Re: [PATCH] sched/rt: use u64 for RT group bandwidth accounting

[Joseph Salisbury]

Thanks for the feedback, Michal!

On 4/1/26 5:48 AM, Michal Koutný wrote:
> Hi.
>
> On Tue, Mar 17, 2026 at 03:34:02PM -0400, Joseph Salisbury <joseph.salisbury@xxxxxxxxxx> wrote:
> > tg_rt_schedulable() sums the bandwidth ratios of all child RT groups
> > and compares the result against the bandwidth available to the parent.
> >
> > On 32-bit builds, the running sum is stored in unsigned long and can
> > overflow when enough child groups are present.
> As per my calculation this'd require at least 4096 sibling RT groups
> (see to_ratio() and BW_SHIFT)...
>
> > In that case, an overcommitted hierarchy may incorrectly pass the
> > schedulability check.
> ...and I expect there'd other issues with so many so little RT groups
> than allowing theoretically unschedulabile configuration.
You are right.  This is mostly a correctness fix for a corner case. This 
patch primarily addresses the case where the arithmetic can silently 
truncate on 32-bit.
> > Use u64 for the local bandwidth totals so the accumulated child sum and
> > the parent limit are compared without 32-bit overflow.
> You may want to fix this also in to_ratio() return type as there are
> other consumers besides this check. (Although, some conditions would
> still be similarly exotic, e.g. normalize_cfs_quota(), like CFQ quota of
> more than 4096 (32b!) CPUs.)
>
> Thanks,
> Michal
That's a good point.   I can send a v2 that returns u64 from to_ratio() 
and keeps the RT totals in u64 as well.