crypto/tls: TLS session resumption re-verifies the client's certificate chain

[marten-seemann]

What version of Go are you using (go version)?

$ go version
1.12.x

Does this issue reproduce with the latest release?

Yes

What did you do?

When using TLS session resumption with client certificates, the server reverifies the client's certificate chain.

What did you expect to see?

When using TLS session resumption the server sends an encrypted (and MACed) ticket to the client, which the client presents during the next handshake. According to the spec, the session ticket is an opaque blob. The Go implementation encodes the client's certificate chain into the session ticket. If the server is able to authenticate and decrypt the session ticket presented by the client, this should serve as proof that the server communicated with this client before, and allow it to skip the (potentially computationally expensive) verification of the certificate chain.

Re-verifying leads to unexpected consequences if the tls.Config.VerifyPeerCertificate callback was changed such that it rejects the certificate in between the original and the resumed connection . Since the server re-verifies the certificate chain, it will reject the connection. The client doesn't perform any verification (other than checking that the certificate didn't expire) and is happy to resume a connection even if its VerifyPeerCertificate callback would have rejected the server's certificate chain on a new connection.

[FiloSottile]

I noticed this while implementing TLS 1.3, and it's not an easy decision. The discrepancy between client and server is bad. On the one hand, if the verification policy has changed, the application probably expects it to apply to all connections, so we should re-verify. On the other hand, verification is expensive and session tickets are a performance measure.

Usually, it's servers that have the most performance concern, though, so I am tempted to resolve the discrepancy towards safety, by making clients reverify as well. Opinions?

(Related to #25352)

[marten-seemann]

Since we're encoding the whole certificate chain in the session ticket, there's no point whatsoever to use session tickets with TLS 1.3 at all: The amount of data transferred during the handshake is (modulo some framing overhead) exactly the same as during a normal handshake, the computational cost of re-verifying the certificate chain is the same, and we're not even saving any roundtrips.

Couldn't one argue that the security properties of a resumed session should be the same as those of the original session? If the session was kept alive (instead of being resumed), the new policy wouldn't apply to that session either, right?

[fiorix]

Landing here to +1 what @marten-seemann said. We've hit that exact issue while debugging slowness in Go servers using TLS 1.3 and ticket resumption. There's at least 140 KB of heap allocation on every handshake, largely due to parsing the certificate chain from the ticket.

We started trying to patch crypto/tls and make ticket generation and validation configurable, by adding the following to tls.Config:

GenerateClientTicket func(Certificate) (ticket []byte, err error)
ValidateClientTicket func(ticket []byte) error

We want to preserve the original behaviour, and only use custom tickets when these functions are provided by the user.

This needs a lot of testing still, and we're far from having anything concrete but I'd like to ask @FiloSottile if this is in the team's radar, and if what I describe above is potentially an acceptable solution to the problem.

[FiloSottile]

Found what's almost certainly a bug related to this while reviewing CL 229122: if the mode is RequestClientCerts or NoClientCerts and the client doesn't provide certificates, VerifyPeerCertificate is not called on a new connection but is called on a resumed connection.

[rolandshoemaker]

@FiloSottile what do you expect the behavior to be here for the VerifyPeerCertificate bug? (side note: it looks like the call is only gated by ClientAuth, if the client doesn't send anything you still get the callback)

It seems reasonable to me that if the server didn't request client auth that VerifyPeerCertificates wouldn't be called, even if the client did provide certificates (because the certificate processing code shouldn't be called) which matches what happens in new connections, but not resumed ones. Is the suggested behavior making resumed connections match new connections?