Re: [PATCH 2/5] sched/core: Keep a sibling's DL server link in the core-wide pick
From: Peter Zijlstra
Date: Wed Nov 18 2026 - 08:58:23 EST
On Wed, Nov 18, 2026 at 02:50:52PM +0100, Peter Zijlstra wrote:
> On Thu, Nov 12, 2026 at 06:15:45PM +0200, Fernand Sieber wrote:
> > The core-wide selection records in rq_i->core_dl_server the DL server
> > each sibling's pick came through, then clears it for every sibling in
> > the cookie matching loop, including when the pick is kept. The sibling
> > takes the fast path, which sets rq->dl_server from rq->core_dl_server,
> > and runs the task its server picked with p->dl_server NULL. Only the
> > selecting CPU keeps the link, as its pick goes through rq->dl_server.
> >
> > A running task is ranked for core scheduling by p->dl_server. While
> > such a task runs on a sibling, each following core-wide selection
> > ranks it as a plain fair or ext task rather than by its server's
> > deadline, so a higher class task on another sibling takes the core
> > back before the server's runtime is used. The RV nomiss monitor, which
> > follows the server through next->dl_server at sched_switch, does not
> > see the server run either.
> >
> > Keep the server that pick_task() left in rq_i->dl_server. With the
> > previous patch it is NULL when the class pick was replaced. Drop the
> > assignment in the first loop, which the matching loop always
> > overwrites.
> >
> > Fixes: bd9bbc96e835 ("sched: Rework dl_server")
> > Signed-off-by: Fernand Sieber <sieberf@xxxxxxxxxx>
> > ---
> > kernel/sched/core.c | 3 +--
> > 1 file changed, 1 insertion(+), 2 deletions(-)
> >
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index f7ac13186ee59..61d4e6e990bba 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -6431,7 +6431,6 @@ pick_next_task(struct rq *rq, struct rq_flags *rf)
> > }
> >
> > rq_i->core_pick = p;
> > - rq_i->core_dl_server = rq_i->dl_server;
> >
> > if (!max || prio_less(max, p, fi_before))
> > max = p;
> > @@ -6462,7 +6461,7 @@ pick_next_task(struct rq *rq, struct rq_flags *rf)
> > }
> >
> > rq_i->core_pick = p;
> > - rq_i->core_dl_server = NULL;
> > + rq_i->core_dl_server = rq_i->dl_server;
> >
> > if (p == rq_i->idle) {
> > if (rq_i->nr_running) {
>
> Isn't this broken in the face of a rq->lock drop through pick_task() ?
>
> Consider this does:
>
> SMT0: pick_task()
> SMT1: pick_task()
> drop rq->lock
> newidle
>
> SMT0: rq_0->core_dl_server = rq_0->dl_server;
>
>
> I am reworking all of that somewhere, but delaying saving rq->dl_server
> like this feels fragile.
>
> Can't we fold the first two patches into something like so?
>
> ---
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 175b52a7ea79..68b980ddf88f 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -6453,10 +6453,16 @@ pick_next_task(struct rq *rq, struct rq_flags *rf)
> p = sched_core_find(rq_i, cookie);
> if (!p)
> p = idle_sched_class.pick_task(rq_i, rf);
> +
> + /*
> + * Neither a core tree pick nor idle comes through a
> + * DL server; drop the one the replaced pick left.
> + */
> + rq_i->dl_server = NULL;
> }
>
> rq_i->core_pick = p;
> - rq_i->core_dl_server = NULL;
> + rq_i->core_dl_server = rq_i->dl_server;
No that isn't quite right either, perhaps just make it:
rq_i->core_dl_server = rq_i->dl_server = NULL;