Le Mon, Jan 19, 2026 at 03:26:07PM +0100, Peter Zijlstra a écrit :
> On Fri, Jan 16, 2026 at 03:51:59PM +0100, Frederic Weisbecker wrote:
>
> > +#ifdef CONFIG_NO_HZ_COMMON
> > +void kcpustat_dyntick_start(void)
> > +{
> > + if (!vtime_generic_enabled_this_cpu()) {
> > + vtime_dyntick_start();
> > + __this_cpu_write(kernel_cpustat.idle_dyntick, 1);
> > + }
> > +}
>
> Why don't we need to make sure steal time is up-to-date at this point?
Yes, there could be steal time since the last tick. It will be included
and accounted in kcpustat_dyntick_stop() and not substracted from system
or idle cputime (but it should!). This wrong behaviour is the same as the
current upstream behaviour. So no known regression.
But check the last patch of the series that tries to fix that:
sched/cputime: Handle dyntick-idle steal time correctly
Thanks.
--
Frederic Weisbecker
SUSE Labs
