WIP: Add support for macvlan and ipvlan networking modes #447
Conversation
Revendor the netlink library to pick up changes required to support macvtap creation within a network namespace. Signed-off-by: Manohar Castelino <[email protected]>
Provide multiple methods to connect the Virtual machine to the container network. The current implementation allows this to be chosen at a node level. In the future we can enhance this to be dynamic, where the container interface type is used to determine the optimal interconnection method. Add support for multi-queue macvtap as an alternate means to connect the container network interface to the virtual machine. Include a workaround for kernel limitation around macvtap index generation logic when interfaces are created within a namespace. The switch from bridge+tap to macvtap brings in support for - multiqueue - lower latency - lower jitter - lower cpu consumption Longer term this need to be configurable and the runtime should be able to choose the networking model at a per pod level. Signed-off-by: Manohar Castelino <[email protected]>
Signed-off-by: Manohar Castelino <[email protected]>
|
@mcastelino all the CRI-O tests are failing... Do they pass on your machine ? |
|
@mcastelino this PR can be closed right ? |
|
@sboeuf no. This is for another new feature. Today we do not support macvlan docker/CNI plugins. This PR will be updated once your PR and @amshinde networking changes land to support macvlan CNM and CNI plugins. This PR is to support plugins like |
|
@sboeuf also this is why I wanted the type field in the endpoint data. This type field will help us determine the right inter-connectivity framework to use based on the type of interface placed in the network namespace. So if we see a
|
mcastelino
changed the title
|
@mcastelino Retriggering the jenkins jobs fixed the CI. Is this ready to be be merged? |
|
This fixes #248 |
|
@mcastelino Need to be closed and moved to https://github.com/kata-containers/runtime/virtcontainers |
This a PoC branch that shows that with minor changes clear containers can support the docker macvlan network plugin
This needs a minor change to how we manage mac address allocations.
It also needs the link discovery logic to be generic.