On Sun, Sep 09, 2018 at 08:29:29PM -0700, Huang, Kai wrote:
> > + */
> > +static int mktme_build_cpumask(void)
> > +{
> > + int online_cpu, mktme_cpu;
> > + int online_pkgid, mktme_pkgid = -1;
> > +
> > + if (!zalloc_cpumask_var(&mktme_cpumask, GFP_KERNEL))
> > + return -ENOMEM;
> > +
> > + for_each_online_cpu(online_cpu) {
> > + online_pkgid = topology_physical_package_id(online_cpu);
> > +
> > + for_each_cpu(mktme_cpu, mktme_cpumask) {
> > + mktme_pkgid =
> > topology_physical_package_id(mktme_cpu);
> > + if (mktme_pkgid == online_pkgid)
> > + break;
> > + }
> > + if (mktme_pkgid != online_pkgid)
> > + cpumask_set_cpu(online_cpu, mktme_cpumask);
> > + }
>
> Could we use 'for_each_online_node', 'cpumask_first/next', etc to simplify the logic?
Kai,
I tried to simplify it and came up with code that looked like this:
int lead_cpu, node;
for_each_online_node(node) {
lead_cpu = cpumask_first(cpumask_of_node(node));
if (lead_cpu < nr_cpu_ids)
cpumask_set_cpu(lead_cpu, mktme_cpumask_NEW);
}
When I test it on an SNC (Sub Numa Cluster) system it gives me too many
CPU's. I get a CPU per Node (just like i asked for;) instead of per Socket.
It has 2 sockets and 4 NUMA nodes.
I kind of remember this when I originally coded it, hence the bottoms up
approach using topology_physical_package_id()
Any ideas?
Alison
