Goal

Sdewan config agent is the controller of Sdewan CRDs. With the config agent, we are able to deploy CNFs. In this page, we have the following terms, let's define them here.

CNF Deployment: A deployment running network function process(openWRT)
Sdewan rule: The rule defines the CNF behaves. We have 3 classes of rules: mwan3, firewall, ipsec. Each class includes several kinds of rules. For example, mwan3 has 2 kinds: mwan3_policy and mwan3_rule. Firewall has 5 kinds: firewall_zone, firewall_snat, firewall_dnat, firewall_forwarding, firewall_rule. Ipsec has xx(ruoyu) kinds: xx, xx.
Sdewan rule CRD: The CRD defines each kind of sdewan rule. For each kind of Sdewan rule, we have a Sdewan rule CRD. Sdewan rule CRD is namespaced resource.
Sdewan rule CR: Instance of Sdewan rule CRD.
Sdewan controller: The controller watching Sdewan rule CRs.
CNF: A network function running in container.

To deploy a CNF, user needs to create one CNF deployment and some Sdewan rule CRs. In a Kubernetes namespace, there could be more than one CNF deployment and many Sdewan rule CRs. We use label to correlate one CNF with some Sdewan rule CRs. The Sdewan controller watches Sdewan rule CRs and applies them onto the correlated CNF by calling CNF REST api.

Sdwan Design Principle

There could be multiple tenants/namespaces in a Kubernetes cluster. User may deploy multiple CNFs in any one or more tenants.
The replica of CNF deployment could be more than one for active/backup purpose. We should apply rules for all the pods under CNF deployment. (This release doesn't implement VRRP between pods)
CNF deployment and Sdewan rule CRs can be created/updated/deleted in any order
The Sdewan controller and CNF process could be crash/restart at anytime for some reasons. We need to handle these scenarios
Each Sdewan rule CR has labels to identify the type it belongs to. 3 types are available at this time: basic, app-intent and k8s-service. We extend k8s user role permission so that we can set user permission at type level of Sdewan rule CR
Sdewan rule CR dependencies are checked on creating/updating/deleting. For example, if we create a mwan3_rule CR which uses policy policy-x, but no mwan3_policy CR named policy-x exists. Then we block the request

CNF Deployment

In this section we describe what the CNF deployment should be like, as well as the pod under the deployment.

CNF pod should has multiple network interfaces attached. We use multus and ovn4nfv CNIs to enable multiple interfaces. So in the CNF pod yaml, we set annotations: k8s.v1.cni.cncf.io/networks, k8s.plugin.opnfv.org/nfn-network.
When user deploys a CNF, she/he most likely want to deploy the CNF on a specified node instead of a random node. Because some nodes may don't have provider network connected. So we set spec.nodeSelector for pod
CNF pod runs Sdewan CNF (based on openWRT in ICN). We use image integratedcloudnative/openwrt:dev
CNF pod should setup with rediness probe. Sdewan controller would check pod readiness before calling CNF REST api.

CNF pod

apiVersion: extensions/v1beta1
kind: Deployment
metadata: 
  name: cnf-1
  namespace: default
  labels:
    sdewanPurpose: cnf-1
spec:
  replicas: 1
  strategy:
    rollingUpdate:
      maxSurge: 25%
      maxUnavailable: 25%
    type: RollingUpdate
  template:
    metadata:
      annotations:
        k8s.plugin.opnfv.org/nfn-network: |-
          { "type": "ovn4nfv", "interface": [
            {
              "defaultGateway": "false",
              "interface": "net0",
              "name": "ovn-priv-net"
            },
            {
              "defaultGateway": "false",
              "interface": "net1",
              "name": "ovn-provider-net1"
            },
            {
              "defaultGateway": "false",
              "interface": "net2",
              "name": "ovn-provider-net2"
            }
          ]}
        k8s.v1.cni.cncf.io/networks: '[{ "name": "ovn-networkobj"}]'
    spec:
      containers:
      - command:
        - /bin/sh
        - /tmp/sdewan/entrypoint.sh
        image: integratedcloudnative/openwrt:dev
        name: sdewan
        readinessProbe:
          failureThreshold: 5
          httpGet:
            path: /
            port: 80
            scheme: HTTP
          initialDelaySeconds: 5
          periodSeconds: 5
          successThreshold: 1
          timeoutSeconds: 1
        securityContext:
          privileged: true
          procMount: Default
        volumeMounts:
        - mountPath: /tmp/sdewan
          name: example-sdewan
          readOnly: true
      nodeSelector:
        kubernetes.io/hostname: ubuntu18

Sdewan rule CRs

CRD defines all properties of a resource, but it's not human friendly. So we paste Sdewan rule CR samples instead of CRDs.

Each Sdewan rule CR has a label named sdewanPurpose to indicate which CNF should the rule be applied onto
Each Sdewan rule CR has the status field which indicates if the latest rule is applied and when it's applied
Mwan3Policy.spec.members[].network should match the networks defined in CNF pod annotation k8s.plugin.opnfv.org/nfn-network. As well as FirewallZone.spec[].network

CR samples of Mwan3 type:

Mwan3Policy CR

apiVersion: batch.sdewan.akraino.org/v1alpha1
kind: Mwan3Policy
metadata:
  name: balance1
  namespace: default
  labels:
    sdewanPurpose: cnf-1
spec:  
  members:
    - network: ovn-net1
      weight: 2
      metric: 2
    - network: ovn-net2
      weight: 3
      metric: 3
status:
  appliedVersion: "2"
  appliedTime: "2020-03-29T04:21:48Z"
  inSync: True

Mwan3Rule CR

apiVersion: batch.sdewan.akraino.org/v1alpha1
kind: Mwan3Rule
metadata:
  name: http_rule
  namespace: default
  labels:
    sdewanPurpose: cnf-1
spec:
  policy: balance1
  src_ip: 192.168.1.2
  dest_ip: 0.0.0.0/0
  dest_port: 80
  proto: tcp
status:
  appliedVersion: "2"
  appliedTime: "2020-03-29T04:21:48Z"
  inSync: True

CR samples of Firewall type:

apiVersion: batch.sdewan.akraino.org/v1alpha1
kind: FirewallZone
metadata:
  name: lan1
  namespace: default
  labels:
    sdewanPurpose: cnf-1
spec:
    newtork:
      - ovn-net1
    input: ACCEPT
    output: ACCEPT
status:
  appliedVersion: "2"
  appliedTime: "2020-03-29T04:21:48Z"
  inSync: True

apiVersion: batch.sdewan.akraino.org/v1alpha1
kind: FirewallRule
metadata:
  name: reject_80
  namespace: default
  labels:
    sdewanPurpose: cnf-1
spec:  
  src: lan1
  src_ip: 192.168.1.2
  src_port: 80
  proto: tcp
  target: REJECT
status:
  appliedVersion: "2"
  appliedTime: "2020-03-29T04:21:48Z"
  inSync: True

apiVersion: batch.sdewan.akraino.org/v1alpha1
kind: FirewallSNAT
metadata:
  name: snat_lan1
  namespace: default
  labels:
    sdewanPurpose: cnf-1
spec:  
  src: lan1
  src_ip: 192.168.1.2
  src_dip: 1.2.3.4
  dest: wan1
  proto: icmp
status:
  appliedVersion: "2"
  appliedTime: "2020-03-29T04:21:48Z"
  inSync: True

apiVersion: batch.sdewan.akraino.org/v1alpha1
kind: FirewallDNAT
metadata:
  name: dnat_wan1
  namespace: default
  labels:
    sdewanPurpose: cnf-1
spec:  
  src: wan1
  src_dport: 19900
  dest: lan1
  dest_ip: 192.168.1.1
  dest_port: 22
  proto: tcp
status:
  appliedVersion: "2"
  appliedTime: "2020-03-29T04:21:48Z"
  inSync: True

apiVersion: batch.sdewan.akraino.org/v1alpha1
kind: FirewallForwarding
metadata:
  name: forwarding_lan_to_wan
  namespace: default
  labels:
    sdewanPurpose: cnf-1
spec:  
  src: lan1
  dest: wan1
status:
  appliedVersion: "2"
  appliedTime: "2020-03-29T04:21:48Z"
  inSync: True

CR samples of IPSec type(ruoyu):

Sdewan rule CRD Reconcile Logic

As we have many kinds of CRDs, they have almost the same reconcile logic. So we only describe the Mwan3Rule logic.

Mwan3Rule Reconcile could be triggered by the following cases:

Create/Update/Delete Mwan3Rule CR
CNF deployment ready status change (With predicate feature, we can only watch CNF deployment readiness status. With enqueueRequestsFromMapFunc, we can enqueue all Mwan3Rule CRs with specified labels.sdewanPurpose, if CNF deployment's ready status changes)
- CNF becomes ready after creating
- CNF becomes ready after restart
- CNF becomes not-ready after crash

Mwan3Rule Reconcile flow:

def Mwan3RuleReconciler.Reconcile(req ctrl.Request):
  rule_cr = k8sClient.get(req.NamespacedName)
  cnf_deployment = k8sClient.get_deployment_with_label(rule_cr.labels.sdewanPurpose)
  if rule_cr DeletionTimestamp exists:
    # The CR is being deleted. finalizer on the CR
    if cnf_deployment exists:
      if cnf_deployment is ready:
        for cnf_pod in cnf_deployment:
          err = openwrt_client.delete_rule(cnf_pod_ip, rule_cr)
          if err:
            return "re-queue req"
        rule_cr.finalizer = nil
        return "ok"
      else:
        return "re-queue req"
    else:
      # Just remove finalizer, because no CNF pod exists
      rule_cr.finalizer = nil
      return "ok"
  else:
    # The CR is not being deleted
    if cnf_deployment not exist:
      return "ok"
    else:
      if cnf_deployment not ready:
        # set appliedVersion = nil if cnf_deployment get into not_ready status
        rule_cr.status.appliedVersion = nil
        return "re-queue req"
      else:
        for cnf_pod in cnf_deployment:
          runtime_cr = openwrt_client.get_rule(cnf_pod_ip)
          if runtime_cr != rule_cr:
            err = openwrt_client.add_or_update_rule(cnf_pod_ip, rule_cr)
            if err:
              # err could be caused by dependencies not-applied or other reason
              return "re-queue req"
        # set appliedVerson only when it's applied for all the cnf pods
        rule_cr.finalizer = cnf_finalizer
        rule_cr.status.appliedVersion = rule_cr.resourceVersion
        rule_cr.status.inSync = True
        return "ok"

Unsual Cases

Controller goes down -> Create CNF Deployment and rule CRs -> Controller goes up
- No reconcile executed before the controller goes up. The rule CRs have empty status, no rules applied to the CNF deployment
- Once the controller goes up, it reconciles every rule CR. In the reconcile function, rules are applied and rule CRs status.appliedVersion are updated
Controller goes down -> delete rule CRs -> Controller goes up
- During the controller are down, rule are not deleted from the CNF Deployment. The rule CRs are not deleted from k8s etcd because of finalizer.
- Once the controller goes up, it reconciles every rule CR. It calls CNF api to delete rules and remove CR finalizer.
CNF deployment goes to not-ready -> after some time -> CNF deployment goes to ready status
- As the CNF deployment goes to not-ready, the controller reconciles every CR which matchs the CNF deployment, to set status.appliedVersion=nil.
- Once the CNF deployment goes to ready status, controller receives the event and reconciles every rule CR. It applies the rule and set status.appliedVersion.
Controller goes down -> CNF deployment pod restart -> Controller goes up
- During the time controller is down, the CNF pod is restarted, the rules do not exist in restarted pod. So all the related rule CRs status.appliedVersion should be set to nil. But the controller is down, it can't receive the CNF down/up event.
- When the controller goes up, it reconciles every rule CR. But it doesn't know the CNF had every restarted. This is a problem, so we can't use CR status.appliedVersion to record if the rule is applied or not. Instead we should already call API to get existing rule and compare, and finally apply if existing rule in CNF is different with CR defination.

Admission Webhook Usage

We use admission webhook to implemention several features.

Prevent creating more than one CNF of the same lable and the same namespace
Validate CR dependencies. For example, mwan3 rule depends on mwan3 policy
Extend user permission to control the operations on rule CRs. For example, we can control that ONAP can't update/delete rule CRs created by platform.

Sdewan rule CR type level Permission Implementation

8s support permission control on namespace level. For example, user1 may be able to create/update/delete one kind of resource(e.g. pod) in namespace ns1, but not namespace ns2. For Sdewan, this can't fit our requirement. We want label level control of Sdewan rule CRs. For example, user_onap can create/update/delete Mwan3Rule CR of label sdewan-bucket-type=app-intent, but not label sdewan-bucket-type=basic.

Let me first describe the extended permission system and then explain how we implement it. In k8s, user or serviceAccount could be bonded to one or more roles. The roles defines the permissions, for example the following role defines that sdewan-test role can create/update Mwan3Rule CRs in default namespace. Also sdewan-testrole can get Mwan3Policy CRs.

apiVersion: rbac.authorization.k8s.io/v1
kind: Role
metadata:
  annotations:
  name: sdewan-test
  namespace: default
rules:
- apiGroups:
  - ""
  resources:
  - mwan3rules
  verbs:
  - create
  - update
- apiGroups:
  - ""
  resources:
  - mwan3policies
  verbs:
  - get

We extend the Role with annotations. In the annotation, we can define labled based permissions. For example, the following role extends sdewan-test role permission: sdewan-test can only create/update Mwan3Rule CRs with label sdewan-bucket-type=app-intent or sdewan-bucket-type=k8s-service. Also it can only get Mwan3Policy CR with label sdewan-bucket-type=app-intent.

apiVersion: rbac.authorization.k8s.io/v1
kind: Role
metadata:
  annotations:
    sdewan-bucket-type-permission: |-
      { "mwan3rules": ["app-intent", "k8s-service"],
      "mwan3policies": ["app-intent"] }
  name: sdewan-test
  namespace: default
rules:
- apiGroups:
  - ""
  resources:
  - mwan3rules
  verbs:
  - create
  - update
- apiGroups:
  - ""
  resources:
  - mwan3policies
  verbs:
  - get

We use admission webhook to implement the type level permission control. Let me describe how admission webhook in simple words. When k8s api receives a request, kube-api call webhook API before save the object into etcd. If the webhook returns allowed=true, kube-api continues to persistent the object into etcd. Otherwise, kube-api reject the request. The webhook can optional tell kube-api to update the object together with allowed=true returned. Webhook request body has a field named userInfo, it indicates who is making the k8s api request. With this field, we can implement the extended permission in webhook.

def mwan3rule_webhook_handle_permission(req admission.Request):
  userinfo = req["userInfo]
  mwan3rule_cr = decode(req)
  roles = k8s_client.get_role_from_user(userinfo)
  for role in roles:
    if mwan3rule_cr.labels.sdewan-bucket-type in role.annotation.sdewan-bucket-type-permission.mwan3rules:
      return {"allowd": True}
  return {"allowd": False}

ServiceRule controller (For next release)

We create a controller watches the services created in the cluster. For each service, it creates a FirewallDNAT CR. On controller startup, it makes a syncup to remove unused CRs.

Sdewan config Agent