Re: [PATCH net-next] net: Treat __napi_schedule_irqoff() as __napi_schedule() on PREEMPT_RT

[Jakub Kicinski]

On Thu, 13 May 2021 00:28:02 +0200 Thomas Gleixner wrote:
> On Wed, May 12 2021 at 23:43, Sebastian Andrzej Siewior wrote:
> > __napi_schedule_irqoff() is an optimized version of __napi_schedule()
> > which can be used where it is known that interrupts are disabled,
> > e.g. in interrupt-handlers, spin_lock_irq() sections or hrtimer
> > callbacks.
> >
> > On PREEMPT_RT enabled kernels this assumptions is not true. Force-
> > threaded interrupt handlers and spinlocks are not disabling interrupts
> > and the NAPI hrtimer callback is forced into softirq context which runs
> > with interrupts enabled as well.
> >
> > Chasing all usage sites of __napi_schedule_irqoff() is a whack-a-mole
> > game so make __napi_schedule_irqoff() invoke __napi_schedule() for
> > PREEMPT_RT kernels.
> >
> > The callers of ____napi_schedule() in the networking core have been
> > audited and are correct on PREEMPT_RT kernels as well.
> >
> > Reported-by: Juri Lelli <juri.lelli@xxxxxxxxxx>
> > Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>  
> 
> Reviewed-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> 
> > ---
> > Alternatively __napi_schedule_irqoff() could be #ifdef'ed out on RT and
> > an inline provided which invokes __napi_schedule().
> >
> > This was not chosen as it creates #ifdeffery all over the place and with
> > the proposed solution the code reflects the documentation consistently
> > and in one obvious place.  
> 
> Blame me for that decision.
> 
> No matter which variant we end up with, this needs to go into all stable
> RT kernels ASAP.

Mumble mumble. I thought we concluded that drivers used on RT can be
fixed, we've already done it for a couple drivers (by which I mean two).
If all the IRQ handler is doing is scheduling NAPI (which it is for
modern NICs) - IRQF_NO_THREAD seems like the right option.

Is there any driver you care about that we can convert to using
IRQF_NO_THREAD so we can have new drivers to "do the right thing"
while the old ones depend on this workaround for now?


Another thing while I have your attention - ____napi_schedule() does
__raise_softirq_irqoff() which AFAIU does not wake the ksoftirq thread.
On non-RT we get occasional NOHZ warnings when drivers schedule napi
from process context, but on RT this is even more of a problem, right?
ksoftirqd won't run until something else actually wakes it up?