Using trace tcpconnect
The trace tcpconnect gadget traces TCP connect calls.
On Kubernetes
In this guide, we will use this gadget to define a restrictive policy for outgoing connections.
Before we start a demo pod that connects to a public HTTP server, we already begin to trace the outgoing connections of our future pod (don't terminate it with Ctrl-C for now).
$ kubectl gadget trace tcpconnect --podname mypod
When we run the pod in a new terminal, we see the output ok since the public HTTP server was reached.
$ kubectl run --restart=Never -ti --image=busybox mypod -- sh -c 'wget -q -O /dev/null -T 3 http://1.1.1.1 && echo ok || echo failed'
ok
In our trace tcpconnect gadget terminal we can now see the logged connection:
$ kubectl gadget trace tcpconnect --podname mypod
K8S.NODE K8S.NAMESPACE K8S.PODNAME K8S.CONTAINERNAME PID COMM IP SRC DST
minikube-docker default mypod mypod 2011630 wget 4 p/default/mypod:46779 r/1.1.1.1:80
minikube-docker default mypod mypod 2011630 wget 4 p/default/mypod:21731 r/1.1.1.1:443
If the pod was started as part of a deployment, the name of the pod is not known
in advance since random characters will be added as suffix.
In that case, it is still possible to trace the connections. We would just
use kubectl gadget trace tcpconnect --selector key=value to filter the pods by
labels instead of names.
There was a HTTP redirect to HTTPS, so we need to allow both ports for our pod. Don't terminate it yet, we will have another look later.
Since we now know which network accesses our pod does, we can define and apply a very restrictive network policy:
# https://raw.githubusercontent.com/inspektor-gadget/inspektor-gadget/main/docs/examples/network-policy.yaml
apiVersion: networking.k8s.io/v1
kind: NetworkPolicy
metadata:
name: restrictive-network-policy
namespace: default
spec:
podSelector:
matchLabels:
run: mypod
policyTypes:
- Ingress
- Egress
ingress:
- from:
- ipBlock:
cidr: 1.1.1.1/32
egress:
- to:
- ipBlock:
cidr: 1.1.1.1/32
ports:
- protocol: TCP
port: 80
- protocol: TCP
port: 443
$ kubectl apply -f https://raw.githubusercontent.com/inspektor-gadget/inspektor-gadget/main/docs/examples/network-policy.yaml
networkpolicy.networking.k8s.io/restrictive-network-policy created
Let's test if the pod still works as expected:
$ kubectl delete pod mypod
$ kubectl run --restart=Never -ti --image=busybox mypod -- sh -c 'wget -q -O /dev/null -T 3 http://1.1.1.1 && echo ok || echo failed'
ok
Switching to the gadget trace tcpconnnect terminal, we see the same connections again (but now with a new PID since it's a new pod):
$ kubectl gadget trace tcpconnect --podname mypod # (still running in old terminal)
K8S.NODE K8S.NAMESPACE K8S.PODNAME K8S.CONTAINERNAME PID COMM IP SRC DST
minikube-docker default mypod mypod 2011630 wget 4 p/default/mypod:46779 r/1.1.1.1:80 # (previous output)
minikube-docker default mypod mypod 2011630 wget 4 p/default/mypod:21731 r/1.1.1.1:443 # (previous output)
minikube-docker default mypod mypod 2011630 wget 4 p/default/mypod:40676 r/1.1.1.1:80
minikube-docker default mypod mypod 2011630 wget 4 p/default/mypod:40630 r/1.1.1.1:443
But what if the pod would connect to other IP addresses which we disallowed? Let's modify our pod to connect to a different address to verify that the connection fails.
$ kubectl delete pod mypod
$ kubectl run --restart=Never -ti --image=busybox mypod -- sh -c 'wget -q -O /dev/null -T 3 http://1.0.0.1 && echo ok || echo failed'
wget: download timed out
failed
Indeed the network policy was applied and we can also see in the gadget output which connection the pod wanted to make in the last line. Since connecting to port 80 failed there is no redirect visible to port 443:
$ kubectl gadget trace tcpconnect --podname mypod # (still running in old terminal)
K8S.NODE K8S.NAMESPACE K8S.PODNAME K8S.CONTAINERNAME PID COMM IP SRC DST
minikube-docker default mypod mypod 2011630 wget 4 p/default/mypod:46779 r/1.1.1.1:80 # (previous output)
minikube-docker default mypod mypod 2011630 wget 4 p/default/mypod:21731 r/1.1.1.1:443 # (previous output)
minikube-docker default mypod mypod 2011630 wget 4 p/default/mypod:40676 r/1.1.1.1:80 # (previous output)
minikube-docker default mypod mypod 2011630 wget 4 p/default/mypod:40630 r/1.1.1.1:443 # (previous output)
minikube-docker default mypod mypod 2011630 wget 4 p/default/mypod:17418 r/1.0.0.1:80
We created a tailored network policy for our (original) demo pod by observing its connection behavior :) Finally, we should delete the demo pod and network policy again:
$ kubectl delete pod mypod
pod "mypod" deleted
$ kubectl delete -f https://raw.githubusercontent.com/inspektor-gadget/inspektor-gadget/main/docs/examples/network-policy.yaml
networkpolicy.networking.k8s.io "restrictive-network-policy" deleted
With ig
Start the gadget on a terminal.
$ sudo ig trace tcpconnect -c test-tcp-connect
Then, create a gadget that performs a TCP connection.
$ docker run -it --rm --name test-tcp-connect busybox /bin/sh -c "wget 1.1.1.1"
Connecting to 1.1.1.1 (1.1.1.1:80)
Connecting to 1.1.1.1 (1.1.1.1:443)
wget: note: TLS certificate validation not implemented
saving to 'index.html'
index.html 100% |********************************| 54361 0:00:00 ETA
'index.html' saved
The gadget will show the connection and related information to it.
$ sudo ig trace tcpconnect -c test-tcp-connect
RUNTIME.CONTAINERNAME PID COMM IP SRC DST
test-tcp-connect 2021739 wget 4 172.17.0.2:4784 1.1.1.1:80
test-tcp-connect 2021739 wget 4 172.17.0.2:14023 1.1.1.1:443
Calculating the latency of a connection
This tools provides a --latency option to show the latency (time) for the connection as measured
from the TCP client perspective: the time from SYN sent to the response packet.
TCP connection latency is a useful performance measure showing the time taken to establish a connection. This typically involves kernel TCP/IP processing and the network round trip time, and not application runtime.
This measures the time from any connection to the response packet, even if the response is a RST (port closed).
When this option is used, the event is only shown when the server replies or the socket is removed.
On Kubernetes
Start the gadget on a terminal:
$ kubectl gadget trace tcpconnect --latency
RUNTIME.CONTAINERNAME PID COMM IP SRC DST LATENCY```
In another terminal, create a nginx service and a pod to send some http requests:
```bash
$ kubectl create service nodeport nginx --tcp=80:80
$ kubectl create deployment nginx --image=nginx
$ kubectl run -ti --privileged --image wbitt/network-multitool myclientpod -- bash
Send some requests to the server:
# curl nginx
# curl nginx
...
The first terminal show all those connections and their latency. In my case both pods are running on the same node, so it's very low:
K8S.NODE K8S.NAMESPACE K8S.PODNAME K8S.CONTAINERNAME PID COMM IP SRC DST LATENCY
minikube-docker default myclientpod myclientpod 2054329 curl 4 p/default/myclientpod:50306 s/default/nginx:80 47.069µs
minikube-docker default myclientpod myclientpod 2054338 curl 4 p/default/myclientpod:53378 s/default/nginx:80 120.017µs
Now, let's use the network emulator to introduce some random delay to the packets on the container interface and send some more requests to the server again:
# tc qdisc add dev eth0 root netem delay 50ms 50ms 25%
# curl nginx
# curl nginx
Now the latency is a lot higher and has some variance because of the emulation configuration:
K8S.NODE K8S.NAMESPACE K8S.PODNAME K8S.CONTAINERNAME PID COMM IP SRC DST LATENCY
...
minikube-docker default myclientpod myclientpod 2056697 curl 4 p/default/myclientpod:32415 s/default/nginx:80 7.820966ms
minikube-docker default myclientpod myclientpod 2056832 curl 4 p/default/myclientpod:32927 s/default/nginx:80 64.388825ms
minikube-docker default myclientpod myclientpod 2056905 curl 4 p/default/myclientpod:46811 s/default/nginx:80 39.244112ms
With ig
Start the trace tcpconnect gadget on a first terminal:
$ sudo ig trace tcpconnect --latency
RUNTIME.CONTAINERNAME PID COMM IP SRC DST LATENCY
Then, start a container and download a web page:
$ docker run -ti --rm --cap-add NET_ADMIN --name=netem wbitt/network-multitool -- /bin/bash
# wget 1.1.1.1
The first terminal will show the connections created and their latency:
RUNTIME.CONTAINERNAME PID COMM IP SRC DST LATENCY
netem 2036550 wget 4 172.17.0.2:47250 1.1.1.1:80 14.149828ms
netem 2036550 wget 4 172.17.0.2:44734 1.1.1.1:443 15.025666ms
In this case, it can be seen that two connections were made by the curl command and their latency
was around 50ms for each of them.
Now, let's configure the network emulator to add a delay of one second to all packets in the container's interface and then download the web page again:
# tc qdisc add dev eth0 root netem delay 1000ms 10ms 100%
# wget 1.1.1.1
In this case the wget command takes way longer to complete the request. We can check that the
latency for those connections is more than one second as expected:
RUNTIME.CONTAINERNAME PID COMM IP SRC DST LATENCY
...
netem 2037935 wget 4 172.17.0.2:38587 1.1.1.1:80 1.006814808s
...
netem 2037935 wget 4 172.17.0.2:10469 1.1.1.1:443 1.010320064s