Re: [RFC PATCH 1/3] veth: implement support for set_channel ethtool op

["Toke Høiland-Jørgensen" <toke@redhat.com>]

Paolo Abeni <pabeni@redhat.com> writes:

> On Fri, 2021-07-09 at 12:49 +0200, Paolo Abeni wrote:
>> On Fri, 2021-07-09 at 12:15 +0200, Toke Høiland-Jørgensen wrote:
>> > > +	if (netif_running(dev))
>> > > +		veth_close(dev);
>> > > +
>> > > +	priv->num_tx_queues = ch->tx_count;
>> > > +	priv->num_rx_queues = ch->rx_count;
>> > 
>> > Why can't just you use netif_set_real_num_*_queues() here directly (and
>> > get rid of the priv members as above)?
>> 
>> Uhm... I haven't thought about it. Let me try ;)
>
> Here there is a possible problem: if the latter
> netif_set_real_num_*_queues() fails, we should not change the current
> configuration, so we should revert the
> first netif_set_real_num_*_queues() change.
>
> Even that additional revert operation could fail. If/when that happen
> set_channel() will leave the device in a different state from both the
> old one and the new one, possibly with an XDP-incompatible number of
> queues.
>
> Keeping the  netif_set_real_num_*_queues() calls in veth_open() avoid
> the above issue: if the queue creation is problematic, the device will
> stay down.
>
> I think the additional fields are worthy, WDYT?

Hmm, wouldn't the right thing to do be to back out the change and return
an error to userspace? Something like:

+	if (netif_running(dev))
+		veth_close(dev);
+
+	old_rx_queues = dev->real_num_rx_queues;
+	err = netif_set_real_num_rx_queues(dev, ch->rx_count);
+	if (err)
+		return err;
+
+	err = netif_set_real_num_tx_queues(dev, ch->tx_count);
+	if (err) {
+		netif_set_real_num_rx_queues(dev, old_rx_queues);
+		return err;
+	}
+
+	if (netif_running(dev))
+		veth_open(dev);
+	return 0;


(also, shouldn't the result of veth_open() be returned? bit weird if you
don't get an error but the device stays down...)

-Toke