Re: [PATCH v4 2/3] arm: shmobile: Add the R9A06G032 SMP enabler driver

From: Rob Herring
Date: Fri Jun 08 2018 - 16:41:44 EST


On Fri, Jun 8, 2018 at 2:47 AM, Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:
> On Fri, Jun 8, 2018 at 8:50 AM, Michel Pollet
> <michel.pollet@xxxxxxxxxxxxxx> wrote:
>> On 07 June 2018 16:55, Rob wrote:
>>> On Thu, Jun 7, 2018 at 1:59 AM, Michel Pollet
>>> <michel.pollet@xxxxxxxxxxxxxx> wrote:
>>> > On 06 June 2018 22:53, Frank wrote:
>>> >> On 06/06/18 14:48, Frank Rowand wrote:
>>> >> > On 06/05/18 23:36, Michel Pollet wrote:
>>> >> >> On 05 June 2018 18:34, Frank wrote:
>>> >> >>> On 06/05/18 04:28, Michel Pollet wrote:
>>> >> >>>> The Renesas R9A06G032 second CA7 is parked in a ROM pen at boot
>>> >> >>>> time, it requires a special enable method to get it started.
>>>
>>> [...]
>>>
>>> >> >>>> + * The second CPU is parked in ROM at boot time. It requires
>>> >> >>>> +waking it after
>>> >> >>>> + * writing an address into the BOOTADDR register of sysctrl.
>>> >> >>>> + *
>>> >> >>>> + * So the default value of the "cpu-release-addr" corresponds
>>> >> >>>> +to
>>> >> >>> BOOTADDR...
>>> >> >>>> + *
>>> >> >>>> + * *However* the BOOTADDR register is not available when the
>>> >> >>>> +kernel
>>> >> >>>> + * starts in NONSEC mode.
>>> >> >>>> + *
>>> >> >>>> + * So for NONSEC mode, the bootloader re-parks the second CPU
>>> >> >>>> +into a pen
>>> >> >>>> + * in SRAM, and changes the "cpu-release-addr" of linux's DT to
>>> >> >>>> +a SRAM address,
>>> >> >>>> + * which is not restricted.
>>> >> >>>
>>> >> >>> The binding document for cpu-release-addr does not have a
>>> >> >>> definition for 32 bit arm. The existing definition is only 64
>>> >> >>> bit arm. Please add the definition for 32 bit arm to patch 1.
>>> >> >>
>>> >> >> Hmmm I do find a definition in
>>> >> >> Documentation/devicetree/bindings/arm/cpus.txt -- just under where
>>> >> >> I added my 'enable-method' -- And it is already used as 32 bits in
>>> >> >> at least arch/arm/boot/dts/stih407-family.dtsi.
>>> >> >
>>> >> > If the correct answer is for cpu-release-addr to be 64 bits in
>>> >> > certain cases (that discussion is ongoing further downthread) then
>>> >> > one approach to maintain compatibility _and_ to fix the devicetree
>>> >> > source files is to change the source code that currently gets
>>> >> > cpu-release-addr as a
>>> >> > 32 bit object to check the size of the property and get it as
>>> >> > either a
>>> >> > 32 bit or 64 bit object, based on the actual size of the property
>>> >> > in the device tree and then change the value in the devicetree
>>> >> > source files to be two cells. BUT this does not consider the
>>> >> > bootloader complication. arch/arm/boot/dts/axm5516-cpus.dtsi has a
>>> >> > note "// Fixed by the boot loader", so the boot loader also has to
>>> >> > be modified to be able to handle the possibility that the property
>>> >> > could be either
>>> >> > 32 bits or 64 bits. I don't know how to maintain compatibility
>>> >> > with the boot loader since we can't force it to change
>>> >> > synchronously with changes in the kernel.
>>> >> >
>>> >> > You can consider this comment to be a drive-by observation. I
>>> >> > think Rob and Geert and people like that are likely to be more
>>> >> > helpful with what to actually do, and you can treat my comment more
>>> >> > as pointing out the issue than as providing the perfect solution.
>>> >>
>>> >> Darn it, hit <send> too quickly.
>>> >>
>>> >> I meant to mention that there are several devicetree source files
>>> >> that have a single cell value for cpu-release-addr, and thus
>>> >> potentially face the same situation, depending on what the final
>>> >> decision is on the proper size for cpu- release-addr. As of v4.17, a git grep
>>> shows one cell values in:
>>> >>
>>> >> arch/arm/boot/dts/axm5516-cpus.dtsi
>>> >> arch/arm/boot/dts/stih407-family.dtsi
>>> >> arch/arm/boot/dts/stih418.dtsi
>>> >
>>> > Yes, I had grepped before I used 32 bits on mine...
>>> >
>>> > Now, what is the decision here? Our bootloader is already modified to
>>> > set it to 32 bits, so I propose that
>>>
>>> And too late to fix the bootloader?
>>
>>
>> Well not too late, but read further on...
>>
>>>
>>> >
>>> > + I change the driver to handle 32 and 64 bits properties
>>>
>>> That's fine if you can't fix the bootloader.
>>>
>>> > + I add this to the cpu.txt, as a separate patch:
>>> > # On other systems, the property can be either
>>> > 32 bits or 64 bits, it is the driver's responsibility
>>> > to deal with either sizes.
>>>
>>> That is definitely not what we want to say. Use of 32-bit should be
>>> considered out of spec. Yes, we have a few platforms in that category, but
>>> they already handle that themselves. Would be nice to fix them, but at least
>>> the STi platforms don't seem too active.
>>>
>>> IMO, we should delete whatever text we can here and at most just refer to
>>> the spec.
>>
>> So actually I didn't use 32 bits by plain chance, I read the cpu.txt file which says
>> that 64 bits systems use 64 bits property, concluded that in my case I ought to
>> use 32 bits, then grepped around and found other systems using 32 bits, therefore
>> I went forward and used it..
>>
>> Nothing said here that it should be 64 bits everywhere -- So the documentation
>> needs fixing somehow. Right now it certainly led me wrong.
>
> Perhaps we should add to Documentation/devicetree/bindings/ the standard
> bindings from ePAPR and successors, too?

I hope you mean *reference* here, not duplicate the bindings here. We
want to move in the other direction and move the common bindings out
of the kernel and into the spec.

The real solution here is validation which I'm working on. I had
already converted cpus.txt. Here's an example of the results of the
validation:

arch/arm/boot/dts/stih410-b2120.dt.yaml:1962:7: cpu@0: 'enable-method'
is a dependency of 'cpu-release-addr'
arch/arm/boot/dts/stih410-b2120.dt.yaml:1965:26:
cpu@0:cpu-release-addr: [155254948] is too short

The schema is up on my yaml-bindings branch. Will send out more details soon.

Rob