Re: [PATCH RFC net-next 2/2] drivers: net: virtio_net: Add tx_timeout function

[Julio Faracco]

Em seg, 7 de out de 2019 Ãs 11:03, Julio Faracco <jcfaracco@xxxxxxxxx> escreveu:
>
> Em seg, 7 de out de 2019 Ãs 04:51, Michael S. Tsirkin <mst@xxxxxxxxxx> escreveu:
> >
> > On Sun, Oct 06, 2019 at 03:45:15PM -0300, jcfaracco@xxxxxxxxx wrote:
> > > From: Julio Faracco <jcfaracco@xxxxxxxxx>
> > >
> > > To enable dev_watchdog, virtio_net should have a tx_timeout defined
> > > (.ndo_tx_timeout). This is only a skeleton to throw a warn message. It
> > > notifies the event in some specific queue of device. This function
> > > still counts tx_timeout statistic and consider this event as an error
> > > (one error per queue), reporting it.
> > >
> > > Signed-off-by: Julio Faracco <jcfaracco@xxxxxxxxx>
> > > Signed-off-by: Daiane Mendes <dnmendes76@xxxxxxxxx>
> > > Cc: Jason Wang <jasowang@xxxxxxxxxx>
> > > ---
> > >  drivers/net/virtio_net.c | 27 +++++++++++++++++++++++++++
> > >  1 file changed, 27 insertions(+)
> > >
> > > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> > > index 27f9b212c9f5..4b703b4b9441 100644
> > > --- a/drivers/net/virtio_net.c
> > > +++ b/drivers/net/virtio_net.c
> > > @@ -2585,6 +2585,29 @@ static int virtnet_set_features(struct net_device *dev,
> > >       return 0;
> > >  }
> > >
> > > +static void virtnet_tx_timeout(struct net_device *dev)
> > > +{
> > > +     struct virtnet_info *vi = netdev_priv(dev);
> > > +     u32 i;
> > > +
> > > +     /* find the stopped queue the same way dev_watchdog() does */
> >
> > not really - the watchdog actually looks at trans_start.
>
> The comments are wrong. It is the negative logic from dev_watchdog.
> Watchdog requires queue stopped AND timeout.
>
> If the queue is not stopped, this queue does not reached a timeout event.
> So, continue... Do not report a timeout.
>
> >
> > > +     for (i = 0; i < vi->curr_queue_pairs; i++) {
> > > +             struct send_queue *sq = &vi->sq[i];
> > > +
> > > +             if (!netif_xmit_stopped(netdev_get_tx_queue(dev, i)))
> > > +                     continue;
> > > +
> > > +             u64_stats_update_begin(&sq->stats.syncp);
> > > +             sq->stats.tx_timeouts++;
> > > +             u64_stats_update_end(&sq->stats.syncp);
> > > +
> > > +             netdev_warn(dev, "TX timeout on send queue: %d, sq: %s, vq: %d, name: %s\n",
> > > +                         i, sq->name, sq->vq->index, sq->vq->name);
> >
> > this seems to assume any running queue is timed out.
> > doesn't look right.
> >
> > also - there's already a warning in this case in the core. do we need another one?
>
> Here, it can be a debug message if the idea is enhance debugging information.
> Other enhancements can be done to enable or disable debug messages.
> Using ethtool methods for instance.

Observation...
Another important point, kernel will thrown WARN_ONCE, only if
ndo_tx_timeout() is implemented.
Even if we are adding an extra/unnecessary netdev_warn() we need this
function to enable dev_watchdog().

>
> >
> > > +             dev->stats.tx_errors++;
> >
> >
> >
> > > +     }
> > > +}
> > > +
> > >  static const struct net_device_ops virtnet_netdev = {
> > >       .ndo_open            = virtnet_open,
> > >       .ndo_stop            = virtnet_close,
> > > @@ -2600,6 +2623,7 @@ static const struct net_device_ops virtnet_netdev = {
> > >       .ndo_features_check     = passthru_features_check,
> > >       .ndo_get_phys_port_name = virtnet_get_phys_port_name,
> > >       .ndo_set_features       = virtnet_set_features,
> > > +     .ndo_tx_timeout         = virtnet_tx_timeout,
> > >  };
> > >
> > >  static void virtnet_config_changed_work(struct work_struct *work)
> > > @@ -3018,6 +3042,9 @@ static int virtnet_probe(struct virtio_device *vdev)
> > >       dev->netdev_ops = &virtnet_netdev;
> > >       dev->features = NETIF_F_HIGHDMA;
> > >
> > > +     /* Set up dev_watchdog cycle. */
> > > +     dev->watchdog_timeo = 5 * HZ;
> > > +
> >
> > Seems to be still broken with napi_tx = false.
> >
> > >       dev->ethtool_ops = &virtnet_ethtool_ops;
> > >       SET_NETDEV_DEV(dev, &vdev->dev);
> > >
> > > --
> > > 2.21.0