[PATCH 1/1] riscv: default to CONFIG_RISCV_SBI_V01=n

[Palmer Dabbelt]

On Fri, 17 Dec 2021 04:09:22 PST (-0800), anup at brainfault.org wrote:
> +Atish
>
> On Fri, Dec 17, 2021 at 4:12 PM Heinrich Schuchardt
> <heinrich.schuchardt at canonical.com> wrote:
>>
>>
>>
>> On 12/16/21 17:51, Jessica Clarke wrote:
>> > On 16 Dec 2021, at 14:17, Heinrich Schuchardt <heinrich.schuchardt at canonical.com> wrote:
>> >>
>> >> On 12/16/21 14:49, Jessica Clarke wrote:
>> >>> On 16 Dec 2021, at 12:35, Heinrich Schuchardt <heinrich.schuchardt at canonical.com> wrote:
>> >>>>
>> >>>> The SBI 0.1 specification is obsolete. The current version is 0.3.
>> >>>> Hence we should not rely by default on SBI 0.1 being implemented.
>> >>> It’s what BBL implements, and some people are still using it,
>> >>> especially given early hardware shipped before OpenSBI grew in
>> >>> popularity.
>> >>> Jess
>> >>
>> >> Do you mean BBL is not developed anymore?
>> >>
>> >> Some people may still be using a 0.1 SBI. But that minority stuck on an outdated software stack does not justify defaulting to deprecated settings in future Linux releases.
>> >
>> > BBL is still actively maintained; its most recent commit was 24 days
>> > ago. Also, the amount of code CONFIG_RISCV_SBI_V01 affects is tiny, so
>> > I see no tangible benefit from making this change, just unnecessary
>> > breakage of perfectly functional systems.
>>
>> Only the default is changed. How could this break any existing system?
>> You can still compile with the deprecated setting.
>>
>> I can not see why we should keep a default that will cause issues on
>> systems complying to the current SBI specification.
>
> I agree with Heinrich.
>
> Almost all SBI implementations (OpenSBI, EDK2, KVM, Xvisor, etc) are
> providing at least SBI v0.2 and we can't endlessly wait for BBL to move
> away from SBI v0.1. We can't totally remove SBI v0.1 but we should
> at least disable it by default.
>
> Same rationale applies to the spinwait CPU operations which are only
> required for systems using BBL. The sparse HART id series from Atish
> can include a patch to have spinwait CPU operations disabled by
> default.
>
> @Atish/Palmer, what do you think ?

I agree with Heinrich: it's just a defconfig change, we're not dropping 
support for SBI-0.1.  The defconfig is generally meant to reflect what 
kernel developers are using, and AFAIK everyone doing upstream 
development moved off the legacy SBI implementations as of a while ago.  
There's some cost associated with supporting the old SBI, it's not like 
this is just some backup calls we're adding.

If vendors are still shipping old firmwares as part of their SDKs that's 
fine, they'll just need to flip on support for it in their configs.  Who 
knows, maybe this will even be the nudge they need to stop shipping 
ancient firmware ;)