Re: [PATCH v4 09/17] xen/arm: introduce a helper to parse device tree NUMA distance map

[Jan Beulich <jbeulich@xxxxxxxx>]

On 26.04.2023 07:33, Henry Wang wrote:
>> -----Original Message-----
>> From: Jan Beulich <jbeulich@xxxxxxxx>
>>
>>> +        /* Get opposite way distance */
>>> +        opposite = __node_distance(to, from);
>>> +        /* The default value in node_distance_map is NUMA_NO_DISTANCE
>> */
>>> +        if ( opposite == NUMA_NO_DISTANCE )
>>
>> And the matrix you're reading from can't hold NUMA_NO_DISTANCE entries?
>> I ask because you don't check this above; you only check against
>> NUMA_LOCAL_DISTANCE.
> 
> My understanding for the purpose of this part of code is to check if the 
> opposite
> way distance has already been set, so we need to compare the opposite way
> distance with the default value NUMA_NO_DISTANCE here.
> 
> Back to your question, I can see your point of the question. However I don't 
> think
> NUMA_NO_DISTANCE is a valid value to describe the node distance in the device
> tree. This is because I hunted down the previous discussions and found [2] 
> about
> we should try to keep consistent between the value used in device tree and 
> ACPI
> tables. From the ACPI spec, 0xFF, i.e. NUMA_NO_DISTANCE means unreachable.
> I think this is also the reason why NUMA_NO_DISTANCE can be used as the 
> default
> value of the distance map, otherwise we won't have any value to use.

The [2] link you provided discusses NUMA_LOCAL_DISTANCE. Looking at
Linux'es Documentation/devicetree/numa.txt, there's no mention of an
upper bound on the distance values. It only says that on the diagonal
entries should be 10 (i.e. matching ACPI, without really saying so).

Jan