Re: virtio_net: can change MTU after installing program

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, Feb 26, 2020 at 05:30:18PM +0800, Jason Wang wrote:
> 
> On 2020/2/26 下午4:39, Michael S. Tsirkin wrote:
> > On Wed, Feb 26, 2020 at 02:37:01AM -0500, Jason Wang wrote:
> > > 
> > > ----- Original Message -----
> > > > On Tue, Feb 25, 2020 at 08:32:14PM -0700, David Ahern wrote:
> > > > > Another issue is that virtio_net checks the MTU when a program is
> > > > > installed, but does not restrict an MTU change after:
> > > > > 
> > > > > # ip li sh dev eth0
> > > > > 2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 xdp qdisc fq_codel
> > > > > state UP mode DEFAULT group default qlen 1000
> > > > >      link/ether 5a:39:e6:01:a5:36 brd ff:ff:ff:ff:ff:ff
> > > > >      prog/xdp id 13 tag c5595e4590d58063 jited
> > > > > 
> > > > > # ip li set dev eth0 mtu 8192
> > > > > 
> > > > > # ip li sh dev eth0
> > > > > 2: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 8192 xdp qdisc fq_codel
> > > > > state UP mode DEFAULT group default qlen 1000
> > > > Well the reason XDP wants to limit MTU is this:
> > > >      the MTU must be less than a page
> > > >      size to avoid having to handle XDP across multiple pages
> > > > 
> > > But even if we limit MTU is guest there's no way to limit the packet
> > > size on host.
> > Isn't this fundamental? IIUC dev->mtu is mostly a hint to devices about
> > how the network is configured. It has to be the same across LAN.  If
> > someone misconfigures it that breaks networking, and user gets to keep
> > both pieces. E.g. e1000 will use dev->mtu to calculate rx buffer size.
> > If you make it too small, well packets that are too big get dropped.
> > There's no magic to somehow make them smaller, or anything like that.
> > We can certainly drop packet > dev->mtu in the driver right now if we want to,
> > and maybe if it somehow becomes important for performance, we
> > could teach host to drop such packets for us. Though
> > I don't really see why we care ...
> > 
> > > It looks to me we need to introduce new commands to
> > > change the backend MTU (e.g TAP) accordingly.
> > > 
> > > Thanks
> > So you are saying there are configurations where host does not know the
> > correct MTU, and needs guest's help to figure it out?
> 
> 
> Yes.
> 
> 
> > I guess it's
> > possible but it seems beside the point raised here.  TAP in particular
> > mostly just seems to ignore MTU, I am not sure why we should bother
> > propagating it there from guest or host. Propagating it from guest to
> > the actual NIC might be useful e.g. for buffer sizing, but is tricky
> > to do safely in case the NIC is shared between VMs.
> 
> 
> Macvlan passthrough mode could be easier I guess.
> 
> Thanks

As usual :) So sure, it's doable for simple configs.
But this seems orthogoal to the question raised in this thread.

-- 
MST