Moved pdkim declaration to satisfy older compilers
As suggested by Dennis Davis to fix an error with gcc 2.95.2
which threw the following error:-
gcc pdkim.c
pdkim.c: In function `pdkim_feed_finish':
pdkim.c:1389: parse error before `*'
pdkim.c:1390: `hdrs' undeclared (first use in this function)
pdkim.c:1390: (Each undeclared identifier is reported only once
pdkim.c:1390: for each function it appears in.)
gmake[2]: *** [pdkim.o] Error 1
See https://lists.exim.org/lurker/message/20120524.094800.89928246.en.html
Phil Pennock [Thu, 24 May 2012 03:43:20 +0000 (23:43 -0400)]
_ISOC99_SOURCE -> _GNU_SOURCE
_ISOC99_SOURCE broke build on Linux (Ubuntu 11.10) because it broke <resolv.h>, <arpa/nameser.h>, etc.
Their u_char and u_int usage relies upon BSD source being enabled too. So use _GNU_SOURCE.
Phil Pennock [Wed, 23 May 2012 16:25:16 +0000 (12:25 -0400)]
Doc: tls_require_ciphers examples
Note how to test strings, provide examples which distinguish port 25 from other ports.
Carefully used short examples, but allows two different strings per implementation
and demonstrates how the strings are very different.
Phil Pennock [Tue, 22 May 2012 02:14:18 +0000 (22:14 -0400)]
Enable PCRE_CONFIG by default
With this, src/EDITME as Local/Makefile *only* needs EXIM_USER to be
set and EXIM_MONITOR commented out for Exim to build on my box.
I think this is a reasonable default; if there are releases of PCRE which
do not include pcre-config, then on those boxes a slight change will be
needed, but only where the file was already having to be edited anyway.
Phil Pennock [Mon, 21 May 2012 04:20:37 +0000 (00:20 -0400)]
features.h; tls_validate_require_cipher: log flag & tests
Pull in <features.h> on Linux.
Switch readconf log from D_all (bug) to D_tls (though D_any would have
worked).
Modified runtest to handle clamped DH bits and
tls_validate_require_cipher added debug logging.
Phil Pennock [Mon, 21 May 2012 02:15:48 +0000 (22:15 -0400)]
Update docs for latest state of TLS affairs.
gnutls-params bits count no longer necessarily what GnuTLS says to use.
The OpenSSL-vs-GnuTLS text needed some updating.
Catches a ChangeLog addition made during the previous commit, so not picked up by it.
Janne Snabb tracked down the GnuTLS 2.12 vs NSS (Thunderbird) interop
problems to a hard-coded limit of 2236 bits for DH in NSS while GnuTLS
was suggesting 2432 bits as normal.
Added new global option tls_dh_max_bits to clamp all DH values (client
or server); unexpanded integer. Default value to 2236. Apply to both
GnuTLS and OpenSSL (which requires tls_dh_params for this).
Tired of debugging "SMTP fails TLS" error messages in mailing-lists
caused by OpenSSL library/include clashes, and of finding out I typo'd
in tls_require_ciphers only at the STARTTLS handshake. During readconf,
fork/drop-privs/initialise-TLS-library. In that, if tls_require_ciphers
is set, then validate it.
The validation child will panic if it can't initialise or if
tls_require_ciphers can't be parsed, else it exits 0. If the child
exits anything other than 0, the main Exim process will exit.
Phil Pennock [Sun, 20 May 2012 23:35:34 +0000 (19:35 -0400)]
Guard TLS SNI callback define better.
Guarded the callback invocation on OpenSSL having TLS extension support.
Failed to guard the callback definition. Fixed.
Problem spotted by Todd Lyons.
Phil Pennock [Sat, 19 May 2012 23:13:51 +0000 (19:13 -0400)]
Portability to HP-UX.
Report and fix from Michael Haardt.
The resolver library change's assumed typedef was absent, but the
underlying struct __res_state is present. Long type issues for the
arithmetic changes.
Phil Pennock [Fri, 18 May 2012 19:52:08 +0000 (15:52 -0400)]
Second SPF fix, moved to where type is correct.
De-initialised "type" var in stack declaration, so a repeat of this mistake
would lead to an uninitialized variable usage warning which would have blocked
the previous incorrect fix from being committed.
Phil Pennock [Fri, 18 May 2012 03:04:36 +0000 (23:04 -0400)]
GnuTLS pretty much passes test suite.
Fixed assumption that tls_certificate non-NULL in server when TLS
advertised.
Weakened an !S_ISREG() to an S_ISDIR() to keep the test-suite happy.
Using:
do { rc = gnutls_handshake(state->session);
} while ((rc == GNUTLS_E_AGAIN) || (rc == GNUTLS_E_INTERRUPTED));
is contra-indicated when you expect SIGALRM to be able to break you out
of the loop. A little _too_ robust there. Switching last part to:
(rc == GNUTLS_E_INTERRUPTED && !sigalrm_seen)
is rather more productive.
Only test not passing is 2025, which makes major assumptions about
cipher suites and needs to be revisited to see what it's trying to
achieve. We fail the test because we successfully deliver the message
without expected errors, because other ciphersuites are available, since
we're no longer limited to a *very* short list embedded in the Exim
code. That sort of failure I can live with.
Phil Pennock [Thu, 17 May 2012 20:18:34 +0000 (16:18 -0400)]
More GnuTLS cleanups/fixes.
Decided "unknown (reason)" in tls_peerdn was wrong, stripped that, added
replacement guard.
Moved cipherbuf construction to where it makes more sense, where peerdn
is extracted, so that setting the exim vars gets back closer to just
some pointer switching.
Fix missing failure check after handshake in client.
Fix tls.c tls_ungetc() and friends by pointing watermark vars at state
content.
Regenerated test-suite D-H params so we don't have too small values,
which was causing connection rejections.
Test-suite output where new test cert info is logged (there will be a
couple more, when I fix a lingering problem with tls_peerdn being unset
in client log-lines).
Phil Pennock [Thu, 17 May 2012 18:05:06 +0000 (14:05 -0400)]
GnuTLS control constants exposed to Makefile.
Mostly care about EXIM_GNUTLS_LIBRARY_LOG_LEVEL for debugging.
If someone screams that we kept the default dh-bits at 1024 for old GnuTLS,
we can point them at EXIM_SERVER_DH_BITS_PRE2_12. The name itself will
tell them to shut up and update their library if they care about security. :)
Phil Pennock [Thu, 17 May 2012 15:17:20 +0000 (11:17 -0400)]
Handle absent tls_require_ciphers correctly.
Fix test-suite certs to not use MD5.
Document that we do not support MD5 certs any longer.
Make test-suite generate probably-correct gnutls-params filename for us.
Phil Pennock [Thu, 17 May 2012 06:53:44 +0000 (02:53 -0400)]
Get TLS SNI server-switching working with GnuTLS.
Registering a cert/key in an x509 credentials *adds* them, and there's
no way to remove them, so we need a shiny new x509_cred each time the
key/cert change.
Since we avoid re-expanding unless tls_sni appears in tls_certificate,
we've mostly avoided the expense unless SNI is in use, and the extra
loading should be minimal, as everything should be in buffer/cache from
a few microseconds beforehand.
This code tested with GnuTLS and OpenSSL clients, without TLS
extensions, with servername, and verifying we do now get the correct
cert.
Phil Pennock [Thu, 17 May 2012 05:32:13 +0000 (01:32 -0400)]
Guards for older releases of GnuTLS.
gnutls_sec_param_to_pk_bits() and gnutls_rnd() are both new as of
GnuTLS 2.12.x. Guard their usage on 2.12.0+ at compile time.
In older versions, the vaguely_random_number() function just immediately
calls the fallback, so it's the same as before this change (just one
extra indirection in the code-path).
Define a constant of 1024 for dh-bits for use in those old releases
where GnuTLS won't tell us how many we should use.
Change the on-disk filename for generated D-H params again, replacing
the -normal with -<bitcount>, so that it's 1024 or whatever, and as
the value changes, Exim will automatically start using the new value.
Phil Pennock [Wed, 16 May 2012 16:15:26 +0000 (12:15 -0400)]
Overhaul of GnuTLS code.
GnuTLS code re-done, using cut&paste for preservation where appropriate.
Stop using deprecated APIs. Stop hard-coding lists of ciphers.
Use gnutls_priority_init() instead.
Turns tls_require_ciphers into a string in the GnuTLS case, not just
OpenSSL case.
Deprecate three gnutls_require_* options; now ignored but not errors.
(No warnings yet).
Added TLS SNI support.
Made the channel binding integration theoretically actually work. I had
it guarded by an #ifdef but the value used was an enum instead. Oops.
Fixed.
New code much more amenable to future work permitting TLS in callouts.
DH param sizes now chosen by GnuTLS maintainers, we use "normal"; that's
suddenly a lot more bits, so the saved filename was changed too.
(GNUTLS_SEC_PARAM_NORMAL).
DH param setup only done for servers now, since clients don't need/use
it.
GnuTLS a lot more robust to library negotiation using stuff we don't
support, error-ing out quickly for other authentication systems (PGP,
etc).
Renamed pseudo_random_number() to vaguely_random_number() which makes
the nature clearer.
GnuTLS now provides a vaguely_random_number() implementation, to match
OpenSSL.
Pull in <inttypes.h> to make the recent arithmetic changes compile on
MacOS.
Nuke test 2011 which related to the gnutls_require_* options now
non-functional.
Phil Pennock [Sun, 6 May 2012 13:32:53 +0000 (06:32 -0700)]
Unbreak build on Solaris.
Solaris needs -lresolv, ie $(LIBRESOLV), for anything referencing DNS.
I added _res stuff to os.c.
os.c is used by multiple tools, but of those only Exim needs the DNS functionality.
So used a !defined(COMPILE_UTILITY) guard.