Re: [patch v1, kernel version 3.2.1] net/ipv4/ip_gre: Ethernet multipoint GRE over IP

Dňa 16. januára 2012 22:22, Jesse Gross [off-list ref] napísal/a:

2012/1/16 Štefan Gula [off-list ref]:

quoted

Dňa 16. januára 2012 17:36, Stephen Hemminger [off-list ref] napísal/a:

quoted

On Mon, 16 Jan 2012 13:13:19 +0100
Štefan Gula [off-list ref] wrote:

quoted

From: Stefan Gula <steweg@gmail.com

This patch is an extension for current Ethernet over GRE
implementation, which allows user to create virtual bridge (multipoint
VPN) and forward traffic based on Ethernet MAC address informations in
it. It simulates the Bridge bahaviour learing mechanism, but instead
of learning port ID from which given MAC address comes, it learns IP
address of peer which encapsulated given packet. Multicast, Broadcast
and unknown-multicast traffic is send over network as multicast
enacapsulated GRE packet, so one Ethernet multipoint GRE tunnel can be
represented as one single virtual switch on logical level and be also
represented as one multicast IPv4 address on network level.

Signed-off-by: Stefan Gula <redacted>

Thanks for the effort, but it is duplicating existing functionality.
It possible to do this already with existing gretap device and the
current bridge.

The same thing is also supported by OpenVswitch.

gretap with bridge will not do the same as gretap allows you to only
encapsulate L2 frames inside the GRE - this one part is actually
utilized in my code. GRE multipoint implementation is also utilized in
my code as well. But what is missing is forwarding logic here, which
prevents the traffic going not optimal way. Scenario one - e.g. if you
connect through 3 sites with using 1 gretap multipoint VPN, it always
forwards frames between site 1 and site 2 even if they are unicast.
That represents waste of bandwidth for site 3. Now assume that there
will be more than 40 sites and I hope you see that single current
multipoint gretap is not also good solution here

The second scenario - e.g. using 3 sites using point-to-point gretap
interfaces between each 2 sites (2 gretap VPN interfaces per site) and
bridging those interfaces with real ones results in looped topology
which needs to utilized STP inside to prevent loops. Once STP
converges the topology will looks like this, traffic from site 1 to
site 2 will go always directly by the way of unicast (on GRE level),
from site 2 to site 3 always directly by the way of unicast (on GRE
level) and from site 1 to site 3 will go indirectly through site 2 due
STP limitations, which results in another not optimalized traffic
flows. Now assume that the number of sites rises, so gretap+standard
bridge code is also not a good solution here.

My code utilizes it that way that I have extended the gretap
multipoint interface with the forwarding logic e.g. using 3 sites,
each site uses only one gretap VPN interface and if destination MAC
address is known to bridge code inside the gretap interface forwarding
logic, it forwards it towards only VPN endpoint that actually need
that by the way of unicasting on GRE level. On the other hand if the
destination MAC address is unknown or destination MAC address is L2
multicast or L2 broadcast than the frame is spread out through
multicasting on GRE level, providing delivery mechanism analogous to
standard switches on top of the multipoint GRE tunnels.

I also get through briefly over OpenVswitch documentation and found
that it is more related to virtualization inside the box like VMware
switches or so and not to such technologies interconnecting two or
more separate segments over routed L3 infrastructure - there is a
mention about the CAPWAP UDP transport but this is more related to
WiFi implementations than generic ones. My patch also doesn't need any
special userspace api to be configured. It utilizes the existing one.

I understand what you're trying to do and I think that the goal makes
sense but I agree with Stephen that this is not the right way to go
about it.  I see two issues:

 * It copies a lot of bridge code, making it unmaintainable and
inflexible to other use cases.
 * The implementation exists in the GRE protocol stack but it applies
equally to other tunneling protocols as well (VXLAN comes to mind).

Open vSwitch doesn't quite do this level of learning yet but it's the
direction that we want to move in (and there's nothing particularly
virtualization specific about it).  What I think makes the most sense
is to create some internal interfaces to the GRE stack that exposes
the information needed to do learning.  That way there is only one
instance of the protocol code for each tunneling mechanism and then
each way of managing those addresses (i.e. the current device-based
mechanism, Open vSwitch, potentially a direct bridge-based mechanism,
etc.) can be reused as well.

I agree with you that using such approach of copying bridge code with
some modifications is maybe not flexible enough, but on the other hand
it does the job needed. It also doesn't breaks any previous
compatibility of usage of gretap interfaces as it modifies the
encapsulation and decapsulation part of codes *how the traffic goes to
and from remote destinations), which doesn't face any change on the
internal linkages coding inside the box (e.g. linking that gretap
interface to standard logical bridge interface is still fully
possible). I would rather see getting some standard set of generic
bridge code, which can be reused anywhere in network stack, but for
now this is the only way I know how to do it. Exposing the GRE
interfaces will not do the job as what I needed was to rebuild the
bridge logic to allow learning endpoint IPs instead of network ports
IDs (it's almost the same as using multiple gretap interfaces inside
one bridge.interface) To obtain back the maintainability I would
assume redesigning the bridge code is the best way here, but I am not
that well coder to do it myself. So if anybody is interested in this
feel free to do it.