Fixes crash introduced by 0ab6dbbc65. It's
probably a regression since it defeats a system designed to stop this
from happening, but I didn't dig through the history.
rehash() closes listeners. If we happen to get a single epoll() result
that wants to first rehash and then accept a connection, the epoll info
will point to a freed rb_fde_t. Other selectors should have similar
problems, but we didn't investigate that.
rb_fde_ts are normally batched up and freed outside the event
processing, but as of the above commit close_listeners() screws that up
by closing pending FDs immediately in order to create new listeners.
I think it might be a bit better to revert this behaviour and simply not
close listeners if we are going to open new ones over them, but have
opted for the smallest reasonable change I can think of.
Helped-by: Eric Mertens <emertens@gmail.com>
Edited by @aaronmdjones:
- Correct some data types and casts
- Minor style fixups (e.g. we put * on the variable name not the type)
- librb/src/openssl.c:
- Defer call of BIO_free(3ssl) to the end of the conditional block
to avoid having calls to it in multiple paths
- Check the return value of SSL_CTX_set0_tmp_dh_pkey(3ssl) because if
it fails then we must use EVP_PKEY_free(3ssl) to avoid a memory leak
This could fail if, for example, the user supplied DSA parameters
in the DH parameters file instead.
- ircd/newconf.c:
- Check whether OSSL_DECODER_CTX_new_for_pkey(3ssl) was able to parse
the given CHALLANGE public key as a valid RSA public key, and then
check whether OSSL_DECODER_from_bio(3ssl) actually loads it
successfully
- ircd/s_newconf.c:
- Use EVP_PKEY_free(3ssl) instead of OPENSSL_free(3ssl) on EVP_PKEY
pointers; this will avoid inadvertent memory leaks if the EVP_PKEY
structure contains any dynamically-allocated child members
- modules/m_challenge.c:
- Unconditionally use EVP(3ssl) to generate the SHA-1 digest of the
random challenge; this API has been around for a very long time and
is available in all supported versions of OpenSSL
- Add lots of error checking to all steps of the process
Tested against 1.1.1 and 3.0; both with missing and provided DH parameters
(which works as you'd expect; the server will not negotiate a DHE cipher
without them), and CHALLENGE, including missing keys or keys of the wrong
type (e.g. when you supply an EdDSA key instead of an RSA key).
This does break compatibility with OpenSSL 1.1.0 and below, which are now
all end-of-life and unsupported anyway.
Closes#357
This code is doing (foo - (char*)0) to convert foo from a pointer
value into a numeric value. Unfortunately, this is undefined
behaviour, which clang-14 is now warning about [1].
Cast to uintptr_t instead. Same result, but well-defined.
[1] cf. commit 0302f1532b
Unlike Linux, Solaris, and Illumos (and probably others), the 2 BSDs that still
support SCTP didn't put SCTP into its own library, they put it into libc.
They, unlike Linux, don't set SOL_SCTP for us. The official method appears to
be calling getprotobyname("sctp") & endprotoent(), with getprotobyname()
returning a struct that has a p_proto entry. This all reads from
/etc/protocols. However, SCTP is assigned 132 by IANA, so it's 132 everywhere,
so I just set SOL_SCTP to 132 if it's not already set.
Sadly, this just sends it to purgatory. It's still around for the
socketpair() emulation and the nanosleep() emulation.
socketpair() obviously only selects() on 2 FDs, so not a huge deal.
nanosleep() only uses it for the timeout, so also not a huge deal.
socketpair() is SUSv3 (2001) and nanosleep() is SUSv2 (1997), so maybe
it's OK to remove those emulations. If so, then we can also remove the
sys/select.h check in configure.ac
The OpenSSL developers decided, during the OpenSSL 1.1.1 development
phase, to use a different API and different set of lists for TLSv1.3
ciphersuites, than for every TLS version preceeding it.
This is stupid, but we have to work with it.
This commit also improves configuration fault resilience. The reason
is that if you don't pass any valid old-style ciphersuites, OpenSSL
will not negotiate an older protocol at all. However, when they
implemented the new API, they decided that lack of any valid
ciphersuites should result in using the defaults. This means that if
you pass a completely invalid ciphersuite list (like "foo"), OR if
you pass a TLSv1.2-only ciphersuite list, TLSv1.3 continues to work.
This is not mirrored; passing a TLSv1.3-only ciphersuite list will
break TLSv1.2 and below.
Therefore we work around this lack of mirroring by falling back to
the default list for each protocol. This means that if
ssl_cipher_list is complete garbage, the default will be used, and
TLS setup will succeed for both protocols. This is logged, so that
administrators can fix their configuration.
I prefer this approach over explicitly disabling the protocols if
their respective ciphersuite lists are invalid, because it will
result in unusable TLSv1.3 if people run newer solanum with their
older charybdis/solanum configuration files that contain custom
ssl_cipher_list definitions. Hindering TLSv1.3 adoption is not an
option, in my opinion.
The downside of this is that it is no longer possible to disable a
protocol family by not including any of its ciphersuites. This could
be remedied by an ssl_protocol_list configuration directive if it is
decided that this functionality is ultimately necessary.
This work is not required for either of the other TLS backends,
because neither of those libraries yet support TLSv1.3, and in the
event that they eventually do, I expect them to allow configuration
of newer ciphersuites with the existing APIs. This can be revisited
if it turns out not to be the case.
Signed-off-by: Aaron Jones <me@aaronmdjones.net>
Tested-by: Aaron Jones <me@aaronmdjones.net>
This only supports two addresses as the intended use is 1 IPv4 and 1 IPv6
address on a single-homed host, and the only supported configuration of
outgoing connections to other servers is to bind a single IPv4 or IPv6
address.
- Add (void) casts for unused function parameters
- Rearrange member in `struct rb_mbedtls_cfg_context' for data alignment
- Document a `clang-4.0 -Weverything' (-Wcast-qual) diagnostic
- Avoid pointless conversions between positive/negative error codes
- Use capital hexadecimals in error codes and properly cast to
(unsigned int) for %x/%X
