Skip to content
Permalink
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
452 lines (348 sloc) 20.4 KB
title canonical_url
Frequently Asked Questions
  • TOC {:toc}

"Why use {{site.prodname}}?"

The problem {{site.prodname}} tries to solve is the networking of workloads (VMs, containers, etc) in a high scale environment. Existing L2 based methods for solving this problem have problems at high scale. Compared to these, we think {{site.prodname}} is more scalable, simpler and more flexible. We think you should look into it if you have more than a handful of nodes on a single site.

{{site.prodname}} also provides a rich network security model that allows operators and developers to declare intent-based network security policy that is automatically rendered into distributed firewall rules across a cluster of containers, VMs, and/or servers.

For a more detailed discussion of this topic, see our blog post at Why Calico?.

"Does {{site.prodname}} work with IPv6?"

Yes! {{site.prodname}}'s core components support IPv6 out-of-the box. However, not all orchestrators that we integrate with support IPv6 yet.

"Why does my container have a route to 169.254.1.1?"

In a {{site.prodname}} network, each host acts as a gateway router for the workloads that it hosts. In container deployments, {{site.prodname}} uses 169.254.1.1 as the address for the {{site.prodname}} router. By using a link-local address, {{site.prodname}} saves precious IP addresses and avoids burdening the user with configuring a suitable address.

While the routing table may look a little odd to someone who is used to configuring LAN networking, using explicit routes rather than subnet-local gateways is fairly common in WAN networking.

Why can't I see the 169.254.1.1 address mentioned above on my host?

{{site.prodname}} tries hard to avoid interfering with any other configuration on the host. Rather than adding the gateway address to the host side of each workload interface, {{site.prodname}} sets the proxy_arp flag on the interface. This makes the host behave like a gateway, responding to ARPs for 169.254.1.1 without having to actually allocate the IP address to the interface.

Why do all cali* interfaces have the MAC address ee:ee:ee:ee:ee:ee?

In some setups the kernel is unable to generate a persistent MAC address and so {{site.prodname}} assigns a MAC address itself. Since {{site.prodname}} uses point-to-point routed interfaces, traffic does not reach the data link layer so the MAC Address is never used and can therefore be the same for all the cali* interfaces.

Can I prevent my Kubernetes pods from initiating outgoing connections?

Yes! The Kubernetes NetworkPolicy API added support for egress policies in v1.8. You can also use calicoctl to configure egress policy to prevent Kubernetes pods from initiating outgoing connections based on the full set of supported {{site.prodname}} policy primitives including labels, Kubernetes namespaces, CIDRs, and ports.

I've heard {{site.prodname}} uses proxy ARP, doesn't proxy ARP cause a lot of problems?

It can, but not in the way that {{site.prodname}} uses it.

In container deployments, {{site.prodname}} only uses proxy ARP for resolving the 169.254.1.1 address. The routing table inside the container ensures that all traffic goes via the 169.254.1.1 gateway so that is the only IP that will be ARPed by the container.

"Is {{site.prodname}} compliant with PCI/DSS requirements?"

PCI certification applies to the whole end-to-end system, of which {{site.prodname}} would be a part. We understand that most current solutions use VLANs, but after studying the PCI requirements documents, we believe that {{site.prodname}} does meet those requirements and that nothing in the documents mandates the use of VLANs.

How do I enable IPIP and NAT Outgoing on an IP Pool?

  1. Retrieve current IP Pool config

    calicoctl get ipPool --export -o yaml > pool.yaml
  2. Modify IP Pool config

    Modify the pool's spec to enable IP-IP and natOutgoing. (See IP Pools for other settings that can be edited.)

    - apiVersion: projectcalico.org/v3
      kind: IPPool
      metadata:
       name: ippool-1
      spec:
        cidr: 192.168.0.0/16
        ipipMode: Always
        natOutgoing: true
  3. Load the modified file.

    calicoctl replace -f pool.yaml

"How does {{site.prodname}} maintain saved state?"

State is saved in a few places in a {{site.prodname}} deployment, depending on whether it's global or local state.

Local state is state that belongs on a single compute host, associated with a single running Felix instance (things like kernel routes, tap devices etc.). Local state is entirely stored by the Linux kernel on the host, with Felix storing it only as a temporary mirror. This makes Felix effectively stateless, with the kernel acting as a backing data store on one side and etcd as a data source on the other.

If Felix is restarted, it learns current local state by interrogating the kernel at start up. It then reads from etcd all the local state which it should have, and updates the kernel to match. This approach has strong resiliency benefits, in that if Felix restarts you don't suddenly lose access to your VMs or containers. As long as the Linux kernel is running, you've still got full functionality.

The bulk of global state is mastered in whatever component hosts the plugin.

  • In the case of OpenStack, this means a Neutron database. Our OpenStack plugin (more strictly a Neutron ML2 driver) queries the Neutron database to find out state about the entire deployment. That state is then reflected to etcd and so to Felix.
  • In certain cases, etcd itself contains the master copy of the data. This is because some Docker deployments have an etcd cluster that has the required resiliency characteristics, used to store all system configuration -and so etcd is configured so as to be a suitable store for critical data.
  • In other orchestration systems, it may be stored in distributed databases, either owned directly by the plugin or by the orchestrator itself.

The only other state storage in a {{site.prodname}} network is in the BGP sessions, which approximate a distributed database of routes. This BGP state is simply a replicated copy of the per-host routes configured by Felix based on the global state provided by the orchestrator.

This makes the {{site.prodname}} design very simple, because we store very little state. All of our components can be shutdown and restarted without risk, because they resynchronize state as necessary. This makes modelling their behaviour extremely simple, reducing the complexity of bugs.

"I heard {{site.prodname}} is suggesting layer 2: I thought you were layer 3! What's happening?"

It's important to distinguish what {{site.prodname}} provides to the workloads hosted in a data center (a purely layer 3 network) with what the {{site.prodname}} project recommends operators use to build their underlying network fabric.

{{site.prodname}}'s core principle is that applications and workloads overwhelmingly need only IP connectivity to communicate. For this reason we build an IP-forwarded network to connect the tenant applications and workloads to each other, and the broader world.

However, the underlying physical fabric obviously needs to be set up too. Here, {{site.prodname}} has discussed how both a layer 2 (see here) or a layer 3 (see here) fabric could be integrated with {{site.prodname}}. This is one of the great strengths of the {{site.prodname}} model: it allows the infrastructure to be decoupled from what we show to the tenant applications and workloads.

We have some thoughts on different interconnect approaches (as noted above), but just because we say that there are layer 2 and layer 3 ways of building the fabric, and that those decisions may have an impact on route scale, does not mean that {{site.prodname}} is "going back to Ethernet" or that we're recommending layer 2 for tenant applications. In all cases we forward on IP packets, no matter what architecture is used to build the fabric.

"How do I control policy/connectivity without virtual/physical firewalls?"

{{site.prodname}} provides an extremely rich security policy model, applying policy at the first and last hop of the routed traffic within the {{site.prodname}} network (the source and destination compute hosts).

This model is substantially more robust to failure than a centralised firewall-based model. In particular, the {{site.prodname}} approach has no single-point-of-failure: if the device enforcing the firewall has failed then so has one of the workloads involved in the traffic (because the firewall is enforced by the compute host).

This model is also extremely amenable to scaling out. Because we have a central repository of policy configuration, but apply it at the edges of the network (the hosts) where it is needed, we automatically ensure that the rules match the topology of the data center. This allows easy scaling out, and gives us all the advantages of a single firewall (one place to manage the rules), but none of the disadvantages (single points of failure, state sharing, hairpinning of traffic, etc.).

Lastly, we decouple the reachability of nodes and the policy applied to them. We use BGP to distribute the topology of the network, telling every node how to get to every endpoint in case two endpoints need to communicate. We use policy to decide if those two nodes should communicate, and if so, how. If policy changes and two endpoints should now communicate, where before they shouldn’t have, all we have to do is update policy: the reachability information does not change. If later they should be denied the ability to communicate, the policy is updated again, and again the reachability doesn’t have to change.

"How does {{site.prodname}} interact with the Neutron API?"

This document document goes into extensive detail about how various Neutron API calls translate into {{site.prodname}} actions.

Why isn't the -p flag on docker run working as expected?

The -p flag tells Docker to set up port mapping to connect a port on the Docker host to a port on your container via the docker0 bridge.

If a host's containers are connected to the docker0 bridge interface, {{site.prodname}} would be unable to enforce security rules between workloads on the same host; all containers on the bridge would be able to communicate with one other.

You can securely configure port mapping by following our guide on Exposing Container Ports to the Internet.

Can {{site.prodname}} containers use any IP address within a pool, even subnet network/broadcast addresses?

Yes! {{site.prodname}} is fully routed, so all IP address within a {{site.prodname}} pool are usable as private IP addresses to assign to a workload. This means addresses commonly reserved in a L2 subnet, such as IPv4 addresses ending in .0 or .255, are perfectly okay to use.

How do I get network traffic into and out of my {{site.prodname}} cluster?

The recommended way to get traffic to/from your {{site.prodname}} network is by peering to your existing data center L3 routers using BGP and by assigning globally routable IPs (public IPs) to containers that need to be accessed from the internet. This allows incoming traffic to be routed directly to your containers without the need for NAT. This flat L3 approach delivers exceptional network scalability and performance.

A common scenario is for your container hosts to be on their own isolated layer 2 network, like a rack in your server room or an entire data center. Access to that network is via a router, which also is the default router for all the container hosts.

If this describes your infrastructure, the External Connectivity tutorial explains in more detail what to do. Otherwise, if you have a layer 3 (IP) fabric, then there are detailed datacenter networking recommendations given in the main this article. We'd also encourage you to get in touch to discuss your environment.

How can I enable NAT for outgoing traffic from containers with private IP addresses?

If you want to allow containers with private IP addresses to be able to access the internet then you can use your data center's existing outbound NAT capabilities (typically provided by the data center's border routers).

Alternatively you can use {{site.prodname}}'s built in outbound NAT capability by enabling it on any {{site.prodname}} IP pool. In this case {{site.prodname}} will perform outbound NAT locally on the compute node on which each container is hosted.

cat << EOF | calicoctl apply -f -
apiVersion: projectcalico.org/v3
kind: IPPool
metadata:
  name: ippool-1
spec:
  cidr: <CIDR>
  natOutgoing: true
EOF

Where <CIDR> is the CIDR of your IP pool, for example 192.168.0.0/16.

Remember: the security profile for the container will need to allow traffic to the internet as well. Refer to the appropriate guide for your orchestration system for details on how to configure policy.

How can I enable NAT for incoming traffic to containers with private IP addresses?

As discussed, the recommended way to get traffic to containers that need to be accessed from the internet is to give them public IP addresses and to configure {{site.prodname}} to peer with the data center's existing L3 routers.

In cases where this is not possible then you can configure incoming NAT (also known as DNAT) on your data centers existing border routers. Alternatively you can configure incoming NAT with port mapping on the host on which the container is running on.

  1. Create a new chain called "expose-ports" to hold the NAT rules.

    iptables -t nat -N expose-ports
  2. Jump to that chain from the OUTPUT and PREROUTING chains.

    iptables -t nat -A OUTPUT -j expose-ports
    iptables -t nat -A PREROUTING -j expose-ports

    Tip: The OUTPUT chain is hit by traffic originating on the host itself; The PREROUTING chain is hit by traffic coming from elsewhere. {: .alert .alert-success}

  3. For each port you want to expose, add a rule to the expose-ports chain, replacing <PUBLIC_IP> with the host IP that you want to use to expose the port and <PUBLIC_PORT> with the host port.

    iptables -t nat -A expose-ports -p tcp --destination <PUBLIC_IP> \
    --dport <PUBLIC_PORT> -j DNAT  --to <CALICO_IP>:<SERVICE_PORT>

For example, you have a container to which you've assigned the CALICO_IP of 192.168.7.4, and you have NGINX running on port 8080 inside the container. If you want to expose this service on port 80 and your host has IP 192.0.2.1, then you could run the following commands:

iptables -t nat -N expose-ports
iptables -t nat -A OUTPUT -j expose-ports
iptables -t nat -A PREROUTING -j expose-ports

iptables -t nat -A expose-ports -p tcp --destination 192.0.2.1 --dport 80 -j DNAT --to 192.168.7.4:8080

{: .alert .alert-success}

The commands will need to be run each time the host is restarted.

Remember: the security profile for the container will need to allow traffic to the exposed port as well. Refer to the appropriate guide for your orchestration system for details on how to configure policy.

Can I run {{site.prodname}} in a public cloud environment?

Yes. If you are running in a public cloud that doesn't allow either L3 peering or L2 connectivity between {{site.prodname}} hosts then you can enable ipip in your {{site.prodname}} IP pool:

cat << EOF | calicoctl apply -f -
apiVersion: projectcalico.org/v3
kind: IPPool
metadata:
  name: ippool-1
spec:
  cidr: <CIDR>
  ipipMode: Always
  natOutgoing: true
EOF

{{site.prodname}} will then route traffic between {{site.prodname}} hosts using IP in IP.

In AWS, you disable Source/Dest. Check instead of using IP in IP as long as all your instances are in the same subnet of your VPC. This will provide the best performance. You can disable this with the CLI, or right click the instance in the EC2 console, and Change Source/Dest. Check from the Networking submenu.

aws ec2 modify-instance-attribute --instance-id <INSTANCE_ID> --source-dest-check "{\"Value\": false}"

cat << EOF | calicoctl apply -f -
apiVersion: projectcalico.org/v3
kind: IPPool
metadata:
  name: ippool-2
spec:
  cidr: <CIDR>
  natOutgoing: true
EOF

On AWS with IP in IP, why do I see no connectivity between workloads or only see connectivity if I ping in both directions?

By default, AWS security groups block incoming IP in IP traffic.

However, if an instance has recently sent some IP in IP traffic out when it receives some incoming IP in IP traffic, then AWS sees that as a response to an outgoing connection and it allows the incoming traffic. This leads to some very confusing behavior where traffic can be blocked and then suddenly start working!

To resolve the issue, add a rule to your security groups that allows inbound and outbound IP in IP traffic (IP protocol number 4) between your hosts.

In Calico for OpenStack, why can't a VM ping its default gateway?

With typical OpenStack networking drivers other than Calico,

  • OpenStack VMs in the same Neutron network appear to be directly connected to each other at layer 2 (Ethernet).

  • When a VM sends to another VM in the same network, there is no routing at all, from the VM point of view. (Of course there may be routing in the underlay network, because compute hosts may be on different subnets.)

  • When a VM sends to something outside its own network, it goes - by simulated layer 2 - to the default gateway first, and then is routed to wherever it is addressed to.

With Calico, this is all different. Any packet sent by a VM is layer-2-terminated and IP-routed by the VM's compute host, whether the VM is sending to another VM in the same network, or to anywhere else. So Calico doesn't need the "default gateway" concept, and it doesn't really make any sense with Calico. If a VM thinks that "my default gateway is the first hop at which the packets I send can be IP-routed", and in any way relies on that, that will be wrong, with Calico networking.

Now, with all that said, for detailed technical reasons to do with the DHCP server (dnsmasq), Calico does actually configure the default gateway IP - i.e. bind it to a Linux network interface - on every compute host with at least one VM in the relevant Neutron network; and that is one of the ingredients needed, in Linux, for a VM to be able to ping that IP.

The reason why it still isn't possible for a VM to ping that IP, is that Calico by default configures iptables rules to block almost all communication to its own host - because in general, of course, a workload should not be able to access and possibly compromise its host. There are a few pinholes here, e.g. for DHCP, but those do not include ping (ICMP Echo). If you start running a command like watch 'sudo iptables-save -c | grep DROP' on a compute host, and then try pinging the default gateway IP from a VM on that host, you will see the DROP count increasing as each ping packet is sent and blocked.

This behaviour is controlled by a config parameter named DefaultEndpointToHostAction, whose default is DROP. For the sake of demonstration, you can change this by adding DefaultEndpointToHostAction = RETURN to /etc/calico/felix.cfg, then use sudo systemctl restart calico-felix to restart Felix, and then you will observe that a VM on that host can ping its default gateway. However we do not recommend routinely operating with DefaultEndpointToHostAction = RETURN, because that potentially allows a malicious VM to compromise its host.

In summary, then, there are two points behind why a VM cannot normally ping its default gateway, with Calico.

  1. The default gateway concept just doesn't really fit, and isn't needed, given how Calico routes everything at the compute node - which is a fundamental aspect of Calico networking for OpenStack.

  2. Calico's iptables rules generally do not allow a VM to contact its host.

You can’t perform that action at this time.