svn commit: r256066 - head/sys/dev/virtio/network

[Bryan Venteicher]

On Sat, Oct 5, 2013 at 4:12 PM, Alexander V. Chernikov <melifaro at freebsd.org
> wrote:

> On 05.10.2013 22:07, Bryan Venteicher wrote:
> > Author: bryanv
> > Date: Sat Oct  5 18:07:24 2013
> > New Revision: 256066
> > URL: http://svnweb.freebsd.org/changeset/base/256066
> >
> > Log:
> >   Do not hold the vtnet Rx queue lock when calling up into the stack
> Do you measure performance penalty for this?

>
>

I wasn't able to measure anything, but I don't have the most update
hardware either. There should hardly ever be any contention and it should
be in its own cacheline, so it is about as cheap as a mutex acquire can get.


> >   This matches other similar drivers and avoids various LOR warnings.
> We're currently trying to eliminate such things in other drivers like
> ixgbe, maybe there can be some other way to eliminate possible LORs?
> >
>

Agreed - I'd prefer not to do this and that's why I originally had it
commented out. I don't think witness was just reporting a false positives
either. vtnet's locking scheme is very similar to ixgbe, so fixes ixgbe
should translate to vtnet.


>  >   Approved by:        re (marius)
> >
> > Modified:
> >   head/sys/dev/virtio/network/if_vtnet.c
> >
> > Modified: head/sys/dev/virtio/network/if_vtnet.c
> >
> ==============================================================================
> > --- head/sys/dev/virtio/network/if_vtnet.c    Sat Oct  5 16:22:33 2013
>      (r256065)
> > +++ head/sys/dev/virtio/network/if_vtnet.c    Sat Oct  5 18:07:24 2013
>      (r256066)
> > @@ -1700,9 +1700,9 @@ vtnet_rxq_input(struct vtnet_rxq *rxq, s
> >       rxq->vtnrx_stats.vrxs_ipackets++;
> >       rxq->vtnrx_stats.vrxs_ibytes += m->m_pkthdr.len;
> >
> > -     /* VTNET_RXQ_UNLOCK(rxq); */
> > +     VTNET_RXQ_UNLOCK(rxq);
> >       (*ifp->if_input)(ifp, m);
> > -     /* VTNET_RXQ_LOCK(rxq); */
> > +     VTNET_RXQ_LOCK(rxq);
> >  }
> >
> >  static int
> > @@ -1782,6 +1782,10 @@ vtnet_rxq_eof(struct vtnet_rxq *rxq)
> >               m_adj(m, adjsz);
> >
> >               vtnet_rxq_input(rxq, m, hdr);
> > +
> > +             /* Must recheck after dropping the Rx lock. */
> > +             if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0)
> > +                     break;
> >       }
> >
> >       if (deq > 0)
> >
>
>
>