On 14/03/2025 12:33, Suzuki K Poulose wrote:
> On 14/03/2025 09:18, Ryan Roberts wrote:
>> On 13/03/2025 18:36, Marc Zyngier wrote:
>>> On Thu, 13 Mar 2025 18:22:00 +0000,
>>> Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
>>>>
>>>> On 13/03/2025 17:34, Marc Zyngier wrote:
>>>>> On Thu, 13 Mar 2025 10:41:10 +0000,
>>>>> Mikołaj Lenczewski <miko.lenczewski@xxxxxxx> wrote:
>>>>>>
>>>>>> diff --git a/arch/arm64/kernel/pi/idreg-override.c b/arch/arm64/kernel/pi/
>>>>>> idreg-override.c
>>>>>> index c6b185b885f7..9728faa10390 100644
>>>>>> --- a/arch/arm64/kernel/pi/idreg-override.c
>>>>>> +++ b/arch/arm64/kernel/pi/idreg-override.c
>>>>>> @@ -209,6 +209,7 @@ static const struct ftr_set_desc sw_features
>>>>>> __prel64_initconst = {
>>>>>> FIELD("nokaslr", ARM64_SW_FEATURE_OVERRIDE_NOKASLR, NULL),
>>>>>> FIELD("hvhe", ARM64_SW_FEATURE_OVERRIDE_HVHE, hvhe_filter),
>>>>>> FIELD("rodataoff", ARM64_SW_FEATURE_OVERRIDE_RODATA_OFF, NULL),
>>>>>> + FIELD("nobbml2", ARM64_SW_FEATURE_OVERRIDE_NOBBML2, NULL),
>>>>>> {}
>>>>>> },
>>>>>> };
>>>>>> @@ -246,6 +247,7 @@ static const struct {
>>>>>> { "rodata=off", "arm64_sw.rodataoff=1" },
>>>>>> { "arm64.nolva", "id_aa64mmfr2.varange=0" },
>>>>>> { "arm64.no32bit_el0", "id_aa64pfr0.el0=1" },
>>>>>> + { "arm64.nobbml2", "arm64_sw.nobbml2=1" },
>>>>>
>>>>> Why is that a SW feature? This looks very much like a HW feature to
>>>>> me, and you should instead mask out ID_AA64MMFR2_EL1.BBM, and be done
>>>>> with it. Something like:
>>>>
>>>> I think this implies that we would expect the BBM field to be advertising BBML2
>>>> support normally and we would check for that as part of the cpufeature
>>>> detection. That's how Miko was doing it in v2, but Yang pointed out that
>>>> AmpereOne, which supports BBML2+NOABORT semantics, doesn't actually advertise
>>>> BBML2 in its MMFR2. So we don't want to check that field, and instead rely
>>>> solely on the MIDR allow-list + a command line override. It was me that
>>>> suggested putting that in the SW feature register, and I think that still
>>>> sounds
>>>> like the right solution for this situation?
>>>
>>> I think this is mixing two different things:
>>>
>>> - preventing BBM-L2 from being visible to the kernel: this is what my
>>> suggestion is doing by nuking an architectural feature in the
>>> relevant register
>>>
>>> - random HW not correctly advertising what they are doing: this is an
>>> erratum workaround
>>>
>>> I'd rather we don't conflate the two things, and make them very
>>> explicitly distinct.
>>
>> It all sounds so obvious when you put it like that! :)
>>
>> I'm guessing there is a layer where the workaround can be applied to the
>> sanitised feature registers on a per-cpu basis and that won't affect this global
>> override which will remain as an overlay on top? If so then that sounds perfect
>> (you can probably tell I find the whole feature management framework rather
>> inpeneterable). That workaround would be added as part of Yang's series anyway.
>
> Unfortunately, there is no easy way to fix this via the normal erratum
> workaround "capability". The sanitised feature registers are handled
> separately (initialised via init_cpu_features() for boot CPU and
> sanitised eachtime via update_cpu_features).
>
> Also we do not "enable" any capability (i.e. calling cpu_enable())
> until the very end, after the CPUs are all brought up (except for boot CPUs).
>
> But it may be possible to "fix up" the BBML2 feature in
> cpuinfo_store_*cpu(), without using the "enable" call back.
>
> Something like:
>
> diff --git a/arch/arm64/kernel/cpuinfo.c b/arch/arm64/kernel/cpuinfo.c
> index 285d7d538342..8c23adbe29f8 100644
> --- a/arch/arm64/kernel/cpuinfo.c
> +++ b/arch/arm64/kernel/cpuinfo.c
> @@ -479,7 +479,11 @@ static void __cpuinfo_store_cpu(struct cpuinfo_arm64 *info)
> info->reg_id_aa64mmfr0 = read_cpuid(ID_AA64MMFR0_EL1);
> info->reg_id_aa64mmfr1 = read_cpuid(ID_AA64MMFR1_EL1);
> info->reg_id_aa64mmfr2 = read_cpuid(ID_AA64MMFR2_EL1);
>
> + /*
> + * The CPU cap is not detected system wide, but we are able to
> + * check if this CPU is affected by the Erratum.
> + */
> + if (this_cpu_has_cap(AMPERE_ONE_ERRATUM_BBML2))
> + // Fixup info->reg_id_aa64_mmfr2 with BBML2.
> +
> info->reg_id_aa64mmfr3 = read_cpuid(ID_AA64MMFR3_EL1);
> info->reg_id_aa64mmfr4 = read_cpuid(ID_AA64MMFR4_EL1);
> info->reg_id_aa64pfr0 = read_cpuid(ID_AA64PFR0_EL1);
> info->reg_id_aa64pfr1 = read_cpuid(ID_AA64PFR1_EL1);
>
This is the type of thing I was imagining, thanks!
> Suzuki
>
>>
>> So sounds like we are back to testing MMFR2.BBM in the matches function, with
>> the addition of Maz's proposal above. Sorry for sending you round the houses,
>> Miko.
>>
>> Thanks,
>> Ryan
>>
>>>
>>> Thanks,
>>>
>>> M.
>>>
>>
>
