Re: [PATCH] psi: Only collect online cpu time in collect_percpu_times

Re: [PATCH] psi: Only collect online cpu time in collect_percpu_times

[Peter Zijlstra]

On Mon, Dec 02, 2019 at 09:09:28PM +0800, Liu Song wrote:
> From: Liu Song <liu.song11@zte.com.cn>
> 
> Tasks can only run on the online cpu, so only need to
> collect the time of the online cpu.
> 
> Signed-off-by: Liu Song <liu.song11@zte.com.cn>
> ---
>  kernel/sched/psi.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/kernel/sched/psi.c b/kernel/sched/psi.c
> index 7acc632c3b82..605f02facb7b 100644
> --- a/kernel/sched/psi.c
> +++ b/kernel/sched/psi.c
> @@ -316,7 +316,7 @@ static void collect_percpu_times(struct psi_group *group,
>  	 * the sampling period. This eliminates artifacts from uneven
>  	 * loading, or even entirely idle CPUs.
>  	 */
> -	for_each_possible_cpu(cpu) {
> +	for_each_online_cpu(cpu) {
>  		u32 times[NR_PSI_STATES];
>  		u32 nonidle;
>  		u32 cpu_changed_states;

And who collects the deltas that remain after offline?

Re: [PATCH] psi: Only collect online cpu time in collect_percpu_times

[Peter Zijlstra]

On Thu, Dec 05, 2019 at 10:05:59AM +0800, liu.song11@zte.com.cn wrote:
> >On Mon, Dec 02, 2019 at 09:09:28PM +0800, Liu Song wrote:
> >> From: Liu Song <liu.song11@zte.com.cn> 
> >>  
> >> Tasks can only run on the online cpu, so only need to
> >> collect the time of the online cpu.
> >>  
> >> Signed-off-by: Liu Song <liu.song11@zte.com.cn> 
> >> ---
> >>  kernel/sched/psi.c | 2 +-
> >>  1 file changed, 1 insertion(+), 1 deletion(-)
> >>  
> >> diff --git a/kernel/sched/psi.c b/kernel/sched/psi.c
> >> index 7acc632c3b82..605f02facb7b 100644
> >> --- a/kernel/sched/psi.c
> >> +++ b/kernel/sched/psi.c
> >> @@ -316,7 +316,7 @@ static void collect_percpu_times(struct psi_group *group,
> >>       * the sampling period. This eliminates artifacts from uneven
> >>       * loading, or even entirely idle CPUs.
> >>       */
> >> -    for_each_possible_cpu(cpu) {
> >> +    for_each_online_cpu(cpu) {
> >>          u32 times[NR_PSI_STATES];
> >>          u32 nonidle;
> >>          u32 cpu_changed_states;
> >
> >And who collects the deltas that remain after offline?
> 
> Hi,
> After the cpu goes offline, there is no activity on the cpu and no time will be updated.

But there might have been activitiy before it went offline.

> So the value of deltas is 0 and it will not contribute to the total time.

No, the value will not change, but it need not be 0.

Re: [PATCH] psi: Only collect online cpu time in collect_percpu_times

[Peter Zijlstra]

On Fri, Dec 06, 2019 at 12:29:28PM +0800, liu.song11@zte.com.cn wrote:

> >No, the value will not change, but it need not be 0.
> 
> Hi,
> 
> Suppose there are 4 cpu online, then we take cpu 3 offline. In "collect_percpu_times",
> because "for_each_possible_cpu", will still collect the time of cpu 3 which is offline.

Correct, someone needs to collect the last deltas.

> However, it is clear that "nonidle" will remain at 0 until cpu 3 comes online again.

How does it become 0?

Consider

	CPU2		CPU3

			// runs crap, nonidle increases

	// offline CPU3

			// goes offline, nonidle is still >0


At this point someone needs to collect the delta from CPU3 to make it 0.
But if you only iterate online CPUs, that will not happen.

> And the value of "deltas[s]" will not change after collecting CPU 2. In "get_recent_times",
> the value of "groupc->times_prev[aggregator][s]" corresponding to cpu 3 will not change.

They will not change after the first collection after hot-un-plug. You
need at least one collection after it stops running crap.

> This is only a case where one CPU goes offline. If there are multiple CPUs offline,
> it will correspond to more meaningless operations. So here should be changed to
> "for_each_online_cpu", the effect is exactly the same as "for_each_possible_cpu",
> and the possible meaningless processes are excluded.

It is not. Or if it is, you failed to explain who collects on hotplug.