[PATCH v3 01/12] sched: fix imbalance flag reset

Preeti U Murthy preeti at linux.vnet.ibm.com
Mon Jul 7 20:13:45 PDT 2014


On 06/30/2014 09:35 PM, Vincent Guittot wrote:
> The imbalance flag can stay set whereas there is no imbalance.
> 
> Let assume that we have 3 tasks that run on a dual cores /dual cluster system.
> We will have some idle load balance which are triggered during tick.
> Unfortunately, the tick is also used to queue background work so we can reach
> the situation where short work has been queued on a CPU which already runs a
> task. The load balance will detect this imbalance (2 tasks on 1 CPU and an idle
> CPU) and will try to pull the waiting task on the idle CPU. The waiting task is
> a worker thread that is pinned on a CPU so an imbalance due to pinned task is
> detected and the imbalance flag is set.
> Then, we will not be able to clear the flag because we have at most 1 task on
> each CPU but the imbalance flag will trig to useless active load balance
> between the idle CPU and the busy CPU.
> 
> We need to reset of the imbalance flag as soon as we have reached a balanced
> state.
> 
> Signed-off-by: Vincent Guittot <vincent.guittot at linaro.org>
> ---
>  kernel/sched/fair.c | 14 +++++++++++---
>  1 file changed, 11 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index d3c73122..0c48dff 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -6615,10 +6615,8 @@ more_balance:
>  		if (sd_parent) {
>  			int *group_imbalance = &sd_parent->groups->sgc->imbalance;
> 
> -			if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0) {
> +			if ((env.flags & LBF_SOME_PINNED) && env.imbalance > 0)
>  				*group_imbalance = 1;
> -			} else if (*group_imbalance)
> -				*group_imbalance = 0;
>  		}
> 
>  		/* All tasks on this runqueue were pinned by CPU affinity */
> @@ -6703,6 +6701,16 @@ more_balance:
>  	goto out;
> 
>  out_balanced:
> +	/*
> +	 * We reach balance although we may have faced some affinity
> +	 * constraints. Clear the imbalance flag if it was set.
> +	 */
> +	if (sd_parent) {
> +		int *group_imbalance = &sd_parent->groups->sgc->imbalance;
> +		if (*group_imbalance)
> +			*group_imbalance = 0;
> +	}
> +
>  	schedstat_inc(sd, lb_balanced[idle]);
> 
>  	sd->nr_balance_failed = 0;
> 
I am not convinced that we can clear the imbalance flag here. Lets take
a simple example. Assume at a particular level of sched_domain, there
are two sched_groups with one cpu each. There are 2 tasks on the source
cpu, one of which is running(t1) and the other thread(t2) does not have
the dst_cpu in the tsk_allowed_mask. Now no task can be migrated to the
dst_cpu due to affinity constraints. Note that t2 is *not pinned, it
just cannot run on the dst_cpu*. In this scenario also we reach the
out_balanced tag right? If we set the group_imbalance flag to 0, we are
ruling out the possibility of migrating t2 to any other cpu in a higher
level sched_domain by saying that all is well, there is no imbalance.
This is wrong, isn't it?

My point is that by clearing the imbalance flag in the out_balanced
case, you might be overlooking the fact that the tsk_cpus_allowed mask
of the tasks on the src_cpu may not be able to run on the dst_cpu in
*this* level of sched_domain, but can potentially run on a cpu at any
higher level of sched_domain. By clearing the flag, we are not
encouraging load balance at that level for t2.

Am I missing something?

Regards
Preeti U Murthy




More information about the linux-arm-kernel mailing list