Message ID | 516D46CC.3080705@ti.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Quoting Andrii Tseglytskyi (2013-04-16 05:40:44) > On 04/16/2013 12:53 AM, Kevin Hilman wrote: > > In addition to Mike's comments (which I completely agree with), it would > > be very helfpul to see how this is actually used. e.g, how the > > regulators are chained together, how the proper ordering is managed, > > etc. etc. > > We would like to handle voltage scaling in the following way: > I expanded the example below to include the SR AVS regulator. > cpufreq_cpu0 > clk_set_rate(cpu0) > | > |-->set_voltage(ABB regulator) > | > |-->set_voltage(AVS) > | > |-->set_voltage(smps123 regulator) Hi Andrii, Why was regulator chaining chosen over a simple sequence of calls to regulator_set_voltage? Instead of nested calls into the regulator framework, why don't you just make the calls serially? E.g: regulator_set_voltage(abb_reg, foo_volt); regulator_set_voltage(avs_reg, bar_volt); regulator_set_voltage(smps123_reg, baz_volt); It is still to be determined where these calls originate from; maybe from clock notifiers, maybe directly from the cpufreq driver's .target() callback, or maybe somewhere else. Regardless, I do not see why regulator chaining is truly necessary here. You are just calling regulator_set_voltage in sequence on a few regulators, right? I think it would help me a lot to understand why regulator chaining is a requirement for this to work properly. Thanks, Mike > > > This simple model will be extended to handle AVS as a part of the chain. > smps123 regulator may be changed to VP/VC regulator. > > Following example is from integration branch, which already has smps123 > regulator. > It demonstrates an example of linkage to chain. ABB regulator is linked > with smps123 and cpu0 inside device tree. > cpu0 calls set_voltage() function for ABB, and then ABB calls > set_voltage() function for smps123 to do actual voltage scaling. > > diff --git a/arch/arm/boot/dts/omap5.dtsi b/arch/arm/boot/dts/omap5.dtsi > index bb5ee70..c8cbbee 100644 > --- a/arch/arm/boot/dts/omap5.dtsi > +++ b/arch/arm/boot/dts/omap5.dtsi > @@ -36,7 +36,7 @@ > cpus { > cpu@0 { > compatible = "arm,cortex-a15"; > - cpu0-supply = <&smps123_reg>; > + cpu0-supply = <&abb_mpu>; > operating-points = < > /* kHz uV */ > /* Only for Nominal Samples */ > @@ -94,6 +94,7 @@ > reg = <0x4ae07cdc 0x8>, > <0x4ae06014 0x4>; > ti,tranxdone_status_mask = <0x80>; > + avs-supply = <&smps123_reg>; > operating-points = < > /* uV ABB */ > 880000 0 > > This RFC patch series is verified together with: > https://patchwork.kernel.org/patch/2445091/ > > Kevin, what do you think about this model in general? Does it fit to > regulator framework? > > Thank you. > > Regards, > Andrii -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Andrii Tseglytskyi <andrii.tseglytskyi@ti.com> writes: > Hi Kevin, > > On 04/16/2013 12:53 AM, Kevin Hilman wrote: >> Andrii Tseglytskyi <andrii.tseglytskyi@ti.com> writes: >> >>> From: "Andrii.Tseglytskyi" <andrii.tseglytskyi@ti.com> >>> >>> Following patch series introduces the Adaptive Body-Bias >>> LDO driver, which handles LDOs voltage during OPP change routine. >>> Current implementation is based on patch series from >>> Mike Turquette: >>> >>> http://marc.info/?l=linux-omap&m=134931341818379&w=2 >>> >>> ABB transition is a part of OPP changing sequence. >>> ABB can operate in the following modes: >>> - Bypass mode: Activated when ABB is not required >>> - FBB mode: Fast Body Bias mode, used on fast OPPs >> Fast? I thought the 'F' was for Forward? > You are right. Should be 'Forward' here. > >>> - RBB mode: Reverse Body Bias mode, used on slow OPPs >>> >>> In current implementation ABB is converted to regulator. >>> Standalone OPP table is used to store ABB mode, it is defined >>> in device tree for each ABB regulator. It has the following format: >>> >>> operating-points = < >>> /* uV ABB (0 - Bypass, 1 - FBB, 2 - RBB) */ >>> 880000 0 >>> 1060000 1 >>> 1250000 1 >>> 1260000 1 >>>> ; >>> ABB regulator is linked to regulator chain >> In addition to Mike's comments (which I completely agree with), it would >> be very helfpul to see how this is actually used. e.g, how the >> regulators are chained together, how the proper ordering is managed, >> etc. etc. > > We would like to handle voltage scaling in the following way: What I meant is that a detailed description of the use case should be included in the changelog. > cpufreq_cpu0 > clk_set_rate(cpu0) > | > |-->set_voltage(ABB regulator) /* all ABB related stuff will be > handled here */ > | > |-->set_voltage(smps123 regulator) /* actual voltage > scaling */ -EASCII_ART_WRAP > > This simple model will be extended to handle AVS as a part of the chain. > smps123 regulator may be changed to VP/VC regulator. > > Following example is from integration branch, which already has > smps123 regulator. I don't know what integration branch you're referring to, and I don't know what the smps123 regulator is. > It demonstrates an example of linkage to chain. ABB regulator is > linked with smps123 and cpu0 inside device tree. > cpu0 calls set_voltage() function for ABB, and then ABB calls > set_voltage() function for smps123 to do actual voltage scaling. > > diff --git a/arch/arm/boot/dts/omap5.dtsi b/arch/arm/boot/dts/omap5.dtsi > index bb5ee70..c8cbbee 100644 > --- a/arch/arm/boot/dts/omap5.dtsi > +++ b/arch/arm/boot/dts/omap5.dtsi > @@ -36,7 +36,7 @@ > cpus { > cpu@0 { > compatible = "arm,cortex-a15"; > - cpu0-supply = <&smps123_reg>; > + cpu0-supply = <&abb_mpu>; > operating-points = < > /* kHz uV */ > /* Only for Nominal Samples */ > @@ -94,6 +94,7 @@ > reg = <0x4ae07cdc 0x8>, > <0x4ae06014 0x4>; > ti,tranxdone_status_mask = <0x80>; > + avs-supply = <&smps123_reg>; > operating-points = < > /* uV ABB */ > 880000 0 > > This RFC patch series is verified together with: > https://patchwork.kernel.org/patch/2445091/ > > Kevin, what do you think about this model in general? Does it fit to > regulator framework? I don't know yet, because I don't think the use case has been described well enough for me to fully understand it the motiviation behind the series. In addition, there are alternative approaches that seem to have been ruled out without describing why. For example, the regulator framework already allows you to override methods with custom hooks (as we do for VC/VP controlled regulators already.) Without thinking about it too deeply, it seems this approach could be used to manage the chain of events you need as well. I can imagine there are limitations to this approach for what you're trying to do, but I don't feel they have been described in the changelog as part of the motivation for this series. So for now, the guidance I have is this: First, write changelogs (and cover letters) assuming your audience has not been staring at the code as long as you have. Even if they have been staring at the same code, assume they've been staring at mainline, and not a random integration branch somewhere. My general advice is to write changelogs in a way that you will understand what you wrote a year from now after having forgotten all the details currently in your brains cache. Even better, write them so that I will understand them in a year since I forget much better than I remember. Second, before inventing something new, start with the existing framework. When the existing framework doesn't work, make an argument for your new approach or extentions to the framework based on why the existing stuff doesn't work. If you don't do this, the reviewers first reaction will almost always be "why don't you use what already exists in the framework." And then you'll have a bunch of back and forth with reviewers when you could've explained the reasoning from the beginning. Hope that helps, Kevin -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 04/16/2013 10:18 PM, Kevin Hilman wrote: > Andrii Tseglytskyi <andrii.tseglytskyi@ti.com> writes: > >> Hi Kevin, >> >> On 04/16/2013 12:53 AM, Kevin Hilman wrote: >>> Andrii Tseglytskyi <andrii.tseglytskyi@ti.com> writes: >>> >>>> From: "Andrii.Tseglytskyi" <andrii.tseglytskyi@ti.com> >>>> >>>> Following patch series introduces the Adaptive Body-Bias >>>> LDO driver, which handles LDOs voltage during OPP change routine. >>>> Current implementation is based on patch series from >>>> Mike Turquette: >>>> >>>> http://marc.info/?l=linux-omap&m=134931341818379&w=2 >>>> >>>> ABB transition is a part of OPP changing sequence. >>>> ABB can operate in the following modes: >>>> - Bypass mode: Activated when ABB is not required >>>> - FBB mode: Fast Body Bias mode, used on fast OPPs >>> Fast? I thought the 'F' was for Forward? >> You are right. Should be 'Forward' here. >> >>>> - RBB mode: Reverse Body Bias mode, used on slow OPPs >>>> >>>> In current implementation ABB is converted to regulator. >>>> Standalone OPP table is used to store ABB mode, it is defined >>>> in device tree for each ABB regulator. It has the following format: >>>> >>>> operating-points = < >>>> /* uV ABB (0 - Bypass, 1 - FBB, 2 - RBB) */ >>>> 880000 0 >>>> 1060000 1 >>>> 1250000 1 >>>> 1260000 1 >>>>> ; >>>> ABB regulator is linked to regulator chain >>> In addition to Mike's comments (which I completely agree with), it would >>> be very helfpul to see how this is actually used. e.g, how the >>> regulators are chained together, how the proper ordering is managed, >>> etc. etc. >> We would like to handle voltage scaling in the following way: > What I meant is that a detailed description of the use case should be > included in the changelog. > >> cpufreq_cpu0 >> clk_set_rate(cpu0) >> | >> |-->set_voltage(ABB regulator) /* all ABB related stuff will be >> handled here */ >> | >> |-->set_voltage(smps123 regulator) /* actual voltage >> scaling */ > -EASCII_ART_WRAP > >> This simple model will be extended to handle AVS as a part of the chain. >> smps123 regulator may be changed to VP/VC regulator. >> >> Following example is from integration branch, which already has >> smps123 regulator. > I don't know what integration branch you're referring to, and I don't > know what the smps123 regulator is. > >> It demonstrates an example of linkage to chain. ABB regulator is >> linked with smps123 and cpu0 inside device tree. >> cpu0 calls set_voltage() function for ABB, and then ABB calls >> set_voltage() function for smps123 to do actual voltage scaling. >> >> diff --git a/arch/arm/boot/dts/omap5.dtsi b/arch/arm/boot/dts/omap5.dtsi >> index bb5ee70..c8cbbee 100644 >> --- a/arch/arm/boot/dts/omap5.dtsi >> +++ b/arch/arm/boot/dts/omap5.dtsi >> @@ -36,7 +36,7 @@ >> cpus { >> cpu@0 { >> compatible = "arm,cortex-a15"; >> - cpu0-supply = <&smps123_reg>; >> + cpu0-supply = <&abb_mpu>; >> operating-points = < >> /* kHz uV */ >> /* Only for Nominal Samples */ >> @@ -94,6 +94,7 @@ >> reg = <0x4ae07cdc 0x8>, >> <0x4ae06014 0x4>; >> ti,tranxdone_status_mask = <0x80>; >> + avs-supply = <&smps123_reg>; >> operating-points = < >> /* uV ABB */ >> 880000 0 >> >> This RFC patch series is verified together with: >> https://patchwork.kernel.org/patch/2445091/ >> >> Kevin, what do you think about this model in general? Does it fit to >> regulator framework? > I don't know yet, because I don't think the use case has been described > well enough for me to fully understand it the motiviation behind the > series. > > In addition, there are alternative approaches that seem to have been > ruled out without describing why. For example, the regulator framework > already allows you to override methods with custom hooks (as we do for > VC/VP controlled regulators already.) Without thinking about it too > deeply, it seems this approach could be used to manage the chain of > events you need as well. I can imagine there are limitations to this > approach for what you're trying to do, but I don't feel they have been > described in the changelog as part of the motivation for this series. > > So for now, the guidance I have is this: > > First, write changelogs (and cover letters) assuming your audience has > not been staring at the code as long as you have. Even if they have > been staring at the same code, assume they've been staring at mainline, > and not a random integration branch somewhere. My general advice is to > write changelogs in a way that you will understand what you wrote a year > from now after having forgotten all the details currently in your brains > cache. Even better, write them so that I will understand them in a year > since I forget much better than I remember. > > Second, before inventing something new, start with the existing > framework. When the existing framework doesn't work, make an argument > for your new approach or extentions to the framework based on why the > existing stuff doesn't work. If you don't do this, the reviewers first > reaction will almost always be "why don't you use what already exists in > the framework." And then you'll have a bunch of back and forth with > reviewers when you could've explained the reasoning from the beginning. > > Hope that helps, > > Kevin > > Hi Kevin, Mike Thanks a lot for your comments. "regulator chain" approach was added to ABB series to demonstrate how this approach works in general, and get more comments on this. It was initially introduced in: https://lkml.org/lkml/2013/4/5/33 <https://lkml.org/lkml/2013/4/5/33> - regulator: query on regulator re-entrance http://www.spinics.net/lists/linux-omap/msg90022.html - [RFC v1 0/1] introduce regulator chain locking scheme Taking in account your comments - I'll split ABB logic and "regulator chain" logic, ABB has a possibility to work standalone. In this case DVFS framework will be responsible for calls like *regulator_set_voltage(abb_regulator)*. Could you please exclude "regulator chain" approach from this review and take a look to ABB handling logic? Regards, Andrii -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 04/16/2013 10:07 PM, Mike Turquette wrote: > Quoting Andrii Tseglytskyi (2013-04-16 05:40:44) >> On 04/16/2013 12:53 AM, Kevin Hilman wrote: >>> In addition to Mike's comments (which I completely agree with), it would >>> be very helfpul to see how this is actually used. e.g, how the >>> regulators are chained together, how the proper ordering is managed, >>> etc. etc. >> We would like to handle voltage scaling in the following way: >> > I expanded the example below to include the SR AVS regulator. > >> cpufreq_cpu0 >> clk_set_rate(cpu0) >> | >> |-->set_voltage(ABB regulator) >> | >> |-->set_voltage(AVS) >> | >> |-->set_voltage(smps123 regulator) > Hi Andrii, > > Why was regulator chaining chosen over a simple sequence of calls to > regulator_set_voltage? Instead of nested calls into the regulator > framework, why don't you just make the calls serially? E.g: > > regulator_set_voltage(abb_reg, foo_volt); > regulator_set_voltage(avs_reg, bar_volt); > regulator_set_voltage(smps123_reg, baz_volt); > > It is still to be determined where these calls originate from; maybe > from clock notifiers, maybe directly from the cpufreq driver's .target() > callback, or maybe somewhere else. Regardless, I do not see why > regulator chaining is truly necessary here. You are just calling > regulator_set_voltage in sequence on a few regulators, right? > > I think it would help me a lot to understand why regulator chaining is a > requirement for this to work properly. > > Thanks, > Mike Hi Mike, Kevin I'd like to provide some explanation regarding proposed TI DVFS design which we would like to follow (regulator chaining is part of it). The following two patch series was intended to prove the possibility of its implementation: - this one 'ARM: OMAP3+: Introduce ABB driver" - and http://www.spinics.net/lists/linux-omap/msg90022.html - [RFC v1 0/1] introduce regulator chain locking scheme but, seems, we started to make it public from "tail" instead of 'head" - sorry for that. While trying to move forward with TI DVFS we've taken into account following major points: - only DT should be used to configure DVFS; - no xxx_initcall, cpu_is_xx() function should be used; - DVFS should be scalable to fit wide SoC/platform’s and multiplatform kernel requirements; - minimize creation of TI specific API; Now, there are two entry points for DVFS in kernel: - CPUFreq - currently it's been decided to use cpufreq-cpu0 for all OMAPs in Main line; - CCF callbacks - have RFC DVFS implementation introduced here http://lwn.net/Articles/540422/. In both cases, the only one regulator need to be provided for CCF or CPUFreq for voltage changing proposes, so DVFS can done in the following way: |------------| |------------| --| RegulatorY |<--| DVFS | |------------| |------------| \ \ \ \_____________________________ \ \ |-------------------| |---------------| |---------| --| RegulatorX (PMIC) |--| Regulator AVS |--| ABB LDO |-- |-------------------| |---------------| |---------| /|\ | |______________________| Voltage adjustment 1) The following use-cases have been taken into account: - SoC/Platform don't need ABB/AVS (supports MPU OPPLOW/OPPNOM for example) - any regulator (VC-VP/I2C/SPI/GPIO ..) may be connected to DVFS - SoC/Platform need ABB - build chain in DT device->CCF->abb_regulator->any regulator(VC-VP/I2C/SPI..) - SoC/Platform need ABB+AVS - build chain in DT device->CCF->abb_regulator->avsX->any regulator(VC-VP/I2C..) 2) Implementation of each part of Voltage scale chain as Regulator will allow: - add each item one by one; - don't expose too much of custom TI API; - handle multi-voltage scaling requests to one rail (ganged rails) automatically (handled by regulator FW already). 3) The "regulator chaining" will allow: - easily configure DVFS form DT depending on current SoC/platform needs (using xxx-supply standard binding in DT); - continue use cpufreq-cpu0 for all OMAP to scale MPU domain; - use RFC DVFS implementation from http://lwn.net/Articles/540422/ for other domains (with some modifications - the most difficult thing will be multi-freq requests handling to one clock). In case, if "regulator chaining" approach is not accepted: - yes, it's possible to create some "Super" TI regulator which will handle ABB+AVS+VC/VP for most of current TI SoCs. - No, for the newest TI SoCs (like DRA7xxx without VC/VP) any regulator can be used as power supply (I2C/SPI/GPIO) and ABB is needed. a) As result, it will be impossible to use cpufreq-cpu0 driver for it, at least, and will need to drop cpufreq-cpu0 support for OMAPs and roll-back to omap-cpufreq. b) for other domains it's possible to create omap_dvfs.c in the similar way as it done in dvfs.c and hack it to handle ABB+AVS+I2C regulator. - will need to add custom TI bindings to handle SoC/Platform dependent configuration; Thanks for you time. Regards -grygorii >> >> This simple model will be extended to handle AVS as a part of the chain. >> smps123 regulator may be changed to VP/VC regulator. >> >> Following example is from integration branch, which already has smps123 >> regulator. >> It demonstrates an example of linkage to chain. ABB regulator is linked >> with smps123 and cpu0 inside device tree. >> cpu0 calls set_voltage() function for ABB, and then ABB calls >> set_voltage() function for smps123 to do actual voltage scaling. >> >> diff --git a/arch/arm/boot/dts/omap5.dtsi b/arch/arm/boot/dts/omap5.dtsi >> index bb5ee70..c8cbbee 100644 >> --- a/arch/arm/boot/dts/omap5.dtsi >> +++ b/arch/arm/boot/dts/omap5.dtsi >> @@ -36,7 +36,7 @@ >> cpus { >> cpu@0 { >> compatible = "arm,cortex-a15"; >> - cpu0-supply = <&smps123_reg>; >> + cpu0-supply = <&abb_mpu>; >> operating-points = < >> /* kHz uV */ >> /* Only for Nominal Samples */ >> @@ -94,6 +94,7 @@ >> reg = <0x4ae07cdc 0x8>, >> <0x4ae06014 0x4>; >> ti,tranxdone_status_mask = <0x80>; >> + avs-supply = <&smps123_reg>; >> operating-points = < >> /* uV ABB */ >> 880000 0 >> >> This RFC patch series is verified together with: >> https://patchwork.kernel.org/patch/2445091/ >> >> Kevin, what do you think about this model in general? Does it fit to >> regulator framework? >> >> Thank you. >> >> Regards, >> Andrii > -- > To unsubscribe from this list: send the line "unsubscribe linux-omap" in > the body of a message to majordomo@vger.kernel.org > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/arch/arm/boot/dts/omap5.dtsi b/arch/arm/boot/dts/omap5.dtsi index bb5ee70..c8cbbee 100644 --- a/arch/arm/boot/dts/omap5.dtsi +++ b/arch/arm/boot/dts/omap5.dtsi @@ -36,7 +36,7 @@ cpus { cpu@0 { compatible = "arm,cortex-a15"; - cpu0-supply = <&smps123_reg>; + cpu0-supply = <&abb_mpu>; operating-points = < /* kHz uV */ /* Only for Nominal Samples */ @@ -94,6 +94,7 @@ reg = <0x4ae07cdc 0x8>, <0x4ae06014 0x4>; ti,tranxdone_status_mask = <0x80>; + avs-supply = <&smps123_reg>; operating-points = < /* uV ABB */ 880000 0