[PATCH] ARM: move firmware_ops to drivers/firmware

[Alex Courbot]

On 11/18/2013 08:58 PM, Catalin Marinas wrote:
> On Mon, Nov 18, 2013 at 03:05:59AM +0000, Alex Courbot wrote:
>> On 11/18/2013 12:59 AM, Catalin Marinas wrote:
>>> On 17 November 2013 08:49, Alexandre Courbot <acourbot at nvidia.com> wrote:
>>>> The ARM tree includes a firmware_ops interface that is designed to
>>>> implement support for simple, TrustZone-based firmwares but could
>>>> also cover other use-cases. It has been suggested that this
>>>> interface might be useful to other architectures (e.g. arm64) and
>>>> that it should be moved out of arch/arm.
>>>
>>> NAK. I'm for code sharing with arm via common locations but this API
>>> goes against the ARMv8 firmware standardisation efforts like PSCI,
>>> encouraging each platform to define there own non-standard interface.
>>
>> I have to say, I pretty much agree with your NAK.
>>
>> The reason for this patch is that the suggestion to move firmware_ops
>> out of arch/arm is the last (I hope) thing that prevents my Trusted
>> Foundation support series from being merged.
>
> Moving it into drivers shouldn't be a workaround. Nice try ;).

Hehe. I thought that just sending a patch would settle the issue one way 
or the other and avoid a huge discussion. Woke up this morning to see 
how wrong I was.

>
>> Now if we can all agree:
>>
>> * that ARMv8 will only use PSCI
>
> Or spin-table (which does not require secure calls). Otherwise, if
> secure firmware is present, SoCs should use PSCI (as the only firmware
> standard currently supported in the arm64 kernel).
>
> However, things evolve and we may have other needs in the future or PSCI
> may not be sufficient or we get newer PSCI revisions. This can be
> extended but my requirement is to decouple booting standard from SoC
> support (together with the aim of having no SoC-specific code under
> arch/arm64). I really don't see why SoCs can't agree on one (or very
> few) standard booting protocol (and legacy argument doesn't work since
> the ARMv8 firmware needs to be converted to AArch64 anyway).
>
>> * that there is no use-case of this interface outside of arch/arm as of
>> today (and none foreseen in the near future)
>
> The firmware_ops are only used under arch/arm so far, I don't see any
> drivers doing anything with it. Also, l2x0_init is ARMv7 only.
>
> On arm64, support for PSCI is handled via cpu_operations in the latest
> kernel. That's an arm64 abstraction and is extensible (but we want to
> keep tight control of this, hence no register_cpu_ops function).
>
>> * that the firmware_ops interface is quite ARMv7-specific anyway,
>
> This was introduced to allow SoC code to enable hooks for SoC-specific
> firmware calls like cpu_idle, l2x0_init. By standardising the interface
> and decoupling it from SoC code on arm64, we don't need per-SoC
> firmware_ops.
>
> Of course, trusted foundations interface could be plugged into cpu_ops
> on arm64 but I will NAK it on the grounds of not using the PSCI API, nor
> the SMC calling convention (and it's easy to fix when porting to ARMv8).
> If a supported standard API is used, then there is no need for
> additional code in the kernel.
>
> BTW, is legacy code the reason for not converting the SMC # to PSCI?
> It's already supported on ARMv7, so you may not have much code left to
> merge in the kernel ;).

The problem here is twofold:

1) we are just consumers of the TrustZone secure monitor who receive a 
binary and do not have any control over its calling conventions. I agree 
that it would be trivial to make it compatible with PSCI, but it's just 
not something we can make by ourselves (TF does not even follow the SMC 
calling convention). If this problem is to be addressed, it should be 
done by forcing the TrustZone secure monitors providers to follow PSCI.

2) devices have already shipped with this firmware. Are we going to just 
renounce supporting them, even though the necessary support is 
lightweight and fits within already existing interfaces?

I certainly do hope that for ARMv8 things will be different and more 
standardized. But that's not something that can be guaranteed unless ARM 
strongly enforces it to firmware vendors. In case such a non-standard 
firmware gets used again, I *do* hope that using cpu_ops will be 
preferred over saying "this device cannot be supported in mainline, ever".

The kernel already supports non-standard hardware, BIOS, ACPI through 
hacks that are *way* more horrible than that. This should certainly not 
be encouraged, but that's not a valid reason to forbid otherwise 
perfectly fine devices to run mainline IMHO.

>
>> * that should a need to move it (for whatever reason) occur later, it
>> will be easy to do (as this patch hopefully demonstrates).
>
> I agree, it's not hard to unify this but so far I haven't seen a good
> reason.

Same here. arm64 has its own cpu_operations. Other archs have different 
needs and if we move this to a common place it will just become a messy 
placeholder for function pointers from which each arch will only use a 
subset.

Not to mention that if we follow the logic completely, we should then 
implement PCSI on top of cpu_ops and cpu_ops on top of firmware_ops (or 
the other way around ;)).

Alex.