TLS1.3

The OpenSSL 1.1.1 release includes support for TLSv1.3. The release is binary and API compatible with OpenSSL 1.1.0. In theory, if your application supports OpenSSL 1.1.0, then all you need to do to upgrade is to drop in the new version of OpenSSL and you will automatically start being able to use TLSv1.3. However there are some issues that application developers and deployers need to be aware of.

Differences with TLS1.2 and below

TLSv1.3 is a major rewrite of the specification. There was some debate as to whether it should really be called TLSv2.0 - but TLSv1.3 it is. There are major changes and some things work very differently. A brief, incomplete, summary of some things that you are likely to notice follows:

• There are new ciphersuites that only work in TLSv1.3. The old ciphersuites cannot be used for TLSv1.3 connections and the new ones cannot be used in TLSv1.2 and below.
• The new ciphersuites are defined differently and do not specify the certificate type (e.g. RSA, DSA, ECDSA) or the key exchange mechanism (e.g. DHE or ECDHE). This has implications for ciphersuite configuration.
• Clients provide a “key_share” in the ClientHello. This has consequences for “group” configuration.
• Sessions are not established until after the main handshake has been completed. There may be a gap between the end of the handshake and the establishment of a session (or, in theory, a session may not be established at all). This could have impacts on session resumption code.
• Renegotiation is not possible in a TLSv1.3 connection
• More of the handshake is now encrypted.
• More types of messages can now have extensions (this has an impact on the custom extension APIs and Certificate Transparency)
• DSA certificates are no longer allowed in TLSv1.3 connections

Note that at this stage only TLSv1.3 is supported. DTLSv1.3 is still in the early days of specification and there is no OpenSSL support for it at this time.

Current status of the TLSv1.3 standard

The TLSv1.3 standard has now been published as [RFC 8446]. During the development of the standard the TLS Working Group published various draft versions. Implementations of draft versions of the standard identify the specific draft version that they are using. This means that implementations based on different draft versions, and also the final RFC version, do not interoperate with each other.

The OpenSSL git master branch (and the 1.1.1-pre9 beta version) contain our development TLSv1.3 code which is based on the final version of RFC8446 and can be used for testing purposes (i.e. it is not for production use). Earlier beta versions of OpenSSL 1.1.1 implemented draft versions of the standard. Those versions contained the macro TLS1_3_VERSION_DRAFT_TXT in the tls1.h header file which identified the specific draft version that was implemented. This macro has been removed from 1.1.1-pre9 and the current master branch.

TLSv1.3 is enabled by default in the latest development versions (there is no need to explicitly enable it). To disable it at compile time you must use the “no-tls1_3” option to “config” or “Configure”.

Although the latest 1.1.1 versions support the final standard version, other applications that support TLSv1.3 may still be using older draft versions. This is a common source of interoperability problems. If two peers supporting different TLSv1.3 draft versions attempt to communicate then they will fall back to TLSv1.2.

Ciphersuites

OpenSSL has implemented support for five TLSv1.3 ciphersuites as follows:

• TLS_AES_256_GCM_SHA384
• TLS_CHACHA20_POLY1305_SHA256
• TLS_AES_128_GCM_SHA256
• TLS_AES_128_CCM_8_SHA256
• TLS_AES_128_CCM_SHA256

Due to the major differences between the way that ciphersuites for TLSv1.2 and below and ciphersuites for TLSv1.3 work, they are configured in OpenSSL differently too.

By default the first three of the above ciphersuites are enabled by default. This means that if you have no explicit ciphersuite configuration then you will automatically use those three and will be able to negotiate TLSv1.3. Note that changing the TLSv1.2 and below cipher list has no impact on the TLSv1.3 ciphersuite configuration.

Applications should use the SSL_CTX_set_ciphersuites() or SSL_set_ciphersuites() functions to configure TLSv1.3 ciphersuites. Note that the functions SSL_CTX_get_ciphers() and SSL_get_ciphers() will return the full list of ciphersuites that have been configured for both TLSv1.2 and below and TLSv1.3.

For the OpenSSL command line applications there is a new "-ciphersuites" option to configure the TLSv1.3 ciphersuite list. This is just a simple colon (":") separated list of TLSv1.3 ciphersuite names in preference order. Note that you cannot use the special characters such as "+", "!", "-" etc, that you can for defining TLSv1.2 ciphersuites. In practice this is not likely to be a problem because there are only a very small number of TLSv1.3 ciphersuites.

For example:

$ openssl s_server -cert mycert.pem -key mykey.pem -cipher ECDHE -ciphersuites "TLS_AES_256_GCM_SHA384:TLS_CHACHA20_POLY1305_SHA256"

This will configure OpenSSL to use any ECDHE based ciphersuites for TLSv1.2 and below. For TLSv1.3 the TLS_AES_256_GCM_SHA384 and TLS_CHACHA20_POLY1305_SHA256 ciphersuites will be available.

Note that all of the above applies to the "ciphers" command line application as well. This can sometimes lead to surprising results. For example this command:

$ openssl ciphers -s -v ECDHE

Will list all the ciphersuites for TLSv1.2 and below that support ECDHE and additionally all of the default TLSv1.3 ciphersuites. Use the "-ciphersuites" option to further configure the TLSv1.3 ciphersuites.

Groups

In TLSv1.3 the client selects a “group” that it will use for key exchange. OpenSSL only supports ECDHE groups for this. The client then sends “key_share” information to the server for its selected group in the ClientHello.

The list of supported groups is configurable. It is possible for a client to select a group that the server does not support. In this case the server requests that the client sends a new key_share that it does support. While this means a connection will still be established (assuming a mutually supported group exists), it does introduce an extra server round trip - so this has implications for performance. In the ideal scenario the client will select a group that the server supports in the first instance.

In practice most clients will use X25519 or P-256 for their initial key_share. For maximum performance it is recommended that servers are configured to support at least those two groups and clients use one of those two for its initial key_share. This is the default case (OpenSSL clients will use X25519).

The group configuration also controls the allowed groups in TLSv1.2 and below. If applications have previously configured their groups in OpenSSL 1.1.0 then you should review that configuration to ensure that it still makes sense for TLSv1.3. The first named (i.e. most preferred group) will be the one used by an OpenSSL client in its intial key_share.

Applications can configure the group list by using SSL_CTX_set1_groups() or a similar function (see here for further details). Alternatively, if applications use SSL_CONF style configuration files then this can be configured using the Groups or Curves command (see here).

Sessions

In TLSv1.2 and below a session is established as part of the handshake. This session can then be used in a subsequent connection to achieve an abbreviated handshake. Applications might typically obtain a handle on the session after a handshake has completed using the SSL_get1_session() function (or similar).

In TLSv1.3 sessions are not established until after the main handshake has completed. The server sends a separate post-handshake message to the client containing the session details. Typically this will happen soon after the handshake has completed, but it could be sometime later (or not at all).

The specification recommends that applications only use a session once (although this may not be enforced). For this reason some servers send multiple session messages to a client. To enforce the “use once” recommendation applications could use SSL_CTX_remove_session() to mark a session as non-resumable (and remove it from the cache) once it has been used. OpenSSL servers that accept "early_data" will automatically enforce single use sessions. Any attempt to resume with a session that has already been used will fallback to a full handshake.

The old SSL_get1_session() and similar APIs may not operate as expected for client applications written for TLSv1.2 and below. Specifically if a client application calls SSL_get1_session() before the server message containing session details has been received then an SSL_SESSION object will still be returned, but any attempt to resume with it will not succeed and a full handshake will occur instead. In the case where multiple sessions have been sent by the server then only the last session will be returned by SSL_get1_session(). Calling SSL_get1_session() after a 2 way shutdown will give a resumable session if any was sent. You can check that a session is resumable with SSL_SESSION_is_resumable().

Client application developers should consider using the SSL_CTX_sess_set_new_cb() API instead. This provides a callback mechanism which gets invoked every time a new session is established. This can get invoked multiple times for a single connection if a server sends multiple session messages.

Note that SSL_CTX_sess_set_new_cb() was also available in previous versions of OpenSSL. Applications that already used that API will still work, but they may find that the callback is invoked at unexpected times, i.e. post-handshake.

An OpenSSL server will immediately attempt to send session details to a client after the main handshake has completed. The number of tickets can be set using SSL_CTX_set_num_tickets. To server applications this post-handshake stage will appear to be part of the main handshake, so calls to SSL_get1_session() should continue to work as before.

If a client sends it's data and directly sends the close notify request and closes the connection, the server will still try to send tickets if configured to do so. Since the connection is already closed by the client, this might result in a write error and receiving the SIGPIPE signal. The write error will be ignored if it's a session ticket. But server applications can still get SIGPIPE they didn't get before.

If the server sends session tickets and you want to be able to get a resumable session, you need to either call SSL_read() after the ticket was sent or do a 2 way shutdown.

Custom Extensions and Certificate Transparency

In TLSv1.2 and below the initial ClientHello and ServerHello messages can contain “extensions”. This allows the base specifications to be extended with additional features and capabilities that may not be applicable in all scenarios or could not be foreseen at the time that the base specifications were written. OpenSSL provides support for a number of “built-in” extensions.

Additionally the custom extensions API provides some basic capabilities for application developers to add support for new extensions that are not built-in to OpenSSL.

Built on top of the custom extensions API is the “serverinfo” API. This provides an even more basic interface that can be configured at run time. One use case for this is Certificate Transparency. OpenSSL provides built-in support for the client side of Certificate Transparency but there is no built-in server side support. However this can easily be achieved using “serverinfo” files. A serverinfo file containing the Certificate Transparency information can be configured within OpenSSL and it will then be sent back to the client as appropriate.

In TLSv1.3 the use of extensions is expanded significantly and there are many more messages that can include them. Additionally some extensions that were applicable to TLSv1.2 and below are no longer applicable in TLSv1.3 and some extensions are moved from the ServerHello message to the EncryptedExtensions message. The old custom extensions API does not have the ability to specify which messages the extensions should be associated with. For that reason a new custom extensions API was required.

The old API will still work, but the custom extensions will only be added where TLSv1.2 or below is negotiated. To add custom extensions that work for all TLS versions application developers will need to update their applications to the new API (see here for details).

The “serverinfo” data format has also been updated to include additional information about which messages the extensions are relevant to. Applications using “serverinfo” files may need to update to the “version 2” file format to be able to operate in TLSv1.3 (see here for details).

Renegotiation

TLSv1.3 does not have renegotiation so calls to SSL_renegotiate() or SSL_renegotiate_abbreviated() will immediately fail if invoked on a connection that has negotiated TLSv1.3.

A common use case for renegotiation is to update the connection keys. The function SSL_key_update() can be used for this purpose in TLSv1.3 (see here for further details).

Another use case is to request a certificate from the client. This can be achieved by using the SSL_verify_client_post_handshake() function in TLSv1.3 (see here for further details).

DSA certificates

DSA certificates are no longer allowed in TLSv1.3. From OpenSSL 1.1.0 and above ciphersuites for TLSv1.2 and below based on DSA are no longer available by default (you must compile OpenSSL with the "enable-weak-ssl-ciphers" option, and explicitly configure the ciphersuites at run time). If your server application is using a DSA certificate and has made the necessary configuration changes to enable the ciphersuites then TLSv1.3 will never be negotiated when that certificate is used for a connection (the maximum version will be TLSv1.2).

Please use an ECDSA or RSA certificate instead.

Middlebox Compatibility Mode

During development of the TLSv1.3 standard it became apparent that in some cases, even if a client and server both support TLSv1.3, connections could sometimes still fail. This is because middleboxes on the network between the two peers do not understand the new protocol and prevent the connection from taking place. In order to work around this problem the TLSv1.3 specification introduced a “middlebox compatibility” mode. This made a few optional changes to the protocol to make it appear more like TLSv1.2 so that middleboxes would let it through. Largely these changes are superficial in nature but do include sending some small but unneccessary messages. OpenSSL has middlebox compatibility mode on by default, so most users should not need to worry about this. However applications may choose to switch it off by calling the function SSL_CTX_clear_options() and passing SSL_OP_ENABLE_MIDDLEBOX_COMPAT as an argument (see here for further details).

If the remote peer is not using middlebox compatibility mode and there are problematic middleboxes on the network path then this could cause spurious connection failures.

Server Name Indication

Server Name Indication (SNI) can be used by the client to select one of several sites on the same host, and so a different X.509 certificate can be sent depending on the hostname that was sent in the SNI extension. If the SNI extension is not sent the server's options are to either disconnect or select a default hostname and matching certificate. The default would typically be the main site.

SNI has been made mandatory to implement in TLS 1.3 but not mandatory to use. Some sites want to encourage the use of SNI and configure a default certificate that fails WebPKI authentication when the client supports TLS 1.3. This is under the assumption that if a hostname is not sent, then it means that the client does not verify the server certificate (unauthenticated opportunistic TLS). For implementation that actually don't send the SNI extension, but do verify the server certificate this can cause connection failures.

To enable SNI you need to use the SSL_set_tlsext_host_name() function. For hostname validation see Hostname validation.

PSKs

Pre-shared Keys work differently in TLSv1.2 and below compared to TLSv1.3.

In TLSv1.2 (and below) special PSK specific ciphersuites are used. A client wishing to use a PSK will offer one (or more) of those ciphersuites to the server in the initial ClientHello message. If the server also wishes to use a PSK, then it will select that ciphersuite and will (optionally) send back an "identity hint" to the client. Finally the client sends back to the server identity details so that the server knows which PSK to use. In OpenSSL 1.1.0 and below this is implemented using a callback mechanism. The callback is called passing in the identity hint (or NULL if there is no hint) and the callback responds by filling in the identity details, as well as the PSK itself.

In TLSv1.3, if a client wishes to use a PSK, then the identity details are sent immediately in the initial ClientHello message. Use of a PSK is independent of any ciphersuite selection. If the server wishes to use the PSK then it will signal this in its response to the client. Otherwise a full (non-PSK) handshake will occur. Note that there is no identity hint in TLSv1.3.

OpenSSL 1.1.1 introduces new TLSv1.3 specific PSK callbacks. See here and here for further details. These are the preferred callbacks to use for TLSv1.3 PSKs. However, if an application has set up the TLSv1.2 PSK callbacks and TLSv1.3 is enabled then OpenSSL will attempt to fallback to using the old style callbacks. In this case, on the client side, the callback will be invoked before any communication with the server has taken place during construction of the initial ClientHello. This is because the identity details must be sent immediately in TLSv1.3. The identity hint value will always be NULL in this case.

Note that the TLSv1.2 callbacks could end up being called twice for the same connection. For example if a client application provides no TLSv1.3 callback and TLSv1.3 is enabled, then it will be called first during the initial ClientHello construction. If the server subsequently selects TLSv1.2 then the callback will be called again later on in the handshake to set up the TLSv1.2 PSK.

TLSv1.3 PSKs must specify a message digest (e.g. such as SHA-256). Where old style TLSv1.2 callbacks are used in a TLSv1.3 context then the message digest will default to SHA-256 (as specified in the standard). A server which has been configured with TLSv1.2 PSK callbacks, but negotiates TLSv1.3 with a client, will prefer ciphersuites based on SHA-256 in order to maximise the chances of a PSK being used.

Non-application data records

TLSv1.3 sends more non-application data records after the handshake is finished. At least the session ticket and possibly a key update is send after the finished message. With TLSv1.2 it happened in case of renegotiation. SSL_read() has always documented that it can return SSL_ERROR_WANT_READ after processing non-application data, even when there is still data that can be read. When SSL_MODE_AUTO_RETRY is set using SSL_CTX_set_mode() OpenSSL will try to process the next record, and so not return SSL_ERROR_WANT_READ while it still has data available. Because many applications did not handle this properly, SSL_MODE_AUTO_RETRY has been made the default. If the application is using blocking sockets and SSL_MODE_AUTO_RETRY is enabled, and select() is used to check if a socket is readable this results in SSL_read() processing the non-application data records, but then try to read an application data record which might not be available and hang.

Conclusion

TLSv1.3 represents a significant step forward and has some exciting new features but there are some hazards for the unwary when upgrading. Mostly these issues have relatively straight forward solutions. Application developers should review their code and consider whether anything should be updated in order to work more effectively with TLSv1.3. Similarly application deployers should review their configuration.