Re: [PATCH] sched/deadline: Make dl-server nohz full aware

[Frederic Weisbecker <frederic@kernel.org>]

Le Wed, May 13, 2026 at 08:38:46AM +0200, Andrea Righi a écrit :
> Hi Juri,
> 
> On Wed, May 13, 2026 at 08:16:21AM +0200, Juri Lelli wrote:
> > On 12/05/26 17:34, Juri Lelli wrote:
> > > Hi Andrea,
> > > 
> > > On 12/05/26 16:55, Andrea Righi wrote:
> > > > Hi Juri,
> > > 
> > > Thanks from the quick review!
> > > 
> > > > On Tue, May 12, 2026 at 11:02:37AM +0200, Juri Lelli wrote:
> > > > > The dl_server_timer() causes spurious IPIs on nohz_full cores, breaking
> > > > > isolation guarantees. The timer executes on a housekeeping core and
> > > > > eventually calls tick_nohz_dep_set_cpu(), sending IPIs to isolated cores
> > > > > even when only a single task is running.
> > > > > 
> > > > > The problem is that dl-servers are not coordinated with nohz_full tick
> > > > > state. Timers can fire and send IPIs to otherwise undisturbed cores.
> > > > > 
> > > > > Fix by managing servers in sched_can_stop_tick():
> > > > > 
> > > > > - When RT tasks run with CFS/SCX tasks, start the appropriate server
> > > > >   and keep the tick running
> > > > > - When only RT tasks remain, stop all servers and allow tick to stop
> > > > >   (except for >1 RR tasks which need the tick for round-robin)
> > > > > - When only CFS/SCX tasks remain, stop all servers before stopping tick
> > > > > 
> > > > > Introduce dl_servers_stop_all() to reduce duplication and abstract
> > > > > server management from core.c. Unify RT handling into one block that
> > > > > handles both RR and FIFO cases.
> > > > > 
> > > > > Fixes: 557a6bfc662c ("sched/fair: Add trivial fair server")
> > > > > Reported-by: David Haufe <dhaufe@simplextrading.com>
> > > > > Closes: https://lore.kernel.org/lkml/CAKJHwtOw_G67edzuHVtL1xC5Vyt6StcZzihtDd0yaKudW=rwVw@mail.gmail.com
> > > > > Signed-off-by: Juri Lelli <juri.lelli@redhat.com>
> > > > > ---
> > > > > I had to modify my first original attempt at fixing this (please take a
> > > > > look at the linked report/discussion) to also take SCX into
> > > > > consideration.
> > > > 
> > > > As mentioned by Frederic, we don't allow to load BPF schedulers when isolcpus=
> > > > is used, so I think we can simplify the sched_can_stop_tick() part.
> > > 
> > > Right! Thanks for confirming.
> > 
> > Ah, but wait. IIUC SCX is incopatible with isolcpus=domain only?
> > scx_can_stop_tick() seems to confirm we need to take care of it when
> > domain flag is not present.
> > 
> > So, maybe we still need to consider SCX in this patch? e.g. in
> > configurations that are not using static domain isolation, but isolate
> > CPUs by configuring tasks affinities.
> 
> Ah! That's right. SCX is incompatible with isolcpus=domain, but we do support
> nohz_full=..., so I think your original approach is correct. It might be worth
> calling out explicitly in the patch description that the SCX handling targets
> nohz_full, so we don't make the same mistake in the future.

Most of the time (I'm actually tempted to say all the time), nohz_full= doesn't
make sense without isolcpus=domain, or the cpuset equivalents. Because nohz_full
is always about workloads that can't bear a few microseconds tick latency, so
why would these workloads want load balancing, kthreads, etc... ? And in fact
why would such a workload want a scheduler at all?

Thanks.

-- 
Frederic Weisbecker
SUSE Labs