Navigating the uncharted waters of SSL/TLS certificates and gRPC with Go


There are different ways to establishing a secure TLS connection with Go and gRPC. Contrary to popular belief, you don’t need to manually provide the Server certificate to your gRPC client in order to encrypt the connection. This post will provide a list of code examples for different scenarios. The source code is available in this repository.

TLS

As stated in RFC 5246, the primary goal of the Transport Layer Security (TLS) protocol is to provide privacy and data integrity between two communicating applications. TLS is one of the authentication mechanisms that are built-in to gRPC. It has TLS integration and promotes the use of TLS to authenticate the server, and to encrypt all the data exchanged between the client and the server [gRPC Authentication].

In order to establishing a TLS Connection, the client must send a Client Hello message to the Server to initiate the TLS Handshake. The TLS Handshake Protocol, allows the server and client to authenticate each other and to negotiate an encryption algorithm and cryptographic keys before the application protocol transmits or receives its first byte of data [RFC 5246].

A Client Hello message includes a list of options the Client supports to establish a secure connection; The TLS Version, a Random number, a Session ID, the Cipher Suites, Compression Methods and Extensions.

The Server replies back with a Server Hello including its preferred TLS Version, a Random number, a Session ID, and the Cipher Suite and Compression Method selected. The Server will also include a signed TLS Certificate. The client⁠ —depending on its configuration⁠— will validate this certificate with a Certificate Authority (CA) to prove the identity of the Server. A CA is a trusted party that issues digital certificates. The certificate could also come on a separate message.

After this negotiation, they start the Client Key exchange over an encrypted channel (Symmetric vs. Asymmetric encryption). Next, they start sending encrypted application data. I’m oversimplifying this part a bit, but I think we already have enough context to evaluate the code snippets to follow.

Certificates

Before we jump into code, let’s talk about certificates. The X.509 v3 certificate format is described in detail in RFC 5280. It encodes, among other things, the server’s public key and a digital signature (to validate the certificate’s authenticity).

1
2
3
4
Certificate ::= SEQUENCE { tbsCertificate TBSCertificate, signatureAlgorithm AlgorithmIdentifier, signatureValue BIT STRING }

Before you ask, TBS implies To-Be-Signed.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
TBSCertificate ::= SEQUENCE { version [0] EXPLICIT Version DEFAULT v1, serialNumber CertificateSerialNumber, signature AlgorithmIdentifier, issuer Name, validity Validity, subject Name, subjectPublicKeyInfo SubjectPublicKeyInfo, ... }

Some of the most relevant fields of a X.509 certificate are:

  • subject: Name of the subject the certificate is issued to.
  • subjectPublicKey: Public Key and algorithm with which the key is used (e.g., RSA, DSA, or Diffie-Hellman). See below.
  • issuer: Name of the CA that has signed and issued the certificate
  • signature: algorithm identifier for the algorithm used by the CA to sign the certificate (same as signatureAlgorithm).

    1
    2
    3
    
    SubjectPublicKeyInfo ::= SEQUENCE {
    algorithm AlgorithmIdentifier,
    subjectPublicKey BIT STRING }

You can see this as Go code in the x.509 library.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
type Certificate struct { ... Signature []byte SignatureAlgorithm SignatureAlgorithm PublicKeyAlgorithm PublicKeyAlgorithm PublicKey interface{} Version int SerialNumber *big.Int Issuer pkix.Name ...

While an SSL Certificate is most reliable when issued by a trusted Certificate Authority (CA), you can create self-signed certificates as decribed in Creating self-signed certificates. You can alternatively run make cert after cloning the repository, which is requiered to executed the following examples.

gRPC

Now, let’s take a look at how we apply and take advantage of all this with Go and gRPC with a very simple gRPC Service. This Service will retrieve usernames by their ID. In the examples, we will query for ID=1, which returns user Nicolas. The protobuf definition is the following.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
syntax = "proto3";

package test;

service gUMI {
 rpc GetByID (GetByIDRequest) returns (User);
}

message GetByIDRequest {
 uint32 id = 1;
}

message User {
 string name = 1;
 string email = 2;
 uint32 id = 3;
}

Insecure gRPC connections

Let’s check a couple of non-recommended practices.

Connection without encryption

If you do NOT want to encrypt the connection, the Go grpc package offers the DialOption WithInsecure() for the Client. This, plus a Server without any ServerOption will result in an unencrypted connection.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
// Client
conn, err := grpc.Dial(address, grpc.WithInsecure())
if err != nil { log.Fatalf("did not connect: %v", err)
}
defer conn.Close() // Server
s := grpc.NewServer()
// ... register gRPC services ...
if err = s.Serve(lis); err != nil { log.Fatalf("failed to serve: %v", err)
}

In order to reproduce this, run make run-server-insecure in one tab and run-client-insecure in another.

1
2
3
4
$ make run-server-insecure
2019/07/05 18:08:03 Creating listener on port: 50051
2019/07/05 18:08:03 Starting gRPC services
2019/07/05 18:08:03 Listening for incoming connections

Second tab.

1
2
$ make run-client-insecure
User found: Nicolas

Client does not authenticate the Server

In this case, we do encrypt the connection using the Server’s public key, however the client won’t validate the integrity of the Server’s certificate, so you can’t make sure you are actually talking to the Server and not to a man in the middle (man-in-the-middle attack).

To do this, we provide the public and private key pair on the server side we created previously. The client needs to set the config flag InsecureSkipVerify from the tls package to true.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
// Client
config := &tls.Config{ InsecureSkipVerify: true,
}
conn, err := grpc.Dial(address, grpc.WithTransportCredentials(credentials.NewTLS(config)))
if err != nil { log.Fatalf("did not connect: %v", err)
}
defer conn.Close() // Server
creds, err := credentials.NewServerTLSFromFile("service.pem", "service.key")
if err != nil { log.Fatalf("Failed to setup TLS: %v", err)
}
s := grpc.NewServer(grpc.Creds(creds))
// ... register gRPC services ...
if err = s.Serve(lis); err != nil { log.Fatalf("failed to serve: %v", err)
}

In order to reproduce this, run make run-server in one tab and run-client in another.

Secure gRPC connections

Let’s look at how we can encrypt the communication channel and validate we are talking to who we think we are.

In order to validate the identity of the Server (authenticate it), the client uses the certification authority (CA) certificate to authenticate the CA signature on the server certificate. You can provide the CA certificate to your client or rely on a set of trusted CA certificates included in your operating system (trusted key store).

Without a CA cert file

In the previous example we didn’t really do anything special on the client side to encrypt the connection, other than setting the InsecureSkipVerify flag to true. In this case we will switch the flag to false to see what happens. The connection won’t be established and the client will log x509: certificate signed by unknown authority.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
// Client
config := &tls.Config{ InsecureSkipVerify: false,
}
conn, err := grpc.Dial(address, grpc.WithTransportCredentials(credentials.NewTLS(config)))
if err != nil { log.Fatalf("did not connect: %v", err)
}
defer conn.Close() // Server
creds, err := credentials.NewServerTLSFromFile("service.pem", "service.key")
if err != nil { log.Fatalf("Failed to setup TLS: %v", err)
}
s := grpc.NewServer(grpc.Creds(creds))
// ... register gRPC services ...
if err = s.Serve(lis); err != nil { log.Fatalf("failed to serve: %v", err)
}

In order to reproduce this, run make run-server in one tab and run-client-noca in another.

With a Certification Authority (CA) cert file

Let’s manually provide the CA cert file (ca.cert) and keep the InsecureSkipVerify option as false.

 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
// Client
b, _ := ioutil.ReadFile("ca.cert")
cp := x509.NewCertPool()
if !cp.AppendCertsFromPEM(b) { return nil, errors.New("credentials: failed to append certificates")
}
config := &tls.Config{ InsecureSkipVerify: false, RootCAs: cp,
}
conn, err := grpc.Dial(address, grpc.WithTransportCredentials(credentials.NewTLS(config)))
if err != nil { log.Fatalf("did not connect: %v", err)
}
defer conn.Close() // Server
creds, err := credentials.NewServerTLSFromFile("service.pem", "service.key")
if err != nil { log.Fatalf("Failed to setup TLS: %v", err)
}
s := grpc.NewServer(grpc.Creds(creds))
// ... register gRPC services ...
if err = s.Serve(lis); err != nil { log.Fatalf("failed to serve: %v", err)
}

In order to reproduce this, run make run-server in one tab and run-client-ca in another.

With CA certificates included in the system (OS/Browser)

An empty tls config (tls.Config{}) will take care of loading your system CA certs. We will validate this scenario in with certificates from Let’s Encrypt for a public domain in a few paragraphs.

You can alternatively manually load the CA certs from the system with SystemCertPool().

1
certPool, err := x509.SystemCertPool()

If you have the Server cert and you trust it

This is most common scenario found on Internet tutorials. If you own the server and client, you could pre-share the server’s certificate (service.pem) with the client and use it directly to encrypt the channel.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
// Client
creds, err := credentials.NewClientTLSFromFile("service.pem", "")
if err != nil { log.Fatalf("could not process the credentials: %v", err)
}
conn, err := grpc.Dial(address, grpc.WithTransportCredentials(creds))
if err != nil { log.Fatalf("did not connect: %v", err)
}
defer conn.Close() // Server
creds, err := credentials.NewServerTLSFromFile("service.pem", "service.key")
if err != nil { log.Fatalf("Failed to setup TLS: %v", err)
}
s := grpc.NewServer(grpc.Creds(creds))
// ... register gRPC services ...
if err = s.Serve(lis); err != nil { log.Fatalf("failed to serve: %v", err)
}

In order to reproduce this, run make run-server in one tab and run-client-file in another.

In the previous examples, we examined different (SSL/TLS) certificate combinations to secure a gRPC channel. As the number of endpoints grows, this process soon gets too complicated to carry out manually. It’s time to look at how to automate the generation of signed certificates our gRPC endpoints can use without our intervention. We will need a Certificate Authority (CA) we can interact with from our Go gRPC endpoints. We will explore alternatives for private and public domains.

For private domains our CA of choice will be the Vault PKI Secrets Engine. In order to generate certificate signing requests (CSR) and renewals from our gRPC endpoints, we will use Certify.

For public certificate generation and distribution, we’ll go with Let’s Encrypt; *a free, automated, and open Certificate Authority*… how cool is that!?. The only thing they require from you is to demonstrate control over the domain with the Automatic Certificate Management Environment (ACME) protocol.

This means we need an ACME client, fortunately there is a list of Go libraries we can chose from for this. In this opportunity, we will use autocert for its ease of use and support for TLS-ALPN-01 challenge.

Private domains: Vault and Certify

Vault

Vault is a secrets management and data protection open source project, which can store and control access to certificates, among other secrets like passwords and tokens. It’s distributed as a binary you can place anywhere in your $PATH. If you want to learn more about Vault, its Getting Started guide is a good place to start. All the details of the setup used for this post are documented here.

First, we run Vault with vault server -config=vault_config.hcl. The config file (vault_config.hcl) provides the storage backend where Vault data is stored. For simplicity, we are just using a local file. You could alternatively choose to store it in-memory, on a cloud provider or else. See all the options in storage Stanza.

1
2
3
storage "file" { path = ".../data"
}

Additionally, we specify the address Vault will bind to. TLS is enabled by default, so we need to provide a certificate and private key pair. If you choose to self-sign these (see these instructions for an example), make sure you keep the Root certificate (ca.cert) handy, you will need it later on to make requests to Vault (*). Other TCP config options are documented in tcp Listener Parameters.

1
2
3
4
5
listener "tcp" { address = "localhost:8200" tls_cert_file = ".../vault.pem" tls_key_file = ".../vault.key"
}

After initializing Vault’s Server and unsealing Vault you can validate is working with an API call.

1
2
3
4
5
6
7
8
$ curl \
 --cacert ca.cert \
 -i https://localhost:8200/v1/sys/health HTTP/1.1 200 OK
... {"initialized":true,"sealed":false,"standby":false, ...}

The next step is to enable Vault PKI Secrets Engine backend with vault secrets enable pki, generate a CA certificate and private key Vault will use to sign certificates, and create a role (my-role) that can make requests for our domain (localhost). See all the details here.

1
2
3
4
vault write pki/roles/my-role \
 allowed_domains=localhost \
 allow_subdomains=true \
 max_ttl=72h

Certify

Now that our Certificate Authority (CA) is ready to go, we can make requests to it, to have our certificates signed. Which certificates you might ask, and how to automatically tell our gRPC endpoints to use them, if we don’t have them yet?. Enter Certify, a Go library to perform certificate distribution and renewal whenever it’s needed, automatically. It not only works with Vault as CA backend, but also with Cloudflare CFSSL and AWS ACM.

The first step to configure Certify is to specify the backend issuer, Vault in this case.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
issuer := &vault.Issuer{ URL: &url.URL{ Scheme: "https", Host: "localhost:8200", }, TLSConfig: &tls.Config{ RootCAs: cp, }, Token: getenv("TOKEN"), Role: "my-role",
}

In this example we identify our Vault instance and access credentials by providing:

  • The listener address we configured for Vault (localhost:8200).
  • The TOKEN we get after initializing Vault’s Server.
  • The role we created (my-role).
  • The CA certificate of the issuer of the certs we provided in Vault’s config. cp is a x509.CertPool that includes ca.cert in this case, as noted in (*).

You can, optionally, provide certificate details via CertConfig. We do it in this case to specify we want to generate private keys for our Certificate Signing Requests (CSR) using the RSA algorithm instead of Certify’s default ECDSA P256.

1
2
3
4
5
6
7
8
cfg := certify.CertConfig{ SubjectAlternativeNames: []string{"localhost"}, IPSubjectAlternativeNames: []net.IP{ net.ParseIP("127.0.0.1"), net.ParseIP("::1"), }, KeyGenerator: RSA{bits: 2048},
}

Certify hooks into the GetCertificate and GetClientCertificate methods of tls.Config via the Certify type, which we now build with; the previously collected information, a Cache method to prevent requesting a new certificate for every incoming connection, and a login plugin (go-kit/log in tis example).

1
2
3
4
5
6
7
8
c := &certify.Certify{ CommonName: "localhost", Issuer: issuer, Cache: certify.NewMemCache(), CertConfig: &cfg, RenewBefore: 24 * time.Hour, Logger: kit.New(logger),
}

The last step is to create a tls.Config pointing to the GetCertificate method of the Certify we just created. Then, use this config in our gRPC Server.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
// Client
// ... as in http://bit.ly/go-grpc-tls-ca ...

// Server
tlsConfig := &tls.Config{ GetCertificate: c.GetCertificate,
} s := grpc.NewServer(grpc.Creds(credentials.NewTLS(tlsConfig)))
// ... register gRPC services ...
if err = s.Serve(lis); err != nil { log.Fatalf("failed to serve: %v", err)
}

You can reproduce this by running make run-server-vault in one tab and make run-client-ca in another after pointing the environmental variable CAFILE to Vault’s certificate file (ca-vault.cert), which you can get as follows:

1
2
3
4
$ curl \
 --cacert ca.cert \
 https://localhost:8200/v1/pki/ca/pem \
 -o ca-vault.cert

Server:

1
2
3
4
5
6
$ make run-server-vault
...
level=debug time=2019-07-15T19:37:12.694833Z caller=logger.go:36 server_name=localhost remote_addr=[::1]:64103 msg="Getting server certificate"
level=debug time=2019-07-15T19:37:12.694936Z caller=logger.go:36 msg="Requesting new certificate from issuer"
level=debug time=2019-07-15T19:37:12.815081Z caller=logger.go:36 serial=451331845556263599050597627925015657462097174315 expiry=2019-07-18T19:37:12Z msg="New certificate issued"
level=debug time=2019-07-15T19:37:12.815115Z caller=logger.go:36 serial=451331845556263599050597627925015657462097174315 took=120.284897ms msg="Certificate found"

Client:

1
2
3
4
$ export CAFILE="ca-vault.cert"
$ make run-client-ca
...
User found: Nicolas

Inspecting the certificate we generated and had signed automatically, will reveal some of the specifics we just configured.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
$ openssl x509 -in grpc-cert.pem -text -noout
Certificate: Data: ... Validity Not Before: Jul 15 19:36:42 2019 GMT Not After : Jul 18 19:37:12 2019 GMT Subject: CN=localhost Subject Public Key Info: Public Key Algorithm: rsaEncryption Public-Key: (2048 bit) Modulus: 00:bf:3c:a3:d8:8c:d8:3c:d0:bd:0c:e0:4c:9d:4d: ... X509v3 extensions: ... Authority Information Access: CA Issuers - URI:https://localhost:8200/v1/pki/ca
X509v3 Subject Alternative Name: DNS:localhost, DNS:localhost, IP Address:127.0.0.1, IP Address:0:0:0:0:0:0:0:1

Public Domains: Let’s Encrypt and autocert

Let’s Encrypt

Can we use Let’s Encrypt for gRPC?. Well, it did work for me. The question might be whether having a public facing gRPC API’s is a good idea or not. Google Cloud seems to be doing it, see Google APIs. However, this is not a very common practice. Anyways, here is how I was able to expose a public gRPC API with certificates we automatically get from Let’s Encrypt.

Is important to emphasize this example is not meant to be replicated for internal/private services. In talking to Jacob Hoffman-Andrews from Let’s Encrypt, he mentioned: In general, I recommend that people don’t use Let’s Encrypt certificates for gRPC or other internal RPC services. In my opinion, it’s both easier and safer to generate a single-purpose internal CA using something like minica and generate both server and client certificates with it. That way you don’t have to open up your RPC servers to the outside internet, plus you limit the scope of trust to just what’s needed for your internal RPCs, plus you can have a much longer certificate lifetime, plus you can get revocation that works.

Let’s Encrypt uses the ACME protocol to verify that an applicant for a certificate legitimately represents the domain name(s) in the certificate. It also provides facilities for other certificate management functions, such as certificate revocation. ACME describes an extensible framework for automating the issuance and domain validation procedure, thereby allowing servers and infrastructure software to obtain certificates without user interaction. [RFC 8555.]

In a nutshell, all we need to do in order to leverage Let’s Encrypt is to run an ACME client. We will use autocert in this example.

autocert

The autocert package provides automatic access to certificates from Let’s Encrypt and any other ACME-based CA. However, keep in mind this package is a work in progress and makes no API stability promises. [Documentation]

In terms of code requirements, the first step to is to declare a Manager with a Prompt that indicate acceptance of the CA’s Terms of Service during account registration, a Cache method* to store and retrieve previously obtained certificates* (directory on the local filesystem in this case), a HostPolicy with the list of domains we can respond to, and optionally and Email address to notify about problems with issued certificates.

1
2
3
4
5
6
manager := autocert.Manager{ Prompt: autocert.AcceptTOS, Cache: autocert.DirCache("golang-autocert"), HostPolicy: autocert.HostWhitelist(host), Email: "test@example.com",
}

This Manager will create a TLS config for us automagically, taking care of the interaction with Let’s Encrypt. The client, on the other hand, just needs a pointer to an empty tls config (&tls.Config{}), which will, by default, load the system CA certificates and therefore trust our CA (Let’s Encrypt).

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
// Client
config := &tls.Config{} conn, err := grpc.Dial(address, grpc.WithTransportCredentials(credentials.NewTLS(config)))
if err != nil { log.Fatalf("did not connect: %v", err)
}
defer conn.Close() // Server
creds := credentials.NewTLS(manager.TLSConfig())
s := grpc.NewServer(grpc.Creds(creds))
// ... register gRPC services ...

// Listener...

If you are paying close attention, you might have noticed we didn’t include the listener section in this example. The reason is how the ACME TLS-based challenge TLS-ALPN-01 works. The TLS with Application Level Protocol Negotiation (TLS ALPN) validation method proves control over a domain name by requiring the client to configure a TLS server to respond to specific connection attempts utilizing the ALPN extension with identifying information. [draft-ietf-acme-tls-alpn-05].

As a side note, autocert added support for TLS-ALPN-01 after Let’s Encrypt announced End-of-Life for all TLS-SNI-01 validation support.

In other words, we need to listen to HTTPS request. The good news is autocert got you covered and can create this special Listener with manager.Listener(). Now, the question is whether HTTPS and gRPC should listen on the same port or not?. Long story short, I couldn’t make it work with independent ports, but if both services listen on 443, it works flawlessly.

gRPC and HTTPS on the same port… say what!?. I know, just because you can doesn’t mean you should. However, the Go gRPC library provides the ServeHTTP method that can help us route incoming requests to the corresponding service. *Note that ServeHTTP uses Go’s HTTP/2 server implementation which is totally separate from grpc-go’s HTTP/2 server. Performance and features may vary between the two paths*. [go-grpc]. You can check some benchmarks in gRPC serveHTTP performance penalty. Having said that, routing would then look like this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
func grpcHandlerFunc(g *grpc.Server, h http.Handler) http.Handler { return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) { ct := r.Header.Get("Content-Type") if r.ProtoMajor == 2 && strings.Contains(ct, "application/grpc") { g.ServeHTTP(w, r) } else { h.ServeHTTP(w, r) } })
}

So we can listen to requests as follows, notice we provide the handler grpcHandlerFunc we just created to http.Serve:

1
2
3
4
5
6
// Listener
lis = manager.Listener() if err = http.Serve(lis, grpcHandlerFunc(s, httpsHandler())); err != nil { log.Fatalf("failed to serve: %v", err))
}

You can reproduce this by running make run-server-public in one tab and make run-client-default in another. For this to work, you need to own a domain (HOST). In my case I used:

1
2
export HOST=grpc.nleiva.com
export PORT=443

Now, I can make gRPC requests from anywhere in the world over the Internet with:

1
2
3
4
$ export HOST=grpc.nleiva.com
$ export PORT=443
$ make run-client-default
User found: Nicolas

Finally, we can take a look at the certificate generated by making an HTTPS request on your browser to https://grpc.nleiva.com/.

Conclusion

There are different ways to go about setting TLS for gRPC. Providing integrity and privacy doesn’t take too much effort, so it’s strongly recommended you stay away of methods like WithInsecure() or setting InsecureSkipVerify flag to true.

Also, managing and distributing certificates for your gRPC endpoints shouldn’t be a hassle if you leverage some of the resources discussed in this post.

If you have any questions, feel free to contact me! I’m nleiva on GitHub and nleiv4 on Twitter.