Re: [PATCH RFC 2/2] veth: propagate bridge GSO to peer

[Solio Sarabia <solio.sarabia@intel.com>]

On Mon, Nov 27, 2017 at 01:15:02PM -0800, Stephen Hemminger wrote:
> On Mon, 27 Nov 2017 12:14:19 -0800
> Solio Sarabia <solio.sarabia@intel.com> wrote:
> 
> > On Sun, Nov 26, 2017 at 11:07:25PM -0800, Stephen Hemminger wrote:
> > > On Sun, 26 Nov 2017 20:13:39 -0700
> > > David Ahern <dsahern@gmail.com> wrote:
> > >   
> > > > On 11/26/17 11:17 AM, Stephen Hemminger wrote:  
> > > > > This allows veth device in containers to see the GSO maximum
> > > > > settings of the actual device being used for output.  
> > > > 
> > > > veth devices can be added to a VRF instead of a bridge, and I do not
> > > > believe the gso propagation works for L3 master devices.
> > > > 
> > > > From a quick grep, team devices do not appear to handle gso changes either.  
> > > 
> > > This code should still work correctly, but no optimization would happen.
> > > The gso_max_size of the VRF or team will
> > > still be GSO_MAX_SIZE so there would be no change. If VRF or Team ever got smart
> > > enough to handle GSO limits, then the algorithm would handle it.  
> > 
> > This patch propagates gso value from bridge to its veth endpoints.
> > However, since bridge is never aware of the GSO limit from underlying
> > interfaces, bridge/veth still have larger GSO size.
> > 
> > In the docker case, bridge is not linked directly to physical or
> > synthetic interfaces; it relies on iptables to decide which interface to
> > forward packets to.
> 
> So for the docker case, then direct control of GSO values via netlink (ie ip link set)
> seems like the better solution.

Adding ioctl support for 'ip link set' would work. I'm still concerned
how to enforce the upper limit to not exceed that of the lower devices.

Consider a system with three NICs, each reporting values in the range
[60,000 - 62,780]. Users could set virtual interfaces' gso to 65,536,
exceeding the limit, and having the host do sw gso (vms settings must
not affect host performance.)

Looping through interfaces?  With the difference that now it'd be
trigger upon user's request, not every time a veth is created (like one
previous patch discussed.)