8834353 enabled again RSA-PSS in TLS 1.2. Unfortunately, there are broken crypto.Signer implementations out there that do the wrong thing when asked to generate RSA-PSS signatures.
TLS 1.3 requires RSA-PSS, so there is no way around ripping this band-aid off, but partially because of this TLS 1.3 was opt-in in Go 1.12 and is opt-out in Go 1.13 (#30055). RSA-PSS in TLS 1.2 would just be enabled with no way to turn it off, and both adding another GODEBUG option, or making GODEBUG=tls13=0 impact TLS 1.2 feel wrong.
What finally tipped the scale is that #28660 provides a nice way for code to opt-out of RSA-PSS in TLS 1.2 if needed. Let's wait to force RSA-PSS in TLS 1.2 until Go 1.14, when TLS 1.3 is also forced on, and there is a code path to disabling it.
Signing with RSA-PSS can uncover faulty crypto.Signer implementations,
and it can fail for (broken) small keys. We'll have to take that
breakage eventually, but it would be nice for it to be opt-out at first.
TLS 1.3 requires RSA-PSS and is opt-out in Go 1.13. Instead of making a
TLS 1.3 opt-out influence a TLS 1.2 behavior, let's wait to add RSA-PSS
to TLS 1.2 until TLS 1.3 is on without opt-out.
Note that since the Client Hello is sent before a protocol version is
selected, we have to advertise RSA-PSS there to support TLS 1.3.
That means that we still support RSA-PSS on the client in TLS 1.2 for
verifying server certificates, which is fine, as all issues arise on the
signing side. We have to be careful not to pick (or consider available)
RSA-PSS on the client for client certificates, though.
We'd expect tests to change only in TLS 1.2:
* the server won't pick PSS to sign the key exchange
(Server-TLSv12-* w/ RSA, TestHandshakeServerRSAPSS);
* the server won't advertise PSS in CertificateRequest
* and the client won't pick PSS for its CertificateVerify
Client-TLSv12-Renegotiate* because "R" requests a client cert).
Client-TLSv13-ClientCert-RSA-RSAPSS was updated because of a fix in the test.
This effectively reverts 8834353.
Testing was made more complex by the undocumented semantics of OpenSSL's
-[client_]sigalgs (see openssl/openssl#9172).
Reviewed-by: Adam Langley <firstname.lastname@example.org>