VRF_DEVICE integration plan

VRF_DEVICE integration plan

[Elluru, Krishna Mohan]

HI Netdev team,

 	Greetings. We have been monitoring the vrf device approach for l3 isolation from cumulus networks and we are currently interested in validating it. We have following questions on them and hoping to get answers from you/concerned team.

1. As per the linux documentation, there are known limits on if_index lookup, as the incoming if_index is changed to vrf_device index and thus an application receiving this packet will perceive this as a vrf_device packet, than right if_index. I saw you mentioned about a special flag to identify the origin, but didn't see the same in the latest linux 4.4.2 version code. Is there a patch expected for it?

2. What are the future additions planned for this approach? Are there any ipv4 and ipv6 known bugs with vrf_device model? 

3. It has been said in the documentation that, with addition of cgroup functionality for vrf device, with net_admin capabilities, we should be able to add an interface to vrf_device, currently it is not so. Any timelines on these?

4. Currently the changes are available and portable from 4.3.x onwards. Is there a plan to port them to previous kernel versions? 

5. Is there a possibility of enabling secondary level lookup, to give a leak functionality to parent route table from device local route table? I tested with veth pair, configured one as default gateway, it is possible to forward traffic b/w the interfaces, looking for cleaner method.

6. With "VRF Device" in place,  please confirm if there are any plans to add VRF support for applications like 

	1.	Ping
	2.	Traceroute
	3.	DNS-Client [glibc]

	In case of DNS-Client, most of the name resolution APIs will have to consider the VRF to do the lookup in  and the way the domain-name/name-server configuration is stored.

Please revert back, if you need more clarification on the questions.

Thanks in Advance
Krishna Mohan

Re: VRF_DEVICE integration plan

[David Ahern]

On 4/23/16 10:07 PM, Elluru, Krishna Mohan wrote:
> HI Netdev team,
>
>   	Greetings. We have been monitoring the vrf device approach for l3 isolation from cumulus networks and we are currently interested in validating it. We have following questions on them and hoping to get answers from you/concerned team.
>
> 1. As per the linux documentation, there are known limits on if_index lookup, as the incoming if_index is changed to vrf_device index and thus an application receiving this packet will perceive this as a vrf_device packet, than right if_index. I saw you mentioned about a special flag to identify the origin, but didn't see the same in the latest linux 4.4.2 version code. Is there a patch expected for it?

you are referring to IP{6}_PKTINFO? I have patches from our 4.1 kernel 
tree that I have rebased to top of tree. I hope to send those out in the 
next few weeks.

>
> 2. What are the future additions planned for this approach? Are there any ipv4 and ipv6 known bugs with vrf_device model?

We have about 20 patches in our tree that I have not sent upstream yet. 
Those patches fix PKTINFO, allow local traffic (e.g, ping in a VRF to a 
local address in a VRF), allow IPv6 multicast and linklocal traffic, and 
the cgroup implementation which has been sent as an RFC.

I posted a few bug fix patches a week or two ago. Not sure what the 
status is with respect to 4.3 - 4.5 trees.

>
> 3. It has been said in the documentation that, with addition of cgroup functionality for vrf device, with net_admin capabilities, we should be able to add an interface to vrf_device, currently it is not so. Any timelines on these?

I don't understand that question. The current implementation allows 
adding interfaces (netdev's) to a VRF. The cgroup allows running a 
process in a VRF context such that AF_INET{6} sockets are automatically 
bound to the VRF device.

>
> 4. Currently the changes are available and portable from 4.3.x onwards. Is there a plan to port them to previous kernel versions?

no. Anyone wanting to use the vrf patches on other kernel versions will 
need to port them.

>
> 5. Is there a possibility of enabling secondary level lookup, to give a leak functionality to parent route table from device local route table? I tested with veth pair, configured one as default gateway, it is possible to forward traffic b/w the interfaces, looking for cleaner method.

Are you referring to inter-vrf routing? See slide 27
http://www.netdevconf.org/1.1/proceedings/slides/ahern-vrf-tutorial.pdf

>
> 6. With "VRF Device" in place,  please confirm if there are any plans to add VRF support for applications like
>
> 	1.	Ping

no need. ping{6} -I <vrf device> ...

> 	2.	Traceroute

no need. traceroute{6} -i <vrf device> ...

> 	3.	DNS-Client [glibc]
>
> 	In case of DNS-Client, most of the name resolution APIs will have to consider the VRF to do the lookup in  and the way the domain-name/name-server configuration is stored.

I have looked into it but no patches worth distributing at the moment.

RE: VRF_DEVICE integration plan

[Elluru, Krishna Mohan]

HI David,
	Thanks a lot for your response. It clarifies few of my questions. Please see inline for with tag MOHAN> for my response.

Thanks
Krishna Mohan.
-----Original Message-----
From: David Ahern [mailto:dsa@cumulusnetworks.com] 
Sent: Monday, April 25, 2016 5:01 AM
To: Elluru, Krishna Mohan <elluru.kri.mohan@hpe.com>; netdev@vger.kernel.org
Subject: Re: VRF_DEVICE integration plan

On 4/23/16 10:07 PM, Elluru, Krishna Mohan wrote:
> HI Netdev team,
>
>   	Greetings. We have been monitoring the vrf device approach for l3 isolation from cumulus networks and we are currently interested in validating it. We have following questions on them and hoping to get answers from you/concerned team.
>
> 1. As per the linux documentation, there are known limits on if_index lookup, as the incoming if_index is changed to vrf_device index and thus an application receiving this packet will perceive this as a vrf_device packet, than right if_index. I saw you mentioned about a special flag to identify the origin, but didn't see the same in the latest linux 4.4.2 version code. Is there a patch expected for it?

you are referring to IP{6}_PKTINFO? I have patches from our 4.1 kernel tree that I have rebased to top of tree. I hope to send those out in the next few weeks.
MOHAN> Yes. Sure. Thanks.

>
> 2. What are the future additions planned for this approach? Are there any ipv4 and ipv6 known bugs with vrf_device model?

We have about 20 patches in our tree that I have not sent upstream yet. 
Those patches fix PKTINFO, allow local traffic (e.g, ping in a VRF to a 
local address in a VRF), allow IPv6 multicast and linklocal traffic, and 
the cgroup implementation which has been sent as an RFC.

I posted a few bug fix patches a week or two ago. Not sure what the 
status is with respect to 4.3 - 4.5 trees.

MOHAN> Sure. Are those patches sent over netdev mailer list?

>
> 3. It has been said in the documentation that, with addition of cgroup functionality for vrf device, with net_admin capabilities, we should be able to add an interface to vrf_device, currently it is not so. Any timelines on these?

I don't understand that question. The current implementation allows 
adding interfaces (netdev's) to a VRF. The cgroup allows running a 
process in a VRF context such that AF_INET{6} sockets are automatically 
bound to the VRF device.

MOHAN> sorry for not being clear. My ask was, to create a namespace we need cap_admin privileges currently, but your earlier mails suggested that we should be able to configure/create vrf device with net_admin capabilities. Is this support present /expected to be added soon?

>
> 4. Currently the changes are available and portable from 4.3.x onwards. Is there a plan to port them to previous kernel versions?

no. Anyone wanting to use the vrf patches on other kernel versions will 
need to port them.
MOHAN> Sure.
>
> 5. Is there a possibility of enabling secondary level lookup, to give a leak functionality to parent route table from device local route table? I tested with veth pair, configured one as default gateway, it is possible to forward traffic b/w the interfaces, looking for cleaner method.

Are you referring to inter-vrf routing? See slide 27
http://www.netdevconf.org/1.1/proceedings/slides/ahern-vrf-tutorial.pdf
Full lookup in VRF table
▪ ip route add table vrf-red 1.1.1.0/24 dev vrf-green
MOHAN> In slide 27 above shows inter vrf routing, requirement is to use current namespace global route table if the ip lookup fails in vrf-device routing table. 
Reference: https://www.juniper.net/techpubs/en_US/junose16.1/topics/task/configuration/mbgp-secondary-routing-table-search.html

>
> 6. With "VRF Device" in place,  please confirm if there are any plans to add VRF support for applications like
>
> 	1.	Ping

no need. ping{6} -I <vrf device> ...

> 	2.	Traceroute

no need. traceroute{6} -i <vrf device> ...

> 	3.	DNS-Client [glibc]
>
> 	In case of DNS-Client, most of the name resolution APIs will have to consider the VRF to do the lookup in  and the way the domain-name/name-server configuration is stored.

I have looked into it but no patches worth distributing at the moment.

MOHAN> Okay, thanks for the inputs.

Re: VRF_DEVICE integration plan

[David Ahern]

On 4/28/16 11:16 AM, Elluru, Krishna Mohan wrote:
>
> I posted a few bug fix patches a week or two ago. Not sure what the
> status is with respect to 4.3 - 4.5 trees.
>
> MOHAN> Sure. Are those patches sent over netdev mailer list?

yes. All patches for VRF - kernel and iproute2 - are sent to netdev.


> MOHAN> sorry for not being clear. My ask was, to create a namespace we need cap_admin privileges currently, but your earlier mails suggested that we should be able to configure/create vrf device with net_admin capabilities. Is this support present /expected to be added soon?

VRF is implemented using a netdevice. As such the ability to create one 
requires the same permissions as creating any other netdevice 
(CAP_NET_ADMIN).


>> 5. Is there a possibility of enabling secondary level lookup, to give a leak functionality to parent route table from device local route table? I tested with veth pair, configured one as default gateway, it is possible to forward traffic b/w the interfaces, looking for cleaner method.
>
> Are you referring to inter-vrf routing? See slide 27
> http://www.netdevconf.org/1.1/proceedings/slides/ahern-vrf-tutorial.pdf
> Full lookup in VRF table
> ▪ ip route add table vrf-red 1.1.1.0/24 dev vrf-green
> MOHAN> In slide 27 above shows inter vrf routing, requirement is to use current namespace global route table if the ip lookup fails in vrf-device routing table.
> Reference: https://www.juniper.net/techpubs/en_US/junose16.1/topics/task/configuration/mbgp-secondary-routing-table-search.html

One solution is to create a VRF device that is associated with the main 
table and then use an inter-vrf route to jump to the main table.

VRF tables do need a default route (e.g., unreachable with high metric 
value) else the FIB lookups will proceed to the next table which is most 
likely not what you want.


David