Skip to contentSkip to conversation

tg: add sched wait_count of a task group

PATCH In review [email protected] 2 participants 2 messages Jan 21, 2021 1 message not in the archive

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 tg: add sched wait_count of a task group

From: Liming Wu <[email protected]>

Now we can rely on PSI to reflect whether there is contention in the task group, but it cannot reflect the details of the contention. Through this metric, we can get avg latency of task group contention
from the dimension of scheduling.
delta(wait_usec)/delta(nr_waits)
Also change wait_sum to wait_usec.

Signed-off-by Liming Wu <[email protected]>

---
 kernel/sched/core.c | 23 +++++++++++++++++++++--
 1 file changed, 21 insertions(+), 2 deletions(-)
kernel/sched/core.c +21 -2
index e7e453492..45e46114e 100644
@@ -8146,12 +8146,17 @@ static int cpu_cfs_stat_show(struct seq_file *sf, void *v)
 
 	if (schedstat_enabled() && tg != &root_task_group) {
 		u64 ws = 0;
+		u64 wc = 0;
 		int i;
 
-		for_each_possible_cpu(i)
+		for_each_possible_cpu(i) {
 			ws += schedstat_val(tg->se[i]->statistics.wait_sum);
+			wc += schedstat_val(tg->se[i]->statistics.wait_count);
+		}
 
-		seq_printf(sf, "wait_sum %llu\n", ws);
+		do_div(ws, NSEC_PER_USEC);
+		seq_printf(sf, "wait_usec %llu\n"
+			   "nr_waits %llu\n", ws, wc);
 	}
 
 	return 0;
@@ -8255,6 +8260,20 @@ static int cpu_extra_stat_show(struct seq_file *sf,
 			   "throttled_usec %llu\n",
 			   cfs_b->nr_periods, cfs_b->nr_throttled,
 			   throttled_usec);
+		if (schedstat_enabled() && tg != &root_task_group) {
+			u64 ws = 0;
+			u64 wc = 0;
+			int i;
+
+			for_each_possible_cpu(i) {
+				ws += schedstat_val(tg->se[i]->statistics.wait_sum);
+				wc += schedstat_val(tg->se[i]->statistics.wait_count);
+			}
+
+			do_div(ws, NSEC_PER_USEC);
+			seq_printf(sf, "wait_usec %llu\n"
+				   "nr_waits %llu\n", ws, wc);
+		}
 	}
 #endif
 	return 0;
Signature · 1 line
2.20.1

A message that is not in the archive (sent off-list or not yet received): 20210121201157.1933-1-wu860403-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#2 lore raw

Your subject is wrong; this is 100% a sched patch.

Quoted 7 lines · On Fri, Jan 22, 2021 at 04:11:57AM +0800…

From: Liming Wu <[email protected]>

Now we can rely on PSI to reflect whether there is contention in the task group, but it cannot reflect the details of the contention. Through this metric, we can get avg latency of task group contention
from the dimension of scheduling.
delta(wait_usec)/delta(nr_waits)

Only if all tasks have the same weight. So in general this is useless.

Quoted 1 line

Also change wait_sum to wait_usec.

Also patches should do 1 thing, also you can't do that, you'll break everybody that expect the current value.