- Links
- BiWeekly Meeting
- Contacts
- Userspace CNI plugin
- Build & Clean
- Network Configuration Reference
- OVS CNI Library Intro
- VPP CNI Library Intro
- Testing
- Unit Testing
The Userspace CNI is a Container Network Interface (CNI) plugin designed to implement userspace networking (as opposed to kernel space networking). An example is any DPDK based applications. It is designed to run with either OVS-DPDK or VPP along with the Multus CNI plugin in Kubernetes deployments. It enhances high performance container Networking solution and Data Plane Acceleration for containers.
Useful Links:
- Source Code: https://github.com/intel/userspace-cni-network-plugin
- Design Document: Userspace Design Document
Join us for project discussions at K8s Network & Resource management meetings
Time: 15:00 - 16:00 GMT / 10:00-11:00 ET / 07:00-08:00 PST every other Monday
For any questions about Userspace CNI, please reach out.
- Report issue via Github: userspace-cni-network-plugin/issues
- Contact via slack: NPWG Slack
- Feel free to contact the developer @garyloug or @bmcfall in slack
- Contact via Google Group: https://groups.google.com/forum/#!forum/userspace-cni
The Userspace CNI is a Container Network Interface (CNI) plugin designed to implement userspace networking (as opposed to kernel space networking), like DPDK based applications. It is designed to run with either OVS-DPDK or VPP running on the host, along with the Multus CNI plugin, which enables additional interfaces in a container.
Userspace networking requires additional considerations. For one, the interface needs to be created/configured on a local vswitch (running on the host). There may also be a desire to add the interface to a specific network on the host through the local vswitch. Second, when the interface is inserted into the container, it is not owned by the kernel, so additional work needs to be done in the container to consume the interface and add to a network within the container. The Userspace CNI is designed to work with these additional considerations by provisioning the local vswitch (OVS-DPDK/VPP), and by pushing data into the container so the interface can be consumed.
The Userspace CNI, based on the input config data, adds interfaces (memif and/or vhost-user interfaces) to a local OVS-DPDK or VPP instance running on the host. Then adds that interface to a local network, like an L2 Bridge. The Userspace CNI then processes config data intended for the container and adds that data to a Database the container can consume.
DPDK Vhostuser is new virtualization technology. Please refer to here for more information.
This plugin is recommended to be built with Go 1.20.1 and either OVS-DPDK 2.9.0-3 or VPP 23.02. Other versions of Go, OVS-DPDK and VPP are theoretically supported, but MIGHT cause unknown issue.
The Userspace CNI requires several files from VPP in-order to build. For this reason we build userspacecni in a container.
userspacecni is built in a container and then transferred to the host. By default Docker is used to build the image
Before building be sure to modify the makefile variable IMAGE_REGISTRY.
IMAGE_REGISTRY should be set to an image registry thats accessible across your cluster.
To build the docker image:
git clone https://github.com/intel/userspace-cni-network-plugin.git
cd userspace-cni-network-plugin
make build
To copy the userspacecni binary to the host directory /opt/cni/bin/
make copy
To push the image to the docker reg defined in the makefile
make push
To deploy the userspacecni across the cluster
make deploy
type(string, required): "userspace"name(string, required): Name of the networkhost(dictionary, required): Host based configurations. Contains userspace interface configuration data as well as host network data userspace interface should be injected into.container(dictionary, optional): Container based configurations. Contains userspace interface configuration data as well as container network data userspace interface should be injected into. Defaults used when data omitted.ipam(dictionary, optional): IPAM configuration to be used for this network.
Given the following network configuration:
sudo cat > /etc/cni/net.d/90-userspace.conf <<EOF
{
"cniVersion": "0.3.1",
"type": "userspace",
"name": "memif-network",
"host": {
"engine": "vpp",
"iftype": "memif",
"netType": "bridge",
"memif": {
"role": "master",
"mode": "ethernet"
},
"bridge": {
"bridgeName": "4"
}
},
"container": {
"engine": "vpp",
"iftype": "memif",
"netType": "interface",
"memif": {
"role": "slave",
"mode": "ethernet"
}
},
"ipam": {
"type": "host-local",
"subnet": "10.56.217.0/24",
"rangeStart": "10.56.217.131",
"rangeEnd": "10.56.217.190",
"routes": [
{
"dst": "0.0.0.0/0"
}
],
"gateway": "10.56.217.1"
}
}
EOF
Integrate with the Multus plugin for a high performance container networking solution for NFV Environments. Refer to Multus (NFV based Multi - Network plugin), DPDK-SRIOV CNI plugins:
Encourage the users/developers to use Multus based Kubernetes CDR/TPR based network objects. Please follow the configuration details in the link: Usage with Kubernetes CRD/TPR based Network Objects
Please refer the Kubernetes Network SIG - Multiple Network PoC proposal for more details refer the link: K8s Multiple Network proposal
# cat > /etc/cni/net.d/10-multus.conf <<EOF
{
"name": "multus-demo-network",
"type": "multus",
"delegates": [
{
"type": "sriov",
"if0": "ens786f1",
"if0name": "net0",
"dpdk": {
"kernel_driver": "ixgbevf",
"dpdk_driver": "igb_uio",
"dpdk_tool": "/path/to/dpdk/tools/dpdk-devbind.py"
}
},
{
"cniVersion": "0.3.1",
"type": "userspace",
"name": "memif-network",
"host": {
"engine": "vpp",
"iftype": "memif",
"netType": "bridge",
"memif": {
"role": "master",
"mode": "ethernet"
},
"bridge": {
"bridgeName": "4"
}
},
"container": {
"engine": "vpp",
"iftype": "memif",
"netType": "interface",
"memif": {
"role": "slave",
"mode": "ethernet"
}
},
"ipam": {
"type": "host-local",
"subnet": "10.56.217.0/24",
"rangeStart": "10.56.217.131",
"rangeEnd": "10.56.217.190",
"routes": [
{
"dst": "0.0.0.0/0"
}
],
"gateway": "10.56.217.1"
}
},
{
"type": "flannel",
"name": "control-network",
"masterplugin": true,
"delegate": {
"isDefaultGateway": true
}
}
]
}
EOF
Note: The Userspace CNI supports different IPAM plugins for the IP addresses management. The generated IP address information will be stored in one configuration file.
You may wish to enable some enhanced logging, especially to understand what is
or isn't working with a particular configuration. Userspace CNI always log via
STDERR, which is the standard method by which CNI plugins communicate errors,
and these errors are logged by the Kubelet. This method is always enabled.
Optionally, Userspace CNI can log to a file on the filesystem. This file will
be written locally on each node where CNI is executed. Configure this via the
LogFile option in the CNI configuration. By default this additional logging
to a flat file is disabled.
For example in your CNI configuration, you may set:
"logFile": "/var/log/userspace-cni.log",
The default logging level is set as warning -- this will log critical errors
and issues detect.
The available logging level values, in increasing order of verbosity are:
panic- Code exiting immediately.
error- Unusual event occurred (invalid input or system issue), so exiting code prematurely.
warning- Unusual event occurred (invalid input or system issue), but continuing with default value.
info- Basic information, indication of major code paths.
debug- Additional information, indication of minor code branches.
verbose- Output of larger variables in code and debug of low level functions.
You may configure the logging level by using the LogLevel option in your
CNI configuration. For example:
"logLevel": "debug",
OVS CNI Library is written in GO and used by UserSpace CNI to interface with the OVS. OVS currently does not have a GO-API, though there are some external packages that are being explored. When the CNI is invoked, OVS CNI library builds up an OVS CLI command (ovs-vsctl) and executes the request.
To install the DPDK-OVS, the source codes contains a document for how to install the DPDK-OVS.
The Userspace CNI plugin builds the OVS CNI Library from the cniovs sub-folder. In order to run with the OVS CNI Library, the OVS python script must be installed on the system. To install the script, run make install as described in Building VPP CNI Library with OVS.
DPDK-OVS is a DPDK based application, so some detailed system requirements can be found at DPDK requirements. Hugepages are the main requirement for the VHOST_USER virtual ports.
echo 'vm.nr_hugepages=2048' > /etc/sysctl.d/hugepages.conf
Or add the following configuration to the grub configuration:
default_hugepagesz=2m hugepagesz=2m hugepages=2048
VPP CNI Library is written in GO and used by UserSpace CNI to interface with the VPP GO-API. When the CNI is invoked, VPP CNI library opens a GO Channel to the local VPP instance and passes gRPC messages between the two.
As mentioned above, to build the Userspace CNI, VPP needs to be installed, or several VPP files to compile against. When VPP is installed, it copies it's json API files to /usr/share/vpp/api/. VPP CNI Libary uses these files to compile against and generate the properly versioned messages to the local VPP Instance. So to build the VPP CNI, VPP must be installed (or the proper json files must be in /usr/share/vpp/api/).
There are several ways to install VPP. This code is based on a fixed release VPP (VPP 23.02), so it is best to install a released version (even though it is possible to build your own).
Below are the VPP prerequisites:
- Hugepages: VPP requires 2M Hugepages. By default, VPP uses 1024 hugepages. If hugepages are not configured, on install VPP will allocate them. This is primarily an issue if you are running in a VM that does not already have hugepage backing, especially when you reboot the VM. If you would like to change the number of hugepages VPP uses, after installing VPP, edit /etc/sysctl.d/80-vpp.conf. However, once VPP has been installed, the default value has been applied. As an example, to reduce the number of hugepages to 512, use:
vm.nr_hugepages=512
vm.max_map_count=2048
kernel.shmmax=1073741824
- SELinux: VPP works with SELinux enabled, but when running with containers, work still needs to be done. Set SELinux to permissive.
https://s3-docs.fd.io/vpp/23.02/gettingstarted/installing/
https://github.com/FDio/vpp/tree/v23.02
https://docs.openvswitch.org/en/latest/intro/install/
https://github.com/openvswitch/ovs/tree/master
Its assumed that VPP has been set up on the host correctly by the user.
On the server that the vpp pods get deployed (not the k8 controller) create the directories /var/run/vpp/app1
and /var/run/vpp/app2 for this example to work
kubectl create ns vpp
kubectl create -f ./examples/vpp-memif-ping/userspace-vpp-netAttach.yaml
kubectl create -n vpp configmap vpp-app-startup-config --from-file=./examples/vpp-memif-ping/startup.conf
kubectl create -n vpp configmap vpp-pod-setup-memif --from-file=./examples/vpp-memif-ping/vpp-pod-setup-memif.sh
kubectl create -f ./examples/vpp-memif-ping/vpp-app-pod-1.yaml
kubectl create -f ./examples/vpp-memif-ping/vpp-app-pod-2.yaml
kubectl exec -itn vpp vpp-app1 -- ./vpp-pod-setup-memif.sh
kubectl exec -itn vpp vpp-app2 -- ./vpp-pod-setup-memif.sh
# ping pod 2 from pod 1 (through vpp)
kubectl exec -itn vpp vpp-app1 -- vppctl "ping 192.168.1.4"
Sample output
116 bytes from 192.168.1.4: icmp_seq=1 ttl=64 time=83.4641 ms
116 bytes from 192.168.1.4: icmp_seq=2 ttl=64 time=72.0057 ms
116 bytes from 192.168.1.4: icmp_seq=3 ttl=64 time=67.9981 ms
116 bytes from 192.168.1.4: icmp_seq=4 ttl=64 time=75.9974 ms
116 bytes from 192.168.1.4: icmp_seq=5 ttl=64 time=72.0021 ms
To verify the local config (on host) in another window:
Before Container Started:
vppctl show interface
Name Idx State Counter Count
GigabitEthernet0/8/0 1 down
GigabitEthernet0/9/0 2 down
local0 0 down
vppctl show mode
l3 local0
l3 GigabitEthernet0/8/0
l3 GigabitEthernet0/9/0
vppctl show memif
sockets
id listener filename
0 no /run/vpp/memif.sock
After the container is started, on the host there should be an additional memif interface created and added to a new L2 bridge, all created by the Userspace CNI.
After vpp pod application Started:
vppctl show interface
Name Idx State Counter Count
GigabitEthernet0/8/0 1 down
GigabitEthernet0/9/0 2 down
local0 0 down
memif1/0 3 up
vppctl show mode
l3 local0
l3 GigabitEthernet0/8/0
l3 GigabitEthernet0/9/0
l2 bridge memif1/0 bd_id 4 shg 0
vppctl show memif
sockets
id listener filename
0 no /run/vpp/memif.sock
1 yes (1) /var/run/vpp/cni/shared/memif-79b661b189b2-net0.sock
interface memif1/0
remote-name "VPP 18.07-16~gca7a68e~b66"
remote-interface "memif1/0"
socket-id 1 id 0 mode ethernet
flags admin-up connected
listener-fd 22 conn-fd 23
num-s2m-rings 1 num-m2s-rings 1 buffer-size 0 num-regions 1
region 0 size 4227328 fd 24
master-to-slave ring 0:
region 0 offset 16512 ring-size 1024 int-fd 26
head 1024 tail 0 flags 0x0001 interrupts 0
slave-to-master ring 0:
region 0 offset 0 ring-size 1024 int-fd 25
head 0 tail 0 flags 0x0001 interrupts 0
The container is setup to start VPP, read the config pushed by the User Space CNI, apply the data and then exit to bash. To verify the container config, in the container, run the following:
vppctl show interface
Name Idx State Counter Count
local0 0 down
memif1/0 1 up
vppctl show interface addr
local0 (dn):
memif1/0 (up):
L3 192.168.210.45/24
vppctl show mode
l3 local0
l3 memif1/0
vppctl show memif
sockets
id listener filename
0 no /run/vpp/memif.sock
1 no /var/run/vpp/cni/shared/memif-05138381d803-net0.sock
interface memif1/0
remote-name "VPP 18.07-rc2~11-g18bde8a"
remote-interface "memif1/0"
socket-id 1 id 0 mode ethernet
flags admin-up slave connected
listener-fd 0 conn-fd 17
num-s2m-rings 1 num-m2s-rings 1 buffer-size 2048 num-regions 1
region 0 size 4227328 fd 18
slave-to-master ring 0:
region 0 offset 0 ring-size 1024 int-fd 19
head 0 tail 0 flags 0x0001 interrupts 0
master-to-slave ring 0:
region 0 offset 16512 ring-size 1024 int-fd 20
head 1024 tail 0 flags 0x0001 interrupts 0
If a second container is started on the same host, the two containers will have L3 connectivity through an L2 bridge on the host. Because the interfaces in the host are owned by VPP, use VPP in one of the containers to ping between the two containers.
vppctl ping 192.168.210.46
64 bytes from 192.168.210.46: icmp_seq=2 ttl=64 time=39.0119 ms
64 bytes from 192.168.210.46: icmp_seq=3 ttl=64 time=37.9991 ms
64 bytes from 192.168.210.46: icmp_seq=4 ttl=64 time=57.0304 ms
64 bytes from 192.168.210.46: icmp_seq=5 ttl=64 time=40.0044 ms
Statistics: 5 sent, 4 received, 20% packet loss
To follow this example you should have a system with kubernetes available and
configured to support native 1 GB hugepages. You should also have multus-cni and
userspace-cni-network-plugin up and running. See examples/crd-userspace-net-ovs-no-ipam.yaml for
example config to use with multus. If using OVS,
check that you have bridge named br0 in your OVS with ovs-vsctl show and if
not, create it with
ovs-vsctl add-br br0 -- set bridge br0 datapath_type=netdev.
Build testpmd container image
make testpmd
Modify the nodeSelector kubernetes.io/hostname: in both ./examples/ovs-vhost/testpmd-pod-1.ymal and ./examples/ovs-vhost/testpmd-pod-2.ymal
kubectl create ns ovs
kubectl apply -f ./examples/ovs-vhost/userspace-ovs-netAttach-1.yaml
kubectl apply -f ./examples/ovs-vhost/testpmd-pod-1.yaml
kubectl apply -f ./examples/ovs-vhost/testpmd-pod-2.yaml
At this point pod 1 should be sending traffic to pod 2. By doing a kubectl logs on pod 2 you will be able to verify that the pod is receiving traffic by looking ath the RX value.
Sample output from a working environment
root@hostname:~/userspace-cni-network-plugin# kubectl logs -n ovs pod/ovs-app2 |tail -11
Port statistics ====================================
######################## NIC statistics for port 0 ########################
RX-packets: 404024960 RX-missed: 0 RX-bytes: 25857597440
RX-errors: 0
RX-nombuf: 0
TX-packets: 0 TX-errors: 0 TX-bytes: 0
Throughput (since last show)
Rx-pps: 1044345 Rx-bps: 534705080
Tx-pps: 0 Tx-bps: 0
############################################################################
The Userspace CNI networking plugin is written in GO and it uses GO Testing package for unit testing. In order to simplify unit test code, a testify toolkit is also used for assertions. Unit test files have suffix _test.go and they are located at the same directory as the production code.
For example userspace package have one source file and one file with unit tests:
userspace/userspace.go
userspace/userspace_test.goIn the file userspace_test.go you can find a set of test functions with the same
name as functions from userspace.go with prefix Test. Each test function
implements a set of tests to test production code thoroughly with different
input data, etc.
Testing binary is built and tests executed by go test command at directory
with test files.
For example, unit tests for userspace package with verbose output can be executed as follows:
cd userspace
go test -vAn example of verbose output of userspace unit tests:
=== RUN TestPrintVersionString
=== RUN TestPrintVersionString/verify_version_string
--- PASS: TestPrintVersionString (0.00s)
--- PASS: TestPrintVersionString/verify_version_string (0.00s)
=== RUN TestLoadNetConf
=== RUN TestLoadNetConf/fail_to_parse_netConf_1
=== RUN TestLoadNetConf/fail_to_parse_netConf_2
=== RUN TestLoadNetConf/fail_to_parse_netConf_3
=== RUN TestLoadNetConf/fail_to_parse_netConf_4
=== RUN TestLoadNetConf/fail_to_parse_netConf_5
=== RUN TestLoadNetConf/fail_to_parse_netConf_6
=== RUN TestLoadNetConf/fail_to_set_default_logging_level
=== RUN TestLoadNetConf/fail_to_set_log_file
=== RUN TestLoadNetConf/load_correct_netConf
--- PASS: TestLoadNetConf (0.00s)
--- PASS: TestLoadNetConf/fail_to_parse_netConf_1 (0.00s)
--- PASS: TestLoadNetConf/fail_to_parse_netConf_2 (0.00s)
--- PASS: TestLoadNetConf/fail_to_parse_netConf_3 (0.00s)
--- PASS: TestLoadNetConf/fail_to_parse_netConf_4 (0.00s)
--- PASS: TestLoadNetConf/fail_to_parse_netConf_5 (0.00s)
--- PASS: TestLoadNetConf/fail_to_parse_netConf_6 (0.00s)
--- PASS: TestLoadNetConf/fail_to_set_default_logging_level (0.00s)
--- PASS: TestLoadNetConf/fail_to_set_log_file (0.00s)
--- PASS: TestLoadNetConf/load_correct_netConf (0.00s)
...It is possible to execute unit tests directly from the host where the code is being developed. However it is strongly recommended to execute them from within a container. It will assure that unit tests will be isolated from the host and thus avoid any collision with the host environment (e.g. k8s or OVS). Another advantage is a possibility to easily execute unit tests at various Linux OS distributions. Support of unit test execution inside containers is implemented by project Makefile and described in following paragraphs.
Project Makefile defines a set of targets suitable for unit testing inside
containers. In order to build and use testing containers, both docker and
gcc (C preprocessor is used to assemble Dockerfiles) have to be installed
at the host.
Example of tools installation at Ubuntu:
apt install docker.io gccAfter the installation is completed, then any of unit test specific targets can be executed in order to:
- build test containers and update them in case that new commit is detected
- execute unit tests
- calculate code coverage
- cleanup unit test container images and generated docker files
Targets for container building, unit testing and code coverage calculation are prepared in three different versions to be executed at:
-
Linux OS distribution defined by
UT_OSenvironment variablee.g. execute build of container image for CentOS8
export UT_OS=centos8 make test-build -
Linux OS distribution specified by target suffix
e.g. calculate code coverage at Ubuntu 20.04
make coverage-ubuntu20.04
-
all supported OS distributions
e.g. execute unit tests at all supported OS distros:
make test-all
The list of supported make targets is also documented as part of make help
as follows:
Make Targets for unit testing inside containers:
make test-clean - Remove test container images and generated Dockerfiles.
make test-build - Build container image for unit tests with OS defined by UT_OS: UT_OS=ubuntu20.04
make test - Run unit tests inside container with OS defined by UT_OS: UT_OS=ubuntu20.04
make coverage - Calculate code coverage in container with OS defined by UT_OS: UT_OS=ubuntu20.04
make test-build-<os> - Build container image for unit tests with <os>, e.g. make test-build-centos8
make test-<os> - Run unit tests inside container with <os>, e.g. make test-centos8
make coverage-<os> - Calculate code coverage inside container with <os>, e.g. make coverage-centos8
make test-build-all - Build container images for unit tests for all supported OS distributions
e.g. make -j 5 test-build-all
make test-all - Run unit tests inside container for all supported OS distributions
e.g. make -j 5 test-all
make coverage-all - Calculate code coverage inside container for all supported OS distributions.
e.g. make -j 5 coverage-all