Expand description
This section collects together goal-oriented documentation.
§Customising private key usage
By default rustls supports PKCS#8-format1 RSA or ECDSA keys, plus PKCS#1-format RSA keys.
However, if your private key resides in a HSM, or in another process, or perhaps another machine, rustls has some extension points to support this:
The main trait you must implement is sign::SigningKey
. The primary method here
is choose_scheme
where you are given a set of SignatureScheme
s the client says
it supports: you must choose one (or return None
– this aborts the handshake). Having
done that, you return an implementation of the sign::Signer
trait.
The sign()
performs the signature and returns it.
(Unfortunately this is currently designed for keys with low latency access, like in a PKCS#11 provider, Microsoft CryptoAPI, etc. so is blocking rather than asynchronous. It’s a TODO to make these and other extension points async.)
Once you have these two pieces, configuring a server to use them involves, briefly:
- packaging your
sign::SigningKey
with the matching certificate chain into asign::CertifiedKey
- making a
ResolvesServerCertUsingSni
and feeding in yoursign::CertifiedKey
for all SNI hostnames you want to use it for, - setting that as your
ServerConfig
’scert_resolver
For a complete example of implementing a custom sign::SigningKey
and sign::Signer
see
the rustls-cng crate.
§Unexpected EOF
TLS has a close_notify
mechanism to prevent truncation attacks2.
According to the TLS RFCs, each party is required to send a close_notify
message before
closing the write side of the connection. However, some implementations don’t send it.
So long as the application layer protocol (for instance HTTP/2) has message length framing
and can reject truncated messages, this is not a security problem.
Rustls treats an EOF without close_notify
as an error of type std::io::Error
with
ErrorKind::UnexpectedEof
. In some situations it’s appropriate for the application to handle
this error the same way it would handle a normal EOF (a read returning Ok(0)
). In particular
if UnexpectedEof
occurs on an idle connection it is appropriate to treat it the same way as a
clean shutdown. And if an application always uses messages with length framing (in other words,
messages are never delimited by the close of the TCP connection), it can unconditionally
ignore UnexpectedEof
errors from rustls.
For PKCS#8 it does not support password encryption – there’s not a meaningful threat model addressed by this, and the encryption supported is typically extremely poor. ↩