Re: [PATCH] sched/cputime: Don't warn when the tick period starts after idle exit

[Frederic Weisbecker <frederic@kernel.org>]

Le Sat, Oct 31, 2026 at 03:11:46PM +0100, Stian Halseth a écrit :
> Hi Frederic,
> 
> 
> On Sat, 2026-10-31 at 14:54 +0100, Frederic Weisbecker wrote:
> > > 
> > > Here is where it goes wrong. One tick has passed in real time, but not
> > > in ktime. The switch happened d after the last tick, jiffies never
> > > counted that d, and the tsc only counts from the switch on, so here
> > > ktime_get() == jiffies + TICK_NSEC - d.
> > 
> > So that assume that during the clocksource switch, the new clock synchronize
> > with the old one so that ktime_get() from the new clock should start from
> > the last value of the old clock. I'm having troubles to verify if
> > tk_setup_internals() does exactly that. 
> 
> Not that one, the timekeeping_forward_now() just before it. It folds
> the old clock into xtime_nsec, and tk_setup_internals() keeps that and
> only resets cycle_last from the new clock. On jiffies that value is the
> last tick.

Ah yes! You got it!

> 
> > 
> > [    1.549194] old= 1352575566
> > [    1.549930] new= 1352578748
> > [    1.550636] clocksource: Switched to clocksource tsc-early
> > 
> > That's 3 us, it could be the simple difference between the 2 printk.
> 
> Look at the timestamps: 736 us passed between the two printks, but
> ktime moved 3 us. The 3 us is tsc-early, from the read in
> tk_setup_internals() to your second ktime_get(). The time from the
> last tick to the first printk, plus the 733 us from there to
> tk_setup_internals(), is never counted. That is the d. If you print
> ktime_get() in tick_periodic() as well, old= will match the previous
> tick exactly.

Makes sense.

> 
> > 
> > But the switch from tsc-early to tsc is much more:
> > 
> > [    2.639818] old= 2445754540
> > [    2.640172] new= 2446107950
> > [    2.640927] clocksource: Switched to clocksource tsc
> > 
> > That's 353 us.
> 
> Here the printks are 354 us apart and ktime moved 353 us. Both clocks
> run, so nothing is lost: tsc-early counts up to the switch and tsc
> continues from there.

Now that's clear!

Thanks.

-- 
Frederic Weisbecker
SUSE Labs