It turns out schedule_setaffinity(2) only sets the current *thread*'s
affinity, not the entire process. Look in /proc to identify all of our
children and map those, too.
Note that this doesn't seem to affect the numa_preferred_nid value in
/proc/$pid/[task/$tid/]sched... at least not immediately. That value does
seem to change for some threads after some load is applied.
Fixes: https://tracker.ceph.com/issues/42054 Signed-off-by: Sage Weil <sage@redhat.com>