The server is in fact rejecting the certificate here, but that error is being thrown out after being printed by your test. You can see the error logged in Travis.

https://github.com/fullstorydev/grpcurl/blob/a90ea9e114ebe5e570fea4fcf44f9351b58b98a5/tls13_bug_test.go#L111-L113

Bad handshake: tls: failed to verify client's certificate: x509: certificate has expired or is not yet valid
Bad handshake: tls: client didn't provide a certificate
--- FAIL: TestTLS13RejectsBadCerts (0.07s)
    tls13_bug_test.go:37: Expecting failure due to use of expired cert!
    tls13_bug_test.go:45: Expecting failure due to use of untrusted cert!
    tls13_bug_test.go:53: Expecting failure due to missing cert!
Bad handshake: tls: client didn't provide a certificate

Your test is checking the error on the client side instead of the server side. In TLS 1.2, there was a whole round-trip of handshake after sending the client cert, so the remote error was making it to the client in the form of an alert before Handshake had to return. In TLS 1.3, the client certificate is sent in the second and last client flight, meaning the client sends the client cert and then returns before hearing back from the server, so it can't return an error in Handshake.

The error will show up on the client first read. This is unavoidable, because the protocol is designed that way, but I can see it causing trouble. Hopefully relatively few users call Handshake directly. We should definitely mention it in the release notes.

@jhump BTW, thank you for that reproduction PR, it was extremely handy.

@FiloSottile, will the server simply close the connection after it sees the cert is bad? If so, does that mean there is no way to distinguish, in the client, between the server rejecting the client's credentials and the server simply being partitioned from the client between the handshake and first read?

I'm trying to figure out how to work around this: for command-line tools (such as grpcurl), it's nice to provide meaningful/useful error messages related to client identity in mutually-authenticated TLS. But it sounds like that may not be feasible with TLS 1.3. I understand the desire to reduce round trips, particularly for TLS over mobile networks, but this is an unfortunate side effect for debugging issues on the client.

One example: it means that using grpc.WithBlock, to dial a gRPC server, won't be useful if there are unrecoverable errors due to TLS certs. The gRPC client will block until the socket is established, the handshake complete, and think the channel is ready for use (just as in my code example). But then every RPC on it will fail due to the transport channel getting closed. IMO, that kind of defeats the purpose of grpc.WithBlock. With gRPC, I guess we can work-around it by sending an HTTP/2 PING frame immediately if we want to confirm the channel is good and the server accepted the cert. But that won't work, for example, with HTTP 1.1 communications.

Anyhow. Sorry for the long-winded comment. I guess, summing up, I'm wondering if you have recommendations for providing better client-side error messages with TLS 1.3 in these situations.

I do agree that it’s unfortunate, but I don’t have useful advice beyond what you already suggested.

Note though that the server won’t just close the connection but send an alert, so the first client read will fail with the same meaningful error as the Handshake used to.

so the first client read will fail with the same meaningful error as the Handshake used to.

Ok, that's useful actually. Thanks for the info.