Re: [PATCH] sched/fair: scale quota and period without losing quota/period ratio precision

From: Xuewei Zhang
Date: Mon Oct 07 2019 - 19:29:19 EST


On Mon, Oct 7, 2019 at 8:14 AM Phil Auld <pauld@xxxxxxxxxx> wrote:
>
> On Thu, Oct 03, 2019 at 05:12:43PM -0700 Xuewei Zhang wrote:
> > quota/period ratio is used to ensure a child task group won't get more
> > bandwidth than the parent task group, and is calculated as:
> > normalized_cfs_quota() = [(quota_us << 20) / period_us]
> >
> > If the quota/period ratio was changed during this scaling due to
> > precision loss, it will cause inconsistency between parent and child
> > task groups. See below example:
> > A userspace container manager (kubelet) does three operations:
> > 1) Create a parent cgroup, set quota to 1,000us and period to 10,000us.
> > 2) Create a few children cgroups.
> > 3) Set quota to 1,000us and period to 10,000us on a child cgroup.
> >
> > These operations are expected to succeed. However, if the scaling of
> > 147/128 happens before step 3), quota and period of the parent cgroup
> > will be changed:
> > new_quota: 1148437ns, 1148us
> > new_period: 11484375ns, 11484us
> >
> > And when step 3) comes in, the ratio of the child cgroup will be 104857,
> > which will be larger than the parent cgroup ratio (104821), and will
> > fail.
> >
> > Scaling them by a factor of 2 will fix the problem.
> >
> > Fixes: 2e8e19226398 ("sched/fair: Limit sched_cfs_period_timer() loop to avoid hard lockup")
> > Signed-off-by: Xuewei Zhang <xueweiz@xxxxxxxxxx>
>
>
> I managed to get it to trigger the second case. It took 50,000 children (20x my initial tests).
>
> [ 1367.850630] cfs_period_timer[cpu11]: period too short, scaling up (new cfs_period_us = 4340, cfs_quota_us = 250000)
> [ 1370.390832] cfs_period_timer[cpu11]: period too short, scaling up (new cfs_period_us = 8680, cfs_quota_us = 500000)
> [ 1372.914689] cfs_period_timer[cpu11]: period too short, scaling up (new cfs_period_us = 17360, cfs_quota_us = 1000000)
> [ 1375.447431] cfs_period_timer[cpu11]: period too short, scaling up (new cfs_period_us = 34720, cfs_quota_us = 2000000)
> [ 1377.982785] cfs_period_timer[cpu11]: period too short, scaling up (new cfs_period_us = 69440, cfs_quota_us = 4000000)
> [ 1380.481702] cfs_period_timer[cpu11]: period too short, scaling up (new cfs_period_us = 138880, cfs_quota_us = 8000000)
> [ 1382.894692] cfs_period_timer[cpu11]: period too short, scaling up (new cfs_period_us = 277760, cfs_quota_us = 16000000)
> [ 1385.264872] cfs_period_timer[cpu11]: period too short, scaling up (new cfs_period_us = 555520, cfs_quota_us = 32000000)
> [ 1393.965140] cfs_period_timer[cpu11]: period too short, but cannot scale up without losing precision (cfs_period_us = 555520, cfs_quota_us = 32000000)
>
> I suspect going higher could cause the original lockup, but that'd be the case with the old code as well.
> And this also gets us out of it faster.
>
>
> Tested-by: Phil Auld <pauld@xxxxxxxxxx>

Thanks a lot for the review and experiment+test Phil! Really appreciate it.

To other scheduler maintainers: Could someone help review and approve
the patch? I'm happy to fix any defect in it :)

Best regards,
Xuewei

>
>
> Cheers,
> Phil
>
>
> > ---
> > kernel/sched/fair.c | 36 ++++++++++++++++++++++--------------
> > 1 file changed, 22 insertions(+), 14 deletions(-)
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 83ab35e2374f..b3d3d0a231cd 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -4926,20 +4926,28 @@ static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
> > if (++count > 3) {
> > u64 new, old = ktime_to_ns(cfs_b->period);
> >
> > - new = (old * 147) / 128; /* ~115% */
> > - new = min(new, max_cfs_quota_period);
> > -
> > - cfs_b->period = ns_to_ktime(new);
> > -
> > - /* since max is 1s, this is limited to 1e9^2, which fits in u64 */
> > - cfs_b->quota *= new;
> > - cfs_b->quota = div64_u64(cfs_b->quota, old);
> > -
> > - pr_warn_ratelimited(
> > - "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us %lld, cfs_quota_us = %lld)\n",
> > - smp_processor_id(),
> > - div_u64(new, NSEC_PER_USEC),
> > - div_u64(cfs_b->quota, NSEC_PER_USEC));
> > + /*
> > + * Grow period by a factor of 2 to avoid lossing precision.
> > + * Precision loss in the quota/period ratio can cause __cfs_schedulable
> > + * to fail.
> > + */
> > + new = old * 2;
> > + if (new < max_cfs_quota_period) {
> > + cfs_b->period = ns_to_ktime(new);
> > + cfs_b->quota *= 2;
> > +
> > + pr_warn_ratelimited(
> > + "cfs_period_timer[cpu%d]: period too short, scaling up (new cfs_period_us = %lld, cfs_quota_us = %lld)\n",
> > + smp_processor_id(),
> > + div_u64(new, NSEC_PER_USEC),
> > + div_u64(cfs_b->quota, NSEC_PER_USEC));
> > + } else {
> > + pr_warn_ratelimited(
> > + "cfs_period_timer[cpu%d]: period too short, but cannot scale up without losing precision (cfs_period_us = %lld, cfs_quota_us = %lld)\n",
> > + smp_processor_id(),
> > + div_u64(old, NSEC_PER_USEC),
> > + div_u64(cfs_b->quota, NSEC_PER_USEC));
> > + }
> >
> > /* reset count so we don't come right back in here */
> > count = 0;
> > --
> > 2.23.0.581.g78d2f28ef7-goog
> >
>
> --