Re: [PATCH v2 3/2] sched/deadline: Check bandwidth overflow earlier for hotplug

[Jon Hunter <jonathanh@nvidia.com>]

Hi Juri,

On 15/11/2024 11:48, Juri Lelli wrote:
> Currently we check for bandwidth overflow potentially due to hotplug
> operations at the end of sched_cpu_deactivate(), after the cpu going
> offline has already been removed from scheduling, active_mask, etc.
> This can create issues for DEADLINE tasks, as there is a substantial
> race window between the start of sched_cpu_deactivate() and the moment
> we possibly decide to roll-back the operation if dl_bw_deactivate()
> returns failure in cpuset_cpu_inactive(). An example is a throttled
> task that sees its replenishment timer firing while the cpu it was
> previously running on is considered offline, but before
> dl_bw_deactivate() had a chance to say no and roll-back happened.
> 
> Fix this by directly calling dl_bw_deactivate() first thing in
> sched_cpu_deactivate() and do the required calculation in the former
> function considering the cpu passed as an argument as offline already.
> 
> By doing so we also simplify sched_cpu_deactivate(), as there is no need
> anymore for any kind of roll-back if we fail early.
> 
> Signed-off-by: Juri Lelli <juri.lelli@redhat.com>
> ---
> Thanks Waiman and Phil for testing and reviewing the scratch version of
> this change. I think the below might be better, as we end up with a
> clean-up as well.
> 
> Please take another look when you/others have time.
> ---
>   kernel/sched/core.c     | 22 +++++++---------------
>   kernel/sched/deadline.c | 12 ++++++++++--
>   2 files changed, 17 insertions(+), 17 deletions(-)
> 
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index d1049e784510..e2c6eacf793e 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -8054,19 +8054,14 @@ static void cpuset_cpu_active(void)
>   	cpuset_update_active_cpus();
>   }
>   
> -static int cpuset_cpu_inactive(unsigned int cpu)
> +static void cpuset_cpu_inactive(unsigned int cpu)
>   {
>   	if (!cpuhp_tasks_frozen) {
> -		int ret = dl_bw_deactivate(cpu);
> -
> -		if (ret)
> -			return ret;
>   		cpuset_update_active_cpus();
>   	} else {
>   		num_cpus_frozen++;
>   		partition_sched_domains(1, NULL, NULL);
>   	}
> -	return 0;
>   }
>   
>   static inline void sched_smt_present_inc(int cpu)
> @@ -8128,6 +8123,11 @@ int sched_cpu_deactivate(unsigned int cpu)
>   	struct rq *rq = cpu_rq(cpu);
>   	int ret;
>   
> +	ret = dl_bw_deactivate(cpu);
> +
> +	if (ret)
> +		return ret;
> +
>   	/*
>   	 * Remove CPU from nohz.idle_cpus_mask to prevent participating in
>   	 * load balancing when not active
> @@ -8173,15 +8173,7 @@ int sched_cpu_deactivate(unsigned int cpu)
>   		return 0;
>   
>   	sched_update_numa(cpu, false);
> -	ret = cpuset_cpu_inactive(cpu);
> -	if (ret) {
> -		sched_smt_present_inc(cpu);
> -		sched_set_rq_online(rq, cpu);
> -		balance_push_set(cpu, false);
> -		set_cpu_active(cpu, true);
> -		sched_update_numa(cpu, true);
> -		return ret;
> -	}
> +	cpuset_cpu_inactive(cpu);
>   	sched_domains_numa_masks_clear(cpu);
>   	return 0;
>   }
> diff --git a/kernel/sched/deadline.c b/kernel/sched/deadline.c
> index 267ea8bacaf6..6e988d4cd787 100644
> --- a/kernel/sched/deadline.c
> +++ b/kernel/sched/deadline.c
> @@ -3505,6 +3505,13 @@ static int dl_bw_manage(enum dl_bw_request req, int cpu, u64 dl_bw)
>   		}
>   		break;
>   	case dl_bw_req_deactivate:
> +		/*
> +		 * cpu is not off yet, but we need to do the math by
> +		 * considering it off already (i.e., what would happen if we
> +		 * turn cpu off?).
> +		 */
> +		cap -= arch_scale_cpu_capacity(cpu);
> +
>   		/*
>   		 * cpu is going offline and NORMAL tasks will be moved away
>   		 * from it. We can thus discount dl_server bandwidth
> @@ -3522,9 +3529,10 @@ static int dl_bw_manage(enum dl_bw_request req, int cpu, u64 dl_bw)
>   		if (dl_b->total_bw - fair_server_bw > 0) {
>   			/*
>   			 * Leaving at least one CPU for DEADLINE tasks seems a
> -			 * wise thing to do.
> +			 * wise thing to do. As said above, cpu is not offline
> +			 * yet, so account for that.
>   			 */
> -			if (dl_bw_cpus(cpu))
> +			if (dl_bw_cpus(cpu) - 1)
>   				overflow = __dl_overflow(dl_b, cap, fair_server_bw, 0);
>   			else
>   				overflow = 1;


I have noticed a suspend regression on one of our Tegra boards and 
bisect is pointing to this commit. If I revert this on top of -next then 
I don't see the issue.

The only messages I see when suspend fails are ...

[   53.905976] Error taking CPU1 down: -16
[   53.909887] Non-boot CPUs are not disabled

So far this is only happening on Tegra186 (ARM64). Let me know if you 
have any thoughts.

Thanks
Jon

-- 
nvpublic