[openstack-dev] [neutron] DVR Tunnel Design Question

Narasimhan, Vivekanandan vivekanandan.narasimhan at hp.com
Thu Sep 18 08:51:56 UTC 2014


Yes... If I recollect, we were using 1.10.x version during that time (wherein discovered this 
as output of ovsapp-ctl fdb-show). After that I didn’t get time to re-verify on later
versions of openvswitch.

BTW, if this is not intended behaviour, then I donot see any particular reason why VLANs
need not be enabled in the current DVR architecture.  

To enable dvr, I need to post a minor patch in L2 agent to bring-in 
DVR rules into Phys bridges (as VLANs are driven by phys bridges 
in OVS L2 Agent).

--
Thanks,

Vivek

-----Original Message-----
From: Kevin Benton [mailto:blak111 at gmail.com] 
Sent: Thursday, September 18, 2014 12:44 PM
To: OpenStack Development Mailing List (not for usage questions)
Subject: Re: [openstack-dev] [neutron] DVR Tunnel Design Question

This sounds like a bug in Openvswitch. The unique constraint should be
VLAN+MAC instead of just VLAN, so observing the same MAC on a new VLAN
should just result in a new entry without the old one being ejected.

Without this behavior, it will also break transparent L2 firewalls.
For example (diagram below), if you divide a switch into two VLANs and then connect the sides together with a firewall in the middle that passes frames without modifying the MAC, the switch will see the same MAC on different VLANs. Based on the behavior you described, this wouldn't work. Is that correct?


---------------------------------
|  x  x  x  x   |   x  x  x  x  |
------------|-------|-------------
 VLAN 1 |         |  VLAN2
          ----------------
          |   Firewall  |
          ----------------

On Wed, Sep 17, 2014 at 11:19 PM, Narasimhan, Vivekanandan <vivekanandan.narasimhan at hp.com> wrote:
> Hi Kevin,
>
>
>
> Typically we noticed that the underlay switches maintained a table 
> like
> this:
>
>
>
> VLAN ID    MAC Address  Learned-Interface
>
>
>
> In the physical underlay, with the current architecture if we enable 
> VLAN, the same DVR Unique
>
> MAC will appear  on different VLANs as the packets get DVR Routed.
>
>
>
> This will result in the rows of the above tables in the switch to be 
> updated very frequently with new
>
> VLANs noted in incoming packets for the same DVR MAC Address, even 
> though they are from the
>
> same physical port.
>
>
>
> We are not sure if all the switches maintained the tables this way , 
> but atleast we
>
> saw Openvswitch implementations did.  So we consciously did not 
> promote VLANs for
>
> initial phase of DVR.
>
>
>
> --
>
> Thanks,
>
>
>
> Vivek
>
>
>
>
>
> From: Kevin Benton [mailto:blak111 at gmail.com]
> Sent: Thursday, September 18, 2014 3:01 AM
> To: OpenStack Development Mailing List (not for usage questions)
> Subject: Re: [openstack-dev] [neutron] DVR Tunnel Design Question
>
>
>
> Can you clarify what you mean with the thrashing condition? MAC 
> addresses only need to be unique per-VLAN so I don't see how the same 
> MAC on multiple VLANs from the same physical port would lead to any issues.
>
>
>
> On Wed, Sep 17, 2014 at 12:41 PM, Armando M. <armamig at gmail.com> wrote:
>
> VLAN is on the radar, vxlan/gre was done to start with.
>
> I believe Vivek mentioned the rationale in some other thread. The gist 
> of it below:
>
> In the current architecture, we use a unique DVR MAC per compute node 
> to forward DVR Routed traffic directly to destination compute node.
> The DVR routed traffic from the source compute node will carry 
> 'destination VMs underlay VLAN' in the frame, but the Source Mac in 
> that same frame will be the DVR Unique MAC. So, same DVR Unique Mac is 
> used for potentially a number of overlay network VMs that would exist 
> on that same source compute node.
>
> The underlay infrastructure switches will see the same DVR Unique MAC 
> being associated with different VLANs on incoming frames, and so this 
> would result in VLAN Thrashing on the switches in the physical cloud 
> infrastructure. Since tunneling protocols carry the entire DVR routed 
> inner frames as tunnel payloads, there is no thrashing effect on 
> underlay switches.
>
> There will still be thrashing effect on endpoints on CNs themselves, 
> when they try to learn that association between inner frame source MAC 
> and the TEP port on which the tunneled frame is received. But that we 
> have addressed in L2 Agent by having a 'DVR Learning Blocker' table, 
> which ensures that learning for DVR routed packets alone is 
> side-stepped.
>
> As a result, VLAN was not promoted as a supported underlay for the 
> initial DVR architecture.
>
> Cheers,
> Armando
>
>
> On 16 September 2014 20:35, 龚永生 <gongysh at unitedstack.com> wrote:
>> I think the VLAN should also be supported later.  The tunnel should 
>> not be the prerequisite for the DVR feature.
>>
>>
>> ------------------ Original ------------------
>> From:  "Steve Wormley"<openstack at wormley.com>;
>> Date:  Wed, Sep 17, 2014 10:29 AM
>> To:  "openstack-dev"<openstack-dev at lists.openstack.org>;
>> Subject:  [openstack-dev] [neutron] DVR Tunnel Design Question
>>
>> In our environment using VXLAN/GRE would make it difficult to keep 
>> some of the features we currently offer our customers. So for a while 
>> now I've been looking at the DVR code, blueprints and Google drive 
>> docs and other than it being the way the code was written I can't 
>> find anything indicating why a Tunnel/Overlay network is required for 
>> DVR or what problem it was solving.
>>
>> Basically I'm just trying to see if I missed anything as I look into 
>> doing a VLAN/OVS implementation.
>>
>> Thanks,
>> -Steve Wormley
>>
>>
>
>> _______________________________________________
>> OpenStack-dev mailing list
>> OpenStack-dev at lists.openstack.org
>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>>
>
> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>
>
>
>
>
> --
>
> Kevin Benton
>
>
> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>



--
Kevin Benton

_______________________________________________
OpenStack-dev mailing list
OpenStack-dev at lists.openstack.org
http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev


More information about the OpenStack-dev mailing list