[PATCH] powerpc/mce: Remove per cpu variables from MCE handlers

[Michael Ellerman]

Mahesh Jagannath Salgaonkar <mahesh at linux.ibm.com> writes:
> On 12/8/20 4:16 PM, Ganesh wrote:
>> 
>> On 12/8/20 4:01 PM, Michael Ellerman wrote:
>>> Ganesh Goudar <ganeshgr at linux.ibm.com> writes:
>>>> diff --git a/arch/powerpc/include/asm/paca.h
>>>> b/arch/powerpc/include/asm/paca.h
>>>> index 9454d29ff4b4..4769954efa7d 100644
>>>> --- a/arch/powerpc/include/asm/paca.h
>>>> +++ b/arch/powerpc/include/asm/paca.h
>>>> @@ -273,6 +274,17 @@ struct paca_struct {
>>>>   #ifdef CONFIG_MMIOWB
>>>>       struct mmiowb_state mmiowb_state;
>>>>   #endif
>>>> +#ifdef CONFIG_PPC_BOOK3S_64
>>>> +    int mce_nest_count;
>>>> +    struct machine_check_event mce_event[MAX_MC_EVT];
>>>> +    /* Queue for delayed MCE events. */
>>>> +    int mce_queue_count;
>>>> +    struct machine_check_event mce_event_queue[MAX_MC_EVT];
>>>> +
>>>> +    /* Queue for delayed MCE UE events. */
>>>> +    int mce_ue_count;
>>>> +    struct machine_check_event  mce_ue_event_queue[MAX_MC_EVT];
>>>> +#endif /* CONFIG_PPC_BOOK3S_64 */
>>>>   } ____cacheline_aligned;
>>> How much does this expand the paca by?
>> 
>> Size of paca is 4480 bytes, these add up another 2160 bytes, so expands
>> it by 48%.
>
> Should we dynamically allocate the array sizes early as similar to that
> of paca->mce_faulty_slbs so that we don't bump up paca size ?

Yeah I think that would be preferable.

That way those allocations can be normal node-local allocations on bare
metal, or when using radix. (Or even on KVM).

In fact what we probably want is a separate struct for all the MCE
related data, eg something like:

struct mce_stuff {
  int nest_count;
  /* Queue for delayed MCE events. */
  int queue_count;
  /* Queue for delayed MCE UE events. */
  int mce_ue_count;

  struct machine_check_event events[MAX_MC_EVT];
  struct machine_check_event event_queue[MAX_MC_EVT];
  struct machine_check_event  ue_event_queue[MAX_MC_EVT];
};

And then you allocate one of those per CPU, inside the RMO for pseries
with hash, and node-local otherwise.

cheers