Re: [tip:sched/fifo 44/45] ERROR: modpost: "sched_setscheduler" undefined!

[Qais Yousef]

On 07/29/20 12:23, Dietmar Eggemann wrote:
> On 21/07/2020 12:13, Qais Yousef wrote:
> > On 07/21/20 10:36, peterz@xxxxxxxxxxxxx wrote:
> >> On Mon, Jul 20, 2020 at 06:19:43PM -0400, Steven Rostedt wrote:
> >>> On Mon, 20 Jul 2020 23:49:18 +0200
> >>> Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >>>
> >>>> Steve, would this work for you, or would you prefer renaming the
> >>>> parameters as well?
> >>>>
> >>>
> >>> Yeah, that's fine. You don't have any sched_fifo_high() ?
> >>
> >> Thanks! and no.
> >>
> >> I'll go write a Changelog and add it to tip/sched/fifo, so that
> >> hopefully, sfr can stop complaining about this build fail ;-)
> >>
> >> I've even argued we should rename fifo_low() to something else, but
> >> failed to come up with a sensible name. The intended case is for when
> >> you want something above normal but don't particularly care about RT at
> >> all.
> >>
> >> The thing is, once you start adding priorities, even low,med,high, we're
> >> back to where we were. And the whole argument is that the kernel cannot
> >> set priorities in any sensible fashion.
> > 
> > Agreed. I am worried about in-kernel users setting random uclamp values too.
> 
> Do we really have to restrict in-kernel user?
> 
> And avoiding module uclamp abuse is covered by 616d91b68cd5 ("sched:
> Remove sched_setscheduler*() EXPORTs").

The worry is not just about modules abuse IMO. We can put a filter in our
emails to catch all patches that try to use this API. I don't think we can
assume we'd catch all.

> 
> > This series should do most of the work but there are more pieces needed on-top.
> > 
> > From what I see we still need to move the sched_setscheduler() from
> > include/linux/sched.h to kernel/sched/sched.h. And sched_setattr() too. The
> > latter has a single user in kernel/trace/trace_selftest.c to create a deadline
> > task. I think that can be easily wrapped with a similar sched_set_dl()
> > function and exported instead.
> 
> But DL does not have the same issue like the FIFO/RR when it comes to
> resource management.
> Not sure if we have to restrict in-kernel user.

I didn't think much about it. But we can relax the wrapper if really needed.
IMO the kernel should present a predictable behavior for userspace. But I don't
know a lot about DL to comment. The easy answer the wrapper could be relaxed to
offer the required tunables without giving direct access to
sched_setscheduler().

Thanks

--
Qais Yousef