Deploy OpenShift 4 on a CentOS/RHEL 7/8 or Fedora KVM using libvirt IPI
CentOS/RHEL 7/8 or Fedora (tested in Fedora 31) host, the one included in the inventory file and that will host and launch the OpenShift cluster.
Subscriptions active in case of RHEL
It has been tested with OpenShift 4.4 (I'm not sure about backwards compatibility)
You need to prepare the install-config.yaml file, including:
- Your Pull secret from https://cloud.redhat.com/
- KVM IP
- Cluster name
- Domain
- Public SSH key
- Number of masters (1 or 3) and workers (0,1,2...or more)
This is a template example:
apiVersion: v1
baseDomain: < my.domain >
compute:
- architecture: amd64
hyperthreading: Enabled
name: worker
platform: {}
replicas: 2
controlPlane:
architecture: amd64
hyperthreading: Enabled
name: master
platform: {}
replicas: 3
metadata:
creationTimestamp: null
name: ocp
networking:
clusterNetwork:
- cidr: 10.128.0.0/14
hostPrefix: 23
machineNetwork:
- cidr: 192.168.126.0/24
networkType: OpenShiftSDN
serviceNetwork:
- 172.30.0.0/16
platform:
libvirt:
URI: qemu+tcp://< kvm ip on the "kvm_interface" interface >/system
network:
if: tt0
publish: External
pullSecret: '< pull secret >'
sshKey: |
< ssh key >
There are some variables that you will need to modify to configure the environment as per your needs:
-
ocp_install_file_path
description: Path to the pre-configured install-config.yaml default: "ocp-config/install-config.yaml" -
ocp_release
description: OCP release number from https://mirror.openshift.com/pub/openshift-v4/clients/ocp default: "4.4.0" -
ocp_master_memory
description: Memory (in MB) for Master nodes default: 16384 -
ocp_master_cpu
description: Number of cores for Master nodes default: 4 -
ocp_master_disk
description: Disk size (in GB) for Master nodes. NOTE: There is a bug in the provider that includes a k when size is big (ie. 500GB), it's better to stick to the default default: 120 -
ocp_worker_memory
description: Memory (in MB) for Master nodes default: 8192 -
ocp_worker_cpu
description: Number of cores for Worker nodes default: 2 -
ocp_worker_disk
description: Disk size (in GB) for Worker nodes default: 120 -
kvm_interface
description: Public interface of the KVM host (from where the OCP VMs will be published) as appears in 'nmcli con show' (it could be "System eno1" instead just eno1) default: "eth0"
There are other variables that shouldn't be modified unless you have
-
kvm_install
description: If "true" the role will install and prepare the KVM service default: "true" -
kvm_configure
description: If "true" the role will configure KVM service so libvirt IPI can use it default: "true" -
kvm_ext_dns
description: External DNS server default: "8.8.8.8" -
kvm_nestedvirtualization
description: Enable nested virtualization support in KVM (for OCP CNV testing) default: "False" -
nfs_storage
description: If "true" NFS storage and a storageclass for dynamic PV provisioning (it’s no supported in Openshift, but it works for testing) will be configured. ou could want to avoid configuring NFS if, for example, you want to install on a laptop and you don’t want to install anything else on your machine. If false only ephemeral storage will be available after the install. default: "true" -
lb
description: If true the role will install and configure haproxy to load balance API among masters and routers between workers. If not true only the first master (API) and to the first worker (APPS) will be getting external requests. default: "true" -
kvm_publish
description: If "true" the role will configure IPTABLES so external request will be forwarded to either the load balancer (if lb = "true") ot to the first master (API) and to the first worker (APPS) (or only to first master if no workers are setup). If "false" the environment will only be available locally to the KVM. default: "true" -
ocp_prepare
description: If "true" the role will prepare the host to launch the OpenShift installation default: "true" -
ocp_install
description: If "true" the role will use the host to launch the OpenShift installation default: "true" -
ocp_create_users
description: If "true" local users will be created. One cluster-wide admin (clusteradmin), 25 users (user1 - user25) included in a group ´developers´ and one cluster wide read only user (viewuser) included in a group called `reviewers`. You can disable it by configuring `ocp_create_users` to `false` or change the usernames or passwords modifying the `ocp_users_password` and `ocp_clusteradmin_password` variables default: "true" -
ocp_users_password
description: Password for the 25 users (userXX) and the cluster wide read only user (viewuser) default: "R3dhat01" -
ocp_clusteradmin_password
description: Password for the cluster-wide admin (clusteradmin) default: "R3dhat01"
First, you can download the role if you don't have it yet in your system:
ansible-galaxy install luisarizmendi.ocp_libvirt_ipi_role
You need to create a playbook (let's say ocp_libvirt_ipi.yaml), an inventory file with the KVM node details and you are good to import the role.
Find below a playbook example where we call the role and include the variables to customize the environment (release name can be found here)
---
- hosts: all
roles:
- role: luisarizmendi.ocp_libvirt_ipi_role
vars:
ocp_install_file_path: "ocp-config/install-config.yaml"
ocp_release: "4.4.0-rc.8"
ocp_master_memory: 16384
ocp_master_cpu: 4
ocp_master_disk: 150
ocp_worker_memory: 20480
ocp_worker_cpu: 4
ocp_worker_disk: 150
kvm_interface: "System eno1"
kvm_nestedvirtualization: "true"
Inventory file does not need any fancy stuff, this is an example:
kvm-node ansible_host=1.2.3.4 ansible_port=22 ansible_user=larizmen
So one probably directory structure that you could have to support the playbook (ocp_libvirt_ipi.yaml in this example) would be this one:
.
├── ansible.cfg
├── ansible.log
├── inventory
├── ocp-config
│ └── install-config.yaml
├── ocp_libvirt_ipi.yaml
...
After the playbook and the inventory are created you can use them to install (tags = install) and configure OpenShift :
ansible-playbook -vv -i <path to inventory> --tags install <path to playbook>
Or to remove (tags = remove) OpenShift and clean up the node:
ansible-playbook -vv -i <path to inventory> --tags remove <path to playbook>
If you still have doubts, you can find an example of all files here.
Luis Javier Arizmendi Alonso
Should I read this? Maybe, it might be interesting if you want to :
- better know how install OpenShift in a Fedora or CentOS/RHEL 7/8 node using libvirt IPI
- know the details about how to run OpenShift 4.4+ All-in-One setup
- know more about how to customize OpenShift installations, including modifying and compiling the OpenShift installer (I should include here 10k+ "not-supported, only for labs" disclaimers)
Let's start why you should be using this Ansible role...
Imagine that you have a baremetal node running (a not too old) Fedora or a CentOS/RHEL 7/8 (or a PC/Laptop with a good amount of RAM and CPU) and you want to run an OpenShift LAB on them, but you don't want to use CodeReady Containers because of multiple reasons, for example because you want to test the latest (or specific) bits, because you need more than one VM running OpenShift, or because you want the flexibility to build the lab as you want. In that case you could use multiple installation paths, for example simulating baremetal UPI, or using libvirt hooks but there is another way to do it that will bring extended capabilities (Machine API): OpenShift libvirt IPI.
OpenShift libvirt IPI is not intended to be used for installing production systems but it's quite helpful simplifying...well..."simplifying" (because you need some tips & tricks to make libvirt IPI work at this moment) the deployment on a KVM node. It's available from OpenShift 4.3 so it's quite new, thus there are multiple aspects that have to be taken into account, and that's why I made this Ansible Role.
The scripts will make use of libvirt IPI installation. The steps made by my scripts are based on https://github.com/openshift/installer/blob/master/docs/dev/libvirt/README.md
By default, masters will be using 16GB of RAM, 120 GB of disk and 4 vcores per node and workers 8GB of RAM, 120 GB of disk and 2 vcores. Those are the minimum requirements according to documentation (although 16GB of disk is enough if you don't want to use ephemeral). Bear in mind that you will need +2GB and 2 cores in your KVM node for bootstrap while installing (remember that OpenShift bootstrap VM is deleted during the installation steps).
You can choose to run a full OpenShift installation (with 3 masters and 2+ nodes), just 3 masters with no workers (masters will run both master and worker roles) or just 1 master (all-in-one). The all-in-one setup would need at least 16GB and 4 cores but put as much RAM and cores you can add, and also take into account the ephemeral storage, depending if you are going to use NFS or not (see below).
You will need to configure just the API and the APPS wildcard to use the environment, although you can always play with the /etc/hosts if you don't have a chance to configure a DNS (or configure a nip.io domain that includes the wildcard that you need).
This IPI installation won't need that you configure an external load balancer (although you can install it with just adjusting lb= "true" in the inventory file), any HTTP server or that you configure SRV in an external DNS.
You won't need to configure a Load Balancer because in the KVM iptables rule will be configured to forward 6443 to the first master and 443 and 80 to the first worker. That's OK if you are thinking about using 1 master and 1 or 1 workers or a all-in-one setup (in case that you don't deploy workers, all traffic will be to the first master), but if you plan to have multiple masters and workers, configuring a load balancer is a good idea because in case than more than 2 workers are deployed there is a chance that the router won't run on the first worker node where the iptables are forwarding. If you want to run HA tests you will need to install including a load balancer.
One last thing is that scripts were tested starting with OCP 4.4 (release candidate) and probably they will work for 4.4+ releases but won't be tested with previous versions.
Inside the playbooks there are some configurations that were needed to make the installation more flexible or even to make it finish. Some of then are in the steps that you can find in the libvirt IPI repo, but others are not...I will describe them in the order that are performed in my playbooks:
Apart from installing libvirt and enabling IP forwarding, libvirt IPI will need to "talk" with the server, so We'll need to accept TCP connections by configuring the appropriate variables in libvirtd.conf and running the service with --listen. This is well explained in the libvirt IPI repo
Libvirt IPI is not included in the installer software by default. In order to "activate" it, you need to build the installer from source including its support, so you have to clone the OpenShift installer GitHub repo and then use the build script (hack/build.sh) but including the variable TAGS setup to TAPS=libvirt
TAGS=libvirt hack/build.sh
..but before running that script to make the build, I applied two changes to the code (shown below).
This is important...wait just for a moment and think about it... you can do all of this because IT IS OPEN SOURCE. It would be impossible if we were using close-source Software...
///////////////////////////
NOTE: Since OpenShift 4.5 this is not needed anymore. I commented out the part that makes these changes in the source code, if you want to keep using older openshift versions you will have to re-enable that part of the playbook.
///////////////////////////
All (supported) IPI providers have a way to modify the created VM resources (CPU, memory and disk). In the libvirt IPI case you can modify the CPU and memory using the manifest (see below) just changing the values that are already there. In order to change the worker nodes disk size you have to include an additional variable that is not in the manifest by default, but it works. The problem is that the code is not (yet?) prepared to allow master nodes disk size changes so the only way to do it is by adding the 'size' variable in the code in file data/data/libvirt/main.tf including the size in bytes.
...
resource "libvirt_volume" "master" {
size = < size >
count = var.master_count
name = "${var.cluster_id}-master-${count.index}"
base_volume_id = module.volume.coreos_base_volume_id
pool = libvirt_pool.storage_pool.name
}
...
The reason why masters are not that flexible is that for LABs, with the default disk size you are good to go....but maybe there is one use case where you want to increase the disk size, and that's when you want to run an All-in-One setup (this is allowed in my playbooks) because the master will need to run the workloads as well (even more important if you don't have configured any storage backend to have PVs).
One more thing to be taken into account is that due an issue with libvirt and the need of a wildcard for the console we need to either forward requests for the APPS URL to the KVM dnsmasq in the openshift-installer code (so we skip the limitation on the libvirt dnsmasq) or, if we don't want to modify (more) the code we could configure a different (from default) APPS URL. We would configure *.apps.< basedomain > instead of *.apps.< CLUSTERNAME >.< basedomain >, in order to let the openshift console being deploy, so bear in mind that change and do not include the cluster name when trying to access your APPs in this cluster. This change should be also done in the manifests, in this case by removing the "cluster name" part (probably ocp if you didn't change it in the install-config.yaml) from the url that appears in the manifests/cluster-ingress-02-config.yml file.
In my case, I decided to change the code in installer/data/data/libvirt/main.tf and forward the *.apps.< clustername >.< basedomain > to the KVM dnsmasq, so we can avoid removing the clustername from the APPS URL making possible to keep using the default one.
When running the install in dedicated hardware with plenty of resources the default timeouts are ok... but this installer is intended to be used even in Laptops, so sometimes it takes longer than that. The only way to modify the timeouts of the openshift installer at this moment (I'm not sure that this will change) is modifying the code.
There different timers:
- Waiting for bootstrap Kubernetes API (default 20 minutes)
- Waiting for bootstrap to complete (40 minutes)
- Waiting for confirmation that the cluster has been initialized (30 minutes, or 60 minutes if it's a baremetal installation)
- Wating for Console URL from the route 'console' in namespace openshift-console (default 10 minutes)
We already review that we could need to change the default disk size (only 16GB) reserved for our OpenShift nodes created by libvirt IPI. Masters can only be changed fixing the size in code. For workers a variable can be added in the manifests, so first we have to create the manifests (openshift-installer create manifests --dir <installation dir>) and then modify the openshift/99_openshift-cluster-api_worker-machineset-0.yaml file, adding the volumeSize with the disk size in bytes
...
spec:
...
template:
...
spec:
...
providerSpec:
value:
...
volume:
volumeSize: <size>
baseVolumeID: ocp-2972r-base
poolName: ocp-2972r
volumeName: ""
...
When building installer from the source code we use images that have OKD content, and we want to install the desired Red Hat OpenShift release, so we need to override the image that the installer will use, otherwise the bootstrap will start but the master will be stuck before showing the login prompt.
You can override the release image by exporting the OPENSHIFT_INSTALL_RELEASE_IMAGE_OVERRIDE variable before launching the openshift installation.
Be aware that you will need to select the right values for ocp_release, ocp_git_branch and ocp_install_install_release_image_override otherwise the installation will fail.
Before OpenShift 4.4, if you configured just one single master node the OpenShift installer will go through with no problems, but starting in OpenShift 4.4 the way that the ETCD cluster is installed and managed has changed. Now the Cluster ETCD operator must allow having just one node as Master (so having no ETCD quorum). This is done using the pretty weird (and auto-explanatory) variable useUnsupportedUnsafeNonHANonProductionUnstableEtcd as you can see in this BUG
You have to wait until the Kubernetes API is ready after launching the install and then apply this patch:
oc patch etcd cluster -p='{"spec": {"unsupportedConfigOverrides": {"useUnsupportedUnsafeNonHANonProductionUnstableEtcd": true}}}' --type=merge
Just in case you didn't notice in the variable name, this is not supported, unstable, not safe and for non HA (of course) environments.
Why not all variables are named as this one? We could just get rid of documentation in that case.
Terraform libvirt plugin does not make dhcp host entries persistent, so if you reboot the KVM node the master nodes won't come up again since they will have the 'api' hostname instead of their proper name , thus the node won't be recognized by ETCD. There are multiple ways to fix this, I decided to do the quick-and-dirty way, just reconfiguring those entries in the libvirt network as persistent.
If you need to add extra manifests to the installation, put them on a folder called extra_manifests and they will be copied over to the manifests folder before the installation begins.
That's all, make responsible use of these playbooks (remember that this is just for LABs) and enjoy the Machine API even when your LAB is running on a single KVM node.