Re: [PATCH v3 05/12] arm64, acpi, numa: NUMA support based on SRAT and SLIT

[Hanjun Guo]

On 2016/1/27 22:01, Robert Richter wrote:
> On 27.01.16 15:12:15, Hanjun Guo wrote:
>> On 2016/1/25 18:21, Robert Richter wrote:
>>> On 23.01.16 17:39:20, Hanjun Guo wrote:
>>>> diff --git a/arch/arm64/kernel/acpi_numa.c b/arch/arm64/kernel/acpi_numa.c
>>>> new file mode 100644
>>>> index 0000000..f7f7533
>>>> --- /dev/null
>>>> +++ b/arch/arm64/kernel/acpi_numa.c
>>>> +/* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
>>>> +int __init acpi_numa_memory_affinity_init(struct acpi_srat_mem_affinity *ma)
>>>> +{
>>>> +	u64 start, end;
>>>> +	int node, pxm;
>>>> +
>>>> +	if (srat_disabled())
>>>> +		return -EINVAL;
>>>> +
>>>> +	if (ma->header.length != sizeof(struct acpi_srat_mem_affinity)) {
>>> Must be:
>>>
>>> 	ma->header.length < sizeof(struct acpi_srat_mem_affinity)) {
>>>
>>> Allow extensions to struct acpi_srat_mem_affinity in newer versions.
>> Hmm, I think we need to remove the check here now.
> No, we might have an out-of-bound access then.
>
>> There are three cases:
>>
>>  - firmware ACPI version is consistent with the ACPICA one, then
>>    ma->header.length == sizeof(struct acpi_srat_mem_affinity )
>>
>>  - firmware ACPI version is not consistent with the ACPICA one,
>>    for example, struct acpi_srat_mem_affinity is extended in
>>    new ACI version, but the formware is using the older one,
>>   then it's ok to use
>>   ma->header.length < sizeof(struct acpi_srat_mem_affinity )
> The check above is ok as we need at least struct
> acpi_srat_mem_affinity as it is now.
>
> If we later change the kernel to support multiple versions of struct
> acpi_srat_mem_affinity, i.e. use data from an extended section, we
> will need to add code to handle that. This will include support of
> data with length < acpi_srat_mem_affinity, in this case we may not use
> extended data.

I checked the ACPI spec about memory affinity structure, it still have 10 bytes
reserved for future use, so I think it's safe as you suggested for next few years.

>
>>  - but if we use the older kernel + updated new firmware,
>>    then
>>    ma->header.length > sizeof(struct acpi_srat_mem_affinity )
>>   will be the case, right?
> Right, and this is a valid case not resulting in an error with my
> suggestion above.

Yes, I just mixed up those two cases.

I will sync with Ganapat to prepare a new version and test it on x86 and
IA64 to make sure this patch set don't break anything.

Thanks
Hanjun