Re: [PATCH 1/2] x86/mce: Disable preemption for CPER decoding

[Yazen Ghannam]

On 6/22/2023 3:42 PM, Luck, Tony wrote:
> > 2) Split mce_setup() into global and per_cpu parts. The memset(), cpuid,
> > etc. would be global, and the cpu_data()* and rdmsr() would be per_cpu.
>
> That sounds good. So global is:
>
>          memset(m, 0, sizeof(struct mce));
>          /* need the internal __ version to avoid deadlocks */
>          m->time = __ktime_get_real_seconds();
>          m->cpuvendor = boot_cpu_data.x86_vendor;
>          m->mcgcap = __rdmsr(MSR_IA32_MCG_CAP);

MCG_CAP would be per_cpu, because the bank count can vary. But I don't 
think this matters in practice. So leaving it global is okay, I think.

>          m->microcode = boot_cpu_data.microcode;
>          m->cpuid = cpuid_eax(1);
>
> Though that last one is perhaps per-cpu if you want to allow for mixed-stepping systems.
> Perhaps m->time also? Questionable whether it is useful to log time this record
> was created, when it refers to something much earlier in the BERT case.

I agree about m->time. It doesn't seem useful in this case.

But I don't know about m->cpuid. Mixing processor revisions is not 
allowed on AMD systems, and I don't know about other vendors. So I'd 
leave m->cpuid as global unless there's a strong case otherwise.

> and per-cpu is:
>
>          m->cpu = m->extcpu = smp_processor_id();
>          m->socketid = cpu_data(m->extcpu).phys_proc_id;
>          m->apicid = cpu_data(m->extcpu).initial_apicid;
>          m->ppin = cpu_data(m->extcpu).ppin;
>
> > Option #2 can also be used in apei_mce_report_mem_error(), I think.
>
> Agreed.

Okay, I'll update that too.

Thanks,
Yazen