[BUG REPORT] sched/rt: Inaccurate numerical calculation in rt_runtime_us constraints

[BUG REPORT] sched/rt: Inaccurate numerical calculation in rt_runtime_us constraints

[zhaowenhui (A)]

Hello,
Recently, we find that the cgroup rt_runtime_us's constraints is not
precise enough in some cases. For example:

(1)
create a father cgroup and a child cgroup, and we exec:
   echo 1048577 > /sys/fs/cgroup/cpu/father/cpu.rt_period_us
   echo 1048577 > /sys/fs/cgroup/cpu/father/child/cpu.rt_period_us
   echo 0 > /sys/fs/cgroup/cpu/father/cpu.rt_runtime_us
   echo 1 > /sys/fs/cgroup/cpu/father/child/cpu.rt_runtime_us

(2)
create a father cgroup and two child cgroups, and we exec:
   echo 20000 > /sys/fs/cgroup/cpu/father/cpu.rt_runtime_us
   echo 10000 > /sys/fs/cgroup/cpu/father/child1/cpu.rt_runtime_us
   echo 10001 > /sys/fs/cgroup/cpu/father/child2/cpu.rt_runtime_us

Logically speaking, the sum of child cgroups' rt_runtime_us should be
less than the fater's rt_runtime_us, but actually both cases above would
work. Because in to_ratio(), "div64_u64(runtime << BW_SHIFT, period)"
ignores the remainders. So if the rt_period_us is big or many child
cgroups' remainders are ignored, it could happen.

But after all, it doesn't damage a lot, and seems not so easy to fix. So
I report this and see what can we do about it.

---
Regards
Zhao Wenhui

Re: [BUG REPORT] sched/rt: Inaccurate numerical calculation in rt_runtime_us constraints

[Tim Chen]

On Thu, 2024-07-18 at 21:02 +0800, zhaowenhui (A) wrote:
> Hello,
> Recently, we find that the cgroup rt_runtime_us's constraints is not
> precise enough in some cases. For example:
> 
> (1)
> create a father cgroup and a child cgroup, and we exec:
>    echo 1048577 > /sys/fs/cgroup/cpu/father/cpu.rt_period_us
>    echo 1048577 > /sys/fs/cgroup/cpu/father/child/cpu.rt_period_us
>    echo 0 > /sys/fs/cgroup/cpu/father/cpu.rt_runtime_us
>    echo 1 > /sys/fs/cgroup/cpu/father/child/cpu.rt_runtime_us
> 
> (2)
> create a father cgroup and two child cgroups, and we exec:
>    echo 20000 > /sys/fs/cgroup/cpu/father/cpu.rt_runtime_us
>    echo 10000 > /sys/fs/cgroup/cpu/father/child1/cpu.rt_runtime_us
>    echo 10001 > /sys/fs/cgroup/cpu/father/child2/cpu.rt_runtime_us
> 1048577
> Logically speaking, the sum of child cgroups' rt_runtime_us should be
> less than the fater's rt_runtime_us, but actually both cases above would
> work. Because in to_ratio(), "div64_u64(runtime << BW_SHIFT, period)"
> ignores the remainders. So if the rt_period_us is big or many child
> cgroups' remainders are ignored, it could happen.
> 
> But after all, it doesn't damage a lot, and seems not so easy to fix. So
> I report this and see what can we do about it.

The loss in precision is about 1/(1<<BW_SHIFT), roughly 1 part per million.
So unless you have tens of thousands of rt cgroups, over-allowing 1 part
per million bandwidth per cgroup is probably not a practical concern.

Is there an actual scenario you encountered where this becomes a problem?

Tim
 
> 
> ---
> Regards
> Zhao Wenhui=
>