Re: [PATCH net] net: vlan: allow vlan device MTU change follow real device from smaller to bigger

[Guillaume Nault]

On Wed, Feb 23, 2022 at 07:17:36AM -0800, Stephen Hemminger wrote:
> On Wed, 23 Feb 2022 12:26:18 +0100
> Guillaume Nault <gnault@xxxxxxxxxx> wrote:
> 
> > On Tue, Feb 22, 2022 at 03:28:15PM -0800, Jakub Kicinski wrote:
> > > On Tue, 22 Feb 2022 11:37:33 +0100 Guillaume Nault wrote:  
> > > > What about an explicit option:
> > > > 
> > > >   ip link add link eth1 dev eth1.100 type vlan id 100 follow-parent-mtu
> > > > 
> > > > 
> > > > Or for something more future proof, an option that can accept several
> > > > policies:
> > > > 
> > > >   mtu-update <reduce-only,follow,...>
> > > > 
> > > >       reduce-only (default):
> > > >         update vlan's MTU only if the new MTU is smaller than the
> > > >         current one (current behaviour).
> > > > 
> > > >       follow:
> > > >         always follow the MTU of the parent device.
> > > > 
> > > > Then if anyone wants more complex policies:
> > > > 
> > > >       follow-if-not-modified:
> > > >         follow the MTU of the parent device as long as the VLAN's MTU
> > > >         was not manually changed. Otherwise only adjust the VLAN's MTU
> > > >         when the parent's one is set to a smaller value.
> > > > 
> > > >       follow-if-not-modified-but-not-quite:
> > > >         like follow-if-not-modified but revert back to the VLAN's
> > > >         last manually modified MTU, if any, whenever possible (that is,
> > > >         when the parent device's MTU is set back to a higher value).
> > > >         That probably requires the possibility to dump the last
> > > >         modified MTU, so the administrator can anticipate the
> > > >         consequences of modifying the parent device.
> > > > 
> > > >      yet-another-policy (because people have a lot of imagination):
> > > >        for example, keep the MTU 4 bytes lower than the parent device,
> > > >        to account for VLAN overhead.
> > > > 
> > > > Of course feel free to suggest better names and policies :).
> > > > 
> > > > This way, we can keep the current behaviour and avoid unexpected
> > > > heuristics that are difficult to explain (and even more difficult for
> > > > network admins to figure out on their own).  
> > > 
> > > My $0.02 would be that if we want to make changes that require new uAPI
> > > we should do it across uppers.  
> > 
> > Do you mean something like:
> > 
> >   ip link set dev eth0 vlan-mtu-policy <policy-name>
> > 
> > that'd affect all existing (and future) vlans of eth0?
> > 
> > Then I think that for non-ethernet devices, we should reject this
> > option and skip it when dumping config. But yes, that's another
> > possibility.
> > 
> > I personnaly don't really mind, as long as we keep a clear behaviour.
> > 
> > What I'd really like to avoid is something like:
> >   - By default it behaves this way.
> >   - If you modified the MTU it behaves in another way
> >   - But if you modified the MTU but later restored the
> >     original MTU, then you're back to the default behaviour
> >     (or not?), unless the MTU of the upper device was also
> >     changed meanwhile, in which case ... to be continued ...
> >   - BTW, you might not be able to tell how the VLAN's MTU is going to
> >     behave by simply looking at its configuration, because that also
> >     depends on past configurations.
> >   - Well, and if your kernel is older than xxx, then you always get the
> >     default behaviour.
> >   - ... and we might modify the heuristics again in the future to
> >     accomodate with situations or use cases we failed to consider.
> > 
> 
> In general these kind of policy choices are done via sysctl knobs.
> They aren't done at netlink/ip link level.

I don't really mind if the configuration is per vlan, per upper device
or per netns, as long as we keep a clear behaviour by default.