On 26/04/23 13:51, Yury Norov wrote:
>> I realized I only wrote half the relevant code - comparing node IDs is
>> meaningless, I meant to compare distances as we walk through the
>> CPUs... I tested the below against a few NUMA topologies and it seems to be
>> sane:
>>
>> @@ -756,12 +773,23 @@ static void __init test_for_each_numa(void)
>> {
>> unsigned int cpu, node;
>>
>> - for (node = 0; node < sched_domains_numa_levels; node++) {
>> - unsigned int hop, c = 0;
>> + for_each_node(node) {
>> + unsigned int start_cpu, prev_dist, hop = 0;
>> +
>> + cpu = cpumask_first(cpumask_of_node(node));
>> + prev_dist = node_distance(node, node);
>> + start_cpu = cpu;
>>
>> rcu_read_lock();
>> - for_each_numa_cpu(cpu, hop, node, cpu_online_mask)
>> - expect_eq_uint(cpumask_local_spread(c++, node), cpu);
>> +
>> + /* Assert distance is monotonically increasing */
>> + for_each_numa_cpu(cpu, hop, node, cpu_online_mask) {
>> + unsigned int dist = node_distance(cpu_to_node(cpu), cpu_to_node(start_cpu));
>
> Interestingly, node_distance() is an arch-specific function. Generic
> implementation is quite useless:
>
> #define node_distance(from,to) ((from) == (to) ? LOCAL_DISTANCE : REMOTE_DISTANCE)
>
> Particularly, arm64 takes the above. With node_distance() implemented
> like that, we can barely test something...
>
riscv and arm64 rely on drivers/base/arch_numa.c to provide
__node_distance() (cf. CONFIG_GENERIC_ARCH_NUMA).
x86, sparc, powerpc and ia64 define __node_distance()
loongarch and mips define their own node_distance().
So all of those archs will have a usable node_distance(), the others won't
and that means the scheduler can't do anything about it - the scheduler
relies on node_distance() to understand the topolgoy!
