Skip to content

Conversation

@neilcook
Copy link

When using the sbr plugin with a CNI plugin such as multus, applications have to bind to the source IP address of the interface for the str plugin to work correctly. They cannot bind to the interface name, because the sbr plugin doesn't currently add a rule for the interface name.

This is problematic because it is straightforward to configure the name of a multus interface, however the IP address is usually assigned from a range, so the exact interface IP address to use is not known in advance unlike the interface name. It would thus be much easier for applications if the sbr plugin added a rule with the interface name in addition to the rule with the interface address.

This PR does exactly that - adds an additional rule with the outbound interface name, so that applications can configure the name of the interface to bind to, rather than having to know the interface IP address.

Obviously, this approach does not work if there are multiple IP addresses for the interface, so this change only adds the rule for the outbound interface name if there is a single IP address on that interface.

I have added tests for the new rule, which all pass.

@neilcook neilcook force-pushed the oif_rule branch 2 times, most recently from a435c72 to 6dfec7f Compare January 26, 2025 14:46
@neilcook neilcook changed the title Add rule for outbound interface when there is a single interface IP sbr: Add rule for outbound interface when there is a single interface IP Jan 26, 2025
neilcook added 7 commits March 4, 2025 14:55
Adding the outbound interface rule allows applications to bind to an interface name
rather than only the interface IP. This allows applications in a multus environment
to be configured with the interface name, which can be configured, rather than the
interface IP address, which is not known in advance.

The outbound interface rule is on;y added if the interface is configured with 1 IP
address, because when there are multiple, only one will be selected, depending on the
rule order, which is non-deterministic.

Signed-off-by: Neil Cook <[email protected]>
The outbound interface rule does not reference anything from ipCfg so should not be
in the loop.

Signed-off-by: Neil Cook <[email protected]>
neilcook added 2 commits March 5, 2025 09:15
…able no is correct

The previous commit moved the rule creation to after the ipCfg loop, but since the loop
increments the table number, the rule gets added to the wrong table.

Signed-off-by: Neil Cook <[email protected]>
Moving the outbound interface rule creation to before the ipCfg loop
means that tests need to reflect the changed order of rules.

Signed-off-by: Neil Cook <[email protected]>
@neilcook
Copy link
Author

neilcook commented May 2, 2025

Any more thoughts on this PR?

@mlguerrero12
Copy link
Member

So, this is intended for packets originating from applications that are bound to a device. After that, a table is selected to continue with the regular routing process. Is that right?

I believe this should be a new parameter instead of applying with there is a single IP (no provided table ID case). It can take precedence over the IPs. Something like useOriginatingIf, when this is set to True by the users, then a rule is added using oif. No need to add rules for IPs (please correct me if I'm wrong here). And move all associated routes of the interface to the table.

Also, why not using the VRF plugin instead?

@neilcook
Copy link
Author

So, this is intended for packets originating from applications that are bound to a device. After that, a table is selected to continue with the regular routing process. Is that right?

Yes. Because binding to IP addresses is problematic, as they are dynamically assigned.

I believe this should be a new parameter instead of applying with there is a single IP (no provided table ID case). It can take precedence over the IPs. Something like useOriginatingIf, when this is set to True by the users, then a rule is added using oif. No need to add rules for IPs (please correct me if I'm wrong here). And move all associated routes of the interface to the table.

Yes, probably setting this via a parameter is better.

Also, why not using the VRF plugin instead?

Doesn't that require the application to be VRF aware?

@mlguerrero12
Copy link
Member

as far as I know, for vrf, you only need to use SO_BINDTODEVICE which I guess you are already using

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants