Re: [PATCH RFC] sched: Disable DL server if sysctl_sched_rt_runtime is -1

[Juri Lelli <juri.lelli@redhat.com>]

Hi Joel,

On 04/03/25 18:47, Joel Fernandes wrote:
> On Tue, Mar 04, 2025 at 03:06:32PM -0500, Steven Rostedt wrote:
> > On Tue,  4 Mar 2025 15:01:16 -0500
> > Joel Fernandes <joelagnelf@nvidia.com> wrote:
> > 
> > > Currently, RCU boost testing in rcutorture is broken because it relies on
> > > having RT throttling disabled. This means the test will always pass (or
> > > rarely fail). This occurs because recently, RT throttling was replaced
> > > by DL server which boosts CFS tasks even when rcutorture tried to
> > > disable throttling (see rcu_torture_disable_rt_throttle()).
> > > 
> > > Therefore this patch prevents DL server from starting when RC torture
> > > sets the sysctl_sched_rt_runtime to -1.
> > > 
> > > With this patch, boosting in TREE09 fails more than 50% of boost attempts
> > > making the test more useful.
> > > 
> > > Also add a check of this to task_non_contending() because otherwise it
> > > throws a warning (in the case when DL server was already started before
> > > rcutorture started).
> > > 
> > 
> > Hmm, I wonder if dl_server caused a regression. That is, disabling rt
> > throttling should allow RT tasks to starve anything it wants. And some RT
> > applications rely on this.
> > 
> > Should this include a Fixes and Cc stable?
> 
> Yeah that makes sense to me, I'll include the Fixes tag in the v2.

Not entirely sure we want to link the (legacy?) sched_rt_runtime
interface to DL server, as it has its own new interface at

/sys/kernel/debug/sched/fair_server/cpuX/*

Admittedly thought the latter is a debug interface, which is not ideal.

I was thinking we might want/need to add a kernel cmdline parameter to
tweak DL server parameters at boot (and possibly disable it), but it is
indeed less flexible than an interface tweakable at runtime.

If we end up using sched_rt_runtime (and _period) to tweak DL server I
believe we should make sure changes are consistent with the debug
interface at least.

Peter, what do you think?

Thanks,
Juri