In the current implementation, even if the task calling execl is bound
to a single CPU (or not allowed to be migrated), it still invokes the
select_task_rq callback to select a CPU. This is unnecessary and
wastes cycles.
Since select_task_rq() already includes checks for the above scenarios
(e.g., tasks bound to a single CPU or forbidden to migrate) and skips
the select_task_rq callback in such cases, we can directly use
select_task_rq() instead of invoking the callback here.
Test environment: 256-CPU X86 server
Test method: Run unixbench's execl test with task bound to a single CPU:
$ numactl -C 10 ./Run execl -c 1
Test results: Average of 5 runs
baseline patched improvement
383.82 436.78 +13.8%
Change Log:
v1->v2
As suggested by Peter, replace manual corner-case checks with
select_task_rq() to align with existing logic.
Additional testing on a 256-CPU server which all sched domains have
SD_BALANCE_EXEC flag, shows that sched_exec now searches all CPUs in the
system (previously, some SD_NUMA sched domains lacked SD_BALANCE_EXEC).
This increased the performance improvement to 13.8%.
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Co-developed-by: Yibin Liu <liuyibin@hygon.cn>
Signed-off-by: Yibin Liu <liuyibin@hygon.cn>
Signed-off-by: Jianyong Wu <wujianyong@hygon.cn>
---
kernel/sched/core.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index f754a60de848..6e4ba3c27e5c 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -5439,10 +5439,11 @@ void sched_exec(void)
{
struct task_struct *p = current;
struct migration_arg arg;
- int dest_cpu;
+ int dest_cpu, wake_flag = WF_EXEC;
scoped_guard (raw_spinlock_irqsave, &p->pi_lock) {
- dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), WF_EXEC);
+ dest_cpu = select_task_rq(p, task_cpu(p), &wake_flag);
+
if (dest_cpu == smp_processor_id())
return;
--
2.43.0