 2020-08-24 16:10, Viresh Kumar wrote: > +Vincent/Saravana/Sibi > > On 21-08-20, 16:00, Ansuel Smith wrote: >> This adds Krait Cache scaling support using the cpufreq notifier. >> I have some doubt about where this should be actually placed (clk or >> cpufreq)? >> Also the original idea was to create a dedicated cpufreq driver (like >> it's done in >> the codeaurora qcom repo) by copying the cpufreq-dt driver and adding >> the cache >> scaling logic but i still don't know what is better. Have a very >> similar driver or >> add a dedicated driver only for the cache using the cpufreq notifier >> and do the >> scale on every freq transition. >> Thanks to everyone who will review or answer these questions. > > Saravana was doing something with devfreq to solve such issues if I > wasn't mistaken. > > Sibi ? IIRC the final plan was to create a devfreq device and devfreq-cpufreq based governor to scale them, this way one can switch to a different governor if required. (I don't see if ^^ applies well for l2 though). In the interim until such a solution is acked on the list we just scale the resources directly from the cpufreq driver. On SDM845/SC7180 SoCs, L3 is modeled as a interconnect provider and is directly scaled from the cpufreq-hw driver.
> -----Messaggio originale----- > Da: sibis=codeaurora.org@mg.codeaurora.org > <sibis=codeaurora.org@mg.codeaurora.org> Per conto di Sibi Sankar > Inviato: lunedì 31 agosto 2020 07:46 > A: Viresh Kumar <viresh.kumar@linaro.org> > Cc: Ansuel Smith <ansuelsmth@gmail.com>; vincent.guittot@linaro.org; > saravanak@google.com; Sudeep Holla <sudeep.holla@arm.com>; Rafael J. > Wysocki <rjw@rjwysocki.net>; Rob Herring <robh+dt@kernel.org>; linux- > pm@vger.kernel.org; devicetree@vger.kernel.org; linux- > kernel@vger.kernel.org > Oggetto: Re: [RFC PATCH v3 0/2] Add Krait Cache Scaling support > > On 2020-08-24 16:10, Viresh Kumar wrote: > > +Vincent/Saravana/Sibi > > > > On 21-08-20, 16:00, Ansuel Smith wrote: > >> This adds Krait Cache scaling support using the cpufreq notifier. > >> I have some doubt about where this should be actually placed (clk or > >> cpufreq)? > >> Also the original idea was to create a dedicated cpufreq driver (like > >> it's done in > >> the codeaurora qcom repo) by copying the cpufreq-dt driver and adding > >> the cache > >> scaling logic but i still don't know what is better. Have a very > >> similar driver or > >> add a dedicated driver only for the cache using the cpufreq notifier > >> and do the > >> scale on every freq transition. > >> Thanks to everyone who will review or answer these questions. > > > > Saravana was doing something with devfreq to solve such issues if I > > wasn't mistaken. > > > > Sibi ? > > IIRC the final plan was to create a devfreq device > and devfreq-cpufreq based governor to scale them, this > way one can switch to a different governor if required. So in this case I should convert this patch to a devfreq driver- Isn't overkill to use a governor for such a task? (3 range based on the cpufreq?) > (I don't see if ^^ applies well for l2 though). In the > interim until such a solution is acked on the list we > just scale the resources directly from the cpufreq In this case for this SoC we can't really scale the L2 freq with the cpu since we observed a bug and we need to switch back to the idle freq sometimes. Also this SoC use the generic cpufreq-dt driver and doesn't have a dedicated driver. So we must use a notifier. > driver. On SDM845/SC7180 SoCs, L3 is modeled as a > interconnect provider and is directly scaled from the > cpufreq-hw driver. > > -- > Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, > a Linux Foundation Collaborative Project.
On 31-08-20, 09:41, ansuelsmth@gmail.com wrote: > On 31-08-20, Sibi wrote: > > On 2020-08-24 16:10, Viresh Kumar wrote: > > > +Vincent/Saravana/Sibi > > > > > > On 21-08-20, 16:00, Ansuel Smith wrote: > > >> This adds Krait Cache scaling support using the cpufreq notifier. > > >> I have some doubt about where this should be actually placed (clk or > > >> cpufreq)? > > >> Also the original idea was to create a dedicated cpufreq driver (like > > >> it's done in > > >> the codeaurora qcom repo) by copying the cpufreq-dt driver and adding > > >> the cache > > >> scaling logic but i still don't know what is better. Have a very > > >> similar driver or > > >> add a dedicated driver only for the cache using the cpufreq notifier > > >> and do the > > >> scale on every freq transition. > > >> Thanks to everyone who will review or answer these questions. > > > > > > Saravana was doing something with devfreq to solve such issues if I > > > wasn't mistaken. > > > > > > Sibi ? > > > > IIRC the final plan was to create a devfreq device > > and devfreq-cpufreq based governor to scale them, this > > way one can switch to a different governor if required. > > So in this case I should convert this patch to a devfreq driver- I think this should happen nevertheless. You are doing DVFS for a device which isn't a CPU and devfreq looks to be the right place of doing so. > Isn't overkill to use a governor for such a task? > (3 range based on the cpufreq?) I am not sure about the governor part here, maybe it won't be required ?
On 2020-09-03 12:23, Viresh Kumar wrote: > On 31-08-20, 09:41, ansuelsmth@gmail.com wrote: >> On 31-08-20, Sibi wrote: >> > On 2020-08-24 16:10, Viresh Kumar wrote: >> > > +Vincent/Saravana/Sibi >> > > >> > > On 21-08-20, 16:00, Ansuel Smith wrote: >> > >> This adds Krait Cache scaling support using the cpufreq notifier. >> > >> I have some doubt about where this should be actually placed (clk or >> > >> cpufreq)? >> > >> Also the original idea was to create a dedicated cpufreq driver (like >> > >> it's done in >> > >> the codeaurora qcom repo) by copying the cpufreq-dt driver and adding >> > >> the cache >> > >> scaling logic but i still don't know what is better. Have a very >> > >> similar driver or >> > >> add a dedicated driver only for the cache using the cpufreq notifier >> > >> and do the >> > >> scale on every freq transition. >> > >> Thanks to everyone who will review or answer these questions. >> > > >> > > Saravana was doing something with devfreq to solve such issues if I >> > > wasn't mistaken. >> > > >> > > Sibi ? >> > >> > IIRC the final plan was to create a devfreq device >> > and devfreq-cpufreq based governor to scale them, this >> > way one can switch to a different governor if required. >> >> So in this case I should convert this patch to a devfreq driver- > > I think this should happen nevertheless. You are doing DVFS for a > device which isn't a CPU and devfreq looks to be the right place of > doing so. > >> Isn't overkill to use a governor for such a task? >> (3 range based on the cpufreq?) > > I am not sure about the governor part here, maybe it won't be required > ? Yeah I don't see it being needed in ^^ case as well. I just mentioned them as an advantage in case you wanted to switch to a different governor in the future. https://lore.kernel.org/lkml/d0bc8877-6d41-f54e-1c4c-2fadbb9dcd0b@samsung.com/ A devfreq governor tracking cpufreq was generally accepted but using a cpufreq notifier to achieve that was discouraged.
> -----Messaggio originale----- > Da: sibis=codeaurora.org@mg.codeaurora.org > <sibis=codeaurora.org@mg.codeaurora.org> Per conto di Sibi Sankar > Inviato: giovedì 3 settembre 2020 09:13 > A: Viresh Kumar <viresh.kumar@linaro.org> > Cc: ansuelsmth@gmail.com; vincent.guittot@linaro.org; > saravanak@google.com; 'Sudeep Holla' <sudeep.holla@arm.com>; 'Rafael J. > Wysocki' <rjw@rjwysocki.net>; 'Rob Herring' <robh+dt@kernel.org>; linux- > pm@vger.kernel.org; devicetree@vger.kernel.org; linux- > kernel@vger.kernel.org > Oggetto: Re: R: [RFC PATCH v3 0/2] Add Krait Cache Scaling support > > On 2020-09-03 12:23, Viresh Kumar wrote: > > On 31-08-20, 09:41, ansuelsmth@gmail.com wrote: > >> On 31-08-20, Sibi wrote: > >> > On 2020-08-24 16:10, Viresh Kumar wrote: > >> > > +Vincent/Saravana/Sibi > >> > > > >> > > On 21-08-20, 16:00, Ansuel Smith wrote: > >> > >> This adds Krait Cache scaling support using the cpufreq notifier. > >> > >> I have some doubt about where this should be actually placed (clk > or > >> > >> cpufreq)? > >> > >> Also the original idea was to create a dedicated cpufreq driver (like > >> > >> it's done in > >> > >> the codeaurora qcom repo) by copying the cpufreq-dt driver and > adding > >> > >> the cache > >> > >> scaling logic but i still don't know what is better. Have a very > >> > >> similar driver or > >> > >> add a dedicated driver only for the cache using the cpufreq notifier > >> > >> and do the > >> > >> scale on every freq transition. > >> > >> Thanks to everyone who will review or answer these questions. > >> > > > >> > > Saravana was doing something with devfreq to solve such issues if I > >> > > wasn't mistaken. > >> > > > >> > > Sibi ? > >> > > >> > IIRC the final plan was to create a devfreq device > >> > and devfreq-cpufreq based governor to scale them, this > >> > way one can switch to a different governor if required. > >> > >> So in this case I should convert this patch to a devfreq driver- > > > > I think this should happen nevertheless. You are doing DVFS for a > > device which isn't a CPU and devfreq looks to be the right place of > > doing so. > > > >> Isn't overkill to use a governor for such a task? > >> (3 range based on the cpufreq?) > > > > I am not sure about the governor part here, maybe it won't be required > > ? > > Yeah I don't see it being needed in ^^ > case as well. I just mentioned them as > an advantage in case you wanted to switch > to a different governor in the future. > > https://lore.kernel.org/lkml/d0bc8877-6d41-f54e-1c4c- > 2fadbb9dcd0b@samsung.com/ > > A devfreq governor tracking cpufreq was > generally accepted but using a cpufreq > notifier to achieve that was discouraged. > I read the patch discussion and it looks like at the very end they lost interest in pushing it. That would very fit what I need here so I'm asking how should I proceed? Keep the cpufreq notifier? Introduce a dedicated governor? Ask them to resume the pushing or try to include the changes to the passive governor by myself? > -- > Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, > a Linux Foundation Collaborative Project.