On 12/08/14 17:26, Jon Medhurst (Tixy) wrote:
On Tue, 2014-08-12 at 14:50 +0100, Chris Redpath wrote:
Frequently in HMP, the big CPUs are only active with one task per CPU and there may be idle CPUs in the big cluster. This patch avoids triggering an idle balance in situations where none of the active CPUs in the current HMP domain have > 1 tasks running.
When packing is enabled, only enforce this behaviour when we are not in the smallest domain - there we idle balance whenever a CPU is over the up_threshold regardless of tasks in case one needs to be moved.
Signed-off-by: Chris Redpath chris.redpath@arm.com
This looks sane to me, though I have one comment about the implementation, see inline comment below.
kernel/sched/fair.c | 27 +++++++++++++++++++++------ 1 file changed, 21 insertions(+), 6 deletions(-)
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 90c8a81..41d0cbd 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -6537,16 +6537,16 @@ static int nohz_test_cpu(int cpu)
- Decide if the tasks on the busy CPUs in the
- littlest domain would benefit from an idle balance
*/ -static int hmp_packing_ilb_needed(int cpu) +static int hmp_packing_ilb_needed(int cpu, int ilb_needed) { struct hmp_domain *hmp;
- /* always allow ilb on non-slowest domain */
- /* allow previous decision on non-slowest domain */ if (!hmp_cpu_is_slowest(cpu))
return 1;
return ilb_needed;/* if disabled, use normal ILB behaviour */ if (!hmp_packing_enabled)
return 1;
return ilb_needed;hmp = hmp_cpu_domain(cpu); for_each_cpu_and(cpu, &hmp->cpus, nohz.idle_cpus_mask) {
@@ -6558,19 +6558,34 @@ static int hmp_packing_ilb_needed(int cpu) } #endif
+DEFINE_PER_CPU(cpumask_var_t, ilb_tmpmask);
- static inline int find_new_ilb(int call_cpu) { int ilb = cpumask_first(nohz.idle_cpus_mask); #ifdef CONFIG_SCHED_HMP
- int ilb_needed = 1;
- int ilb_needed = 0;
- int cpu;
- struct cpumask* tmp = per_cpu(ilb_tmpmask, smp_processor_id());
Why do we need a percpu static variable ilb_tmpmask? It seems to only be used once, here in this function, so could we not instead just have a local stack based temporary variable like:
struct cpumask tmp;
or have I missed something?
We could do that, but this is called during sched tick so I wanted to avoid creating the cpumask on the stack. Do you think that's a reasonable thing to do or do you think it'd be just as quick as calling smp_processor_id()?
The per-cpu variable stems from having a cached mask and not wanting to share it during ticks.
/* restrict nohz balancing to occur in the same hmp domain */ ilb = cpumask_first_and(nohz.idle_cpus_mask, &((struct hmp_domain *)hmp_cpu_domain(call_cpu))->cpus);
- /* check to see if it's necessary within this domain */
- cpumask_andnot(tmp,
&((struct hmp_domain *)hmp_cpu_domain(call_cpu))->cpus,nohz.idle_cpus_mask);- for_each_cpu(cpu, tmp) {
if (cpu_rq(cpu)->nr_running > 1) {ilb_needed = 1;break;}- }
- #ifdef CONFIG_SCHED_HMP_LITTLE_PACKING if (ilb < nr_cpu_ids)
ilb_needed = hmp_packing_ilb_needed(ilb);
ilb_needed = hmp_packing_ilb_needed(ilb, ilb_needed);#endif
if (ilb_needed && ilb < nr_cpu_ids && idle_cpu(ilb))