We have been seeing some failures in CKI project, where the kernels on some virtual machines errored out on attempting to do taskset on non-existing CPUs.

The root cause is that fallback topology is used on those nodes, since some /proc or /sysfs data made little sense. Fallback topology faked some (consecutive) CPU numbers, and then jcstress used taskset to hook forked JVMs to it. But that is not guaranteed to work: the CPU numbers are not guaranteed to be consecutive.

When using fallback topology, we should force NONE affinity mode.