Re: [PATCH netdev] virtio-net: support XDP_TX when not more queues

[Jesper Dangaard Brouer <brouer@redhat.com>]

On Wed, 10 Feb 2021 16:40:41 -0500
"Michael S. Tsirkin" <mst@redhat.com> wrote:

> On Wed, Jan 13, 2021 at 04:08:57PM +0800, Xuan Zhuo wrote:
> > The number of queues implemented by many virtio backends is limited,
> > especially some machines have a large number of CPUs. In this case, it
> > is often impossible to allocate a separate queue for XDP_TX.
> > 
> > This patch allows XDP_TX to run by reuse the existing SQ with
> > __netif_tx_lock() hold when there are not enough queues.

I'm a little puzzled about the choice of using the netdevice TXQ
lock __netif_tx_lock() / __netif_tx_unlock().
Can you explain more about this choice?

> > 
> > Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
> > Reviewed-by: Dust Li <dust.li@linux.alibaba.com>  
> 
> I'd like to get some advice on whether this is ok from some
> XDP experts - previously my understanding was that it is
> preferable to disable XDP for such devices than
> use locks on XDP fast path.

I think it is acceptable, because the ndo_xdp_xmit / virtnet_xdp_xmit
takes a bulk of packets (currently 16).

Some drivers already does this.

It would have been nice if we could set a feature flag, that allow
users to see that this driver uses locking in the XDP transmit
(ndo_xdp_xmit) function call... but it seems like a pipe dream :-P

Code related to the locking

> > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> > index ba8e637..7a3b2a7 100644
> > --- a/drivers/net/virtio_net.c
> > +++ b/drivers/net/virtio_net.c
[...]
> > @@ -481,14 +484,34 @@ static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
> >  	return 0;
> >  }
> >  
> > -static struct send_queue *virtnet_xdp_sq(struct virtnet_info *vi)
> > +static struct send_queue *virtnet_get_xdp_sq(struct virtnet_info *vi)
> >  {
> >  	unsigned int qp;
> > +	struct netdev_queue *txq;
> > +
> > +	if (vi->curr_queue_pairs > nr_cpu_ids) {
> > +		qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
> > +	} else {
> > +		qp = smp_processor_id() % vi->curr_queue_pairs;
> > +		txq = netdev_get_tx_queue(vi->dev, qp);
> > +		__netif_tx_lock(txq, raw_smp_processor_id());
> > +	}
> >  
> > -	qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
> >  	return &vi->sq[qp];
> >  }
> >  
> > +static void virtnet_put_xdp_sq(struct virtnet_info *vi)
> > +{
> > +	unsigned int qp;
> > +	struct netdev_queue *txq;
> > +
> > +	if (vi->curr_queue_pairs <= nr_cpu_ids) {
> > +		qp = smp_processor_id() % vi->curr_queue_pairs;
> > +		txq = netdev_get_tx_queue(vi->dev, qp);
> > +		__netif_tx_unlock(txq);
> > +	}
> > +}
> > +
> >  static int virtnet_xdp_xmit(struct net_device *dev,
> >  			    int n, struct xdp_frame **frames, u32 flags)
> >  {
> > @@ -512,7 +535,7 @@ static int virtnet_xdp_xmit(struct net_device *dev,
> >  	if (!xdp_prog)
> >  		return -ENXIO;
> >  
> > -	sq = virtnet_xdp_sq(vi);
> > +	sq = virtnet_get_xdp_sq(vi);
> >  
> >  	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
> >  		ret = -EINVAL;
> > @@ -560,12 +583,13 @@ static int virtnet_xdp_xmit(struct net_device *dev,
> >  	sq->stats.kicks += kicks;
> >  	u64_stats_update_end(&sq->stats.syncp);
> >  
> > +	virtnet_put_xdp_sq(vi);
> >  	return ret;
> >  }
> >  



-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer