Message-ID: <1494267188.8248.21.camel@redhat.com>
Date:   Mon, 08 May 2017 13:13:08 -0500
From:   Dan Williams <dcbw@...hat.com>
To:     "Chiappero, Marco" <marco.chiappero@...el.com>,
        Jiri Benc <jbenc@...hat.com>
Cc:     "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        "David S . Miller" <davem@...emloft.net>,
        "Kirsher, Jeffrey T" <jeffrey.t.kirsher@...el.com>,
        "Duyck, Alexander H" <alexander.h.duyck@...el.com>,
        "Grandhi, Sainath" <sainath.grandhi@...el.com>,
        Mahesh Bandewar <maheshb@...gle.com>
Subject: Re: [PATCH net-next 9/9] ipvlan: introduce individual MAC addresses

On Mon, 2017-05-08 at 15:29 +0000, Chiappero, Marco wrote:
> > -----Original Message-----
> > From: Jiri Benc [mailto:jbenc@...hat.com]
> > Sent: Thursday, May 4, 2017 5:44 PM
> > To: Chiappero, Marco <marco.chiappero@...el.com>
> > Cc: Dan Williams <dcbw@...hat.com>; netdev@...r.kernel.org; David S
> > .
> > Miller <davem@...emloft.net>; Kirsher, Jeffrey T
> > <jeffrey.t.kirsher@...el.com>; Duyck, Alexander H
> > <alexander.h.duyck@...el.com>; Grandhi, Sainath
> > <sainath.grandhi@...el.com>; Mahesh Bandewar <maheshb@...gle.com>
> > Subject: Re: [PATCH net-next 9/9] ipvlan: introduce individual MAC
> > addresses
> > 
> > On Thu, 4 May 2017 09:37:00 +0000, Chiappero, Marco wrote:
> > > This looks conceptually wrong. Yes, ipvlan works at L3 (which is
> > > an
> > > implementation detail anyway), but slaves are Ethernet interfaces
> > > and
> > > should behave as much as possible as such regardless, with an
> > > individual MAC address assigned.
> > 
> > Isn't the proper fix then converting ipvlan interfaces to be L3
> > only interfaces?
> > I.e., ARPHRD_NONE? There's not much ipvlan can do with arbitrary
> > Ethernet
> > frames anyway. Of course, a flag to switch to the new behavior
> > would be
> > needed in order to preserve backwards compatibility.
> 
> Yes, L3 only interfaces would be a valid solution and a major
> differentiation with respect to macvlan. In fact it's been considered
> but abandoned because:
> - it would be a break from the past
> - VMs and containers usually expect Ethernet devices
> 
> Having that said, I'm ok with this solution if deemed preferable.
>  
> > This patchset looks very wrong. For proper support of multiple MAC
> > addresses,
> > we have macvlan and it's pointless to add that to ipvlan.
> 
> Ipvlan will never have proper support for L2, it's a L3 thing and
> doesn't aim at replacing macvlan. So, the options are:
> 1) remove it and provide L3 only interfaces - as you are proposing
> 2) fully emulate it preserving the single unicast MAC rule - my
> proposal
> 
> I'm open to both solutions, to me the important thing is to address
> the problems with the current implementation, one way or another,
> making it comply with basic networking concepts and expectations.
> 
> > And doing some kind of weird MAC NAT in ipvlan just to satisfy
> > broken tools that
> > can't cope with multiple interfaces with the same MAC address is
> > wrong, too.
> 
> Ipvlan has always had the MAC issue, regardless, these tools simply
> make it more apparent. And as I said already, whether they are broken

If we're talking about the slaves being unable to change when the
parent MAC changes, everyone agrees that was a bug.

If we are talking about the "all slaves have the same MAC" then that's
not an "issue", that's the way it's designed.  Whether that design is
the best design possible is debatable, but that's the way it currently
is.

> is debatable (yet I have to read a reasonable motivation). At the 

Existing tools are already broken for bond slaves and a couple existing
drivers, see below.

Note that making the MACs unique would break DHCP functionality,
because now the MAC address encoded in the 'chaddr' field of the DHCP
packet would no longer correspond to the MAC address of the device that
DHCP replies should be received on.  The userspace process writes the
'chaddr' field, and the userspace process would see the unique (and
incorrect) MAC address.

Misrepresenting the MAC address of the device, as this would do, is
simply not going to work.  The tools that expect the unique MAC
addresses are already broken.

> very least their expectation to have unique addresses on the same
> broadcast domain is hardly arguable. Should ipvlan considered
> special? Again, questionable.

bond slaves
drivers/net/ethernet/toshiba/ps3_gelic_net.c
drivers/s390/net/qeth_l3_main.c

already all have the same MAC address for different netdevices.  Yeah,
not many people have PS3s anymore, but s390 qeth is fairly widely used.
 Just pointing out that ipvlan is hardly the first device to have
encountered this issue, or to have solved it this way.  qeth does
pretty much the same thing as ipvlan.

I'm not arguing that ipvlan is perfect.  I'm just arguing that in its
current form (eg, virtualized L2 device) making this change is
incorrect.

Dan

> > Those tools are already broken anyway, there's nothing preventing
> > anyone to
> > set the same MAC address to multiple interfaces.
> 
> What are you trying to demonstrate, that you can shoot yourself in
> the foot? That you can interfere with the controller? That you can
> intentionally break your network? Yes, of course you can, so what?
> As long as network components comply with expected behaviors,
> including the ability to change the MAC address, every orchestrator
> will do its job of managing the network and will do it correctly.
> 
> > I suppose those tools don't work with bonding and bridge, either?
> 
> NaaS software and SDN controllers are built around bridges, a bridge
> is a well-defined, coherent, fundamental network component. As soon
> as ipvlan will behave coherently as well, upper layers will handle it
> properly.
> By the way, none of them has the same limitations of ipvlan.
>  
> > > So, either we fix this by forcing slaves to stay in sync with
> > > master,
> > 
> > Yes, that's the correct behavior. Well, at least as correct as one
> > can get with the
> > ipvlan broken design of pretending that an interface is L2 when in
> > fact, it is not.
> 
> Eventually we got there:  "ipvlan broken design". Let's fix it then.
> Would moving to L3 only interfaces be accepted instead?
> --------------------------------------------------------------
> Intel Research and Development Ireland Limited
> Registered in Ireland
> Registered Office: Collinstown Industrial Park, Leixlip, County
> Kildare
> Registered Number: 308263
> 
> 
> This e-mail and any attachments may contain confidential material for
> the sole
> use of the intended recipient(s). Any review or distribution by
> others is
> strictly prohibited. If you are not the intended recipient, please
> contact the
> sender and delete all copies.
> 

Powered by blists - more mailing lists