Re: [PATCH] tick-sched: fix idle and iowait sleeptime accounting vs CPU hotplug

[Frederic Weisbecker]

Le Mon, Jan 15, 2024 at 05:35:55PM +0100, Heiko Carstens a écrit :
> When offlining and onlining CPUs the overall reported idle and iowait
> times as reported by /proc/stat jump backward and forward:
> 
> > cat /proc/stat
> cpu  132 0 176 225249 47 6 6 21 0 0
> cpu0 80 0 115 112575 33 3 4 18 0 0
> cpu1 52 0 60 112673 13 3 1 2 0 0
> 
> > chcpu -d 1
> > cat /proc/stat
> cpu  133 0 177 226681 47 6 6 21 0 0
> cpu0 80 0 116 113387 33 3 4 18 0 0
> 
> > chcpu -e 1
> > cat /proc/stat
> cpu  133 0 178 114431 33 6 6 21 0 0 <---- jump backward
> cpu0 80 0 116 114247 33 3 4 18 0 0
> cpu1 52 0 61 183 0 3 1 2 0 0        <---- idle + iowait start with 0
> 
> > chcpu -d 1
> > cat /proc/stat
> cpu  133 0 178 228956 47 6 6 21 0 0 <---- jump forward
> cpu0 81 0 117 114929 33 3 4 18 0 0
> 
> Reason for this is that get_idle_time() in fs/proc/stat.c has different
> sources for both values depending on if a CPU is online or offline:
> 
> - if a CPU is online the values may be taken from its per cpu
>   tick_cpu_sched structure
> 
> - if a CPU is offline the values are taken from its per cpu cpustat
>   structure
> 
> The problem is that the per cpu tick_cpu_sched structure is set to zero on
> CPU offline. See tick_cancel_sched_timer() in kernel/time/tick-sched.c.
> 
> Therefore when a CPU is brought offline and online afterwards both its idle
> and iowait sleeptime will be zero, causing a jump backward in total system
> idle and iowait sleeptime. In a similar way if a CPU is then brought
> offline again the total idle and iowait sleeptimes will jump forward.
> 
> It looks like this behavior was introduced with commit 4b0c0f294f60
> ("tick: Cleanup NOHZ per cpu data on cpu down").
> 
> This was only noticed now on s390, since we switched to generic idle time
> reporting with commit be76ea614460 ("s390/idle: remove arch_cpu_idle_time()
> and corresponding code").
> 
> Fix this by preserving the values of idle_sleeptime and iowait_sleeptime
> members of the per-cpu tick_sched structure on CPU hotplug.
> 
> Fixes: 4b0c0f294f60 ("tick: Cleanup NOHZ per cpu data on cpu down")
> Reported-by: Gerald Schaefer <gerald.schaefer@xxxxxxxxxxxxx>
> Signed-off-by: Heiko Carstens <hca@xxxxxxxxxxxxx>
> ---
>  kernel/time/tick-sched.c | 5 +++++
>  1 file changed, 5 insertions(+)
> 
> diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
> index a17d26002831..d2501673028d 100644
> --- a/kernel/time/tick-sched.c
> +++ b/kernel/time/tick-sched.c
> @@ -1576,13 +1576,18 @@ void tick_setup_sched_timer(void)
>  void tick_cancel_sched_timer(int cpu)
>  {
>  	struct tick_sched *ts = &per_cpu(tick_cpu_sched, cpu);
> +	ktime_t idle_sleeptime, iowait_sleeptime;
>  
>  # ifdef CONFIG_HIGH_RES_TIMERS
>  	if (ts->sched_timer.base)
>  		hrtimer_cancel(&ts->sched_timer);
>  # endif
>  
> +	idle_sleeptime = ts->idle_sleeptime;
> +	iowait_sleeptime = ts->iowait_sleeptime;
>  	memset(ts, 0, sizeof(*ts));
> +	ts->idle_sleeptime = idle_sleeptime;
> +	ts->iowait_sleeptime = iowait_sleeptime;

And this is safe because it is in global stop machine. So we are
guaranteed that nobody sees the transitionning state. In the worst
case ts->idle_sleeptime_seq is observed as changed to 0 in read_seqcount_retry()
and the values are simply fetched again.

Reviewed-by: Frederic Weisbecker <frederic@xxxxxxxxxx>

This makes me think that we should always use cpustat[CPUTIME_IDLE] instead of
maintaining this separate ts->idle_sleeptime field. kcpustat even has a seqcount
that would make ts->idle_sleeptime_seq obsolete. Then the tick based idle accounting
could disappear on nohz, along with a few hacks. Instead of that we are
currently maintaining two different idle accounting that are roughly the same.

But anyway this is all a different story, just mumbling to myself for the next
nohz cleanups.

Thanks!