Re: [RFC PATCH v5 20/29] sched/deadline: Allow deeper hierarchies of RT cgroups

[Tejun Heo]

Hello, Luca.

Yes, the admission rule prevents subtrees from escaping ancestors'
cpu.rt.max, which addresses the main concern.

Two interface simplifications worth considering on top:

1. Drop cpu.rt.min. The parent already owns the partitioning of its
   cpu.rt.max via the children's cpu.rt.max values, so the local share
   is just the leftover. Admission collapses to Sum(children.max.util)
   <= max.util, and the cgroup's own DL server runs at cpu.rt.max's
   period with the remaining utilization. One fewer knob, no
   information lost.

2. Use "max" as the off/inherit sentinel instead of "root". Matches the
   existing cpu.max convention. At the root cgroup, "max" means the
   feature is off and behavior matches today; at non-root cgroups,
   "max" means tasks bubble up to the nearest configured ancestor.
   "root" reads oddly at the root cgroup, where it really just means
   "no DL server".

Auto-revert falls out for free: writing "max" at the root cascades
descendants back to "max", since manual tear-down isn't reachable
anyway (you can't set a leaf to "max" while its parent isn't).

Thanks.

--
tejun