| Message ID | 20220504075153.185208-1-marcan@marcan.st (mailing list archive) |
|---|---|
| Headers | show |
| Series | Apple SoC cpufreq driver | expand |
On 04-05-22, 16:51, Hector Martin wrote: > Hi folks, > > Here's a second take on the cpufreq driver for Apple SoCs. This is a > complete rewrite using a stand-alone cpufreq driver instead of using the > cpufreq-dt infrastructure. > > Since v1 we ran some experiments on the memory controller performance > switching and it turns out it doesn't make a huge difference, so it > makes sense to punt that feature to the future (perhaps once a proper > memory controller driver exists for other reasons, e.g. for error > handling). > > One advantage of having a standalone cpufreq driver is that we can > support fast switching. This also means any future interaction with > the memory controller will probably use some bespoke mechanism instead > of the genpd infrastructure, so we can keep the fast path without > allowing sleeps/etc. > > The driver is based on scpi-cpufreq.c, with some bits (e.g. the > apple,freq-domain stuff) inspired by how cpufreq-qcom-hw does it. > I'm not sure if that particular property should be described > in a binding, since it goes in the cpu nodes (qcom doesn't have it > anywhere...). Hi Mani, I can see that Rob asked you to add this somewhere, maybe in arm/cpu stuff, but I don't think you ever sent a patch with that. What happened ? https://lore.kernel.org/lkml/20201013171800.GA3716411@bogus/
Hi Viresh, On Wed, May 04, 2022 at 03:57:45PM +0530, Viresh Kumar wrote: > On 04-05-22, 16:51, Hector Martin wrote: > > Hi folks, > > > > Here's a second take on the cpufreq driver for Apple SoCs. This is a > > complete rewrite using a stand-alone cpufreq driver instead of using the > > cpufreq-dt infrastructure. > > > > Since v1 we ran some experiments on the memory controller performance > > switching and it turns out it doesn't make a huge difference, so it > > makes sense to punt that feature to the future (perhaps once a proper > > memory controller driver exists for other reasons, e.g. for error > > handling). > > > > One advantage of having a standalone cpufreq driver is that we can > > support fast switching. This also means any future interaction with > > the memory controller will probably use some bespoke mechanism instead > > of the genpd infrastructure, so we can keep the fast path without > > allowing sleeps/etc. > > > > The driver is based on scpi-cpufreq.c, with some bits (e.g. the > > apple,freq-domain stuff) inspired by how cpufreq-qcom-hw does it. > > I'm not sure if that particular property should be described > > in a binding, since it goes in the cpu nodes (qcom doesn't have it > > anywhere...). > > Hi Mani, > > I can see that Rob asked you to add this somewhere, maybe in arm/cpu > stuff, but I don't think you ever sent a patch with that. What > happened ? > > https://lore.kernel.org/lkml/20201013171800.GA3716411@bogus/ > Oops. Looks like that one slipped through the cracks. I did add it to my todo list for qcom-cpufreq but missed it completely. I will look into it. Thanks, Mani > -- > viresh
On Wed, May 04, 2022 at 09:30:26PM +0530, Manivannan Sadhasivam wrote: > Hi Viresh, > > On Wed, May 04, 2022 at 03:57:45PM +0530, Viresh Kumar wrote: > > On 04-05-22, 16:51, Hector Martin wrote: > > > Hi folks, > > > > > > Here's a second take on the cpufreq driver for Apple SoCs. This is a > > > complete rewrite using a stand-alone cpufreq driver instead of using the > > > cpufreq-dt infrastructure. > > > > > > Since v1 we ran some experiments on the memory controller performance > > > switching and it turns out it doesn't make a huge difference, so it > > > makes sense to punt that feature to the future (perhaps once a proper > > > memory controller driver exists for other reasons, e.g. for error > > > handling). > > > > > > One advantage of having a standalone cpufreq driver is that we can > > > support fast switching. This also means any future interaction with > > > the memory controller will probably use some bespoke mechanism instead > > > of the genpd infrastructure, so we can keep the fast path without > > > allowing sleeps/etc. > > > > > > The driver is based on scpi-cpufreq.c, with some bits (e.g. the > > > apple,freq-domain stuff) inspired by how cpufreq-qcom-hw does it. > > > I'm not sure if that particular property should be described > > > in a binding, since it goes in the cpu nodes (qcom doesn't have it > > > anywhere...). > > > > Hi Mani, > > > > I can see that Rob asked you to add this somewhere, maybe in arm/cpu > > stuff, but I don't think you ever sent a patch with that. What > > happened ? > > > > https://lore.kernel.org/lkml/20201013171800.GA3716411@bogus/ > > > > Oops. Looks like that one slipped through the cracks. I did add it to my todo > list for qcom-cpufreq but missed it completely. > > I will look into it. > So I did send a patch for adding the qcom specific property [1], but Rob asked for a common one and that's where it got lost. But in the last revision [2], you've asked for converting the qcom cpufreq driver to support the generic performance domains instead. For maintaining compatibility with the old dts files, we need to support the qcom specific property as well. I will send a series for that. Thanks, Mani [1] https://patchwork.kernel.org/project/linux-pm/patch/20201020153944.18047-1-manivannan.sadhasivam@linaro.org/#23718065 [2] https://patchwork.kernel.org/project/linux-pm/patch/20210701105730.322718-4-angelogioacchino.delregno@somainline.org/ > Thanks, > Mani > > > -- > > viresh
Hi folks, Here's a second take on the cpufreq driver for Apple SoCs. This is a complete rewrite using a stand-alone cpufreq driver instead of using the cpufreq-dt infrastructure. Since v1 we ran some experiments on the memory controller performance switching and it turns out it doesn't make a huge difference, so it makes sense to punt that feature to the future (perhaps once a proper memory controller driver exists for other reasons, e.g. for error handling). One advantage of having a standalone cpufreq driver is that we can support fast switching. This also means any future interaction with the memory controller will probably use some bespoke mechanism instead of the genpd infrastructure, so we can keep the fast path without allowing sleeps/etc. The driver is based on scpi-cpufreq.c, with some bits (e.g. the apple,freq-domain stuff) inspired by how cpufreq-qcom-hw does it. I'm not sure if that particular property should be described in a binding, since it goes in the cpu nodes (qcom doesn't have it anywhere...). Changes since v1: - Complete rewrite - Reports current frequency to userspace properly (incl. if different from requested due to hardware constraints) - Supports fast switching - MCC latency control stuff no longer included, punted for later - Supports exposing higher states as turbo states Hector Martin (4): MAINTAINERS: Add entries for Apple SoC cpufreq driver dt-bindings: cpufreq: apple,soc-cpufreq: Add binding for Apple SoC cpufreq cpufreq: apple-soc: Add new driver to control Apple SoC CPU P-states arm64: dts: apple: Add CPU topology & cpufreq nodes for t8103 .../bindings/cpufreq/apple,soc-cpufreq.yaml | 121 +++++++ MAINTAINERS | 2 + arch/arm64/boot/dts/apple/t8103.dtsi | 203 ++++++++++- drivers/cpufreq/Kconfig.arm | 9 + drivers/cpufreq/Makefile | 1 + drivers/cpufreq/apple-soc-cpufreq.c | 330 ++++++++++++++++++ drivers/cpufreq/cpufreq-dt-platdev.c | 2 + 7 files changed, 658 insertions(+), 10 deletions(-) create mode 100644 Documentation/devicetree/bindings/cpufreq/apple,soc-cpufreq.yaml create mode 100644 drivers/cpufreq/apple-soc-cpufreq.c