Re: [RFC PATCH v3 21/21] sched/cputime: Cope with steal time going backwards or negative

[Paul Durrant]

On 22/05/2024 01:17, David Woodhouse wrote:
> From: David Woodhouse <dwmw@xxxxxxxxxxxx>
>
> In steal_account_process_time(), a delta is calculated between the value
> returned by paravirt_steal_clock(), and this_rq()->prev_steal_time which
> is assumed to be the *previous* value returned by paravirt_steal_clock().
>
> However, instead of just assigning the newly-read value directly into
> ->prev_steal_time for use in the next iteration, ->prev_steal_time is
> *incremented* by the calculated delta.
>
> This used to be roughly the same, modulo conversion to jiffies and back,
> until commit 807e5b80687c0 ("sched/cputime: Add steal time support to
> full dynticks CPU time accounting") started clamping that delta to a
> maximum of the actual time elapsed. So now, if the value returned by
> paravirt_steal_clock() jumps by a large amount, instead of a *single*
> period of reporting 100% steal time, the system will report 100% steal
> time for as long as it takes to "catch up" with the reported value.
> Which is up to 584 years.
>
> But there is a benefit to advancing ->prev_steal_time only by the time
> which was *accounted* as having been stolen. It means that any extra
> time truncated by the clamping will be accounted in the next sample
> period rather than lost. Given the stochastic nature of the sampling,
> that is more accurate overall.
>
> So, continue to advance ->prev_steal_time by the accounted value as
> long as the delta isn't egregiously large (for which, use maxtime * 2).
> If the delta is more than that, just set ->prev_steal_time directly to
> the value returned by paravirt_steal_clock().
>
> Fixes: 807e5b80687c0 ("sched/cputime: Add steal time support to full dynticks CPU time accounting")
> Signed-off-by: David Woodhouse <dwmw@xxxxxxxxxxxx>
> ---
>   kernel/sched/cputime.c | 20 ++++++++++++++------
>   1 file changed, 14 insertions(+), 6 deletions(-)

Reviewed-by: Paul Durrant <paul@xxxxxxx>