Daiki Ueno [Wed, 12 Aug 2020 05:27:17 +0000 (07:27 +0200)]
cert-session: fail hard if mandatory stapling is not honored
According to the documentation, the GNUTLS_CERT_INVALID flag must
always be set in case of verification failure, together with the flag
indicating the actual error cause.
Petr Pavlu [Wed, 8 Jul 2020 08:12:30 +0000 (10:12 +0200)]
pubkey: avoid spurious audit messages from _gnutls_pubkey_compatible_with_sig()
When checking in _gnutls_pubkey_compatible_with_sig() whether a public
key is compatible with a signature algorithm, run first
pubkey_supports_sig() before performing weaker checks that can accept
the given algorithm but with an audit-log warning. This avoids an issue
when a weaker check would log an audit message for some signature
algorithm that would then be determined as incompatible by the
pubkey_supports_sig() check anyway.
For instance, a GnuTLS server might have a certificate with a SECP384R1
public key and a client can report that it supports
ECDSA-SECP256R1-SHA256 and ECDSA-SECP384R1-SHA384. In such a case, the
GnuTLS server will eventually find that it must use
ECDSA-SECP384R1-SHA384 with this public key. However, the code would
first run _gnutls_pubkey_compatible_with_sig() to check if SECP384R1 is
compatible with ECDSA-SECP256R1-SHA256. The function would report the
audit warning "The hash size used in signature (32) is less than the
expected (48)" but then reject the signature algorithm in
pubkey_supports_sig() as incompatible because it has a different curve.
Since the algorithm gets rejected it is not necessary to inform about
its hash size difference in the audit log.
SP800-56A rev3 section 5.7.1.2 step 2 mandates that the validity of
the calculated shared secret is verified before the data is returned
to the caller. This patch adds the validation check.
SP800-56A rev3 section 5.7.1.1 step 2 mandates that the validity of the
calculated shared secret is verified before the data is returned to the
caller. This patch adds the validation check.
Split up system-override-sig-hash.sh
so that the errors won't get swallowed or conflated.
Also correct unused `srcdir` to `builddir`,
which I believe was meant to be set there.
Signed-off-by: Alexander Sosedkin <asosedkin@redhat.com>
Steve Lhomme [Mon, 22 Jun 2020 07:09:05 +0000 (09:09 +0200)]
fix connectx not available on older macOS SDK
Fixes this compilation error:
system/fastopen.c:134:9: error: 'connectx' is only available on macOS 10.11 or newer [-Werror,-Wunguarded-availability]
ret = connectx(fd, &endpoints, SAE_ASSOCID_ANY, CONNECT_RESUME_ON_READ_WRITE | CONNECT_DATA_IDEMPOTENT, NULL, 0, NULL, NULL);
^~~~~~~~
/Applications/Xcode9.2.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.13.sdk/usr/include/sys/socket.h:713:5: note: 'connectx' has been marked as being introduced in macOS 10.11 here, but the deployment target is macOS 10.7.0
The detection is the same as found in curl [1].
If HAVE_BUILTIN_AVAILABLE is not available we fallback to the code without
TCP_FASTOPEN_OSX.
The OS values match exactly the values found in
https://opensource.apple.com/source/xnu/xnu-4570.41.2/bsd/sys/socket.h
James Bottomley [Sun, 28 Jun 2020 19:33:09 +0000 (21:33 +0200)]
build: use $(LIBPTHREAD) rather than non-existent $(LTLIBPTHREAD)
On a very recent openSUSE build, libgnutls is getting built without
libpthread. This caused a thread related error when trying to load a
pkcs11 module that uses threading. The reason is rather convoluted:
glibc actually controls all the pthread_ function calls, but it
returns success without doing anything unless -lpthread is in the link
list. What's happening is that gnutls_system_mutex_init() is being
called on _gnutls_pkcs11_mutex before library pthreading is
initialized, so the pthread_mutex_init ends up being a nop. Then, when
the pkcs11 module is loaded, pthreads get initialized and the call to
pthread_mutex_lock is real, but errors out on the uninitialized mutex.
The problem seems to be that nothing in the gnulib macros gnutls
relies on for threading support detection actually sets LTLIBPTHREAD,
they only set LIBPTHREAD. The fix is to use LIBPTHREAD in
lib/Makefile.in
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Daiki Ueno [Fri, 26 Jun 2020 08:21:26 +0000 (10:21 +0200)]
dhe: check if DH params in SKE match the FIPS approved algorithms
SP800-56A rev. 3 restricts the FIPS compliant clients to use only
approved DH parameters, defined in RFC 7919 and RFC 3526. This adds a
check in the handling of ServerKeyExchange if DHE is negotiated.
Fiona Klute [Tue, 16 Jun 2020 19:23:14 +0000 (21:23 +0200)]
Update tlsfuzzer to get Python interpreter detection
Tlsfuzzer also assumed the Python interpreter would be called
"python", this update is necessary to get a fixed version (see
https://github.com/tomato42/tlsfuzzer/pull/671).
Fiona Klute [Sun, 14 Jun 2020 10:52:46 +0000 (12:52 +0200)]
Wipe session ticket keys before releasing the session structure
This includes both a copy of the master key and one or two derived
keys, all of which could be used to decrypt session tickets if
stolen. The derived keys could only be used for tickets issued within
a certain time frame (by default several hours).
The documentation for gnutls_session_ticket_enable_server() already
states that the master key should be wiped before releasing it, and
the same should apply to internal copies.
Daiki Ueno [Tue, 9 Jun 2020 08:41:18 +0000 (10:41 +0200)]
tests: check_for_datefudge: don't exit the test programs
This makes check_for_datefudge not to immediately exit the program,
but to return non-zero to allow the tests by themselves to control the
behavior when "datefudge" is not found.
Vitezslav Cizek [Tue, 9 Jun 2020 11:54:04 +0000 (13:54 +0200)]
configure: improve nettle, gmp, and hogweed soname detection
Some linkers might optimize away the libraries passed on the
command line if they aren't actually needed, such as gnu ld with
--as-needed.
The ldd output then won't list the shared libraries and the
detection will fail.
Make sure nettle and others are really used.
Daiki Ueno [Mon, 8 Jun 2020 04:45:24 +0000 (06:45 +0200)]
configure.ac: prefer the latest version of build infrastructure
AM_GNU_GETTEXT_REQUIRE_VERSION tells autopoint to copy the latest
possible build infrastructure installed on the system, rather than the
fixed version from the archive.dir.tar.xz. This makes the
bootstrapping slightly faster and allows us not to stick with the
ancient gettext version.
Steve Lhomme [Sat, 6 Jun 2020 12:05:16 +0000 (14:05 +0200)]
Merge the extra libraries to link dynamically in GNUTLS_LIBS_PRIVATE
This should fix #1020 where bcrypt is missing from thirdparty_libadd.
Ultimately it would be good to add libraries that always need to be linked in
one variable that is shared between the Makefile and the pkg-config file.
Daiki Ueno [Fri, 5 Jun 2020 14:26:33 +0000 (16:26 +0200)]
crypto-api: always allocate memory when serializing iovec_t
The AEAD iov interface falls back to serializing the input buffers if
the low-level cipher doesn't support scatter/gather encryption.
However, there was a bug in the functions used for the serialization,
which causes memory leaks under a certain condition (i.e. the number
of input buffers is 1).
This patch makes the logic of the functions simpler, by removing a
micro-optimization that tries to minimize the number of calls to
malloc/free.
The original problem was reported by Marius Steffen in:
https://bugzilla.samba.org/show_bug.cgi?id=14399
and the cause was investigated by Alexander Haase in:
https://gitlab.com/gnutls/gnutls/-/merge_requests/1277
Daiki Ueno [Tue, 2 Jun 2020 18:53:11 +0000 (20:53 +0200)]
stek: differentiate initial state from valid time window of TOTP
There was a confusion in the TOTP implementation in stek.c. When the
mechanism is initialized at the first time, it records the timestamp
but doesn't initialize the key. This removes the timestamp recording
at the initialization phase, so the key is properly set later.
Daiki Ueno [Tue, 2 Jun 2020 03:38:28 +0000 (05:38 +0200)]
gnutls_cipher_init: fix potential memleak
Upon failure this function returns without freeing memory allocated
internally. This makes sure that it is released and do not touch the
output handle argument.
Daiki Ueno [Mon, 1 Jun 2020 15:23:59 +0000 (17:23 +0200)]
serv: omit upper bound of --maxearlydata option definition
It turned out that AutoGen treats numbers that exceed INT_MAX in a
platform dependent way. In this case, 4294967295 (UINT_MAX) is
treated as is on 64-bit platforms, while it is interpreted as "-1" on
32-bit platforms. This causes a problem when the program
documentation is compiled under multilib environment.
Reported by Ivan Molodetskikh in:
https://bugzilla.redhat.com/show_bug.cgi?id=1841844
and the cause was identified by Anderson Toshiyuki Sasaki.
Daiki Ueno [Sun, 31 May 2020 11:59:53 +0000 (13:59 +0200)]
x509: trigger fallback verification path when cert is expired
gnutls_x509_trust_list_verify_crt2 use the macro SIGNER_OLD_OR_UNKNOWN
to trigger the fallback verification path if the signer of the last
certificate is not in the trust store. Previously, it doesn't take
into account of the condition where the certificate is expired.
Daiki Ueno [Sun, 31 May 2020 10:39:14 +0000 (12:39 +0200)]
_gnutls_pkcs11_verify_crt_status: check validity against system cert
To verify a certificate chain, this function replaces known
certificates with the ones in the system trust store if possible.
However, if it is found, the function checks the validity of the
original certificate rather than the certificate found in the trust
store. That reveals a problem in a scenario that (1) a certificate is
signed by multiple issuers and (2) one of the issuers' certificate has
expired and included in the input chain.
This patch makes it a little robuster by actually retrieving the
certificate from the trust store and perform check against it.
projects / thirdparty / gnutls.git / log
