# Sanity checking Istio's mTLS on permissive mode

One of the activities in Istio’s Mutual TLS Deep-Dive task is the verification of the mTLS connection using curl requests inside the Envoy sidecar (instead of having Envoy do it for us). Everything works just fine but things start to get strange when we try the same curl requests with the server set to permissive mode (accept both HTTP and mTLS connection) as shown below:

The output above was made possible by istioctl authn tls-check. See aside for more information.

CRDs applied are (see aside for more information):

Interestingly enough, when I forced the curl request to use mTLS I got:

however the plain HTTP ones work. I reverted back to strict mTLS mode just to make sure I didn’t bork my request and it worked. Does this mean permissive mode is equivalent to only accepting HTTP requests?

### Oh-so-magical Envoy

I played around with Istio a bit and see if I misconfigured something but to no avail so I hopped on the internet and started searching. I then came across this issue which kind of resembles my problem but instead of using ISTIO_MUTUAL as the mode for the destination rule, they were using MUTUAL and a custom CA to boot. I quickly skimmed the thread and they started mentioning about potential problems with how Istio builds Envoy’s configuration. So I took a peek inside the sidecar’s configuration and sure enough, there lies the explanation to the problem I’m having (spoiler alert: mTLS works in permissive mode).

The following JSON file shows the server’s sidecar listener configuration when set to use strict mTLS:

Now, this is what the configuration looks like when set to use permissive mode:

Summarizing the listeners gives us:

• On strict mode:

• Listener - mTLS support
• On permissive mode:

• Listener 1 - mTLS support, application protocol should be istio
• Listener 2 - HTTP support

So when we try our mTLS request, we get funneled to the plain HTTP listener (that does not accept mTLS) since our target listener is expecting istio as the application protocol (which we lack), causing the request to fail.

Our curl request works however in strict mTLS mode since Envoy only expects an mTLS connection but doesn’t bother looking for istio as the application protocol.

Our observations are further confirmed when we take a look at the cluster configuration of the client service:

Envoy sets the protocol to istio via the alpnProtocols configuration setting when the client is set to use mTLS.

### Actual proof of mTLS over permissive mode

Everything that we’ve gone through so far doesn’t prove that Istio indeed uses mTLS in permissive mode, it only explains why it fails. Luckily, I was playing around with integrating Open Policy Agent with Istio (tackled on a future entry) which gives our smoking gun.

Sending the request:

yielded an OPA input of:

This is enough proof that Istio uses mTLS on permissive mode as well. Case closed!

### [Aside] Checking mTLS mappings

Istio comes with a command-line utility tool to help users debug/diagnose their mesh called istioctl. One of the commands available to istioctl is authn tls-check which basically checks what kind of connection a particular service can receive (as a server) and send (as a client).

For example, running:

istio-1.3.2/bin/istioctl authn tls-check sleep-69c766786-662j7.mtls-pg

which can be interpreted as “check the mTLS setting for the pod named sleep-69c766786-662j7 in the mtls-pg namespace concerning all of the pods in the mesh” yields:

Focusing on the SERVER and CLIENT columns, we explain the output as follows:

• sleep-69c766786-662j7.mtls-pg sees the service sleep.default as a server that can accept either plaintext (HTTP) or mTLS connections (permissive mode). We can tweak this using Policies which will show up in the AUTHN POLICY column.

• When sleep-69c766786-662j7.mtls-pg sends a request to sleep.default, the Envoy sidecar attached to it will send the request over a plaintext connection (HTTP). We can tweak this as well using DestinationRules which will show up in the DESTINATION RULE column.

### [Aside] Policies and Destination Rules

Let’s run through how we can change the mTLS mappings between services using Policies and Destination Rules.

Running:

istio-1.3.2/bin/istioctl authn tls-check sleep-69c766786-662j7.mtls-pg httpbin.mtls-pg.svc.cluster.local

(we show the mTLS setting with respect to the httpbin.mtls-pg.svc.cluster.local service only) yields:

This tells us that httpbin.mtls-pg is running in permissive mode. We can change it to strict mode by applying the following CRD:

Which gives us:

We’ve now successfully configured httpbin.mtls-pg to only accept mTLS connections. The problem here now is that all the pods in the mesh acting as a client to httpbin.mtls-pg will only send their request over HTTP which httpbin.mtls-pg won’t accept, hence the CONFLICT status. To fix this, we apply the following CRD:

Which in turn yields:

All pods acting as clients to httpbin.mtls-pg will now only send their traffic via mTLS. It is important to note though that the previously applied CRD will only affect client requests to httpbin.mtls-pg but not to other services, as shown below: