[load-balancer] migrate fn, handle affine change.
Handle the case, exercised in the set_migrate_fn_test, where the
affinity of a thread changes when a migrate function is set. When
determining if we could stay on the old cpu until delayed migration
happened, in the old path we compared against available_mask in the
new path we also included affinity in that decision which was incorrect.
Also we should be comparing against last_cpu, not initial_cpu. It is not
correct to schedule a migration on a new thread so we add in the bypass
if the last_cpu is unset, which is the new thread case.
This brings us inline with scheduler.cc:650 which is the main flow.
Test: Ran k ut with the load-balancer enabled.
Change-Id: Ia83f47412b3bca861abfb2a4d19c8796a57bd5bd
Reviewed-on: https://fuchsia-review.googlesource.com/c/fuchsia/+/405375
Reviewed-by: Nick Maniscalco <maniscalco@google.com>
Testability-Review: Nick Maniscalco <maniscalco@google.com>
Commit-Queue: Ed Coyne <edcoyne@google.com>
diff --git a/zircon/kernel/lib/load_balancer/include/lib/load_balancer_percpu.h b/zircon/kernel/lib/load_balancer/include/lib/load_balancer_percpu.h
index 0c34b49..365e1a9 100644
--- a/zircon/kernel/lib/load_balancer/include/lib/load_balancer_percpu.h
+++ b/zircon/kernel/lib/load_balancer/include/lib/load_balancer_percpu.h
@@ -159,10 +159,11 @@
return initial_cpu;
}
- if (unlikely(thread->has_migrate_fn() && initial_cpu_available)) {
+ if (unlikely(last_cpu != INVALID_CPU && last_cpu != lowest_cpu &&
+ thread->has_migrate_fn() && (mp_get_active_mask() & ToMask(last_cpu)))) {
// Stay where we are, the migrate_fn_ will migrate us later.
thread->scheduler_state().set_next_cpu(lowest_cpu);
- return initial_cpu;
+ return last_cpu;
}
return lowest_cpu;
