On Wed, May 06, 2020 at 12:30:52AM +0200, Thomas Gleixner wrote:
> "Michael S. Tsirkin" <m...@redhat.com> writes:
> > On Tue, May 05, 2020 at 02:08:56PM +0200, Thomas Gleixner wrote:
> >>
> >> The following lockdep splat happens reproducibly on 5.7-rc4
> >
> >> ================================
> >> WARNING: inconsistent lock state
> >> 5.7.0-rc4+ #79 Not tainted
> >> --------------------------------
> >> inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W} usage.
> >> ip/356 [HC0[0]:SC1[1]:HE1:SE0] takes:
> >> f3ee4cd8 (&syncp->seq#2){+.?.}-{0:0}, at: net_rx_action+0xfb/0x390
> >> {SOFTIRQ-ON-W} state was registered at:
> >> lock_acquire+0x82/0x300
> >> try_fill_recv+0x39f/0x590
> >
> > Weird. Where does try_fill_recv acquire any locks?
>
> u64_stats_update_begin(&rq->stats.syncp);
>
> That's a 32bit kernel which uses a seqcount for this. sequence counts
> are "lock" constructs where you need to make sure that writers are
> serialized.
>
> Actually the problem at hand is that try_fill_recv() is called from
> fully preemptible context initialy and then from softirq context.
>
> Obviously that's for the open() path a non issue, but lockdep does not
> know about that. OTOH, there is other code which calls that from
> non-softirq context.
Well to be more specific, try_fill_recv is either called from napi,
or from preemptible contexts with napi disabled and rtnl lock taken.
Two try_fill_recv calls in parallel would cause ring corruption
and no end of mischief, we certainly don't want that even on 64 bit
kernels.
> The hack below made it shut up. It's obvioulsy not ideal, but at least
> it let me look at the actual problem I was chasing down :)
>
> Thanks,
>
> tglx
I'll post a hack based on Eric's suggestion since that is at least
0-overhead on 64 bit. It would be nice if we could teach
lockdep that it's about napi...
> 8<-----------
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -1243,9 +1243,11 @@ static bool try_fill_recv(struct virtnet
> break;
> } while (rq->vq->num_free);
> if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
> + local_bh_disable();
> u64_stats_update_begin(&rq->stats.syncp);
> rq->stats.kicks++;
> u64_stats_update_end(&rq->stats.syncp);
> + local_bh_enable();
> }
>
> return !oom;
