Re: [PATCH 11/23] sched,spi: Convert to sched_set_fifo*()

[Guenter Roeck]

On Wed, Apr 22, 2020 at 9:41 AM Doug Anderson <dianders@xxxxxxxxxxxx> wrote:
>
> Hi,
>
> On Wed, Apr 22, 2020 at 8:48 AM Guenter Roeck <groeck@xxxxxxxxxx> wrote:
> >
> > On Wed, Apr 22, 2020 at 7:35 AM Doug Anderson <dianders@xxxxxxxxxxxx> wrote:
> > >
> > > Hi,
> > >
> > > On Wed, Apr 22, 2020 at 6:56 AM Mark Brown <broonie@xxxxxxxxxx> wrote:
> > > >
> > > > On Wed, Apr 22, 2020 at 01:27:30PM +0200, Peter Zijlstra wrote:
> > > > > Because SCHED_FIFO is a broken scheduler model (see previous patches)
> > > > > take away the priority field, the kernel can't possibly make an
> > > > > informed decision.
> > > > >
> > > > > No effective change.
> > > >
> > > > Copying Doug who did this change and Guenter who reviewed it.  This
> > > > looks fine to me but I've no particular involvement with the code or
> > > > platforms that are affected here.
> > >
> > > Thanks!  Probably the maintainers of cros_ec_spi.c (Benson and Enric)
> > > should be aware of it, too.  CCing them.
> > >
> > > From my point of view, my response is pretty much identical to the one
> > > I wrote when the priority was reduced from "MAX_RT_PRIO - 1" to
> > > "MAX_RT_PRIO / 2" [1].  Basically, any priority that keeps us from
> > > being preempted by tasks that are only high priority for performance
> > > reasons (like dm crypt and loopback did when I last analyzed) is fine.
> > > Our priority needs to be high not for performance reasons but for
> > > correctness reasons (the other side will drop our data if we don't
> > > respond in time).
> > >
> > The crypto engine ends up running at the same priority level, so I am
> > a bit concerned that this patch series will re-introduce the problem
> > that Doug's initial patch tried to solve.
>
> Do you have a pointer to the code you're looking at?  Digging through

I was looking at crypto/crypto_engine.c:crypto_engine_alloc_init().
Maybe that is different code and I misunderstand its use, though.

Guenter

> my old investigation for dm-crypt showed the problem to be the code
> touched by commit a1b89132dc4f ("dm crypt: use
> WQ_HIGHPRI for the IO and crypt workqueues").  Interestingly enough,
> that's been reverted in commit f612b2132db5 ("Revert "dm crypt: use
> WQ_HIGHPRI for the IO and crypt workqueues"").
>
> Even if something is using WQ_HIGHPRI, last I checked WQ_HIGHPRI was
> the highest non-relatime priority.  Looking quickly I see
> "HIGHPRI_NICE_LEVEL" which is MIN_NICE.  I don't think that implies
> realtime, but I assume sched_fifo() still does.
>
> -Doug