This crate contains FFI bindings for the rustls TLS library, so you can use the library in any language that supports FFI (C, C++, Python, etc). It also contains demo C programs that use those bindings to run an HTTPS server, and to make an HTTPS request.
Rustls is a modern TLS library written in Rust, meaning it is less likely to have memory safety vulnerabilities than equivalent TLS libraries written in memory unsafe languages.
If you are using rustls-ffi to replace OpenSSL, note that OpenSSL provides cryptographic primitives in addition to a TLS library. Rustls-ffi only provides the TLS library. If you use the cryptographic primitives from OpenSSL you may need to find another library to provide the cryptographic primitives.
You'll need to install the Rust toolchain (version 1.64
or above) and a C compiler (gcc and clang should both work).
In its current form rustls-ffi's Makefile infrastructure will generate a static
system library (e.g. --crate-type=staticlib), producing a .a or .lib file
(depending on the OS).
We recommend using rustls-ffi as a static library as we make no guarantees of ABI stability across versions at this time, and dynamic library support is considered experimental.
To build a static library in optimized mode:
make
To install in /usr/local/:
sudo make install
To build a static library in debug mode:
make PROFILE=debug
To link against the resulting static library, on Linux:
-lrustls -lgcc_s -lutil -lrt -lpthread -lm -ldl -lc
To link against the resulting static library, on macOS:
-lrustls -liconv -lSystem -lc -l
If the linking instructions above go out of date, you can get an up-to-date list via:
RUSTFLAGS="--print native-static-libs" cargo build
Using rustls-ffi as a static library has some downsides. Notably each application that links the static library will need to be rebuilt for each update to rustls-ffi, and duplicated copies of rustls-ffi will be included in each application.
Building rustls-ffi as a dynamic library (--crate-type=cdylib) can resolve these
issues, however this approach comes with its own trade-offs. We currently consider
this option experimental.
At this time rustls-ffi makes no guarantees about ABI stability. Each release of rustls-ffi may introduce breaking changes to the ABI and so the built library should use the exact rustls-ffi version as the dynamic library SONAME.
Since building a useful dynamic library is more complex than building a static
library, rustls-ffi uses cargo-ci in place
of the Makefile system used for the static library.
This takes care of:
- Generating the
rustls.hheader file. - Building a
.soor.dylibfile (depending on the OS). - Generating a pkg-config
.pcfile. - Installing the library and header files in the appropriate location.
If your operating system doesn't package cargo-c natively
(see package availability),
you can install it with:
cargo install cargo-c
To build a dynamic library in optimized mode:
cargo capi build --release
To install in /usr/local/:
sudo cargo capi install
To build a static library in debug mode:
cargo capi build
To link against the resulting dynamic library, use pkg-config to populate your
LDLIBS and CFLAGS as appropriate:
LDLIBS="$(pkg-config --libs rustls)"
CFLAGS="$(pkg-config --cflags rustls)"
Rustls doesn't do any I/O on its own. It provides the protocol handling, and leaves it up to the user to send and receive bytes on the network. Because of that it can be used equally well in a blocking or non-blocking I/O context. See the rustls documentation for a diagram of its input and output methods, along with a description of the TLS features it supports.
This library defines an enum, rustls_result, to indicate success or failure of
a function call. All fallible functions return a rustls_result. If a function
has other outputs, it provides them using output parameters (pointers to
caller-provided objects). For instance:
rustls_result rustls_connection_read(const rustls_connection *conn,
uint8_t *buf,
size_t count,
size_t *out_n);In this example, buf and out_n are output parameters.
For a given struct, all functions that start with the name of that struct are
either associated functions or methods of that struct. For instance,
rustls_connection_read is a method of rustls_connection. A function
that takes a pointer to a struct as the first parameter is considered a method
on that struct. Structs in this library are always created and destroyed by
library code, so the header file only gives a declaration of the structs, not
a definition.
As a result, structs are always handled using pointers. For each struct, there
is generally a function ending in _new() to create that struct. Once you've
got a pointer to a struct, it's your responsibility to (a) ensure no two
threads are concurrently mutating that struct, and (b) free that struct's
memory exactly once. Freeing a struct's memory will usually be accomplished
with a function starting with the struct's name and ending in _free().
You can tell if a method will mutate a struct by looking at the first
parameter. If it's a const*, the method is non-mutating. Otherwise, it's
mutating.
Input parameters will always be either a const pointer or a primitive type
(int, size_t, etc). Output parameters will always be a non-const pointer.
The caller is responsible for ensuring that the memory pointed to by output parameters is not being concurrently accessed by other threads. For primitive types and pointers-to-pointers this is most commonly accomplished by passing the address of a local variable on the stack that has no references elsewhere. For buffers, stack allocation is also a simple way to accomplish this, but if the buffer is allocated on heap and references to it are shared among threads, the caller will need to take additional steps to prevent concurrent access (for instance mutex locking, or single-threaded I/O).
When an output parameter is a pointer to a pointer (e.g.
rustls_connection **conn_out, the function will set its argument
to point to an appropriate object on success. The caller is considered to take
ownership of that object and must be responsible for the requirements above:
preventing concurrent mutation, and freeing it exactly once.
For a method, the first parameter will always be a pointer to the struct being operated on. Next will come some number of input parameters, then some number of output parameters.
As a minor exception to the above: When an output parameter is a byte buffer
(*uint8_t), the next parameter will always be a size_t denoting the size of
the buffer. This is considered part of the output parameters even though it is
not directly modified.
There are no in/out parameters. When an output buffer is passed, the library only writes to that buffer and does not read from it.
For fallible functions, values are only written to the output arguments if the function returns success. There are no partial successes or partial failures. Callers must check the return value before relying on the values pointed to by output arguments.
Rustls supports various types of user customization via callbacks. All callbacks
take a void *userdata parameter as their first parameter. Unless otherwise
specified, this will receive a value that was associated with a
rustls_connection via rustls_connection_set_userdata. If no such value was
set, they will receive NULL. The read and write callbacks are a particular
exception to this rule - they receive a userdata value passed through from the
current call to rustls_connection_{read,write}_tls.
The library checks all pointers in arguments for NULL and will return an error
rather than dereferencing a NULL pointer. For some methods that are infallible
except for the possibility of NULL (for instance
rustls_connection_is_handshaking), the library returns a convenient
type (e.g. bool) and uses a suitable fallback value if an input is NULL.
In case of a bug (e.g. exceeding the bounds of an array), Rust code may
emit a panic. Panics are treated like exceptions in C++, unwinding the stack.
Unwinding past the FFI boundary is undefined behavior, so this library catches
all unwinds and turns them into RUSTLS_RESULT_PANIC (when the function is
fallible).
Functions that are theoretically infallible don't return rustls_result, so we
can't return RUSTLS_RESULT_PANIC. In those cases, if there's a panic, we'll
return a default value suitable to the return type: NULL for pointer types,
false for bool types, and 0 for integer types.
Several features of the C bindings are marked as EXPERIMENTAL as they are
need further evaluation and will most likely change significantly in the future.
The rustls_server_config_builder_set_hello_callback and its provided information
in rustls_client_hello will change. The current design is a snapshot of the
implementation efforts in
mod_tls to provide
rustls-based TLS as module for the Apache webserver.
For a webserver hosting multiple domains on the same endpoint, it is highly desirable to have individual TLS settings, depending on the domain the client wants to talk to. Most domains have their own TLS certificates, some have configured restrictions on other features as well, such as TLS protocol versions, ciphers or client authentication.
The approach to this taken with the current rustls_client_hello is as follows:
If you have a single site and one certificate, you can preconfigure the
rustls_server_config accordingly and do not need to register any callback.
If you need to support multiple rustls_server_configs on the same connection
endpoint, you can start the connection with a default rustls_server_config
and register a client hello callback. The callback inspects the SNI/ALPN/cipher
values announced by the client and selects the appropriate configuration
to use.
When your callback returns, the handshake of rustls will fail, as no
certificate was configured. This will be noticeable as an error returned
from rustls_connection_write_tls(). You can then free this connection and
create the one with the correct setting for the domain chosen.
For this to work, your connection needs to buffer the initial data from the client, so these bytes can be replayed to the second connection you use. Do not write any data back to the client while you are in the initial connection. The client hellos are usually only a few hundred bytes.
By default, rustls does not load any trust anchors (root certificates), not even the system trust anchor store, which means that TLS certificate verification will fail by default. You are responsible for loading certificates using one of the following methods:
-
rustls_root_cert_store_add_pem, which adds a single certificate to a root store. -
rustls_client_config_builder_load_roots_from_file, which loads certificates from a file. -
A custom method for finding certificates where they are stored and then added to the rustls root store.