Linux Virtualization - Apr 2023 - [PATCH net] virtio-net: reject small vring sizes

> > > > > Actually, I think that all you need to do is disable
NETIF_F_SG,
> > > > > and things will work, no?
> > > >
> > > > I think that this is not so simple, if I understand
correctly, by disabling NETIF_F_SG we will never receive a chained skbs to
transmit, but we still have more functionality to address, for example:
> > > > * The TX timeouts.
> > >
> > > I don't get it. With a linear skb we can transmit it as long
as there's
> > > space for 2 entries in the vq: header and data. What's the
source of the
> > > timeouts?
> > >
> >
> > I'm not saying that this is not possible, I meant that we need
more changes to virtio-net.
> > The source of the timeouts is from the current implementation of
virtnet_poll_tx.
> >
> > if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
> >       netif_tx_wake_queue(txq);
> 
> Oh right. So this should check NETIF_F_SG then.
> BTW both ring size and s/g can be tweaked by ethtool, also
> needs handling.
> 
Good point.
> >
> > > > * Guest GSO/big MTU (without VIRTIO_NET_F_MRG_RXBUF?), we
can't chain page size buffers anymore.
> > >
> > > I think we can.  mergeable_min_buf_len will just be large.
> > >
> >
> > I meant that we can't just by clearing NETIF_F_SG, we'll need
to change virtio-net a little bit more, for example, the virtnet_set_big_packets
function.
> >
> 
> Right - for RX, big_packets_num_skbfrags ignores ring size and that's
> probably a bug if mtu is very large.
> 
So, what do you think, we should fix virtio-net to work with smaller rings? we
should fail probe?

I think that since this never came up until now, there is no big demand to such
small rings.

On Mon, Apr 17, 2023 at 07:33:58AM +0000, Alvaro Karsz
wrote:
> > > > > > Actually, I think that all you need to do is
disable NETIF_F_SG,
> > > > > > and things will work, no?
> > > > >
> > > > > I think that this is not so simple, if I understand
correctly, by disabling NETIF_F_SG we will never receive a chained skbs to
transmit, but we still have more functionality to address, for example:
> > > > > * The TX timeouts.
> > > >
> > > > I don't get it. With a linear skb we can transmit it as
long as there's
> > > > space for 2 entries in the vq: header and data. What's
the source of the
> > > > timeouts?
> > > >
> > >
> > > I'm not saying that this is not possible, I meant that we
need more changes to virtio-net.
> > > The source of the timeouts is from the current implementation of
virtnet_poll_tx.
> > >
> > > if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
> > >       netif_tx_wake_queue(txq);
> > 
> > Oh right. So this should check NETIF_F_SG then.
> > BTW both ring size and s/g can be tweaked by ethtool, also
> > needs handling.
> > 
> 
> Good point.
> 
> > >
> > > > > * Guest GSO/big MTU (without VIRTIO_NET_F_MRG_RXBUF?),
we can't chain page size buffers anymore.
> > > >
> > > > I think we can.  mergeable_min_buf_len will just be large.
> > > >
> > >
> > > I meant that we can't just by clearing NETIF_F_SG, we'll
need to change virtio-net a little bit more, for example, the
virtnet_set_big_packets function.
> > >
> > 
> > Right - for RX, big_packets_num_skbfrags ignores ring size and
that's
> > probably a bug if mtu is very large.
> > 
> 
> So, what do you think, we should fix virtio-net to work with smaller rings?
we should fail probe?
> 
> I think that since this never came up until now, there is no big demand to
such small rings.
The worry is that once we start failing probe there's just a tiny chance
hosts begin to rely on us failing probe then we won't be able to fix it.
So it depends on the size of the patch I think. So far it seems small enough
that wasting code on failing probe isn't worth it.

-- 
MST