Introduction

This guide provides an overview of how to use the s2n-tls library. Chapters in this guide go over important concepts in the s2n-tls library and include explanations for how to use s2n-tls APIs to configure specific TLS features.

s2n-tls API

The API exposed by s2n-tls is the set of functions and declarations that are in the s2n.h header file. Any functions and declarations that are in the s2n.h file are intended to be stable (API and ABI) within major version numbers of s2n-tls releases. Other functions and structures used in s2n-tls internally can not be considered stable and their parameters, names, and sizes may change.

In general, s2n-tls APIs are not thread safe unless explicitly specified otherwise.

Read Error Handling for information on processing API return values safely.

The VERSIONING.rst document contains more details about s2n's approach to versions and API changes.

API Reference

s2n-tls uses Doxygen to document its public API. The latest s2n-tls documentation can be found on GitHub pages.

The doxygen documentation should be used in conjunction with this guide.

Examples

To understand the API it may be easiest to see examples in action. s2n-tls's bin directory includes an example client (s2nc) and server (s2nd).

Note: s2nc and s2nd are intended for testing purposes only, and should not be used in production.

Initialization and Teardown

Initialization

The s2n-tls library must be initialized with s2n_init() before calling most library functions. s2n_init() will error if it is called more than once, even when an application uses multiple threads.

Initialization can be modified by calling s2n_crypto_disable_init() or s2n_disable_atexit() before s2n_init().

An application can override s2n-tls’s internal memory management by calling s2n_mem_set_callbacks() before calling s2n_init().

If you are trying to use FIPS mode, you must enable FIPS in your libcrypto library (probably by calling FIPS_mode_set(1)) before calling s2n_init().

Teardown

Thread-local Memory

We recommend calling s2n_cleanup() from every thread created after s2n_init() to ensure there are no memory leaks. s2n-tls has thread-local memory that it attempts to clean up automatically at thread-exit. However, this is done using pthread destructors and may not work if you are using a threads library other than pthreads.

Library Cleanup

A full cleanup and de-initialization of the library can be done by calling s2n_cleanup_final(). s2n-tls allocates some memory at initialization that is intended to live for the duration of the process, but can be cleaned up earlier with s2n_cleanup_final(). Not calling this method may cause tools like ASAN or valgrind to detect memory leaks, as the memory will still be allocated when the process exits.

Error Handling

s2n-tls functions that return 'int' return 0 to indicate success and -1 to indicate failure. s2n-tls functions that return pointer types return NULL in the case of failure. When an s2n-tls function returns a failure, s2n_errno will be set to a value corresponding to the error. This error value can be translated into a string explaining the error in English by calling s2n_strerror(s2n_errno, "EN"). A string containing human readable error name, can be generated with s2n_strerror_name(). A string containing internal debug information, including filename and line number, can be generated with s2n_strerror_debug(). This string is useful to include when reporting issues to the s2n-tls development team.

Example:

if (s2n_config_set_cipher_preferences(config, prefs) < 0) {
    printf("Setting cipher prefs failed! %s : %s", s2n_strerror(s2n_errno, "EN"), s2n_strerror_debug(s2n_errno, "EN"));
    return -1;
}

NOTE: To avoid possible confusion, s2n_errno should be cleared after processing an error: s2n_errno = S2N_ERR_T_OK

When using s2n-tls outside of C, the address of the thread-local s2n_errno may be obtained by calling the s2n_errno_location() function. This will ensure that the same TLS mechanisms are used with which s2n-tls was compiled.

Error Types

s2n-tls organizes errors into different "types" to allow applications to handle error values without catching all possibilities. Applications using non-blocking I/O should check the error type to determine if the I/O operation failed because it would block or for some other error. To retrieve the type for a given error use s2n_error_get_type(). Applications should perform any error handling logic using these high level types:

Here's an example that handles errors based on type:

#define SUCCESS 0
#define FAILURE 1
#define RETRY 2

s2n_errno = S2N_ERR_T_OK;
if (s2n_negotiate(conn, &blocked) < 0) {
    switch(s2n_error_get_type(s2n_errno)) {
        case S2N_ERR_T_BLOCKED:
            /* Blocked, come back later */
            return RETRY;
        case S2N_ERR_T_CLOSED:
            return SUCCESS;
        case S2N_ERR_T_IO:
            handle_io_err(errno);
            return FAILURE;
        case S2N_ERR_T_PROTO:
            handle_proto_err();
            return FAILURE;
        case S2N_ERR_T_ALERT:
            log_alert(s2n_connection_get_alert(conn));
            return FAILURE;
        /* Everything else */
        default:
            log_other_error();
            return FAILURE;
    }
}

Blinding

Blinding is a mitigation against timing side-channels which in some cases can leak information about encrypted data. This blog post includes a good description of blinding and how it mitigates timing side-channels like Lucky13.

By default s2n-tls will cause a thread to sleep between 10 and 30 seconds whenever an error triggered by the peer occurs. The default implementation of blinding blocks the thread.

For non-blocking blinding, an application can handle the delay itself. This is useful for applications that are handling many connections in a single thread. The application should set the S2N_SELF_SERVICE_BLINDING option with s2n_connection_set_blinding(). After s2n_recv() or s2n_negotiate() return an error, the application must call s2n_connection_get_delay() and pause activity on the connection for the specified number of nanoseconds before calling s2n_shutdown(), close(), or shutdown(). s2n_shutdown() will fail if called before the blinding delay elapses. To correctly implement self-service blinding, the application must have nanosecond-level resolution on its implementation of the delay. Not properly implementing self-service blinding (such as by not waiting for the full delay before calling close() on the underlying socket) makes an application potentially vulnerable to timing side-channel attacks.

The maximum blinding delay is configurable via s2n_config_set_max_blinding_delay(). However, setting a maximum delay lower than the recommended default (30s) will make timing side-channel attacks easier. The lower the delay, the fewer requests and less total time an attacker will need to execute a side-channel attack. If a lower delay is required for reasons such as client timeouts, then the highest value practically possible should be chosen to limit risk. Do not lower the blinding delay without fully understanding the risks.

Stacktraces

s2n-tls has an mechanism to capture stacktraces when errors occur. This mechanism is off by default, but can be enabled in code by calling s2n_stack_traces_enabled_set(). It can be enabled globally by setting the environment variable S2N_PRINT_STACKTRACE=1.

Call s2n_print_stacktrace() to print your stacktrace.

Note: Enabling stacktraces can significantly slow down unit tests, causing failures on tests (such as s2n_cbc_verify) that measure the timing of events.

TLS Connections

In general, s2n-tls works by operating on s2n_connection structs. A user should first create a connection by calling s2n_connection_new(). Then a TLS handshake can be performed on the connection. Finally, the connection can be used to send and receive encrypted data. An s2n_config struct can be associated with a connection to configure additional options.

Call s2n_connection_free() to free the memory allocated for connection when no longer needed.

Connection Memory

The connection struct is roughly 4KB with some variation depending on how it is configured. Maintainers of the s2n-tls library carefully consider increases to the size of the connection struct as they are aware some users are memory-constrained.

A connection struct has memory allocated specifically for the TLS handshake. Memory-constrained users can free that memory by calling s2n_connection_free_handshake() after the handshake is successfully negotiated. Note that the handshake memory can be reused for another connection if s2n_connection_wipe() is called, so freeing it may result in more memory allocations later. Additionally some functions that print information about the handshake may not produce meaningful results after the handshake memory is freed.

The input and output buffers consume the most memory on the connection after the handshake. It may not be necessary to keep these buffers allocated when a connection is in a keep-alive or idle state. Call s2n_connection_release_buffers() to wipe and free the in and out buffers associated with a connection to reduce memory overhead of long-lived connections.

Connection Reuse

Connection objects can be re-used across many connections to reduce memory allocation. Calling s2n_connection_wipe() will wipe an individual connection's state and allow the connection object to be re-used for a new TLS connection.

Connection Info

s2n-tls provides many methods to retrieve details about the handshake and connection, such as the parameters negotiated with the peer. For a full list, see our doxygen guide.

Protocol Version

s2n-tls provides multiple different methods to get the TLS protocol version of the connection. They should be called after the handshake has completed.

• s2n_connection_get_actual_protocol_version(): The actual TLS protocol version negotiated during the handshake. This is the primary value referred to as "protocol_version", and the most commonly used.
• s2n_connection_get_server_protocol_version(): The highest TLS protocol version the server supports.
• s2n_connection_get_client_protocol_version(): The highest TLS protocol version the client advertised.

Configuring the Connection

s2n_config objects are used to change the default settings of a s2n-tls connection. Use s2n_config_new() to create a new config object. To associate a config with a connection call s2n_connection_set_config(). A config should not be altered once it is set on a connection as this will produce undefined behavior. It is not necessary to create a config object per connection; one config object should be used for many connections. Call s2n_config_free() to free the object when no longer needed. Only free the config object when all connections using it have been freed.

Calling s2n_config_new() can have a performance cost during config creation due to loading default system certificates into the trust store (see Configuring the Trust Store). For increased performance, use s2n_config_new_minimal() when system certificates are not needed for certificate validation.

Most commonly, a s2n_config object is used to set the certificate key pair for authentication and change the default security policy. See the sections for certificates and security policies for more information on those settings.

Overriding the Config

Some s2n_config settings can be overridden on a specific connection if desired. For example, s2n_config_append_protocol_preference() appends a list of ALPN protocols to a s2n_config. Calling the s2n_connection_append_protocol_preference() API will override the list of ALPN protocols for an individual connection. Not all config APIs have a corresponding connection API so if there is one missing contact us with an explanation on why it is required for your use-case.

Security Policies

s2n-tls uses pre-made security policies to help avoid common misconfiguration mistakes for TLS.

s2n_config_set_cipher_preferences() sets a security policy, which includes the cipher/kem/signature/ecc preferences and protocol version.

Supported TLS Versions

Currently TLS 1.2 is our default version, but we recommend TLS 1.3 where possible. To use TLS 1.3 you need a security policy that supports TLS 1.3.

SSL 3.0, TLS 1.0, and TLS 1.1

s2n-tls supports older versions, but their use is not recommended.

Compatibility with older versions of TLS may also require support for older ciphersuites. Although it is possible to negotiate older versions of TLS with more modern options like SHA256, AES, ECDSA, and ECDHE, older clients and servers may only support older options like SHA1, 3DES, RSA certs, and either RSA or DHE key exchange. If a security policy allows older TLS versions but does not allow older ciphersuites, handshakes with older clients and servers may fail.

SSL 2.0

s2n-tls will not negotiate SSL 2.0, but will accept SSLv2 ClientHellos advertising a higher protocol version like TLS1.2. See the "Compatibility with SSL 2.0" section in the TLS 1.2 RFC.

Compatibility with SSLv2 ClientHellos advertising TLS1.2 may require similar support for older ciphersuites as compatibility with older versions of TLS does. In particular, SSLv2 ClientHellos are likely to require support for SHA1 and either RSA or DHE key exchange. This is due to technical limitations of the SSLv2 ClientHello, which does not include TLS extensions.

Chart: Security Policy Version To Protocol Version And Ciphersuites

The following chart maps the security policy version to protocol version and ciphersuites supported.

The "default", "default_tls13", and "default_fips" versions are special in that they will be updated with future s2n-tls changes to keep up-to-date with current security best practices. Ciphersuites, protocol versions, and other options may be added or removed, or their internal order of preference might change. Warning: this means that the default policies may change as a result of library updates, which could break peers that rely on legacy options.

In contrast, numbered or dated versions are fixed and will never change. The numbered equivalents of the default policies are currently:

• "default": "20240501"
• "default_fips": "20240502"
• "default_tls13": "20240503" For previous defaults, see the "Default Policy History" section below.

"default_fips" does not currently support TLS1.3. If you need a policy that supports both FIPS and TLS1.3, choose "20230317". We plan to add TLS1.3 support to both "default" and "default_fips" in the future.

"rfc9151" is derived from Commercial National Security Algorithm (CNSA) Suite Profile for TLS and DTLS 1.2 and 1.3. This policy restricts the algorithms allowed for signatures on certificates in the certificate chain to RSA or ECDSA with sha384, which may require you to update your certificates. Like the default policies, this policy may also change if the source RFC definition changes.

s2n-tls does not expose an API to control the order of preference for each ciphersuite or protocol version. s2n-tls follows the following order:

1. Always prefer the highest protocol version supported
2. Always use forward secrecy where possible. Prefer ECDHE over DHE.
3. Prefer encryption ciphers in the following order: AES128, AES256, ChaCha20, 3DES, RC4.
4. Prefer record authentication modes in the following order: GCM, Poly1305, SHA256, SHA1, MD5.

NOTE: ChaCha20-Poly1305 cipher suites will not be available if s2n-tls is built with an older libcrypto like openssl-1.0.2. The underlying encrypt/decrypt functions are not available in older versions.

ChaCha20 Boosting

s2n-tls usually prefers AES over ChaCha20. However, some clients-- particularly mobile or IOT devices-- do not support AES hardware acceleration, making AES less efficient and performant than ChaCha20. In this case, clients will indicate their preference for ChaCha20 by listing it first during cipher suite negotiation. Usually s2n-tls servers ignore client preferences, but s2n-tls offers "ChaCha20 boosted" security policies that will choose ChaCha20 over AES if the client indicates a preference for ChaCha20. This is available in the "CloudFront-TLS-1-2-2021-ChaCha20-Boosted" policy, which is identical to the "CloudFront-TLS-1-2-2021" policy listed above but with ChaCha20 Boosting enabled.

Chart: Security Policy Version To Supported Signature Schemes

NOTE: Legacy SHA-1 algorithms are not supported in TLS1.3. Legacy SHA-1 algorithms will be supported only if TLS1.2 has been negotiated and the security policy allows them.

NOTE: RSA-PSS certificates will not be available if s2n-tls is built with an older libcrypto like openssl-1.0.2. RSA certificate signatures with PSS padding (RSA-PSS-RSAE) will still be available, but RSA-PSS certificate signatures with PSS padding (RSA-PSS-PSS) will not be available.

Chart: Security policy version to supported curves/groups

Default Policy History

Sending and Receiving

Basic IO setup

By default, s2n-tls sends and receives data using a provided file descriptor (usually a socket) and the read/write system calls. The file descriptor can be set with s2n_connection_set_fd(), or separate read and write file descriptors can be set with s2n_connection_set_read_fd() and s2n_connection_set_write_fd(). The provided file descriptor should be active and connected.

In general the application is free to configure the file descriptor as preferred, including socket options. s2n-tls itself sets a few socket options:

• If available, TCP_QUICKACK is used during the TLS handshake
• If available and enabled via s2n_connection_use_corked_io(), TCP_CORK is used during the TLS handshake. TCP_NOPUSH or TCP_NODELAY may be used if TCP_CORK is not available.

Important Note: If the read end of the pipe is closed unexpectedly, writing to the pipe will raise a SIGPIPE signal. s2n-tls does NOT handle SIGPIPE. A SIGPIPE signal will cause the process to terminate unless it is handled or ignored by the application. See the signal man page for instructions on how to handle C signals, or simply ignore the SIGPIPE signal by calling signal(SIGPIPE, SIG_IGN) before calling any s2n-tls IO methods.

Blocking or Non-Blocking?

s2n-tls supports both blocking and non-blocking I/O.

• In blocking mode, each s2n-tls I/O function will not return until it has completed the requested IO operation.
• In non-blocking mode, s2n-tls I/O functions will immediately return, even if the socket couldn't send or receive all the requested data. In this case, the I/O function will return S2N_FAILURE, and s2n_error_get_type() will return S2N_ERR_T_BLOCKED. The I/O operation will have to be called again in order to send or receive the remaining requested data.

Some s2n-tls I/O functions take a blocked argument. If an I/O function returns an S2N_ERR_T_BLOCKED error, the blocked argument will be set to a s2n_blocked_status value, indicating what s2n-tls is currently blocked on. Note that unless an I/O function returns S2N_FAILURE with an S2N_ERR_T_BLOCKED error, the blocked argument is meaningless, and should not be used in any application logic.

Servers in particular usually prefer non-blocking mode. In blocking mode, a single connection blocks the thread while waiting for more IO. In non-blocking mode, multiple connections can make progress by returning control while waiting for more IO using methods like poll or select.

To use s2n-tls in non-blocking mode, set the underlying file descriptors as non-blocking. For example:

int flags = fcntl(fd, F_GETFL, 0);
if (flags < 0) return S2N_FAILURE;
if (fcntl(fd, F_SETFL, flags | O_NONBLOCK) < 0) return S2N_FAILURE;

Note: If an application requires non-blocking IO, it likely also requires self-service blinding. See Blinding.

Errors and Alerts

If the peer sends an alert, the next call to a read IO method will report S2N_FAILURE and s2n_error_get_type() will return S2N_ERR_T_ALERT. The specific alert received is available by calling s2n_connection_get_alert().

In TLS1.3, all alerts are fatal. s2n-tls also treats all alerts as fatal in earlier versions of TLS by default. s2n_config_set_alert_behavior() can be called to force s2n-tls to treat pre-TLS1.3 warning alerts as not fatal, but that behavior is not recommended unless required for compatibility. In the past, attacks against TLS have involved manipulating the alert level to disguise fatal alerts as warnings.

If s2n-tls encounters a fatal error, the next call to a write IO method will send a close_notify alert to the peer. Except for a few exceptions, s2n-tls does not send specific alerts in order to avoid leaking information that could be used for a sidechannel attack. To ensure that the alert is sent, s2n_shutdown() should be called after an error.

Performing the TLS Handshake

Before application data can be sent or received, an application must perform a handshake to establish a TLS connection with the peer.

To perform the handshake, call s2n_negotiate() until it either returns S2N_SUCCESS or returns S2N_FAILURE without a S2N_ERR_T_BLOCKED error.

For an example of how to perform a basic handshake, see examples/s2n_negotiate.c

Application Data

After the TLS handshake, an application can send and receive encrypted data.

Although most s2n-tls APIs are not thread-safe, s2n_send() and s2n_recv() may be called simultaneously from two different threads. This means that an application may have one thread calling s2n_send() and one thread calling s2n_recv(), but NOT multiple threads calling s2n_recv() or multiple threads calling s2n_send().

Even if an application only intends to send data or only intends to receive data, it should implement both send and receive in order to handle alerts and post-handshake TLS control messages like session tickets.

Sending Application Data

s2n_send() and its variants encrypt and send application data to the peer. The sending methods return the number of bytes written and may indicate a partial write. Partial writes are possible not just for non-blocking I/O, but also for connections aborted while active.

A single call to s2n_send() may involve multiple system calls to write the provided application data. s2n-tls breaks the application data into fixed-sized records before encryption, then calls write for each record. See the record size documentation for how record size may impact performance.

In non-blocking mode, s2n_send() will send data from the provided buffer and return the number of bytes sent, as long as the socket was able to send at least 1 byte. If no bytes could be sent on the socket, s2n_send() will return S2N_FAILURE, and s2n_error_get_type() will return S2N_ERR_T_BLOCKED. To ensure that all the provided data gets sent, applications should continue calling s2n_send() until the return values across all calls have added up to the length of the data, or until s2n_send() returns an S2N_ERR_T_BLOCKED error. After an S2N_ERR_T_BLOCKED error is returned, applications should call s2n_send() again only after the socket is able to send more data. This can be determined by using methods like poll or select.

Unlike OpenSSL, repeated calls to s2n_send() should not duplicate the original parameters, but should update the inputs per the indication of size written. s2n-tls will attempt to sanity check that the inputs are properly updated between calls. Because of those sanity checks, a zero-length send call cannot be used as a flushing mechanism.

s2n_sendv_with_offset() behaves like s2n_send(), but supports vectorized buffers. The offset input should be updated between calls to reflect the data already written.

s2n_sendv() also supports vectorized buffers, but assumes an offset of 0. Because of this assumption, a caller would have to make sure that the input vectors are updated to account for a partial write. Therefore s2n_sendv_with_offset() is preferred.

For examples of how to send data of length data_size with s2n_send() or s2n_sendv_with_offset(), see examples/s2n_send.c

Receiving Application Data

s2n_recv() reads and decrypts application data from the peer, copying it into the application-provided output buffer. It returns the number of bytes read, and may indicate a partial read even if blocking IO is used. It returns "0" to indicate that the peer has shutdown the connection.

By default, s2n_recv() will return after reading a single TLS record. s2n_recv() can be called repeatedly to read multiple records. To allow s2n_recv() to read multiple records with a single call, use s2n_config_set_recv_multi_record().

In non-blocking mode, s2n_recv() will read data into the provided buffer and return the number of bytes read, as long as at least 1 byte was read from the socket. If no bytes could be read from the socket, s2n_recv() will return S2N_FAILURE, and s2n_error_get_type() will return S2N_ERR_T_BLOCKED. To ensure that all data on the socket is properly received, applications should continue calling s2n_recv() until it returns an S2N_ERR_T_BLOCKED error. After an S2N_ERR_T_BLOCKED error is returned, applications should call s2n_recv() again only after the socket has received more data. This can be determined by using methods like poll or select.

Unlike OpenSSL, repeated calls to s2n_recv() should not duplicate the original parameters, but should update the inputs per the indication of size read.

For an example of how to read all the data sent by the peer into one buffer, see s2n_example_recv() in examples/s2n_recv.c

For an example of how to echo any data sent by the peer, see s2n_example_recv_echo() in examples/s2n_recv.c

s2n_peek() can be used to check if more application data may be returned from s2n_recv() without performing another read from the file descriptor. This is useful when using select() on the underlying s2n-tls file descriptor, because a call to s2n_recv() may read more data into s2n-tls's internal buffer than was requested or can fit into the application-provided output buffer. This extra application data will be returned by the next call to s2n_recv(), but select() will be unable to tell the application that there is more data available and that s2n_recv() should be called again. An application can solve this problem by calling s2n_peek() to determine if s2n_recv() needs to be called again.

Closing the Connection

s2n_shutdown() attempts a graceful closure at the TLS layer. It does not close the underlying transport. The call may block on either reading or writing.

s2n_shutdown() should be called after an error in order to ensure that s2n-tls sends an alert to notify the peer of the failure.

s2n_shutdown() will discard any application data received from the peer. This can lead to data truncation, so s2n_shutdown_send() may be preferred for TLS1.3 connections where the peer continues sending after the application initiates shutdown. See Closing the connection for writes below.

Because s2n_shutdown() attempts a graceful shutdown, it will not return success unless a close_notify alert is successfully both sent and received. As a result, s2n_shutdown() may fail when interacting with a non-conformant TLS implementation or if called on a connection in a bad state.

Once s2n_shutdown() is complete:

• The s2n_connection handle cannot be used for reading or writing.
• The underlying transport can be closed, most likely via shutdown() or close().
• The s2n_connection handle can be freed via s2n_connection_free() or reused via s2n_connection_wipe()

Closing the connection for writes

TLS1.3 supports closing the write side of a TLS connection while leaving the read side unaffected. This indicates "end-of-data" to the peer without preventing future reads. This feature is usually referred to as "half-close".

s2n-tls offers the s2n_shutdown_send() method to close the write side of a connection. Unlike s2n_shutdown(), it does not wait for the peer to respond with a close_notify alert and does not discard any incoming application data. An application can continue to call s2n_recv() after a call to s2n_shutdown_send().

s2n_shutdown_send() may still be called for earlier TLS versions, but most TLS implementations will react by immediately discarding any pending writes and closing the connection.

If s2n_shutdown_send() is used, the application should still call s2n_shutdown() or wait for s2n_recv() to return 0 to indicate end-of-data before cleaning up the connection or closing the read side of the underlying transport.

Custom IO Callbacks

By default, s2n-tls sends and receives data using a provided file descriptor (usually a socket) and the read/write system calls. To change this default behavior, an application can implement custom send and receive methods using s2n_connection_set_recv_cb() and s2n_connection_set_send_cb(). The application can pass inputs (such as a file descriptor or destination buffer) to the custom IO methods by using s2n_connection_set_recv_ctx() and s2n_connection_set_send_ctx(). s2n-tls will call the custom IO methods with the custom context instead of calling the default implementation.

The custom IO methods may send or receive less than the requested length. They should return the number of bytes sent/received, or set errno and return an error code < 0. s2n-tls will interpret errno set to EWOULDBLOCK or EAGAIN as indicating a retriable blocking error, and set s2n_errno and the s2n_blocked_status appropriately. s2n-tls will interpret a return value of 0 as a closed connection.

TLS Record Sizes

Throughput vs Latency

When sending data, s2n-tls uses a default maximum record size which experimentation has suggested provides a reasonable balance of performance and throughput.

s2n_connection_prefer_throughput() can be called to increase the record size, which minimizes overhead. It also increases s2n-tls's memory usage.

s2n_connection_prefer_low_latency() can be called to decrease the record size, which allows the receiver to decrypt the data faster. It also decreases s2n-tls's memory usage.

These options only affect the size of the records that s2n-tls sends, not the behavior of the peer.

Maximum Fragment Length

The maximum number of bytes that can be sent in a TLS record is called the "maximum fragment length", and is set to 2^14 bytes by default. Regardless of the maximum record size that s2n-tls uses when sending, it may receive records containing up to 2^14 bytes of plaintext.

A client can request a lower maximum fragment length by calling s2n_config_send_max_fragment_length(), reducing the size of TLS records sent and providing benefits similar to s2n_connection_prefer_low_latency(). However, many TLS servers either ignore these requests or handle them incorrectly, so a client should never assume that a lower maximum fragment length will be honored. If a server accepts the requested maximum fragment length, the client will respect that maximum when sending.

By default, an s2n-tls server will ignore a client's requested maximum fragment length. If s2n_config_accept_max_fragment_length() is called, the server will respect the client's requested maximum fragment length when sending, but will not reject client records with a larger fragment size.

If a maximum fragment length is negotiated during the connection, it will override the behavior configured by s2n_connection_prefer_throughput() and s2n_connection_prefer_low_latency().

Dynamic Record Sizing

Sending smaller records at the beginning of a connection can decrease first byte latency, particularly if TCP slow start is used.

s2n_connection_set_dynamic_record_threshold() can be called to initially send smaller records. The connection will send the first resize_threshold bytes in records small enough to fit in a single standard 1500 byte ethernet frame. Whenever timeout_threshold seconds pass without sending data, the connection will revert to this behavior and send small records again.

Dynamic record sizing doesn't completely override s2n_connection_prefer_throughput(), s2n_connection_prefer_low_latency(), or the negotiated maximum fragment length. Once resize_threshold is hit, records return to the maximum size configured for the connection. And if the maximum fragment length negotiated with the peer is lower than what dynamic record sizing would normally produce, the lower value will be used.

Certificates and Authentication

TLS uses certificates to authenticate the server (and optionally the client). The handshake will fail if the client cannot verify the server’s certificate.

Authentication is usually the most expensive part of the handshake. To avoid the cost, consider using session resumption or pre-shared keys.

Configuring the Trust Store

To validate the peer’s certificate, a config's local “trust store” must contain a certificate that can authenticate the peer’s certificate. To add certificates to the trust store, call s2n_config_set_verification_ca_location() or s2n_config_add_pem_to_trust_store().

s2n_config_new() initializes the trust store with the default system certificates, which may include public CAs or other unexpected certificates. If s2n-tls is verifying certificates and does not require these default system certificates, they should be cleared with s2n_config_wipe_trust_store() before calling s2n_config_set_verification_ca_location() or s2n_config_add_pem_to_trust_store().

Note that the s2n_verify_host_fn callback must be implemented to validate the identity of all received certificates. A client implementation is set by default. If the identity of the received certificates are implicit (i.e. the certificates are self-signed, or signed by a privately owned CA), the s2n_verify_host_fn callback need not perform any additional validation. However, s2n_config_wipe_trust_store() should be called before adding certificates to the trust store in this case, in order to avoid implicitly trusting the identity of peers that may present certificates trusted via an unexpected default system certificate.

Configs created with s2n_config_new_minimal() are initialized with empty trust stores. To add default system certificates to these configs, call s2n_config_load_system_certs().

Server Authentication

A server must have a certificate and private key pair to prove its identity. s2n-tls supports RSA, RSA-PSS, and ECDSA certificates, and allows one of each type to be added to a config for a given domain name.

Create a new certificate and key pair by calling s2n_cert_chain_and_key_new(), then load the pem-encoded data with s2n_cert_chain_and_key_load_pem_bytes(). Call s2n_config_add_cert_chain_and_key_to_store() to add the certificate and key pair to the config. When a certificate and key pair is no longer needed, it must be cleaned up with s2n_cert_chain_and_key_free().

A client can add restrictions on the certificate’s hostname by setting a custom s2n_verify_host_fn with s2n_config_set_verify_host_callback() or s2n_connection_set_verify_host_callback(). The default behavior is to require that the hostname match the server name set with s2n_set_server_name().

SNI

TLS servers will often serve multiple domains from a single IP address, with each domain potentially requiring its own certificate. When a TLS client receives a server's certificate, it will check to ensure that it was issued for the domain that the client is connecting to. As such, the server needs to know which domain the client is connecting to in order to send the correct certificate. This information is communicated to the server via the Server Name Indication (SNI) extension.

Client configuration

s2n_set_server_name() is used on client connections to specify a domain name in the SNI extension.

Server configuration

Each certificate loaded with s2n_config_add_cert_chain_and_key_to_store() is automatically associated with every domain name listed in its Subject Alternative Name (SAN) field. The domain names listed in the Common Name (CN) field are used instead if the certificate doesn't contain an SAN field. s2n-tls servers automatically send a certificate that matches the value of the client's SNI extension during the TLS handshake.

s2n-tls servers allow a maximum of 1 certificate to be loaded per domain, per certificate type (RSA, RSA-PSS, ECDSA). By default, a newly loaded certificate with an overlapping domain name will not replace the existing certificate associated with that domain. The s2n_cert_tiebreak_callback may be implemented to customize which certificate is selected when an overlapping domain is encountered.

When selecting a certificate to send to the client, s2n-tls servers prefer an exact SNI match before falling back to a certificate with an associated wildcard domain (a domain starting with "*.", covering all subdomains). A default certificate is selected if no SNI match exists, or if the client doesn't send an SNI extension. A default certificate is set when s2n_config_add_cert_chain_and_key_to_store() is called for the first time for a given certificate type (RSA, RSA-PSS, ECDSA). s2n_config_set_cert_chain_and_key_defaults() may be used to override the default certificates.

Client / Mutual Authentication

Client authentication is not enabled by default. However, the server can require that the client also provide a certificate, if the server needs to authenticate clients before accepting connections.

Client authentication can be configured by calling s2n_config_set_client_auth_type() or s2n_connection_set_client_auth_type() for both the client and server. Additionally, the client will need to load a certificate and key pair as described for the server in Server Authentication and the server will need to configure its trust store as described in Configuring the Trust Store.

When using client authentication, the server MUST implement the s2n_verify_host_fn, because the default behavior will likely reject all client certificates.

When using client authentication with TLS1.3, s2n_negotiate will report a successful handshake to clients before the server validates the client certificate. If the server then rejects the client certificate, the client may later receive an alert while calling s2n_recv, potentially after already having sent application data with s2n_send. This is a quirk of the TLS1.3 protocol message ordering: the server does not send any more handshake messages after the client sends the client certificate (see the TLS1.3 state machine). There is no security risk, since the client has already authenticated the server, but it could make handshake failures and authentication errors more difficult to handle.

Certificate Inspection

Applications may want to know which certificate was used by a server for authentication during a connection, since servers can set multiple certificates. s2n_connection_get_selected_cert() will return the local certificate chain object used to authenticate. s2n_connection_get_peer_cert_chain() will provide the peer's certificate chain, if they sent one. Use s2n_cert_chain_get_length() and s2n_cert_chain_get_cert() to parse the certificate chain object and get a single certificate from the chain. Use s2n_cert_get_der() to get the DER encoded certificate if desired.

Additionally s2n-tls has functions for parsing certificate extensions on a certificate. Use s2n_cert_get_x509_extension_value_length() and s2n_cert_get_x509_extension_value() to obtain a specific DER encoded certificate extension from a certificate. s2n_cert_get_utf8_string_from_extension_data_length() and s2n_cert_get_utf8_string_from_extension_data() can be used to obtain a specific UTF8 string representation of a certificate extension instead. These functions will work for both RFC-defined certificate extensions and custom certificate extensions.

Certificate Revocation

Certificate revocation is how CAs inform validators that an active certificate should not be trusted. This commonly occurs when a private key has been leaked and the identity of the certificate's owner can no longer be trusted.

s2n-tls supports two methods of certificate revocation: OCSP stapling and CRLs. A fundamental difference between the two is that with OCSP stapling, the peer offering the certificate validates the revocation status of its own certificate. This peer can choose not to send a certificate status response, and applications will have to decide whether or not to fail certificate validation in this case. With CRLs, the application checks the revocation status of the certificate itself, without relying on the peer. However, CRLs must be retrieved and stored by the application, which requires more network and memory utilization than OCSP stapling.

Users who want certificate revocation should look closely at their use-case and decide which method is more appropriate. We suggest using OCSP stapling if you're sure your peer supports OCSP stapling. CRLs should be used if this assumption can't be made. However, s2n-tls does not enable applications to fetch CRLs for received certificates in real-time. This method should only be used if you're able to obtain CRLs in advance for all certificates you expect to encounter.

OCSP Stapling

Online Certificate Status Protocol (OCSP) is a protocol to establish whether or not a certificate has been revoked. The requester (usually a client), asks the responder (usually a server), to ‘staple’ the certificate status information along with the certificate itself. The certificate status sent back will be either expired, current, or unknown, which the requester can use to determine whether or not to accept the certificate.

OCSP stapling can be applied to both client and server certificates when using TLS1.3, but only to server certificates when using TLS1.2.

To use OCSP stapling, the requester must call s2n_config_set_status_request_type() with S2N_STATUS_REQUEST_OCSP. The responder will need to call s2n_cert_chain_and_key_set_ocsp_data() to set the raw bytes of the OCSP stapling data.

The OCSP stapling information will be automatically validated if the underlying libcrypto supports OCSP validation. s2n_config_set_check_stapled_ocsp_response() can be called with "0" to turn this off. Call s2n_connection_get_ocsp_response() to retrieve the received OCSP stapling information for manual verification.

CRL Validation

Note: the CRL validation feature in s2n-tls is currently considered unstable, meaning the CRL APIs are subject to change in a future release. To access the CRL APIs, include api/unstable/crl.h.

Certificate Revocation Lists (CRLs) are lists of issued, unexpired certificates that have been revoked by the CA. CAs publish updated versions of these lists periodically. A validator wishing to verify a certificate obtains a CRL from the CA, validates the CRL, and checks to ensure the certificate is not contained in the list, and therefore has not been revoked by the CA.

The s2n CRL lookup callback must be implemented and set via s2n_config_set_crl_lookup_cb() to enable CRL validation in s2n-tls. This callback will be triggered once for each certificate in the certificate chain.

The CRLs for all certificates received in the handshake must be obtained in advance of the CRL lookup callback, outside of s2n-tls. It is not possible in s2n-tls to obtain CRLs in real-time. Applications should load these CRLs into memory, by creating s2n_crls via s2n_crl_new(), and adding the obtained CRL data via s2n_crl_load_pem(). The s2n_crl should be freed via s2n_crl_free() when no longer needed.

The application must implement a way to look up the correct CRL for a given certificate. This can be done by comparing the hash of the received certificate's issuer with the hash of the CRL's issuer. The certificate's issuer hash is retrieved via s2n_crl_lookup_get_cert_issuer_hash(), and the CRL's issuer hash is retrieved via s2n_crl_get_issuer_hash(). Once a CRL is found with a matching issuer hash, call s2n_crl_lookup_set() to provide s2n-tls with this CRL.

Call s2n_crl_lookup_ignore() to ignore a received certificate if its CRL can't be found. This will cause the certificate validation logic to fail with a S2N_ERR_CRL_LOOKUP_FAILED error if the certificate is needed in the chain of trust. The certificate validation logic will not fail if the ignored certificate ends up not being included in the chain of trust.

By default, the CRL validation logic will not fail on CRLs that are not yet active, or are expired. Timestamp validation can optionally be performed in the CRL lookup callback by calling s2n_crl_validate_active() and s2n_crl_validate_not_expired().

Certificate Transparency

Certificate transparency is a framework to store public logs of CA-issued certificates. If requested, certificate owners can send a signed certificate timestamp (SCT) to prove that their certificate exists in these logs. The requester can choose whether or not to accept a certificate based on this information.

Certificate transparency information can be applied to both client and server certificates when using TLS1.3, but only to server certificates when using TLS1.2.

To use certificate transparency, the requester (usually the client) must call s2n_config_set_ct_support_level() with S2N_CT_SUPPORT_REQUEST. The responder (usually the server) must call s2n_cert_chain_and_key_set_sct_list() to set the raw bytes of the transparency information.

Call s2n_connection_get_sct_list() to retrieve the received certificate transparency information. The format of this data is the SignedCertificateTimestampList structure defined in section 3.3 of RFC 6962.

Examining the Client Hello

Getting a Client Hello

From a connection

s2n-tls stores the received Client Hello and makes it available to the application. Call s2n_connection_get_client_hello() to get a pointer to the s2n_client_hello struct storing the Client Hello message. A NULL value will be returned if the connection has not yet received the Client Hello. The earliest point in the handshake when this struct is available is during the Client Hello Callback. The stored Client Hello message will not be available after calling s2n_connection_free_handshake().

From raw bytes

s2n-tls can parse a Client Hello from raw bytes. Call s2n_client_hello_parse_message() with raw bytes representing a Client Hello message (including the message header, but excluding the record header). The returned pointer to a s2n_client_hello struct behaves the same as a pointer returned from s2n_connection_get_client_hello(), except that the memory is owned by the application and must be freed with s2n_client_hello_free().

Examining the message

Call s2n_client_hello_get_raw_message() to retrieve the complete Client Hello message with the random bytes on it zeroed out.

Call s2n_client_hello_get_cipher_suites() to retrieve the list of cipher suites sent by the client.

Call s2n_client_hello_get_session_id() to retrieve the session ID sent by the client in the ClientHello message. Note that this value may not be the session ID eventually associated with this particular connection since the session ID can change when the server sends the Server Hello. The official session ID can be retrieved with s2n_connection_get_session_id()after the handshake completes.

Call s2n_client_hello_get_extensions() to retrieve the entire list of extensions sent in the Client Hello. Call s2n_client_hello_get_extension_by_id() to retrieve a specific extension. Because s2n_client_hello_get_extension_by_id() doesn't distinguish between zero-length extensions and missing extensions, s2n_client_hello_has_extension() should be used to check for the existence of an extension.

Call s2n_client_hello_get_supported_groups() to retrieve the entire list of supported groups sent by the client.

SSLv2

s2n-tls will not negotiate SSLv2, but will accept SSLv2 ClientHellos advertising a higher protocol version like SSLv3 or TLS1.0. This was a backwards compatibility strategy used by some old clients when connecting to a server that might only support SSLv2.

You can determine whether an SSLv2 ClientHello was received by checking the value of s2n_connection_get_client_hello_version(). If an SSLv2 ClientHello was received, then s2n_connection_get_client_protocol_version() will still report the real protocol version requested by the client.

SSLv2 ClientHellos are formatted differently than ClientHellos in later versions. s2n_client_hello_get_raw_message() and s2n_client_hello_get_cipher_suites() will produce differently formatted data. See the documentation for those methods for details about proper SSLv2 ClientHello parsing.

Client Hello Callback

Users can access the Client Hello during the handshake by setting the callback s2n_config_set_client_hello_cb(). A possible use-case for this is to modify the s2n_connection based on information in the Client Hello. This should be done carefully, as modifying the connection in response to untrusted input can be dangerous. In particular, switching from an s2n_config that supports TLS1.3 to one that does not opens the server up to a possible version downgrade attack.

s2n_connection_server_name_extension_used() MUST be invoked before exiting the callback if any of the connection properties were changed on the basis of the Server Name extension. If desired, the callback can return a negative value to make s2n-tls terminate the handshake early with a fatal handshake failure alert.

Callback Modes

The callback can be invoked in two modes: S2N_CLIENT_HELLO_CB_BLOCKING or S2N_CLIENT_HELLO_CB_NONBLOCKING. Use s2n_config_set_client_hello_cb_mode() to set the desired mode.

The default mode, "blocking mode", will wait for the Client Hello callback to succeed and then continue the handshake. Use this mode for light-weight callbacks that won't slow down the handshake or block the main thread, like logging or simple configuration changes.

In contrast, "non-blocking mode" will wait for the ClientHello callback to succeed and then pause the handshake, immediately returning from s2n_negotiate with an error indicating that the handshake is blocked on application input. This allows the application to do expensive or time-consuming work like network calls outside of the callback without blocking the main thread. Only when the application calls s2n_client_hello_cb_done() will the handshake be able to resume.

Session Resumption

TLS handshake sessions are CPU-heavy due to the calculations involved in authenticating a certificate. These calculations can be skipped after the first connection by turning on session resumption. This mechanism stores state from the previous session and uses it to establish the next session, allowing the handshake to skip the costly authentication step while keeping the same cryptographic guarantees. The authentication step can be skipped because both the server and client will use their possession of the key from the previous session to prove who they are. We usually refer to the stored session state as a "session ticket". Note that this session ticket is encrypted by the server, so a server will have to set up an external key in order to do session resumption.

Session Ticket Key

The key that encrypts and decrypts the session state is not related to the keys negotiated as part of the TLS handshake and has to be set by the server by calling s2n_config_add_ticket_crypto_key(). Unlike other methods that modify s2n_config objects, s2n_config_add_ticket_crypto_key() can be called after setting an s2n_config object on a connection. See RFC5077 for guidelines on securely generating keys.

Each key has two different expiration dates. The first expiration date signifies the time that the key can be used for both encryption and decryption. The second expiration date signifies the time that the key can be used only for decryption. This mechanism is to ensure that a session ticket can be successfully decrypted if it was encrypted by a key that was about to expire. The full lifetime of the key is therefore the encrypt-decrypt lifetime plus the decrypt-only lifetime. To alter the default key lifetime call s2n_config_set_ticket_encrypt_decrypt_key_lifetime() and s2n_config_set_ticket_decrypt_key_lifetime().

The server will stop issuing session resumption tickets if a user doesn't set up a new key before the previous key passes through its encrypt-decrypt lifetime. Therefore it is recommended to add a new key when half of the previous key's encrypt-decrypt lifetime has passed.

Stateless Session Resumption

In stateless session resumption the server sends a session ticket to a client after a successful handshake, and the client can send that ticket back to the server during a new connection to skip the authentication step. This mechanism allows servers to avoid storing individual state for each client, and for that reason is the preferred method for resuming a session.

Servers should call s2n_config_set_session_tickets_onoff() to enable stateless session resumption. Additionally the server needs to set up an encryption key using s2n_config_add_ticket_crypto_key().

Clients should call s2n_config_set_session_tickets_onoff() to enable stateless session resumption and set a session ticket callback function using s2n_config_set_session_ticket_cb(), which will allow clients to receive a session ticket when it arrives. Then s2n_connection_set_session() should be called with that saved ticket when attempting to resume a new connection.

Stateful Session Resumption

In stateful session resumption, also known as session caching, the server caches the session state per client and resumes a session based on the client's session ID. Note that session caching has not been implemented for > TLS1.2. If stateful session resumption is turned on and a TLS1.3 handshake is negotiated, the caching mechanism will not store that session and resumption will not be available the next time the client connects.

Servers should set the three caching callback functions: s2n_config_set_cache_store_callback(), s2n_config_set_cache_retrieve_callback(), and s2n_config_set_cache_delete_callback() and then call s2n_config_set_session_cache_onoff() to enable stateful session resumption. Session caching will not be turned on unless all three session cache callbacks are set prior to calling s2n_config_set_session_cache_onoff(). Additionally, the server needs to set up an encryption key using s2n_config_add_ticket_crypto_key().

Clients should call s2n_connection_get_session() to retrieve some serialized state about the session. Then s2n_connection_set_session() should be called with that saved state when attempting to resume a new connection.

The cache_delete_callback is called when a connection encounters a fatal error. This allows a server to delete a potentially corrupted or faulty session from its cache. Because an unexpected end-of-stream is considered a fatal error, an application should ensure that it performs a graceful TLS shutdown when using session caching. For more information on how to close connections, see Closing the Connection.

Session Resumption in TLS1.2 and TLS1.3

In TLS1.2, session ticket messages are sent during the handshake and are automatically received as part of calling s2n_negotiate(). They will be available as soon as negotiation is complete.

In TLS1.3, session ticket messages are sent after the handshake as "post-handshake" messages, and may not be received as part of calling s2n_negotiate(). A s2n-tls server will send tickets immediately after the handshake, so clients can receive them by calling s2n_recv() immediately after the handshake completes. However, other server implementations may send their session tickets later, at any time during the connection.

Additionally, in TLS1.3, multiple session tickets may be issued for the same connection. Servers can call s2n_config_set_initial_ticket_count() to set the number of tickets they want to send and s2n_connection_add_new_tickets_to_send() to increase the number of tickets to send during a connection.

Session Resumption Forward Secrecy

In TLS1.2, the secret stored inside the ticket is the original session's master secret. Because of this, TLS1.2 session tickets are not forward secret, meaning that compromising the resumed session's secret exposes the original session's encrypted data.

In contrast, in TLS1.3 the secret stored inside the ticket is derived from the original session's master secret. The derivation uses a cryptographic operation that can't be reversed by an attacker to retrieve the original master secret. Therefore, TLS1.3 session tickets are forward secret, meaning compromising the resumed session's secret will not expose the original session's encrypted data.

Keying Material Lifetimes in TLS1.2 and TLS1.3

In TLS1.2, a full handshake can issue a session ticket encrypted with a specific session ticket encryption key. Connections that resume using that session ticket will not issue new session tickets. Therefore, the lifetime of the original "keying material"-- meaning the lifetime of any secret derived from the original full handshake-- is limited by the lifetime of the session ticket encryption key. Applications can set the session ticket encryption key lifetime with s2n_config_set_ticket_encrypt_decrypt_key_lifetime().

In TLS1.3, connections that resume using a session ticket CAN issue new session tickets. This is because TLS1.3 tickets are intended to be single-use, and each ticket contains a different secret: see Session Resumption Forward Secrecy. These new session tickets may be encrypted with newer session ticket encryption keys, allowing the original "keying material" to outlive the original session ticket encryption key. However, TLS1.3 enforces a specific separate "keying material" lifetime, which servers can configure with s2n_connection_set_server_keying_material_lifetime(). This effectively places a limit on how long sessions can be resumed before a new full handshake is required.

Offloading Private Key Operations

By default, s2n-tls automatically uses the configured private key to synchronously perform the signature and decryption operations required for a tls handshake. However, this default behavior may not work for some situations.

For example:

• An application may want to perform the CPU-expensive signature and decryption operations asynchronously to avoid blocking the main event loop. See Asynchronous private key operations
• An application may not have direct access to the private key, such as when using PKCS#11. See Offloading private key operations

To handle these use cases, s2n-tls provides a callback to allow users to control how these operations are performed. The callback is set via s2n_config_set_async_pkey_callback() and is triggered every time s2n_negotiate() performs an action involving the private key. The callback is passed op, an opaque object representing the private key operation. To avoid memory leaks, op must always eventually be freed by calling s2n_async_pkey_op_free().

The private key operation can be performed by calling s2n_async_pkey_op_perform() (or s2n_async_pkey_op_set_output(): see Offloading private key operations). The required private key can be retrieved using the s2n_connection_get_selected_cert() and s2n_cert_chain_and_key_get_private_key() calls. The operation can then be finalized with s2n_async_pkey_op_apply() to continue the handshake.

Asynchronous Private Key Operations

When s2n-tls is used in non-blocking mode, private key operations can be completed asynchronously. This model can be useful to move execution of CPU-heavy private key operations out of the main event loop, preventing s2n_negotiate() from blocking the loop for a few milliseconds each time the private key operation needs to be performed.

To handle private key operations asynchronously, return from the callback without calling s2n_async_pkey_op_perform() or s2n_async_pkey_op_apply(). Usually the user would do this by spawning a separate thread to perform op and immediately returning S2N_SUCCESS from the callback without waiting for that separate thread to complete. In response, s2n_negotiate() will return S2N_FAILURE with an error of type S2N_ERR_T_BLOCKED and s2n_blocked_status set to S2N_BLOCKED_ON_APPLICATION_INPUT. All subsequent calls to s2n_negotiate() will produce the same result until s2n_async_pkey_op_apply() is called to finalize the op.

Note: It is not safe to call multiple functions on the same conn or op objects from 2 different threads at the same time. Doing so will produce undefined behavior. However it is safe to have a call to a function involving only conn at the same time as a call to a function involving only op, as those objects are not coupled with each other. It is also safe to free conn or op at any moment with respective function calls, with the exception that conn cannot be freed inside the s2n_async_pkey_fn() callback.

Synchronous Private Key Operations

Despite the "async" in the function names, private key operations can also be completed synchronously using the callback. To complete an operation synchronously, simply call s2n_async_pkey_op_perform() and s2n_async_pkey_op_apply() inside the callback. If the callback succeeds, the handshake will continue uninterrupted. If the callback fails, s2n_negotiate() will fail with an error of type S2N_ERR_T_INTERNAL.

Offloading Private Key Operations

The s2n_async_pkey_op_perform() call used to perform a private key operation requires direct access to the private key. In some cases, like when using PKCS#11, users may not have access to the private key. In these cases, we can substitute s2n_async_pkey_op_set_output() for s2n_async_pkey_op_perform() to tell s2n-tls the result of the operation rather than having s2n-tls perform the operation itself.

s2n-tls provides a number of calls to gather the information necessary for an outside module or library to perform the operation. The application can query the type of private key operation by calling s2n_async_pkey_op_get_op_type(). In order to perform an operation, the application must ask s2n-tls to copy the operation's input into an application supplied buffer. The appropriate buffer size can be determined by calling s2n_async_pkey_op_get_input_size(). Once a buffer of the proper size is allocated, the application can request the input data by calling s2n_async_pkey_op_get_input(). After the operation is completed, the finished output can be copied back to S2N by calling s2n_async_pkey_op_set_output(). Once the output is set, the private key operation can be completed by calling s2n_async_pkey_op_apply() as usual.

Offloading can be performed either synchronously or asynchronously. If the offloaded operation fails synchronously, simply return S2N_FAILURE from the callback. If the offloaded operation fails asynchronously, s2n-tls does not provide a way to communicate that result. Instead, simply shutdown and cleanup the connection as you would for any other error.

Pre-shared Keys

s2n-tls supports pre-shared keys (PSKs) as of TLS1.3. PSKs allow users to establish secrets outside of the handshake, skipping certificate exchange and authentication.

Benefits of Using Pre-Shared Keys

Using pre-shared keys can avoid the need for public key operations. This is useful in performance-constrained environments with limited CPU power. PSKs may also be more convenient from a key management point of view: If the system already has a mechanism for sharing secrets, that mechanism can be reused for TLS PSKs.

Security Considerations

A PSK must not be shared between more than one server and one client. An entity that acts as both a server and a client should not use the same PSK for both roles. For more information see: Selfie: reflections on TLS 1.3 with PSK.

Configuring External Pre-Shared Keys

Both clients and servers will need to create and append a PSK to a connection in order to negotiate that PSK. Call s2n_external_psk_new() to allocate memory for a PSK object. Call s2n_psk_set_identity() to set a unique identifier for the PSK. Note that this identity will be transmitted over the network unencrypted, so do not include any confidential information in it. Call s2n_psk_set_secret() to set the secret value for a given PSK. Deriving a shared secret from a password or other low-entropy source is not secure and is subject to dictionary attacks. See this RFC for more guidelines on creating a secure PSK secret. Call s2n_psk_set_hmac() to change the hmac algorithm (defaults to S2N_PSK_HMAC_SHA256) for a given PSK. Note that the hmac algorithm may influence server cipher-suite selection.

Call s2n_connection_append_psk() to append the newly created PSK to the connection. Both the server and client should call this API to add PSKs to their connection. PSKs that are appended first will be preferred over PSKs appended last unless custom PSK selection logic is implemented. Use s2n_psk_free() to free the memory allocated for a PSK once you have associated it with a connection.

External PSKs and Session Resumption cannot both be used in TLS13. Therefore, users must pick which mode they are using by calling s2n_config_set_psk_mode() prior to the handshake. Additionally, s2n_connection_set_psk_mode() overrides the PSK mode set on the config for a particular connection.

Selecting a Pre-Shared Key

By default, a server chooses the first identity in its PSK list that also appears in the client's PSK list. The s2n_psk_selection_callback is available if you would like to implement your own PSK selection logic. Currently, this callback is not asynchronous. Call s2n_config_set_psk_selection_callback() to associate your s2n_psk_selection_callback with a config.

The s2n_psk_selection_callback will provide the list of PSK identities sent by the client in the psk_list input parameter. You will need to create an offered PSK object by calling s2n_offered_psk_new() and pass this object as a parameter in s2n_offered_psk_list_next() in order to populate the offered PSK object. Call s2n_offered_psk_list_has_next() prior to calling s2n_offered_psk_list_next() to determine if there exists another PSK in the psk_list. Call s2n_offered_psk_get_identity() to get the identity of a particular s2n_offered_psk.

Call s2n_offered_psk_list_choose_psk() to choose a particular s2n_offered_psk to be used for the connection. Note that the server must have already configured the corresponding PSK on the connection using s2n_connection_append_psk(). To disable PSKs for the connection and perform a full handshake instead, set the PSK identity to NULL. Call s2n_offered_psk_free() once you have chosen a particular PSK to free the memory allocated.

If desired, s2n_offered_psk_list_reread() returns the offered PSK list to its original read state. After s2n_offered_psk_list_reread() is called, the next call to s2n_offered_psk_list_next() will return the first PSK in the offered PSK list.

Use s2n_connection_get_negotiated_psk_identity() to retrieve the PSK identity selected by the server for the connection.

In the following example, s2n_psk_selection_callback chooses the first client offered PSK identity present in an external store.

int s2n_psk_selection_callback(struct s2n_connection *conn, void *context,
                               struct s2n_offered_psk_list *psk_list)
{
    struct s2n_offered_psk *offered_psk = s2n_offered_psk_new();

    while (s2n_offered_psk_list_has_next(psk_list)) {
        uint8_t *client_psk_id = NULL;
        uint16_t client_psk_id_len = 0;

        s2n_offered_psk_list_next(psk_list, offered_psk);
        s2n_offered_psk_get_identity(offered_psk, &client_psk_id, &client_psk_id_len);
        struct s2n_psk *psk = user_lookup_identity_db(client_psk_id, client_psk_id_len);

        if (psk) {
            s2n_connection_append_psk(conn, psk);
            s2n_offered_psk_list_choose_psk(psk_list, offered_psk);
            break;
        }
    }
    s2n_offered_psk_free(&offered_psk);
    return S2N_SUCCESS;
}

Early Data or 0-RTT Data

TLS1.3 introduced the ability for clients to send data before completing the handshake when using external pre-shared keys or session resumption.

WARNING: Early data does not have the same security properties as regular data sent after a successful handshake.

• It is not forward secret. If the PSK or session resumption secret is compromised, then the early data is also compromised.
• It is susceptible to replay attacks unless proper precautions are taken. Early data can be captured and successfully resent by an attacker. See the TLS1.3 RFC section on replay attacks for more details, and Adding anti-replay protection for how to implement counter measures.

Do not enable early data for your application unless you have understood and mitigated the risks.

Configuring Session Resumption for Early Data

To use early data with session tickets, early data must be enabled on a server by setting the maximum early data allowed to a non-zero value with s2n_config_set_server_max_early_data_size() or s2n_connection_set_server_max_early_data_size(). The server then begins issuing tickets that support early data, and clients can use early data when they use those tickets.

Configuring External Pre-shared Keys for Early Data

To use early data with pre-shared keys, individual pre-shared keys must support early data. In addition to configuring the maximum early data allowed, each pre-shared key needs an associated cipher suite and if applicable, application protocol. The server only accepts early data if the pre-shared key's associated cipher suite and application protocol match the cipher suite and the application protocol negotiated during the handshake.

The maximum early data allowed and cipher suite can be set with s2n_psk_configure_early_data(). If the connection will negotiate an application protocol then the expected application protocol can be set with s2n_psk_set_application_protocol().

Sending Early Data

To send early data, your application should call s2n_send_early_data() before it calls s2n_negotiate().

s2n_connection_get_remaining_early_data_size() can be called to check how much more early data the client is allowed to send. If s2n_send_early_data() exceeds the allowed maximum, s2n-tls returns a usage error.

Like other IO functions, s2n_send_early_data() can potentially fail repeatedly with a blocking error before it eventually succeeds: see I/O Functions for more information. An application can stop calling s2n_send_early_data() at any time, even if the function has not returned success yet. If s2n_send_early_data() does return success, the connection is ready to complete the handshake and begin sending normal data. However, s2n_send_early_data() can continue to be called to send more early data if desired.

Once a client finishes sending early data, you should call s2n_negotiate() to complete the handshake just as you would for a handshake that did not include early data.

For example:

uint8_t early_data[] = "early data to send";
ssize_t total_data_sent = 0, len = sizeof(early_data);
while (total_data_sent < len) {
    ssize_t data_sent = 0;
    int r = s2n_send_early_data(client_conn, early_data + total_data_sent,
            len - total_data_sent, &data_sent, &blocked);
    total_data_sent += data_sent;
    if (r == S2N_SUCCESS) {
        break;
    } else if (s2n_error_get_type(s2n_errno) != S2N_ERR_T_BLOCKED) {
        exit(1);
    }
}
while (s2n_negotiate(client_conn, &blocked) != S2N_SUCCESS) {
    if (s2n_error_get_type(s2n_errno) != S2N_ERR_T_BLOCKED) {
        exit(1);
    }
}

Receiving Early Data

To receive early data, your application should call s2n_recv_early_data() before it calls s2n_negotiate().

Like other S2N IO functions, s2n_recv_early_data() can potentially fail repeatedly with a blocking error before it eventually succeeds: see I/O Functions for more information. Once s2n_recv_early_data() has been called, it must be called until it returns success. If an application stops calling s2n_recv_early_data() early, some early data may be left unread and cause later calls to s2n_negotiate() to return fatal errors. Calling s2n_recv_early_data() again after it returns success is possible but has no effect on the connection.

Once a server has read all early data, you should call s2n_negotiate() to complete the handshake just as you would for a handshake that did not include early data.

For example:

uint8_t early_data[MAX_EARLY_DATA] = { 0 };
ssize_t total_data_recv = 0, data_recv = 0;
while (s2n_recv_early_data(conn, early_data + total_data_recv, MAX_EARLY_DATA - total_data_recv,
        &data_recv, &blocked) != S2N_SUCCESS) {
    total_data_recv += data_recv;
    if (s2n_error_get_type(s2n_errno) != S2N_ERR_T_BLOCKED) {
        exit(1);
    }
}
while (s2n_negotiate(conn, &blocked) != S2N_SUCCESS) {
    if (s2n_error_get_type(s2n_errno) != S2N_ERR_T_BLOCKED) {
        exit(1);
    }
}

Adding Anti-replay Protection

s2n-tls does not include anti-replay protection automatically. Effective anti-replay protection for a multi-server application requires an external state shared by all servers. Without shared state, an attacker can capture early data originally sent to server A and successfully replay it against server B.

The TLS1.3 specification suggests two possible anti-replay solutions that a user can implement:

1. Single-Use Tickets: Valid tickets are stored in a shared database and deleted after use. s2n_connection_get_negotiated_psk_identity_length() and s2n_connection_get_negotiated_psk_identity() can be used to get the ticket identifier, or "pre-shared key identity", associated with offered early data.
2. Client Hello Recording: Instead of recording outstanding valid tickets, unique values from recent ClientHellos can be stored. The client hello message can be retrieved with s2n_connection_get_client_hello() and the pre-shared key identity can be retrieved with s2n_connection_get_negotiated_psk_identity_length() and s2n_connection_get_negotiated_psk_identity(), but s2n-tls does not currently provide methods to retrieve the validated binders or the ClientHello.random.

The s2n_early_data_cb() can be used to hook an anti-replay solution into s2n-tls. The callback can be configured by using s2n_config_set_early_data_cb(). Using the s2n_offered_early_data pointer offered by the callback, s2n_offered_early_data_reject() or s2n_offered_early_data_accept() can accept or reject the client request to use early data.

An example implementation:

int s2n_early_data_cb_impl(struct s2n_connection *conn, struct s2n_offered_early_data *early_data)
{
    uint16_t identity_size = 0;
    s2n_connection_get_negotiated_psk_identity_length(conn, &identity_size);
    uint8_t *identity = malloc(identity_size);
    s2n_connection_get_negotiated_psk_identity(conn, identity, identity_size);

    if (user_verify_single_use_ticket(identity)) {
        s2n_offered_early_data_accept(early_data);
    } else {
        s2n_offered_early_data_reject(early_data);
    }

    free(identity);
    return S2N_SUCCESS;
}

The callback can also be implemented asynchronously by returning S2N_SUCCESS without either accepting or rejecting the early data. The handshake will then fail with an S2N_ERR_T_BLOCKED error type and s2n_blocked_status set to S2N_BLOCKED_ON_APPLICATION_INPUT until s2n_offered_early_data_reject() or s2n_offered_early_data_accept() is called asynchronously.

An example asynchronous implementation:

void *user_accept_or_reject_early_data(void *arg)
{
    struct s2n_offered_early_data *early_data = (struct s2n_offered_early_data *) arg;
    if (user_slowly_verify_early_data(early_data)) {
        s2n_offered_early_data_accept(early_data);
    } else {
        s2n_offered_early_data_reject(early_data);
    }
    return NULL;
}

int s2n_early_data_cb_async_impl(struct s2n_connection *conn, struct s2n_offered_early_data *early_data)
{
    pthread_t thread_id;
    pthread_create(&thread_id, NULL, user_accept_or_reject_early_data, (void *) early_data);
    return S2N_SUCCESS;
}

s2n_offered_early_data_get_context_length() and s2n_offered_early_data_get_context() can be called to examine the optional user context associated with the early data. Unlike most s2n-tls callbacks, the context is not configured when the callback is set. Instead, the context is associated with the specific pre-shared key or session ticket used for early data. The context can be set for external pre-shared keys by calling s2n_psk_set_early_data_context(). For session tickets, s2n_connection_set_server_early_data_context() can be used to set the context the server includes on its new session tickets. Because the server needs to serialize the context when creating a new session ticket, the context is a byte buffer instead of the usual void pointer.

Post Quantum (PQ) Support

s2n-tls supports post-quantum key exchange for TLS1.3. Currently, only Kyber is supported. See the draft IETF standard: https://datatracker.ietf.org/doc/html/draft-ietf-tls-hybrid-design

Specifically, s2n-tls supports hybrid key exchange. PQ hybrid key exchange involves performing both classic ECDH key exchange and post-quantum Kyber key exchange, then combining the two resultant secrets. This strategy combines the high assurance of the classical key exchange algorithms with the quantum-resistance of the new post-quantum key exchange algorithms. If one of the two algorithms is compromised, either because advances in quantum computing make the classic algorithms insecure or because cryptographers find a flaw in the relatively new post-quantum algorithms, the secret is still secure. Hybrid post-quantum key exchange is more secure than standard key exchange, but is slower and requires more processing and more network bandwidth.

Careful: if an s2n-tls server is configured to support post-quantum key exchange, the server will require that any client that advertises support ultimately uses post-quantum key exchange. That will result in a retry and an extra round trip if the client does not intially provide a post-quantum key share.

Requirements

AWS-LC

s2n-tls must be built with aws-lc to use post-quantum key exchange. See the s2n-tls build documentation for how to build with aws-lc.

If you're unsure what cryptography library s2n-tls is built against, trying running s2nd or s2nc:

> s2nd localhost 8000
libcrypto: AWS-LC
Listening on localhost:8000

Security Policy

Post-quantum key exchange is enabled by configuring a security policy (see Security Policies) that supports post-quantum key exchange algorithms.

"default_pq" is the equivalent of "default_tls13", but with PQ support. Like the other default policies, "default_pq" may change as a result of library updates. The fixed, numbered equivalent of "default_pq" is currently "20240730". For previous defaults, see the "Default Policy History" section below.

Other available PQ policies are compared in the tables below.

Chart: Security Policy Version To PQ Hybrid Key Exchange Methods

Chart: Security Policy Version To Classic Key Exchange

If the peer doesn't support a PQ hybrid key exchange method, s2n-tls will fall back to a classical option.

Chart: Security Policy Version To Ciphers

Chart: Security Policy Version To Signature Schemes

Chart: Security Policy Version To TLS Protocol Version

Default Policy History