A Common Lisp interface to OpenSSL / LibreSSL.
Distinguishing features: CL+SSL is portable code based on CFFI and gray streams. It defines its own libssl BIO_METHOD, so that TLS I/O can be written over portable Lisp streams instead of bypassing the streams and giving OpenSSL a socket file descriptor to send data over. (But the file descriptor approach is still used if possible.)
License: MIT-style.
The library is available via Quicklisp.
The Git repository: https://github.com/cl-plus-ssl/cl-plus-ssl.
Send bug reports to the GitHub issue tracker. The old mailing list [email protected] is also still available (list information).
Usually OpenSSL / LibreSSL shared libraries are provided by your package manager and very likely are already installed.
https://wiki.openssl.org/index.php/Binaries lists several sources of binary distributions. For example, https://www.slproweb.com/products/Win32OpenSSL.html (slproweb.com is a 3rd party; if you have questions about the OpenSSL installer they provide, please ask in the mailing list specified on the linked page).
If you chose to install the DLLs into the OpenSSL installation's "bin" directory (recommended), then be sure to add the bin directory to your PATH environment variable and restart your session. e.g. "C:\Program Files\OpenSSL-Win64\bin"
Browable package definitions - click on a symbol to navigate to the definition and read docstring:
- https://cl-plus-ssl.github.io/cl-plus-ssl/package.html
- https://cl-plus-ssl.github.io/cl-plus-ssl/config.html
A single HTML page dump of all exported symbols with their docstrings and parameters: https://cl-plus-ssl.github.io/cl-plus-ssl/cl-plus-ssl-api.html
Basically, after TCP connection is created, we wrap the TCP socket stream
into a TLS encrypted stream using cl+ssl:make-ssl-client-stream
,
or cl+ssl:make-ssl-server-stream
. See how it's done in the
examples/example.lisp. That's a
self-contained file, you can load it or copy-paste into your
Slime session and try the examples as suggested in the comments at the
top of the file.
For more comfortable use learn some of OpenSSL API. In particular that SSL object represents a TLS session, SSL_CTX object is a context multiple SSL objects can derive from thus sharing common parameters. BIO is a stream-like input/output abstraction OpenSSL uses for actual data transfer. That OpenSSL API uses reference counting; therefore, for example, if an SSL object is derived from an SSL_CTX, we can apply SSL_CTX_free to the SSL_CTX when the SSL is still being used - our call will just decrement the reference count so if the SSL holds another reference to the SSL_CTX, the SSL_CTX will remain alive and will only be destroyed after the SSL releases this reference.
Knowing OpenSSL will also allow for more flexibility and control, as cl+ssl high-level functions do not cover all possible approaches.
OpenSSL comes with several BIO implementations, like file BIO, socket BIO, memory BIO, etc. Also OpenSSL API allows user to create custom BIO methods by providing a number of callbacks.
cl+ssl uses either socket BIO, or a custom BIO that implements
all input / output with Lisp functions like cl:write-byte
,
cl:read-byte
.
When a Lisp stream is passed to cl+ssl:make-ssl-client-stream
or cl+ssl:make-ssl-server-stream
, the choice of BIO is made
based on the :unwrap-stream-p
parameter.
If :unwrap-stream-p
is true, a socket file descriptor is extracted
from the Lisp stream and passed to OpenSSL using the SSL_set_fd
OpenSSL function, which results in socket BIO created for
that SSL session.
If :unwrap-stream-p
is false, a Lisp BIO is created and
passed to OpenSSL with the SSL_set_bio
OpenSSL funcion.
The default value of :unwrap-stream-p
is special variable
cl+ssl:*default-unwrap-stream-p*
which is initialized to t
,
meaning socket BIO is used by default.
This allows to dynamically change the mode of operation of the
code that omits the :unwrap-stream-p
parameter.
For the test-https-client
function from the example.lisp:
;; use socket BIO
(let ((cl+ssl:*default-unwrap-stream-p* t))
(tls-example::test-https-client "www.google.com"))
;; use Lisp BIO
(let ((cl+ssl:*default-unwrap-stream-p* nil))
(tls-example::test-https-client "www.google.com"))
If cl+ssl:make-ssl-*-stream
functions receive
a file descriptor instead of a Lisp stream,
they unconditionally use socket BIO.
By default cl+ssl searches for OpenSSL shared libraries in platform-dependent default locations.
To explicitly specify what to load use cl+ssl/config module before loading cl+ssl:
(ql:quickload :cl+ssl/config)
(cl+ssl/config:define-libssl-path "/opt/local/lib/libssl.dylib")
(cl+ssl/config:define-libcrypto-path "/opt/local/lib/libcrypto.dylib")
(ql:quickload :cl+ssl)
Note, the path
parameter of those two macros is not evaluated.
You can only use literal values. This is dictated by CFFI.
In network communications it is desirable to be able to interrupt reading or writing if their duration exceeds some limit.
For example, when making an HTTP request, if it is stuck, we want to interrupt it instead of leaving the thread hanging forever. Similarly, a server can encourage some very slow or stuck clients, who do not timely send their requests or read responses, and keep connection, and possibly a connection handling thread, occupied. We want to interrupt server handling for such clients.
In the description below, by deadline we mean an absolute time, by which all IO operations have to be completed. And by timeout we mean longest duration every individual IO operation should resume in (or at least have some progress).
For example:
(json-parser:read-object gzipped-chunked-tls-stream)
Such a call can internally perform many primitive IO
operations like cl:read-byte
or cl:read-sequence
.
If executed with IO timeouts configured, every individual
operation can take less then the the timeout, but overall
duration may still be unpredictable. In contrast, when
executed with deadline configured, the overall duration
is constrained by the deadline.
Some lisp implementations have implementation-specific ways to set timeouts or deadlines for socket communications.
For example, on SBCL:
(with-deadline (:seconds 3)
(write-string "something" socket-stream)
(read-line socket-stream)
...
)
On CCL:
(setf (ccl:stream-deadline socket-stream)
(+ (get-internal-real-time)
(* 3 internal-time-units-per-second))
(write-string "something" socket-stream)
(read-line socket-stream)
...
(setf (ccl:stream-deadline socket-stream) nil)
;; or
(setf (ccl:stream-intput-timeout socket-stream) 3
(ccl:stream-output-timeout socket-stream) 3)
(write-string "somemthig" socket-stream)
(read-line socket-stream)
...
In the examples, if time runs out, the current IO operation will be interrupted by signalling an implementation-specific error.
When using cl+ssl with LispBIO, such implementation-specific features remain working. You configure timeout / deadline in implementation-specific way (through the stream as on CCL or globally as on SBCL) and LispBIO, when performing standard CL IO functions, receives an error.
BIO captures the error condition, and returns an error status to OpenSSL. Also BIO saves the error info in OpenSSL error queue (tries at least).
OpenSSL returns an error status to cl+ssl stream code,
which signals an error condition (a subtype of cl+ssl::ssl-error
)
to the application code.
Application
(cl:read-line ssl-stream) ^ error
------------------------- |
-->
cl+ssl::ssl-stream
^ return -1
------------------------- |
-->
OpenSSL native code
^ return -1
------------------------- |
-->
Lisp BIO
^ ccl:deadline-timeout
-------------------------- |
-->
(cl:read-sequence socket-stream)
Note, in the original design of cl+ssl the error signaled in the BIO code by standard Lisp IO function was not captured by BIO, but instead it was passed through to the application code, so the app saw the implementation specific condition.
Application ^ ccl:deadline-timeout
--------------------- |
cl+ssl stream |
--------------------- |
OpenSSL native code |
--------------------- |
Lisp BIO -->
But in this case, the non-local exit goes across C call stack of the OpenSSL native code, likely preventing proper stack cleanup. CFFI manual explicitly advises against that.
That's why BIO handles the timeout error.
For the future we consider changing the behavior to save the original timeout error condition, let the C stack unwind normally through error codes, but then resignaling the the saved error, instead of a cl+ssl defined condition. That will be closer to the old cl+ssl behavior. Although not sure if that's really the best approach.
In any case, expect an error to be signaled if your implementation can arrange for timeout / deadline error on the socket stream.
Note, such implementation extensions do not always work reliably, for example https://groups.google.com/g/sbcl-devel/c/-eLw-Wv3Prc.
In case of socket BIO, cl+ssl emulates on CCL and SBCL
the behavior of plain socket streams on these implementations:
it will signal ccl::communication-deadline-expired
or
sb-sys:deadline-timeout
. (This is implemented by setting
a non-block flag on the socket file descriptor so that
SSL_Read / SSL_Write only transfer as many data as
possible without blocking. Then cl+ssl waits
for file descriptor to be ready for more IO,
making sure the waiting is no longer than the implementation
specific deadline. In case of CCL the deadline is taken from
the stream originally passed to cl+ssl:make-client / -server-stream
,
from which the socket file descriptor was extracted. In
case of SBCL the waiting operation used is aware of the global
deadline).
For other Lisp implementations deadlines are not currently supported in the socket BIO mode.
There are tickets / pull requests open to implement deadlines or timeouts for socket BIO in all implementations: #146 #69
One more idea for timeout implementation is to set socket option SO_RCVTIMEO and SO_SNDTIMEO for the TCP socket. OpenSSL will probably respect those options. But that is not tested. (Even if supported by OpenSSL, we may have problems, for example, because on several Lisps the socket file descriptor is used in non-blocking mode; if OpenSSL functions signall the sockopt timeout by returning SSL_ERROR_WANT_READ / SSL_ERROR_WANT_WRITE, cl+ssl code will simply repeat the function call.)
If you save your application as a Lisp image, call (cl+ssl:reload)
when that image is loaded to make sure necessary re-initialization
is performed.
This should work fine if the location and version of the OpenSSL shared libraries have not changed. If they have changed, you may get errors, as users report: #167
CL+SSL requires CFFI with callback support.
CL Test Grid results: https://common-lisp.net/project/cl-test-grid/library/cl+ssl.html
- session caching (what about it?)
- The FFI code for all platforms except clisp needs to be rewritten. (update 2017-07-05: does it? why?)
This library is a fork of SSL-CMUCL. The original SSL-CMUCL source code was written by Eric Marsden and includes contributions by Jochen Schmidt.
Jochen Schmidt also has his own portable CL-SSL bindings (Gray streams based), available as a part of the acl-compat portability layer of his https://portableaserve.sourceforge.net/.
Development into CL+SSL was done by David Lichteblau. After that many people contributed patches, as can be seen in the git history.
2017-07-03
- Hostname verification added, thanks to Ilya Khaprov. Default mode for make-ssl-client-stream is to verify the connection. New keywrd argument verify is added to make-ssl-client-stream with the same possible values as Drakma uses for http request verification.
201?-??-??
- See commits.
2011-05-22
- Added new public function RANDOM-BYTES.
2011-05-22
- The source code repository is moved to Git.
2011-03-25
- OpenSSL libraries names for OpenBSD, thanks to Thomas de Grivel.
2010-05-26
- Fixed two bugs in LISTEN, thanks to Ron Garret.
2009-09-17
- libssl loading on FreeBSD 7.2 fixed, thanks to Stian Sletner.
2008-xx-yy
- Support for I/O deadlines (Clozure CL and SBCL).
- Support for encrypted keys, thanks to Vsevolod Dyomkin.
- Chained certificates support, thanks to Juhani Ränkimies.
- More secure initialization of OpenSSL random number generator.
- Minor CLISP-specific fixes.
2007-xx-yy
- Fixed windows support, thanks to Matthew Kennedy and Anton Vodonosov.
2007-07-07
- Improved CLISP support, thanks to Pixel // pinterface, as well as client certificate support.
- Re-introduced support for direct access to file descriptors as an optimization. New function stream-fd. New keyword argument close-callback.
2007-01-16: CL+SSL is now available under an MIT-style license.