Re: [PATCH 2/2] ipvlan: always allow the broadcast MAC address

[Dan Williams <dcbw@redhat.com>]

On Mon, 2015-03-30 at 10:56 -0700, Mahesh Bandewar wrote:
> On Mon, Mar 30, 2015 at 10:44 AM, Dan Williams <dcbw@redhat.com> wrote:
> > On Mon, 2015-03-30 at 09:54 -0700, Mahesh Bandewar wrote:
> >> On Mon, Mar 30, 2015 at 7:37 AM, Dan Williams <dcbw@redhat.com> wrote:
> >> > On Sat, 2015-03-28 at 19:32 +0100, Jiri Benc wrote:
> >> >> On Fri, 27 Mar 2015 22:56:15 -0700, Mahesh Bandewar wrote:
> >> >> > The current logic disables broadcast by default and enables only when
> >> >> > an IPv4 address is added. If this is inverted and -
> >> >> > enables broadcast by default but disables it when only IPv6
> >> >> > address(es) is / are added. These links can have multiple addresses
> >> >> > and hence have to be careful if any one of those is IPv4 then
> >> >> > broadcast bit has to be set.
> >> >>
> >> >> You'd have to be careful and ignore IPv6 link local addresses.
> >> >> Those are added automatically whenever IPv6 is enabled and their
> >> >> presence does not mean the network is not IPv4 only.
> >> >>
> >> >> But I don't like such magic behavior. It would lead to DHCP sometimes
> >> >> working and sometimes not in mixed v4/v6 environment depending on
> >> >> whether DHCPv4 or SLAAC was faster.
> >> >>
> >> >> Could we perhaps add a flag when creating ipvlan interface stating
> >> >> whether IPv4 broadcast should be always enabled? Or, rather, the other
> >> >> way round - whether it should be disabled by default. Call it "nodhcp"
> >> >> or so.
> >> >>
> >> >> Btw, speaking about IPv6 link local addresses, these actually do not
> >> >> work with ipvlan correctly. I'm getting DAD failures if I have more
> >> >> than one ipvlan interface, which is no wonder. This means that ipvlan
> >> >> cannot work with IPv6 reliably by default (unless you take care of ll
> >> >> address assignment and ensure all ipvlan interfaces get a different
> >> >> one).
> >> >
> >> > ipvlan doesn't set dev_id.  Once dev_id is set the kernel's IPv6LL
> >> > address generation code will assign a different LL address to each
> >> > ipvlan interface created from the same physical interface, despite that
> >> > they have the same MAC address.
> >> >
> >> Yes, that was what my plan was but never got around fixing that
> >>
> >> > But of course you'd have to be careful to assign a *different* dev_id
> >> > than any of that physical interface's non-ipvlan children too, and I
> >> > have no idea how that would work since dev_id is currently done
> >> > per-driver.  eg, if you have a physical interface with dev_id=1 which
> >> > you then create an ipvlan from, that ipvlan must not use dev_id=1 or it
> >> > will be assigned the same IPv6LL address as the parent.
> >> >
> >> The description is very clear for dev_id (in netdevice.h). So the idea
> >> of using the subsequent numbers after master's id should be possible.
> >> After all these logical devices are going to share the same link. Most
> >> physical drivers don't assign dev-id so the beginning is 0x0 (for the
> >> physical driver) and from 0x1 can be assigned to the logical links.
> >> The definition is not clear in terms of what is the beginning (0x0 or
> >> 0x1) but from the code that generates the IPv6LL it's common that it's
> >> 0x0 hence logical links on top of these links can use 0x1 onward.
> >> However a check to see if the master-link has dev-id and staying clear
> >> of that should be sufficient.
> >
> > My point was that if you have a parent with a non-zero dev_id, there can
> > be other siblings of the parent that have a different dev_id and share
> > the same MAC address.  So creating an ipvlan with parent->dev_id + 1
> > doesn't work, because the parent may have a sibling with parent->dev_id
> > + 1 and the same MAC address already.
> >
> May be I'm missing something but is there a scenario where sibling
> (physical / port) will be sharing the same LL-address? The definition
> / description in netdevice.h is  -
> 
>                *  @dev_id:        Used to differentiate devices that share
>                *                          the same link layer address
> 
> So I's assuming the layered / stacked devices (children) rather than
> ports etc (siblings). What am I missing?

I don't think that distinction matters since you can create an ipvlan
interface on top of any other interface except a macvlan.  So any driver
that sets dev_id could be the parent of an ipvlan interface.  That
appears to include some CAN devices and s390's qeth driver at the
moment.

Dan