Re: [net-next PATCH] net: rtnetlink: Support alias interfaces with RTM_GETLINK

[Phil Sutter <phil@nwl.cc>]

On Thu, Feb 07, 2019 at 02:21:56PM +0100, Michal Kubecek wrote:
> On Thu, Feb 07, 2019 at 02:02:15PM +0100, Phil Sutter wrote:
> > On Thu, Feb 07, 2019 at 01:39:39PM +0100, Michal Kubecek wrote:
> > 
> > > But I still don't think it would be a good idea. It's bad enough that
> > > (as I just learned to my surprise) "ip link del dummy1:0" deletes dummy1
> > > without any complaint because ip uses SIOCGIFINDEX ioctl for ifindex
> > > lookup.
> > 
> > I'm struggling a bit with all this. The original problem is iproute2
> > commit 50b9950dd9011 ("link dump filter") which changed 'ip link show'
> > to not use if_indextoname() when given just an interface name. So lookup
> > happens by name (via RTM_GETLINK) and consequently 'ip link show eth0:1'
> > doesn't give link stats of eth0 anymore.
> 
> I would rather consider it a bug that it ever did. It's quite harmless
> with "show" but with "set" or "delete", the effect can be quite
> disastrous.

Yes, you're right. And unless maintainers care about this compatibility,
I guess these monsters will slowly disappear as things progress towards
netlink.

> We want to preserve backward compatibility in general but iproute2
> commit 50b9950dd901 is 4.5 years old and nobody seems to have complained
> about the change in behaviour until now.

Well, enterprise distributions deliberately try to prevent those changes
from hitting their customers. And (perfect match), these users are the
least tolerating ones. :)

> > Given that iproute2 is supposed to be backwards compatible, the only
> > valid option I see is to make sure netlink API calls like the above
> > behave identical to the ioctl ones they replace. Which means allowing
> > for 'ip link show eth0:42' even if there's no address with that label
> > assigned to eth0 as well as your example above.
> 
> One reason why I don't like this idea is that iproute2 is not the only
> user of rtnetlink interface. There is wicked and glibc for sure, most
> likely also NetworkManager (don't remember) and systemd-networkd (didn't
> check) and certainly many others I never heard of. Changing the logic
> in kernel rtnetlink implementation would affect all of them.
> 
> If we want to restore the old behaviour of ip (which I'm not convinced
> of), it would make more sense to me to strip the :* suffix in iproute2.

I just sent a patch, let's see what userspace has to say about it.

Thanks, Phil