iptables: How Kubernetes Services Direct Traffic to Pods

This is the third part of a series on Docker and Kubernetes networking. We’ll be tackling how Kubernetes’s kube-proxy component uses iptables to direct service traffic to pods randomly. We’ll focus on the ClusterIP type of Kubernetes services.

The goal of this post is to implement the iptables rules needed for a service like:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
apiVersion: v1
kind: Service
metadata:
  name: app-service
spec:
  clusterIP: 10.100.100.100
  selector:
    component: app
  ports:
    - protocol: TCP
      port: 8080
      targetPort: 8080

The previous posts so far are:

  1. How do Kubernetes and Docker create IP Addresses?!
  2. iptables: How Docker Publishes Ports

Like the first two articles, we won’t use Docker or Kubernetes in this post. Instead, we’ll learn the underlying tools used.

Recall that Kubernetes creates a network namespace for each pod. We’ll be manually creating network namespaces with python HTTP servers running, which will be treated as our “pods.”

Note: This post only works on Linux. I’m using Ubuntu 19.10, but this should work on other Linux distributions.

create virtual devices and run HTTP servers in network namespaces

We’re going to quickly set up an environment as we did in the previous post.

If a refresher is needed on any of the following, please take a look at How do Kubernetes and Docker create IP Addresses?! and How Docker Publishes Ports.

Let’s get started. Enable IP forwarding by running:

1
sudo sysctl --write net.ipv4.ip_forward=1

Now we need to

  • create a virtual bridge (named bridge_home)
  • create two network namespaces (named netns_dustin and netns_leah)
  • configure 8.8.8.8 for DNS in the network namespaces
  • create two veth pairs connected to bridge_home
  • assign 10.0.0.11 to the veth running in netns_dustin
  • assign 10.0.0.21 to the veth running in netns_leah
  • setup default routing in our network namespaces
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
sudo ip link add dev bridge_home type bridge
sudo ip address add 10.0.0.1/24 dev bridge_home

sudo ip netns add netns_dustin
sudo mkdir -p /etc/netns/netns_dustin
echo "nameserver 8.8.8.8" | sudo tee -a /etc/netns/netns_dustin/resolv.conf
sudo ip netns exec netns_dustin ip link set dev lo up
sudo ip link add dev veth_dustin type veth peer name veth_ns_dustin
sudo ip link set dev veth_dustin master bridge_home
sudo ip link set dev veth_dustin up
sudo ip link set dev veth_ns_dustin netns netns_dustin
sudo ip netns exec netns_dustin ip link set dev veth_ns_dustin up
sudo ip netns exec netns_dustin ip address add 10.0.0.11/24 dev veth_ns_dustin

sudo ip netns add netns_leah
sudo mkdir -p /etc/netns/netns_leah
echo "nameserver 8.8.8.8" | sudo tee -a /etc/netns/netns_leah/resolv.conf
sudo ip netns exec netns_leah ip link set dev lo up
sudo ip link add dev veth_leah type veth peer name veth_ns_leah
sudo ip link set dev veth_leah master bridge_home
sudo ip link set dev veth_leah up
sudo ip link set dev veth_ns_leah netns netns_leah
sudo ip netns exec netns_leah ip link set dev veth_ns_leah up
sudo ip netns exec netns_leah ip address add 10.0.0.21/24 dev veth_ns_leah

sudo ip link set bridge_home up
sudo ip netns exec netns_dustin ip route add default via 10.0.0.1
sudo ip netns exec netns_leah ip route add default via 10.0.0.1

Next, create iptables rules to allow traffic in and out of the bridge_home device:

1
2
sudo iptables --table filter --append FORWARD --in-interface bridge_home --jump ACCEPT
sudo iptables --table filter --append FORWARD --out-interface bridge_home --jump ACCEPT

Then, create another iptables rule to masquerade requests from our network namespaces:

1
sudo iptables --table nat --append POSTROUTING --source 10.0.0.0/24 --jump MASQUERADE

Moving on, start an HTTP server in the netns_dustin network namespace:

1
sudo ip netns exec netns_dustin python3 -m http.server 8080

Finally, open another terminal and start an HTTP server in the netns_leah network namespace:

1
sudo ip netns exec netns_leah python3 -m http.server 8080

At this point, our environment will look like:

diagram showing virtual ethernet devices, physical ethernet device, and network namespaces

Note: Your IP address may differ from the 192.168.0.100 and the interface may have a different name than enp4s0.

For a sanity check, the following commands should work:

1
2
3
4
curl 10.0.0.11:8080
curl 10.0.0.21:8080
sudo ip netns exec netns_dustin curl 10.0.0.21:8080
sudo ip netns exec netns_leah curl 10.0.0.11:8080

add a virtual IP in iptables

When a Kubernetes Service is created a ClusterIP is assigned for that new service. Conceptually, a ClusterIP is a virtual IP. kube-proxy in iptables-mode is responsible for creating iptables rules to handle these virtual IP addresses as described in Virtual IPs and service proxies.

Let’s make a simple iptables rule to see what it takes to handle a virtual IP address. Later we’ll refactor to align our rules with how kube-proxy creates rules.

Note: I’m going to assume some familiarity with iptables. Check out How Docker Publishes Ports if you’re not comfortable with the following sections.

Create a new chain named DUSTIN-SERVICES in the nat table by running:

1
sudo iptables --table nat --new DUSTIN-SERVICES

Next, we’ll want the PREROUTING and OUTPUT chains to look through the DUSTIN-SERVICES chain via:

1
2
3
4
5
6
7
8
9
sudo iptables \
  --table nat \
  --append PREROUTING \
  --jump DUSTIN-SERVICES

sudo iptables \
  --table nat \
  --append OUTPUT \
  --jump DUSTIN-SERVICES

At this point, we can then create a rule in the DUSTIN-SERVICES chain to handle a virtual IP. Our virtual IP will be 10.100.100.100. Let’s create a rule that directs traffic for 10.100.100.100:8080 to 10.0.0.11:8080. Recall, that 10.0.0.11:8080 is the python HTTP server running in the netns_dustin namespace.

1
2
3
4
5
6
7
8
9
sudo iptables \
  --table nat \
  --append DUSTIN-SERVICES \
  --destination 10.100.100.100 \
  --protocol tcp \
  --match tcp \
  --dport 8080 \
  --jump DNAT \
  --to-destination 10.0.0.11:8080

This looks very familiar to a rule we created in How Docker Publishes Ports! This time we’re specifying a destination of 10.100.100.100 instead of a local address type.

Let’s request our virtual IP by executing:

1
curl 10.100.100.100:8080

Nice! We’ve just handled traffic for a virtual IP!

Now for some bad news. Let’s try requesting the virtual IP address from netns_dustin.

1
sudo ip netns exec netns_dustin curl 10.100.100.100.8080

This command may succeed for some and will fail for others. What gives?!

enable hairpin mode (and promiscuous mode)

If the last command failed for you, I’m going to bet you have Docker running. That was the case for me at least. So why is Docker interfering? Well, it technically isn’t, but Docker enables a little setting called net.bridge.bridge-nf-call-iptables. This configures bridges to consider iptables when handling traffic. This also causes issues with a request leaving a device that is destined for the same device, which is exactly the scenario we hit in the last command!

To be super clear, we have a request leaving veth_dustin which has a source IP address of 10.0.0.11. The request is destined for 10.100.100.100. Our iptables rule then performs a DNAT on 10.100.100.100 to 10.0.0.11. This is where the problem happens. The request needs to be directed to where the request came from!

Let’s get everyone’s environment configured the same way. This means that if the last command worked for you, we’re going to break it here pretty soon. Fun stuff.

First, check if net.bridge.bridge-nf-call-iptable is enabled.

1
sysctl net.bridge.bridge-nf-call-iptables

If you get the following error:

sysctl: cannot stat /proc/sys/net/bridge/bridge-nf-call-iptables: No such file or directory

then run the following command:

1
sudo modprobe br_netfilter

This will load the br_netfilter module. After run sysctl net.bridge.bridge-nf-call-iptables again.

I think everyone should be seeing net.bridge.bridge-nf-call-iptables is enabled (1 output). If for some reason it’s disabled (0) then run the following:

1
sudo sysctl --write net.bridge.bridge-nf-call-iptables=1

Now everyone should see the following command fail:

1
sudo ip netns exec netns_dustin curl 10.100.100.100.8080

Now for the fix! We need to enable hairpin mode on veth_dustin connected to bridge_home. Hairpin mode enables a request leaving a device to be received by the same device.

Fun fact: veth_dustin is called a port on bridge_home. Similar to having a physical ethernet cable plugged into a port on a physical bridge and the other end is plugged into a physical computer.

To enable hairpin mode on veth_dustin, run:

1
sudo brctl hairpin bridge_home veth_dustin on

Try the following command again:

1
sudo ip netns exec netns_dustin curl 10.100.100.100.8080

It’s a success!

Since we’ll want our network namespaces to be able to talk to themselves via our virtual IPs we’ll need hairpin mode enabled on each port of the bridge device. Fortunately, there’s a way to configure this on the bridge device instead of each port.

Start by undoing what we did earlier and disable hairpin mode:

1
sudo brctl hairpin bridge_home veth_dustin off

Note: This previous step isn’t technically required, but it’ll help to demonstrate the next step works.

Bridges can be in promiscuous mode, which will treat all attached ports (veths in our case) as if they all had hairpin mode enabled. We can enable promiscuous mode on bridge_home by running:

1
sudo ip link set bridge_home promisc on

I don’t know why promiscuous is shortened to promisc. I do know I’ve spelled promiscuous wrong so many times while researching. Maybe that’s why?

Run the following beloved command again:

1
sudo ip netns exec netns_dustin curl 10.100.100.100.8080

Success again! With promiscuous mode enabled on bridge_home, we won’t have to worry about enabling hairpin mode on each veth, such as veth_leah, in the future!

align iptables rules with kube-proxy

So far we’ve created a single iptables rule to handle one service (10.100.100.100) with one backend (10.0.0.11). We created this rule in a chain named DUSTIN-SERVICES, which is named similarly to kube-proxy’s KUBERNETES-SERVICES. kube-proxy creates a chain per service and has KUBERNETES-SERVICES jump to the respective service chain based on the destination.

Let’s start by creating a new chain for our service. We’re going to name our service HTTP. kube-proxy uses hashes in its chain names, but we’ll stick with HTTP to help with understanding. Create a new chain by running:

1
2
3
sudo iptables \
  --table nat \
  --new DUSTIN-SVC-HTTP

Let’s add a rule to our DUSTIN-SVC-HTTP chain that will direct traffic to our backend (10.0.0.11).

1
2
3
4
5
6
7
sudo iptables \
  --table nat \
  --append DUSTIN-SVC-HTTP \
  --protocol tcp \
  --match tcp \
  --jump DNAT \
  --to-destination 10.0.0.11:8080

Finally, we’ll want DUSTIN-SERVICES to use the DUSTIN-SVC-HTTP chain. Delete the previous rule we created in DUSTIN-SERVICES via:

1
2
3
4
5
6
7
8
9
sudo iptables \
  --table nat \
  --delete DUSTIN-SERVICES \
  --destination 10.100.100.100 \
  --protocol tcp \
  --match tcp \
  --dport 8080 \
  --jump DNAT \
  --to-destination 10.0.0.11:8080

and add a rule in DUSTIN-SERVICES to jump to DUSTIN-SVC-HTTP on matching destination via:

1
2
3
4
5
6
7
8
sudo iptables \
  --table nat \
  --append DUSTIN-SERVICES \
  --destination 10.100.100.100 \
  --protocol tcp \
  --match tcp \
  --dport 8080 \
  --jump DUSTIN-SVC-HTTP

At this point, the following commands will remain successful:

1
2
curl 10.100.100.100:8080
sudo ip netns exec netns_dustin curl 10.100.100.100:8080

In the future, adding a new service consists of:

  • create a new chain for the service, such DUSTIN-SVC-HTTP
  • create a rule in the service chain to direct traffic to a backend, such as 10.0.0.11
  • add a rule to DUSTIN-SERVICES to jump to the service chain, such as DUSTIN-SVC-HTTP

refactor service chain to support multiple backends

We just refactored our DUSTIN-SERVICES chain to jump to individual service chains. Now, we want to refactor our service chain (DUSTIN-SVC-HTTP) to jump to other chains for directing traffic to backends.

Note: I’ve been using the word backend here, but these are also referred to as endpoints in Kubernetes. Typically, the endpoints are IP addresses of pods.

Let’s create a new chain for our 10.0.0.11 endpoint. kube-proxy also uses a hash for these chain names, but we’ll create a chain named DUSTIN-SEP-HTTP1 representing the first service endpoint (SEP) for HTTP. Create the new chain via:

1
2
3
sudo iptables \
  --table nat \
  --new DUSTIN-SEP-HTTP1

And we’ll add a familiar-looking rule to the new DUSTIN-SEP-HTTP1 chain:

1
2
3
4
5
6
7
sudo iptables \
  --table nat \
  --append DUSTIN-SEP-HTTP1 \
  --protocol tcp \
  --match tcp \
  --jump DNAT \
  --to-destination 10.0.0.11:8080

We’ll then delete the rule we added to DUSTIN-SVC-HTTP and add a rule in DUSTIN-SVC-HTTP to jump to DUSTIN-SEP-HTTP1.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
sudo iptables \
  --table nat \
  --delete DUSTIN-SVC-HTTP \
  --protocol tcp \
  --match tcp \
  --jump DNAT \
  --to-destination 10.0.0.11:8080

sudo iptables \
  --table nat \
  --append DUSTIN-SVC-HTTP \
  --jump DUSTIN-SEP-HTTP1

The following commands should still work:

1
2
curl 10.100.100.100:8080
sudo ip netns exec netns_dustin curl 10.100.100.100:8080

Now we’re ready to start adding additional backends.

use iptables to serve random backends for virtual IPs

As mentioned in the Kubernetes documentation, kube-proxy directs traffic to backends randomly. How does it do that? iptables of course!

iptables support directing traffic to a backend based on probability. This is a super cool concept to me because I previously thought iptables was very deterministic!

Let’s start by adding a new chain and rule for our second HTTP backend (10.0.0.21) running in the netns_leah network namespace.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
sudo iptables \
  --table nat \
  --new DUSTIN-SEP-HTTP2

sudo iptables \
  --table nat \
  --append DUSTIN-SEP-HTTP2 \
  --protocol tcp \
  --match tcp \
  --jump DNAT \
  --to-destination 10.0.0.21:8080

We’ll then need to add another rule to the DUSTIN-SVC-HTTP chain to randomly jump to the DUSTIN-SEP-HTTP2 chain we just created. We can add this rule by running:

1
2
3
4
5
6
7
sudo iptables \
  --table nat \
  --insert DUSTIN-SVC-HTTP 1 \
  --match statistic \
  --mode random \
  --probability 0.5 \
  --jump DUSTIN-SEP-HTTP2

It’s very important to notice that we are inserting this rule to be first in the DUSTIN-SVC-HTTP chain. iptables goes down the list of rules in order. So by having this rule first, we’ll have a 50% chance of jumping to this chain. If it’s a hit, iptables will jump to DUSTIN-SEP-HTTP2. If it’s a miss, then iptables will go to the next rule, which will always jump to DUSTIN-SEP-HTTP1.

A common misconception is that each rule should have a probability of 50%, but this will cause problems in the following scenario:

  1. iptables looks at the first rule (the jump to DUSTIN-SEP-HTTP2) and let’s say it’s a miss on the 50%
  2. iptables looks at the next rule (the jump to DUSTIN-SEP-HTTP1) and let’s say it’s also a miss on the 50%

Now our virtual IP wouldn’t direct to any backend! So the probability is based on the number of remaining backends to choose from. If we were to insert a third backend, that rule would have a probability of 33%.

Anyways, if we then run the following command:

1
curl 10.100.100.100:8080

We’ll see requests being made randomly to our python HTTP servers running in netns_leah and netns_dustin network namespaces. This is load balancing via iptables!

closing thoughts

After these three posts on container and pod networking, I’ve learned more about networking than I ever thought I would. The remaining topics I’d like to learn are:

  • how does kube-proxy work in IPVS mode?
  • conntrack and how it’s used in iptables rules by kube-proxy
  • how are virtual tunnels and BGP optionally used in multi-node Kubernetes clusters?

Have any knowledge to share about the above topics? Or any other additional questions? Then please feel free to reach out and let me know on Twitter, LinkedIn, or GitHub.

updatedupdated2020-08-132020-08-13