2018-10-22 03:10:11

by Yi Wang

Subject: [PATCH] sched/numa: fix choosing isolated CPUs when task_numa_migrate()

When trying to migrate to a CPU in task_numa_migrate(), we invoke
task_numa_find_cpu() to choose a spot, in which function we skip
the CPU which is not in cpus_allowed, but forgot to concern the
isolated CPUs, and this may cause the task would run on the isolcpus.

This patch fixes this issue by checking the load_balance_mask.

Signed-off-by: Yi Wang <[email protected]>
Reviewed-by: Yi Liu <[email protected]>
---
kernel/sched/fair.c | 10 ++++++++++
1 file changed, 10 insertions(+)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 908c9cd..0fa0cee 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -1709,6 +1709,7 @@ static void task_numa_compare(struct task_numa_env *env,
rcu_read_unlock();
}

+static int is_cpu_load_balance(int cpu);
static void task_numa_find_cpu(struct task_numa_env *env,
long taskimp, long groupimp)
{
@@ -1731,6 +1732,9 @@ static void task_numa_find_cpu(struct task_numa_env *env,
if (!cpumask_test_cpu(cpu, &env->p->cpus_allowed))
continue;

+ if (!is_cpu_load_balance(cpu))
+ continue;
+
env->dst_cpu = cpu;
task_numa_compare(env, taskimp, groupimp, maymove);
}
@@ -8528,6 +8532,12 @@ static int should_we_balance(struct lb_env *env)
return balance_cpu == env->dst_cpu;
}

+static int is_cpu_load_balance(int cpu)
+{
+ struct cpumask *cpus = this_cpu_cpumask_var_ptr(load_balance_mask);
+
+ return cpumask_test_cpu(cpu, cpus);
+}
+
/*
* Check this_cpu to ensure it is balanced within domain. Attempt to move
* tasks if there is an imbalance.
--
1.8.3.1

2018-10-23 17:37:46

by Srikar Dronamraju

Subject: Re: [PATCH] sched/numa: fix choosing isolated CPUs when task_numa_migrate()

Hi Yi Wang,

> When trying to migrate to a CPU in task_numa_migrate(), we invoke
> task_numa_find_cpu() to choose a spot, in which function we skip
> the CPU which is not in cpus_allowed, but forgot to concern the
> isolated CPUs, and this may cause the task would run on the isolcpus.
>
> This patch fixes this issue by checking the load_balance_mask.
>
> Signed-off-by: Yi Wang <[email protected]>
> Reviewed-by: Yi Liu <[email protected]>
> ---
> kernel/sched/fair.c | 10 ++++++++++
> 1 file changed, 10 insertions(+)
>

I had proposed something similar
http://lkml.kernel.org/r/[email protected]

but at that time, Peter felt we should fix it at different level.
http://lkml.kernel.org/r/[email protected]

and I think he is right.

I have system with 32 cpus, have passed isolcpus as a kernel parameter.

$ grep -o "isolcpus=[,,1-9]*" /proc/cmdline
isolcpus=1,5,9,13
$ grep -i cpus_allowed /proc/$$/status
Cpus_allowed: ffffdddd
Cpus_allowed_list: 0,2-4,6-8,10-12,14-31


So for a task running on top_cpuset will not have isolcpus as part of the
cpus_allowed.

However if a said task were to call sched_setaffinity, then there is every
likely hood of the cpus being passed being a mix of isolcpus and
nonisolcpus.

For example
perf bench numa mem --no-data_rand_walk -p 4 -t 8 -G 0 -P 3072 -T 0 -l 50 -c -s 1000

$ for i in $(pgrep -f perf); do grep -i cpus_allowed_list /proc/$i/task/*/status ; done | head -n 10
Cpus_allowed_list: 0,2-4,6-8,10-12,14-31
/proc/2107/task/2107/status:Cpus_allowed_list: 0-31
/proc/2107/task/2196/status:Cpus_allowed_list: 0-31
/proc/2107/task/2197/status:Cpus_allowed_list: 0-31
/proc/2107/task/2198/status:Cpus_allowed_list: 0-31
/proc/2107/task/2199/status:Cpus_allowed_list: 0-31
/proc/2107/task/2200/status:Cpus_allowed_list: 0-31
/proc/2107/task/2201/status:Cpus_allowed_list: 0-31
/proc/2107/task/2202/status:Cpus_allowed_list: 0-31
/proc/2107/task/2203/status:Cpus_allowed_list: 0-31

So the cpus_allowed has a mix of isolcpus and nonisolcpus.

While this patch fixes the problem, there is a risk of missing other places
like update_numa_stats also iterates and accounts for tasks stats running on
isolcpus.

I will send a patch with a slightly different approach.
Request you to review and verify the same.

--
Thanks and Regards
Srikar Dronamraju