Re: [PATCH v2 1/2] ACPI/PPTT: Add support for ACPI 6.3 thread flag

[Jeremy Linton <jeremy.linton@arm.com>]

Hi,

On 6/19/19 4:15 AM, John Garry wrote:
> On 18/06/2019 22:28, Jeremy Linton wrote:
>> Hi,
>>
>> On 6/18/19 12:23 PM, John Garry wrote:
>>> On 18/06/2019 15:40, Valentin Schneider wrote:
>>>> On 18/06/2019 15:21, Jeremy Linton wrote:
>>>> [...]
>>>>>>> + * Return: -ENOENT if the PPTT doesn't exist, the CPU cannot be
>>>>>>> found or
>>>>>>> + *       the table revision isn't new enough.
>>>>>>> + * Otherwise returns flag value
>>>>>>> + */
>>>>>>
>>>>>> Nit: strictly speaking we're not returning the flag value but its 
>>>>>> mask
>>>>>> applied to the flags field. I don't think anyone will care about
>>>>>> getting
>>>>>> the actual flag value, but it should be made obvious in the doc:
>>>>>
>>>>> Or I clarify the code to actually do what the comments says. Maybe
>>>>> that is what John G was also pointing out too?
>>>>>
>>>
>>> No, I was just saying that the kernel topology can be broken without
>>> this series.
>>>
>>>>
>>>> Mmm I didn't find any reply from John regarding this in v1, but I
>>>> wouldn't
>>>> mind either way, as long as the doc & code are aligned.
>>>>
>>>
>>> BTW, to me, function acpi_pptt_cpu_is_thread() seems to try to do too
>>> much, i.e. check if the PPTT is new enough to support the thread flag
>>> and also check if it is set for a specific cpu. I'd consider separate
>>> functions here.
>>
> 
> Hi,
> 
>> ? Your suggesting replacing the
>>
> 
> I am not saying definitely that this should be changed, it's just that 
> acpi_pptt_cpu_is_thread() returning false, true, or "no entry" is not a 
> typical API format.
> 
> How about acpi_pptt_support_thread_info(cpu) and 
> acpi_pptt_cpu_is_threaded(cpu), both returning false/true only?

I'm not sure we want to be exporting what is effectively a version check 
into the rest of the code. Plus, AFAIK it doesn't really simplify 
anything except the case of ACPI machines with revision 1 PPTTs, because 
those would only be doing a single check and assuming the state of the 
MT bit. That MT check is suspect anyway, although AFAIK it gets the 
right answer on all machines that predate ACPI 6.3..


> 
> None of this is ideal.
> 
> BTW, Have you audited which arm64 systems have MT bit set legitimately?

Not formally, given I don't have access to everything available.

> 
>>
>> if (table->revision >= rev)
> 
> I know that checking the table revision is not on the fast path, but it 
> seems unnecessarily inefficient to always read it this way, I mean 
> calling acpi_table_get().
> 
> Can you have a static value for the table revision? Or is this just how 
> other table info is accessed in ACPI code?

Yes caching the revision internally would save a get/put per core, for 
older machines. I don't think its a big deal in normal operation but its 
a couple extra lines so...

I will post it with an internally cached saved_pptt_rev. That will save 
CPU count get/puts in the case where the revision isn't new enough.


> 
>>     cpu_node = acpi_find_processor_node(table, acpi_cpu_id);
>>
>> check with
>>
>> if (revision_check(table, rev))
>>     cpu_node = acpi_find_processor_node(table, acpi_cpu_id);
>>
>>
>> and a function like
>>
>> static int revision_check(acpixxxx *table, int rev)
>> {
>>     return (table->revision >= rev);
>> }
>>
>> Although, frankly if one were to do this, it should probably be a macro
>> with the table type, and used in the dozen or so other places I found
>> doing similar checks (spcr, iort, etc).
>>
>> Or something else?
>>
>>
>>
>>
> 
> thanks,
> John