Re: [RFC PATCH v3 05/10] sched/fair: Don't consider paravirt CPUs for wakeup and load balance

From
Shrikanth Hegde <sshegde@linux.ibm.com>
Date
2025-11-08 12:05:21
Also in
lkml
Reply to
K Prateek Nayak, 2025-09-11

On 9/11/25 10:53 AM, K Prateek Nayak wrote:
Hello Shrikanth,

On 9/10/2025 11:12 PM, Shrikanth Hegde wrote:
quoted
@@ -8563,7 +8563,7 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
  		if (!is_rd_overutilized(this_rq()->rd)) {
  			new_cpu = find_energy_efficient_cpu(p, prev_cpu);
  			if (new_cpu >= 0)
-				return new_cpu;
+				goto check_new_cpu;
Should this fallback to the overutilized path if the most energy
efficient CPU is found to be paravirtualized or should
find_energy_efficient_cpu() be made aware of it?

While thinking about this, are there any such system which has vCPUs and
overcommits and still has energy model backing it?

Highly unlikely. So, I am planning to put a warning there and see if any
such usage exists there.
quoted
  			new_cpu = prev_cpu;
  		}
  
@@ -8605,7 +8605,12 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
  	}
  	rcu_read_unlock();
  
-	return new_cpu;
+	/* If newly found or prev_cpu is a paravirt cpu, use current cpu */
+check_new_cpu:
+	if (is_cpu_paravirt(new_cpu))
+		return cpu;
+	else
nit. redundant else.
quoted
+		return new_cpu;
  }
  
  /*
@@ -11734,6 +11739,12 @@ static int sched_balance_rq(int this_cpu, struct rq *this_rq,
  
  	cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
  
+#ifdef CONFIG_PARAVIRT
+	/* Don't spread load to paravirt CPUs */
+	if (static_branch_unlikely(&cpu_paravirt_push_tasks))
+		cpumask_andnot(cpus, cpus, cpu_paravirt_mask);
+#endif
Can something similar be also be done in select_idle_sibling() and
sched_balance_find_dst_cpu() for wakeup path?
We have this pattern in select_task_rq_fair

cpu = smp_processor_id();
new_cpu = prev_cpu;

task is waking up after a while, so likely prev_cpu is marked as paravirt
and in such cases we should return current cpu. if current cpu is paravirt(unlikely),
and prev_cpu is also paravirt, then should return current cpu.
In next sched tick it will be pushed out.

select_idle_sibling(p, prev_cpu, new_cpu); - (new_cpu will remain prev_cpu if wake_affine doesn't change it)
Will have to change the prototype to send current cpu as well.
Thread (33 messages) 33 messages, 5 authors, last reply 2025-11-10
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help