On Mon, 14 Mar 2022 09:09:49 -0600
Kristof Provost <k...@freebsd.org> wrote:
> On 14 Mar 2022, at 7:44, Michael Gmelin wrote:
> > On Sun, 13 Mar 2022 17:53:44 +0000
> > "Bjoern A. Zeeb" <bzeeb-li...@lists.zabbadoz.net> wrote:
> >
> >> On 13 Mar 2022, at 17:45, Michael Gmelin wrote:
> >>
> >>>> On 13. Mar 2022, at 18:16, Bjoern A. Zeeb
> >>>> <bzeeb-li...@lists.zabbadoz.net> wrote:
> >>>>
> >>>> On 13 Mar 2022, at 16:33, Michael Gmelin wrote:
> >>>>> It's important to point out that this only happens with
> >>>>> kern.ncpu>1. With kern.ncpu==1 nothing gets stuck.
> >>>>>
> >>>>> This perfectly fits into the picture, since, as pointed out by
> >>>>> Johan,
> >>>>> the first commit that is affected[0] is about multicore
> >>>>> support.
> >>>>
> >>>> Ignore my ignorance, what is the default of net.isr.maxthreads
> >>>> and net.isr.bindthreads (in stable/13) these days?
> >>>>
> >>>
> >>> My tests were on CURRENT and I’m afk, but according to cgit[0][1],
> >>> max is 1 and bind is 0.
> >>>
> >>> Would it make sense to repeat the test with max=-1?
> >>
> >> I’d say yes, I’d also bind, but that’s just me.
> >>
> >> I would almost assume Kristof running with -1 by default (but he
> >> can chime in on that).
> >
> > I tried various configuration permutations, all with ncpu=2:
> >
> > - 14.0-CURRENT #0 main-n253697-f1d450ddee6
> > - 13.1-BETA1 #0 releng/13.1-n249974-ad329796bdb
> > - net.isr.maxthreads: -1 (which results in 2 threads), 1, 2
> > - net.isr.bindthreads: -1, 0, 1, 2
> > - net.isr.dispatch: direct, deferred
> >
> > All resulting in the same behavior (hang after a few seconds). They
> > all
> > work ok when running on a single core instance (threads=1 in this
> > case).
> >
> > I also ran the same test on 13.0-RELEASE-p7 for
> > comparison (unsurprisingly, it's ok).
> >
> > I placed the script to reproduce the issue on freefall for your
> > convenience, so running it is as simple as:
> >
> > fetch https://people.freebsd.org/~grembo/hang_epair.sh
> > # inspect content
> > sh hang_epair.sh
> >
> > or, if you feel lucky
> >
> > fetch -o - https://people.freebsd.org/~grembo/hang_epair.sh | sh
> >
> With that script I can also reproduce the problem.
>
> I’ve experimented with this hack:
>
> diff --git a/sys/net/if_epair.c b/sys/net/if_epair.c
> index c39434b31b9f..1e6bb07ccc4e 100644
> --- a/sys/net/if_epair.c
> +++ b/sys/net/if_epair.c
> @@ -415,7 +415,10 @@ epair_ioctl(struct ifnet *ifp, u_long
> cmd, caddr_t data)
>
> case SIOCSIFMEDIA:
> case SIOCGIFMEDIA:
> + printf("KP: %s() SIOCGIFMEDIA\n", __func__);
> sc = ifp->if_softc;
> + taskqueue_enqueue(epair_tasks.tq[0],
> &sc->queues[0].tx_task);
> +
> error = ifmedia_ioctl(ifp, ifr, &sc->media,
> cmd); break;
>
> That kicks the receive code whenever I `ifconfig epair0a`, and I see
> a little more traffic every time I do so.
> That suggests pretty strongly that there’s an issue with how we
> dispatch work to the handler thread. So presumably there’s a race
> between epair_menq() and epair_tx_start_deferred().
>
> epair_menq() tries to only enqueue the receive work if there’s
> nothing in the buf_ring, on the grounds that if there is the previous
> packet scheduled the work. Clearly there’s an issue there.
>
> I’ll try to dig into that in the next few days.
>
Hi Kristof,
This sounds plausible. I spent a few hours getting familiar with the
epair code and came up with a patch that seems to fix the issue at hand
(both with and without RSS). I'm not certain that it is a good
solution, especially in terms of performance, but I wanted to share it
with you anyway, maybe it helps:
https://people.freebsd.org/~grembo/epair.patch
Best
Michael
--
Michael Gmelin
