This avoids a compiler warning regarding casting a file descriptor to a
pointer (as input to gnutls_transport_set_ptr()), and also ensures that
the pointer is valid for the lifetime of the session.
If gnutls_bye() fails with a fatal error, we would reattempt it again
and again, even though this may then go on to e.g. cause a segmentation
fault.
Now we just keep retrying if it was interrupted, in line with the other
backends, up to a maximum of 3 retries.
Yeah, I know, I said I was happy with it and wouldn't be changing it.
However, the new GNUTLS backend I'm working on has prompted this.
E.g. MbedTLS error codes and GNUTLS error codes are both negative ints,
but GNUTLS will not tolerate positive input values. Let's treat
both backends the same.
* Move file/data assignments to the top of the function
* Don't attempt to set a hardcoded P-384 ECDH key if we have the new
SSL_CTX_set1_curves_list() function (OpenSSL 1.0.2+)
* Rename variables consistent with other backends and wrap the function
arguments.
* Disable OpenSSL's disabling of TLS 1/n-1 record splitting.
In other words, enable TLS 1/n-1 record splitting.
* Other misc cleanups.
OpenSSL is perfectly capable of having a single context that is shared
by both client and server sessions alike; one simply needs to call
SSL_set_accept_state (for server) or SSL_set_connect_state (for client)
before attempting handshaking.
OpenSSL uses `unsigned long' type for its error codes, so that's
what (lib)ratbox used to store the error values.
Unfortunately, MbedTLS uses int, and its error codes are negative.
On machines where `int' and `long' are the same size, this could
result in storing a truncated error code.
This patch inverts the sign bit on error codes and then casts them
to unsigned long for storage.
MbedTLS itself (specifically, `mbedtls_strerror()') will function
properly with negative or positive input values. It even converts
negative input values to positive before checking them against the
list of known error codes!
See also: `library/error.c' in the MbedTLS 2.1+ distribution.
All 3 backends (MbedTLS, OpenSSL, GNUTLS) are going to have the same
function name for returning error strings. This will help to reduce the
diffs between them.
* Add generic direction enum for negotiation setup.
* Rename a rather long wrapper function to a shorter one consistent with
what it does.
* Rework context setup function.
* Don't check for handshake state before beginning handshaking.
The old backend began a handshake and then stepped into the callback
function if it was interrupted; the current one just jumps right into
it, so there is no need to check if it has previously succeeded,
because it hasn't been attempted yet.
* Add missing errno assignment to one of the handshake wrappers.
* Don't bother checking if SSL_P(F) is NULL when we already checked if
F->ssl is NULL -- this should be impossible.
* Don't bother checking if SSL_C(F) is NULL -- this was a no-op.
* Change the socket send and recv functions to not peer into a foreign
ratbox structure -- use the correct function to get the socket fd.
* Rewrap some lines and function arguments.
Other backends will be brought into line with this backend soon.
This will enable easier maintenance of the backends, by reducing the
diffs between them, which should make different behaviour easier to
spot.
At the moment, if a link quits in just the right (wrong [1]) way,
the quit reason will resemble:
<-- foo (~bar@baz) has quit (Read error: (-0x0) )
This should resolve that.
[1] Peers should send a close_notify alert before abruptly shutting
down their socket. This will result in a sane quit message:
<-- foo (~bar@baz) has quit (Read error: (-0x7880) SSL -
The peer notified us that the connection is going to be closed)
[ci skip]
As a client, require all peers (i.e. other IRC servers) to support secure
renegotiation. Break handshakes with servers that don't. We do not
renegotiate our sessions, but this is the most secure option regardless.
As a client, disable TLS Session Tickets. The server side MbedTLS code
does not have any ticket callbacks configured, so an MbedTLS IRC Server
will not issue tickets -- however, others could. Server connections are
not expected to be short-lived enough to benefit from the usage of tickets,
and their issuance harms forward secrecy.
I have removed all non-MbedTLS-specific code from this backend and
copied the non-OpenSSL-specific portions of the OpenSSL backend code
over; as it seems to be more reliable.
If initialising the server context fails, but the client one succeeds,
we will not only leak memory, but the error message reported for
initialising the server context might not make sense, because we
initialise the client context after and that could erase or change the
list of queued errors.
This scenario is considered rare. Nevertheless, we now initialise the
client context after *successfully* initialising the server context.
Commit 5c8da48 introduced a fix for issue #186 by freeing the old SSL_CTX
structure before constructing a new one, which could disconnect existing
clients otherwise.
Unfortunately, the freeing is done first, which means that if setting up
a new structure fails for any reason, there will be no usable structures
left, but they are still referenced.
This fix moves the freeing to the end of the function, using intermediate
new variables in the meantime. This problem was discovered while testing
against OpenSSL 1.1.0 RC6.
LibreSSL does not have the new version macros & functions that OpenSSL
1.1.0 implements. This causes a compile-time failure against LibreSSL.
Further, the runtime function for returning the library version returns
the wrong number (the hardcoded constant number SSLEAY_VERSION_NUMBER
aka OPENSSL_VERSION_NUMBER, instead of LIBRESSL_VERSION_NUMBER).
Add more ifdef soup to remedy the situation.
The code already assumes the presence of fopen(3) and errno, and, by
extension, fclose(3) and strerror(3), so just use those instead of the
BIO wrappers.
Additionally, don't fail to initialise if the DH file does exist but
parsing it fails, as per the pre-existing comment about them being
optional.
* Move certificate, key, DH parameters and configuration to heap
(Documentation states that setting new configuration, e.g.
during a rehash, is unsupported while connections using that
configuration are active)
This is the same approach as the fix for #186
Refcount these structures so as to not introduce a memory leak
On rehash, it will use new structures only if there are no
errors in constructing them
* Add better error-reporting (strings in addition to numbers)
where possible
* Coalesce several connection memory allocations into one function
* Reduce boilerplate where possible (Charybdis targets C99)
* Support private key being in certificate file, and having no
DH parameters file
* Correct erroneous closing comment
openssl:
* Don't manually initialise libssl 1.1.0 -- it does this automatically
* SSL_library_init() should be called first otherwise
* Move SSL_CTX construction to rb_setup_ssl_server()
* Test for all required files (certificate & key) before doing anything
* Free the old CTX before constructing a new one (Fixes#186)
* Properly abort rb_setup_ssl_server() on CTX construction failures
* Support ECDHE on more than one curve on OpenSSL 1.0.2 and above
* Clean up ifdef indentation
* Fix DH parameters memory leak
mbedtls:
* Fix certificate fingerprint generation
* Fix library linking order
* Fix incorrect printf()-esque argument count
* Return digest length for fingerprints instead of 1, consistent
with the other backends
sslproc / ssld:
* Fingerprint methods have no assocated file descriptors
* Send TLS information (cipher, fingerprint) before data
* Use correct header length for fingerprint method
Authored-by: Aaron Jones <aaronmdjones@gmail.com>
Authored-by: William Pitcock <nenolod@dereferenced.org>
Authored-by: Simon Arlott <sa.me.uk>
