[Linux-stm32] [PATCH v2 00/14] Introduce STM32MP1 RCC in secured mode

[Alex G.]

On 3/11/21 12:10 PM, Alexandre TORGUE wrote:
> Hi Guys
> 
> On 3/11/21 5:11 PM, Marek Vasut wrote:
>> On 3/11/21 3:41 PM, Ahmad Fatoum wrote:
>>> Hello,
>>
>> Hi,
>>
>>> On 11.03.21 15:02, Alexandre TORGUE wrote:
>>>> On 3/11/21 12:43 PM, Marek Vasut wrote:
>>>>> On 3/11/21 9:08 AM, Alexandre TORGUE wrote:
>>>>>> 1- Break the current ABI: as soon as those patches are merged, 
>>>>>> stm32mp157c-dk2.dtb will impose to use
>>>>>> A tf-a for scmi clocks. For people using u-boot spl, the will have 
>>>>>> to create their own "no-secure" devicetree.
>>>>>
>>>>> NAK, this breaks existing boards and existing setups, e.g. DK2 that 
>>>>> does not use ATF.
>>>>>
>>>>>> 2-As you suggest, create a new "secure" dtb per boards (Not my 
>>>>>> wish for maintenance perspectives).
>>>>>
>>>>> I agree with Alex (G) that the "secure" option should be opt-in.
>>>>> That way existing setups remain working and no extra requirements 
>>>>> are imposed on MP1 users. Esp. since as far as I understand this, 
>>>>> the "secure" part isn't really about security, but rather about 
>>>>> moving clock configuration from Linux to some firmware blob.
>>>>>
>>>>>> 3- Keep kernel device tree as they are and applied this secure 
>>>>>> layer (scmi clocks phandle) thanks to dtbo in
>>>>>> U-boot.
>>>>>
>>>>> Is this really better than
>>>>> #include "stm32mp15xx-enable-secure-stuff.dtsi"
>>>>> in a board DT ? Because that is how I imagine the opt-in "secure" 
>>>>> option could work.
>>>>>
>>>>
>>>> Discussing with Patrick about u-boot, we could use dtbo application 
>>>> thanks to extlinux.conf. BUT it it will not prevent other case (i.e. 
>>>> TF-A which jump directly in kernel@). So the "least worst" solution 
>>>> is to create a new "stm32mp1257c-scmi-dk2 board which will overload 
>>>> clock entries with a scmi phandle (as proposed by Alex).
>>>
>>> I raised this issue before with your colleagues. I still believe the 
>>> correct way
>>> would be for the TF-A to pass down either a device tree or an overlay 
>>> with the
>>> actual settings in use, e.g.:
>>>
>>>    - Clocks/Resets done via SCMI
>>>    - Reserved memory regions
>>>
>>> If TF-A directly boots Linux, it can apply the overlay itself, 
>>> otherwise it's
>>> passed down to SSBL that applies it before booting Linux.
>>
>> That sounds good and it is something e.g. R-Car already does, it 
>> merges DT fragment from prior stages at U-Boot level and then passes 
>> the result to Linux.
>>
>> So on ST hardware, the same could very well happen and it would work 
>> for both non-ATF / ATF / ATF+TEE options.
> 
> Even this solution sounds good but we are currently not able to do it in 
> our TF-A/u-boot so not feasible for the moment. So we have to find a 
> solution for now. Create a new dtb can be this solution. Our internal 
> strategy is to use scmi on our official ST board. It will be a really 
> drawback to include a "no-scmi.dtsi" in DH boards (for example) and to 
> create a stm32mp157c-noscmi-dk2.dts ?

It could work, as long as all users are reminded to change their build 
scripts to pick up a "-noscmi.dtb". I suspect that if this were the case 
we'll see quite a few bug reports saying "stm32mp1 no longer boots with 
kernel v5.13".

I didn't think of this originally, though u-boot already does the DTB 
patching for OPTEE reserved memory regions. It's not too hard to also 
patch in the SCMI clocks at boot. In u-boot's case, runtime detection 
might even be feasible.

Alex