[PATCH] sched: fix newly idle load balance in case of SMT

In the presence of SMT, newly idle balance was never happening for
multi-core and SMP domains (even when both the logical siblings are
idle).

If thread 0 is already idle and when thread 1 is about to go to idle,
newly idle load balance always think that one of the threads is not idle
and skips doing the newly idle load balance for multi-core and SMP
domains.

This is because of the idle_cpu() macro, which checks if the current
process on a cpu is an idle process. But this is not the case for the
thread doing the load_balance_newidle().

Fix this by using runqueue's nr_running field instead of idle_cpu(). And
also skip the logic of 'only one idle cpu in the group will be doing
load balancing' during newly idle case.

Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

diff --git a/kernel/sched.c b/kernel/sched.c
index 645256b..e36d99d 100644 (file)
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -2235,7 +2235,7 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
 
                        rq = cpu_rq(i);
 
-                       if (*sd_idle && !idle_cpu(i))
+                       if (*sd_idle && rq->nr_running)
                                *sd_idle = 0;
 
                        /* Bias balancing toward cpus of our domain */
@@ -2257,9 +2257,11 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
                /*
                 * First idle cpu or the first cpu(busiest) in this sched group
                 * is eligible for doing load balancing at this and above
-                * domains.
+                * domains. In the newly idle case, we will allow all the cpu's
+                * to do the newly idle load balance.
                 */
-               if (local_group && balance_cpu != this_cpu && balance) {
+               if (idle != CPU_NEWLY_IDLE && local_group &&
+                   balance_cpu != this_cpu && balance) {
                        *balance = 0;
                        goto ret;
                }