X509_get_subject_name() in OpenSSL 1.0.2 does not mark its argument as a
const pointer, so need to type cast this to avoid a build warning.
Fixes: 841205a1ce ("OpenSSL: Add 'check_cert_subject' support for TLS server")
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
This patch added 'check_cert_subject' support to match the value of
every field against the DN of the subject in the client certificate. If
the values do not match, the certificate verification will fail and will
reject the user.
This option allows hostapd to match every individual field in the right
order, also allow '*' character as a wildcard (e.g OU=Development*).
Note: hostapd will match string up to 'wildcard' against the DN of the
subject in the client certificate for every individual field.
Signed-off-by: Paresh Chaudhary <paresh.chaudhary@rockwellcollins.com>
Signed-off-by: Jared Bents <jared.bents@rockwellcollins.com>
Signed-off-by: Jouni Malinen <j@w1.fi>
Pass the group order (if known/specified) to crypto_dh_derive_secret()
(and also to OpenSSL DH_generate_key() in case of Group 5) and verify
that the public key received from the peer meets 1 < pubkey < p and
pubkey^q == 1 mod p conditions.
While all these use cases were using only ephemeral DH keys, it is
better to use more explicit checks while deriving the shared secret to
avoid unexpected behavior.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
The flags variable needs to be initialized to 0 if check_crl is 0 in
the updated configuration.
Fixes: 159a7fbdea ("crl_reload_interval: Add CRL reloading support")
Signed-off-by: Jouni Malinen <j@w1.fi>
Add test-tls program that can be used for fuzzing the internal TLS
client and server implementations. This tool can write client or server
messages into a file as an initialization step and for the fuzzing step,
that file (with potential modifications) can be used to replace the
internally generated message contents.
The TEST_FUZZ=y build parameter is used to make a special build where a
hardcoded random number generator and hardcoded timestamp are used to
force deterministic behavior for the TLS operations.
Signed-off-by: Jouni Malinen <j@w1.fi>
This patch adds a new flag 'crl_reload_interval' to reload CRL
periodically. This can be used to reload ca_cert file and the included
CRL information on every new TLS session if difference between the last
reload and the current time in seconds is greater than
crl_reload_interval.
This reloading is used for cases where check_crl is 1 or 2 and the CRL
is included in the ca_file.
Signed-off-by: Paresh Chaudhary <paresh.chaudhary@rockwellcollins.com>
Signed-off-by: Jared Bents <jared.bents@rockwellcollins.com>
When using LibreSSL build fails with:
../src/crypto/tls_openssl.o: in function `tls_connection_client_cert':
../src/crypto/tls_openssl.c:2817: undefined reference to `SSL_use_certificate_chain_file'
collect2: error: ld returned 1 exit status
make: *** [Makefile:1901: wpa_supplicant] Error 1
There is no such function in LibreSSL.
Signed-off-by: Stefan Strogin <stefan.strogin@gmail.com>
Some distributions (e.g., Debian) have started introducting systemwide
OpenSSL policies to disable older protocol versions and ciphers
throughout all programs using OpenSSL. This can result in significant
number of interoperability issues with deployed EAP implementations.
Allow explicit wpa_supplicant (EAP peer) and hostapd (EAP server)
parameters to be used to request systemwide policies to be overridden if
older versions are needed to be able to interoperate with devices that
cannot be updated to support the newer protocol versions or keys. The
default behavior is not changed here, i.e., the systemwide policies will
be followed if no explicit override configuration is used. The overrides
should be used only if really needed since they can result in reduced
security.
In wpa_supplicant, tls_disable_tlsv1_?=0 value in the phase1 network
profile parameter can be used to explicitly enable TLS versions that are
disabled in the systemwide configuration. For example,
phase1="tls_disable_tlsv1_0=0 tls_disable_tlsv1_1=0" would request TLS
v1.0 and TLS v1.1 to be enabled even if the systemwide policy enforces
TLS v1.2 as the minimum version. Similarly, openssl_ciphers parameter
can be used to override systemwide policy, e.g., with
openssl_ciphers="DEFAULT@SECLEVEL=1" to drop from security level 2 to 1
in Debian to allow shorter keys to be used.
In hostapd, tls_flags parameter can be used to configure similar
options. E.g., tls_flags=[ENABLE-TLSv1.0][ENABLE-TLSv1.1]
Signed-off-by: Jouni Malinen <j@w1.fi>
SSL_use_certificate_chain_file() was added in OpenSSL 1.1.0, so need to
maintain the old version using SSL_use_certificate_file() for backwards
compatibility.
Fixes: 658c39809b ("OpenSSL: Load chain certificates from client_cert file")
Signed-off-by: Jouni Malinen <j@w1.fi>
The PADDING array used when adding padding bits in MD4 never change
so can be made const. Making it const puts the array in .rodata
section and can save a few bytes of RAM for systems running without
virtual memory.
Signed-off-by: Mikael Kanstrup <mikael.kanstrup@sony.com>
The function sha512_compress() has a local variable that consumes 640
bytes. This is very heavy for embedded devices that have limited stack
resources. Handle this by replacing the static allocation with a dynamic
one.
Signed-off-by: Ilan Peer <ilan.peer@intel.com>
According to random(4) manual, /dev/random is essentially deprecated on
Linux for quite some time:
"The /dev/random interface is considered a legacy interface, and
/dev/urandom is preferred and sufficient in all use cases, with the
exception of applications which require randomness during early boot
time; for these applications, getrandom(2) must be used instead, because
it will block until the entropy pool is initialized."
An attempt to use it would cause unnecessary blocking on machines
without a good hwrng even when it shouldn't be needed. Since Linux 3.17,
a getrandom(2) call is available that will block only until the
randomness pool has been seeded.
It is probably not a good default yet as it requires a fairly recent
kernel and glibc (3.17 and 2.25 respectively).
Signed-off-by: Lubomir Rintel <lkundrak@v3.sk>
This helps the server to build the chain to trusted CA when PEM encoding
of client_cert is used with multiple listed certificates. This was
already done for the server certificate configuration, but the client
certificate was limited to using only the first certificate in the file.
Signed-off-by: Isaac Boukris <iboukris@gmail.com>
Add the ability to ignore time-based CRL errors from OpenSSL by
specifying a new configuration parameter, check_crl_strict=0.
This causes the following:
- This setting does nothing when CRL checking is not enabled.
- When CRL is enabled, "strict mode" will cause CRL time errors to not
be ignored and will continue behaving as it currently does.
- When CRL is enabled, disabling strict mode will cause CRL time
errors to be ignored and will allow connections.
By default, check_crl_strict is set to 1, or strict mode, to keep
current functionality.
Signed-off-by: Sam Voss <sam.voss@rockwellcollins.com>
Build configurations with CONFIG_TLS=internal and NEED_SHA512 failed due
to missing sha512.c file. Add that file even though this is not really
used in the currently available configuration combinations since DPP and
OWE are the only users of it and the internal crypto implementation
supports neither.
Signed-off-by: Jouni Malinen <j@w1.fi>
Some versions of OpenSSL need server support for ECDH to be explicitly
enabled, so provide a new parameter for doing so and all
SSL_{,CTX_}set_ecdh_auto() for versions that need it to enable automatic
selection.
Signed-off-by: Hristo Venev <hristo@venev.name>
When using LibreSSL instead of OpenSSL, linkage of hostapd executable
fails with the following error when using some LibreSSL versions
../src/crypto/tls_openssl.o: In function `tls_verify_cb':
tls_openssl.c:(.text+0x1273): undefined reference to `ASN1_STRING_get0_data'
../src/crypto/tls_openssl.o: In function `tls_connection_peer_serial_num':
tls_openssl.c:(.text+0x3023): undefined reference to `ASN1_STRING_get0_data'
collect2: error: ld returned 1 exit status
make: *** [Makefile:1278: hostapd] Error 1
ASN1_STRING_get0_data is present in recent OpenSSL, but absent in some
versions of LibreSSL (confirmed for version 2.6.5), so fallback needs to
be defined in this case, just like for old OpenSSL.
This patch was inspired by similar patches to other projects, such as
spice-gtk, pjsip.
Link: https://bugs.gentoo.org/672834
Signed-off-by: Andrey Utkin <andrey_utkin@gentoo.org>
This can be used to fetch the serial number of the peer certificate in
the EAP server. For now, this is implemented only with OpenSSL.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
SSL_session_reused() is the same as the deprecated SSL_cache_hit(). The
engine load stuff is now handled by OPENSSL_init().
Signed-off-by: Rosen Penev <rosenp@gmail.com>
The n argument to this function is number of bits, not bytes, to shift.
As such, need to use mp_rshb() instead of mp_rshd(). This fixes EAP-pwd
with P-521 curve.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
Implement tls_connection_get_eap_fast_key() using cryptographic
primitives as wolfSSL implements different spec.
Signed-off-by: Sean Parkinson <sean@wolfssl.com>
Depend on proper wolfSSL configuration instead of trying to define these
build configuration values externally.
Signed-off-by: Sean Parkinson <sean@wolfssl.com>
Use the correct intermediate result from mp_sqrmod() in the following
mp_mulmod() call (t is not initialized here; it is used only after this
step).
Signed-off-by: Sean Parkinson <sean@wolfssl.com>
Provide full uncompressed DER data length to wc_ecc_import_point_der()
even though a compressed form is used here. In addition, use
ECC_POINT_COMP_* defined values to make this more readable.
Signed-off-by: Sean Parkinson <sean@wolfssl.com>
This may be needed to avoid interoperability issues with the new
protocol version and significant changes for EAP use cases in both key
derivation and handshake termination.
Signed-off-by: Jouni Malinen <j@w1.fi>
This makes this more easily available throughout the handshake
processing, if needed, compared to having to pass through the function
argument through the full path from
tls_connection{,_server}_handshake().
Signed-off-by: Jouni Malinen <j@w1.fi>
OpenSSL 1.1.1 added cases where ClientHello generation may fail due to
"no ciphers available". There is no point in sending out the resulting
TLS Alert message to the server since the server does not know what to
do with it before ClientHello. Instead, simply terminate the TLS
handshake locally and report EAP failure to avoid getting stuck waiting
for a timeout.
Signed-off-by: Jouni Malinen <j@w1.fi>
LibreSSL v2.7 claims an OPENSSL_VERSION_NUMBER value that would indicate
that SSL_OP_NO_TLSv1_3 is available, but that does not seem to be the
case with LibreSSL. As such, skip this step based on whether
SSL_OP_NO_TLSv1_3 is defined to avoid build issues.
Signed-off-by: Jouni Malinen <j@w1.fi>
This commit takes care of the sigalg configuration using the relatively
recent SSL_CTX_set_verify_algorithm_prefs() addition from April 2017 to
address the functionality that was already there with OpenSSL using
SSL_set1_sigalgs_list().
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
BoringSSL removed the special OpenSSL cipher suite value "SUITEB192", so
need to map that to the explicit ciphersuite
(ECDHE-ECDSA-AES256-GCM-SHA384), curve (P-384), and sigalg
(SSL_SIGN_ECDSA_SECP384R1_SHA384) to allow 192-bit level Suite B with
ECDSA to be used.
This commit takes care of the sigalg configuration using the relatively
recent SSL_CTX_set_verify_algorithm_prefs() addition from April 2017.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
BoringSSL removed the special OpenSSL cipher suite value "SUITEB192", so
need to map that to the explicit ciphersuite
(ECDHE-ECDSA-AES256-GCM-SHA384), curve (P-384), and sigalg
(SSL_SIGN_ECDSA_SECP384R1_SHA384) to allow 192-bit level Suite B with
ECDSA to be used.
This commit takes care of the ciphersuite and curve configuration.
sigalg change is in a separate commit since it requires a newer
BoringSSL API function that may not be available in all builds.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
In practice, this does the same thing (i.e., allows only the P-384 curve
to be used), but using an older API function that happens to be
available in some BoringSSL builds while the newer one is not.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
TLS v1.3 needs to be explicitly disabled to allow cipher suite selection
for EAP-FAST to work with OpenSSL builds that include TLS v1.3 support.
Without this, OpenSSL refuses to generate ClientHello due to the cipher
suite list including only ciphers allowed with older versions than TLS
v1.3.
Signed-off-by: Jouni Malinen <jouni@codeaurora.org>
GnuTLS-based builds can now be done using either libnettle or libgcrypt
for crypto functionality:
CONFIG_TLS=gnutls
CONFIG_CRYPTO=nettle
CONFIG_TLS=gnutls
CONFIG_CRYPTO=gnutls
Signed-off-by: Jouni Malinen <j@w1.fi>
This allows OpenSSL-style configuration of Suite B parameters to be used
in the wpa_supplicant network profile. 128-bit and 192-bit level
requirements for ECDHE-ECDSA cases are supported. RSA >=3K case is
enforced using GnuTLS %PROFILE_HIGH special priority string keyword.
Signed-off-by: Jouni Malinen <j@w1.fi>
This extends GnuTLS support for tls_disable_v1_{0,1,2}=1 flags in the
phase1 network profile parameter in wpa_supplicant.
Signed-off-by: Jouni Malinen <j@w1.fi>
Indicate more clearly when the parsing succeeds to avoid ending the
debug prints with various internal GnuTLS internal error messages even
when the parsing actually succeeded in the end.
Signed-off-by: Jouni Malinen <j@w1.fi>
Replace the internal HMAC MD5, SHA-1, and SHA256 implementations with
the ones from libgcrypt and also add the SHA384 and SHA512 versions.
Signed-off-by: Jouni Malinen <j@w1.fi>
Replace the internal SHA256 implementation with the one from libgcrypt
and also add the SHA384 and SHA512 versions.
Signed-off-by: Jouni Malinen <j@w1.fi>
Use a shared helper function instead of implementing practically same
sequence separately for each hash function.
Signed-off-by: Jouni Malinen <j@w1.fi>
The patch offers alternate implementations of some functions using the
abstract cryptographic API.
This work was done in preparation for the changes to allow hostap to be
compiled with the wolfSSL cryptography and TLS library.
Signed-off-by: Sean Parkinson <sean@wolfssl.com>
This implements crypto_dh_init() and crypto_dh_derive_secret() using
os_get_random() and crypto_mod_exp() for all crypto_*.c wrappers that
include crypto_mod_exp() implementation.
Signed-off-by: Jouni Malinen <j@w1.fi>
The BoringSSL version of crypto_ecdh_init() and dpp_gen_keypair() works
fine with OpenSSL as well, so use that same implementation for both to
avoid unnecessary maintanence of multiple versions.
Signed-off-by: Jouni Malinen <j@w1.fi>
There is no need to go through EC_GROUP_new_by_curve_name(),
EC_KEY_new(), and EC_KEY_set_group() when a single call to
EC_KEY_new_by_curve_name() takes care of all that.
Signed-off-by: Jouni Malinen <j@w1.fi>
This allows wpa_supplicant configuration with phase1="tls_suiteb=1" to
use openssl_ciphers="ECDHE-RSA-AES256-GCM-SHA384" to further limit the
possible TLS cipher suites when using Suite B with RSA >3K keys. This
combination disables use of DHE and as such, mandates ECDHE to be used.
Signed-off-by: Jouni Malinen <j@w1.fi>
These functions are a bit awkward to use for one-off file loads, as
suggested by the tls_clear_default_passwd_cb() logic. There was also
some historical mess with OpenSSL versions and either not having per-SSL
settings, having per-SSL settings but ignoring them, and requiring the
per-SSL settings.
Instead, loading the key with the lower-level functions seems a bit
tidier and also allows abstracting away trying both formats, one after
another.
Signed-off-by: David Benjamin <davidben@google.com>
There's no need to make an extra copy of private_key_passwd for
SSL_{CTX_,}set_default_passwd_cb().
Signed-off-by: David Benjamin <davidben@google.com>
BoringSSL restored the previously removed AES-192 ECB support in ("Add
AES-192 ECB.") commit. Since this is needed for DPP with the P-384
curve, restore support for this through EVP_aes_192_ecb().
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
This makes it clearer why some AES operations fail especially with
BoringSSL where the 192-bit case is not supported.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
BoringSSL does not provide some of the OpenSSL API that was used here,
so update this to use similar design to what was already done with DPP
key derivation.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
It looks like BoringSSL claims to have OPENSSL_VERSION_NUMBER for a
1.1.0 version, but it does not provide SSL_set_default_passwd_cb*(). For
now, comment out this regardless of the version BoringSSL claims to be.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
It looks like BoringSSL claims to have OPENSSL_VERSION_NUMBER for a
1.1.0 version, but it does not provide SSL_set1_sigalgs_list(). For now,
comment out this regardless of the version BoringSSL claims to be.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
It looks like BoringSSL claims to have OPENSSL_VERSION_NUMBER for a
1.1.0 version, but it does not provide RSA_bits(). For now, add this
backwards compatibility wrapper for BoringSSL regardless of the version
it claims to be.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
Reject a DHE handshake if the server uses a DH prime that does not have
sufficient length to meet the Suite B 192-bit level requirement (<= 3k
(3072) bits).
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
The hostapd.conf tls_flags=[SUITEB-NO-ECDH] and wpa_supplicant network
profile phase1="tls_suiteb_no_ecdh=1" can now be used to configure Suite
B RSA constraints with ECDHE disabled. This is mainly to allow
the DHE TLS cipher suite to be tested.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
Reject a peer certificate chain if it includes an RSA public key that
does not use sufficient key length to meet the Suite B 192-bit level
requirement (<= 3k (3072) bits).
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
This adds phase1 parameter tls_suiteb=1 into wpa_supplicant
configuration to allow TLS library (only OpenSSL supported for now) to
use Suite B 192-bit level rules with RSA when using >= 3k (3072) keys.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
ubsan reported:
../src/crypto/random.c:69:30: runtime error: shift exponent 32 is too large for 32-bit type 'unsigned int'
Explicitly check for the ROL32(x, 0) case which is supposed to be a
no-op.
Signed-off-by: Michael Braun <michael-dev@fami-braun.de>
Previously, the pointer to strdup passwd was left in OpenSSL library
default_passwd_cb_userdata and even the default_passwd_cb was left set
on an error path. To avoid unexpected behavior if something were to
manage to use there pointers, clear them explicitly once done with
loading of the private key.
Signed-off-by: Jouni Malinen <j@w1.fi>
Since OpenSSL version 1.1.0f, SSL_use_PrivateKey_file() uses the
callback from the SSL object instead of the one from the CTX, so let's
set the callback on both SSL and CTX. Note that
SSL_set_default_passwd_cb*() is available only in 1.1.0.
Signed-off-by: Beniamino Galvani <bgalvani@redhat.com>
Add a build option to select different default ciphers for OpenSSL
instead of the hardcoded default "DEFAULT:!EXP:!LOW".
This new option is useful on distributions where the security level
should be consistent for all applications, as in Fedora [1]. In such
cases the new configuration option would be set to "" or
"PROFILE=SYSTEM" to select the global crypto policy by default.
[1] https://fedoraproject.org/wiki/Changes/CryptoPolicy
Signed-off-by: Beniamino Galvani <bgalvani@redhat.com>
The additional SHA-384 and SHA-512 functionality is needed to support
DPP with various ECC curves.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
Previously, only SHA1 hash -based server certificate matching was used,
but the OCSP response may use SHA256 instead of SHA1, so check the match
with both hash functions, if needed.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
The KDF define in RFC 5295 is very similar to HKDF-Expand() defined in
RFC 5869. Allow a NULL label to be used to select the RFC 5869 version
with arbitrary seed (info in RFC 5869) material without forcing the
label and NULL termination to be included. HKDF-Expand() will be needed
for OWE.
Signed-off-by: Jouni Malinen <j@w1.fi>
These allow ECDH to be used with compressed public key encoding (only
x-coordinate). This is needed for FILS PFS and OWE.
Signed-off-by: Jouni Malinen <j@w1.fi>
This leads to cleaner code overall, and also reduces the size
of the hostapd and wpa_supplicant binaries (in hwsim test build
on x86_64) by about 2.5 and 3.5KiB respectively.
The mechanical conversions all over the code were done with
the following spatch:
@@
expression SIZE, SRC;
expression a;
@@
-a = os_malloc(SIZE);
+a = os_memdup(SRC, SIZE);
<...
if (!a) {...}
...>
-os_memcpy(a, SRC, SIZE);
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
CONFIG_TLS=linux can now be used to select the crypto implementation
that uses the user space socket interface (AF_ALG) for the Linux kernel
crypto implementation. This commit includes some of the cipher, hash,
and HMAC functions. The functions that are not available through AF_ALG
(e.g., the actual TLS implementation) use the internal implementation
(CONFIG_TLS=internal).
Signed-off-by: Jouni Malinen <j@w1.fi>
These operations may fail with some crypto wrappers, so allow the
functions to report their results to the caller.
Signed-off-by: Jouni Malinen <j@w1.fi>
This is a copy of the internal HMAC-SHA256 implementation with the hash
block size and output length updated to match SHA384 parameters.
Signed-off-by: Jouni Malinen <j@w1.fi>
The output length was incorrect (32 from the copy-pasted SHA256
version). Fix this to return the correct number of octets (48) for
SHA384. This fixes incorrect key derivation in FILS when using the
SHA384-based AKM.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
If the keychain holds additional certificates other than the end
certificate, read them into the certificate chain.
Signed-off-by: Paul Stewart <pstew@google.com>
Number of deployed use cases assume the default OpenSSL behavior of auto
chaining the local certificate is in use. BoringSSL removed this
functionality by default, so we need to restore it here to avoid
breaking existing use cases.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
This simplifies the implementation since the SSL_clear_options() and
SSL_CTX_clear_options() are available in all supported versions of
OpenSSL. These were previously needed with older (now obsolete) versions
of OpenSSL, but the ifdefs were missed when removing the more explicit
version macro based backwards compatibility sections.
In practice, this reverts commit
d53d2596e4.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
The previous implementation was hardcoded to use 128-bit AES key
(AEAD_AES_SIV_CMAC_256). Extend this by allowing AEAD_AES_SIV_CMAC_384
and AEAD_AES_SIV_CMAC_512 with 192-bit and 256-bit AES keys.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
This basically just follows commit
587b0457e0 ('LibreSSL: Fix build with
LibreSSL') with the same pattern, which was missed here.
Signed-off-by: Julian Ospald <hasufell@hasufell.de>
Due to a missing guard for old OpenSSL code, SSL_library_init() was not
called, which is required for LibreSSL. Likewise for cleanup.
Signed-off-by: Christian Neukirchen <chneukirchen@gmail.com>
Add LibreSSL check to old OpenSSL #ifdef guard as DH_{get0,set0}_key()
is not implemented in LibreSSL.
Signed-off-by: Christian Neukirchen <chneukirchen@gmail.com>
To be consistent with OpenSSL 1.1.0, the free functions should
internally check for NULL. EVP_MD_CTX_free also was missing an
EVP_MD_CTX_cleanup, so this leaked a little.
OpenSSL 1.1.0 also has given get_rfc3526_prime_1536 a better namespace
with get_rfc3526_prime_1536 as a compatibility-only name. Use that
instead in 1.1.0.
Signed-off-by: David Benjamin <davidben@google.com>
Commit 49fe2ada20 ('OpenSSL: Support
OpenSSL 1.1.0 DH opacity') started using the new accessor functions, but
used incorrect success check for the DH_set0_key() call. This resulted
in dh5_init_fixed() failures and double-free on error path if the build
was linked against OpenSSL 1.1.0. Fix this by checking DH_set0_key()
return value to be 1 for the success case.
Signed-off-by: Jouni Malinen <j@w1.fi>
OpenSSL 1.1.0 (master branch) apparently ended up modifying the API
after the beta 2 release that was supposed to complete the work. Mark
the variables const to fix the compilation with the modified OpenSSL
API.
Signed-off-by: Jouni Malinen <j@w1.fi>
Since crypto_openssl.c is now implementing couple of functions
internally, pull in the relevant header files md5.h and aes_wrap.h to
make sure the function declaration are consistent.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
Recent versions of engine_pkcs11 are set up to be autoloaded on demand
with ENGINE_by_id() because they don't need explicit configuration.
But if we *do* want to explicitly configure them with a PKCS#11 module
path, we should still do so.
We can't tell whether it was already initialised, but it's harmless to
repeat the MODULE_PATH command if it was.
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Tested-by: Michael Schaller <misch@google.com>
This avoids internal access of structs and also removes the dependency
on the reimplemented TLS PRF functions when EAP-FAST support is not
enabled. Notably, BoringSSL doesn't support EAP-FAST, so there is no
need to access its internals with openssl_get_keyblock_size().
Signed-Off-By: David Benjamin <davidben@google.com>
Most protocols extracting keys from TLS use RFC 5705 exporters which is
commonly implemented in TLS libraries. This is the mechanism used by
EAP-TLS. (EAP-TLS actually predates RFC 5705, but RFC 5705 was defined
to be compatible with it.)
EAP-FAST, however, uses a legacy mechanism. It reuses the TLS internal
key block derivation and derives key material after the key block. This
is uncommon and a misuse of TLS internals, so not all TLS libraries
support this. Instead, we reimplement the PRF for the OpenSSL backend
and don't support it at all in the GnuTLS one.
Since these two are very different operations, split
tls_connection_prf() in two. tls_connection_export_key() implements the
standard RFC 5705 mechanism that we expect most TLS libraries to
support. tls_connection_get_eap_fast_key() implements the
EAP-FAST-specific legacy mechanism which may not be implemented on all
backends but is only used by EAP-FAST.
Signed-Off-By: David Benjamin <davidben@google.com>
Commit 68ae4773a4 ('OpenSSL: Use library
wrapper functions to access cert store') fixed most of these, but missed
a few.
Signed-Off-By: David Benjamin <davidben@google.com>
Commit 4104267e81 ('Fix memory leak on NFC
DH generation error path') modified the generic (non-OpenSSL)
implementation of dh5_init() to free the previously assigned public key,
if any. However, that commit did not modify the OpenSSL specific version
of this function. Add the same change there to maintain consistent
behavior between these two implementations of the same function.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
BoringSSL added OpenSSL 1.1.0's SSL_get_client_random() and friends in
working towards opaquifying the SSL struct. But it, for the moment,
still looks more like 1.0.2 than 1.1.0 and advertises
OPENSSL_VERSION_NUMBER as such. This means that there is no need to
define those in BoringSSL and defining them causes conflicts. (C does
not like having static and non-static functions with the same name.)
As requested, this is conditioned on defined(BORINGSSL_API_VERSION) so
wpa_supplicant may continue to support older BoringSSLs for a time.
(BoringSSL revisions without the accessors predate BoringSSL maintaining
a BORINGSSL_API_VERSION.)
Also add a missing opensslv.h include. tls_openssl.c is sensitive to
OPENSSL_VERSION_NUMBER, so it should include the header directly rather
than rely on another header to do so.
Signed-off-by: David Benjamin <davidben@google.com>
The OpenSSL 1.1.0 Beta 2 release made DH opaque and that broke
compilation of crypto_openssl.c. Fix this by using the new accessor
functions when building against OpenSSL 1.1.0 or newer.
Signed-off-by: Jouni Malinen <j@w1.fi>
While this function could return NULL if the parameter issued to it were
NULL, that does not really happen here. Anyway, since this can result in
a warning from a static analyzer that does can see the return NULL
without fully understanding what it means here, check the return value
explicitly against NULL to avoid false warnings.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
Commit 8bcf8de827 ('OpenSSL: Fix memory
leak in PKCS12 additional certificate parsing') tried to fix a memory
leak in both the 1.0.2(and newer) and 1.0.1 branches of PKCS12 parsing.
However, the 1.0.1 case was not properly tested and freeing of the
certificate after a successful SSL_CTX_add_extra_chain_cert() call
resulted in use of freed memory when going through the TLS handshake.
Fix this by not freeing the certificate in that specific case.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
The additional PKCS12 certificates were not freed properly in the loop
added in commit de2a7b796d ('OpenSSL: Use
connection certificate chain with PKCS#12 extra certs').
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
Commit 5c9a33702f ('OpenSSL: Clean up
crypto_hash_*() to use a single implementation') added a wrapper
function to allow the new OpenSSL API to be used with older OpenSSL
versions. However, the HMAC_CTX_free() wrapper was incorrectly skipping
the call to HMAC_CTX_cleanup() which is still needed to free the
resources OpenSSL allocated internally.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
SSL_library_init() does not work properly after EVP_cleanup() starting
from OpenSSL 1.1.0 pre release 3. The automated library init/deinit
functions in that pre release are supposed to handle all initialization
and deinitialiation, so comment out the explicit calls to these function
with OpenSSL 1.1.0 and newer.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
Commit 1eb87ae48d ('OpenSSL: Use
EVP_CIPHER_CTX_new() to work with OpenSSL 1.1.0') started using
EVP_CIPHER_CTX_new() to allocate EVP_CIPHER_CTX from heap instead of
using stack memory. This commit used incorrect EVP_CIPHER_CTX_reset()
function in number of cases when the allocated memory was supposed to be
freed instead of just reset for reuse. Fix this by using
EVP_CIPHER_CTX_free() properly.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
If SSL_CTX_new(SSLv23_method()) fails, tls_init() error path did not
free the allocated struct tls_data instance.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
Fix a possible null pointer dereference in tls_parse_pkcs12() when
loading a PKCS#12 file for the server keys and the file includes extra
certificates.
Signed-off-by: Ayala Beker <ayala.beker@intel.com>
In the Android-specific case, make ca_cert directive parse a
space-separated list of hex-encoded CA certificate aliases following the
"keystores://" prefix. Server certificate validation should succeed as
long as the chain ends with one of them.
Signed-off-by: Rubin Xu <rubinxu@google.com>
Use compatibility wrapper functions to allow a single implementation
based on the latest OpenSSL API to be used to implement these functions
instead of having to maintain two conditional implementation based on
the library version.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
Use compatibility wrapper functions to allow a single implementation
based on the latest OpenSSL API to be used to implement these functions
instead of having to maintain two conditional implementation based on
the library version.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
The changes needed for OpenSSL 1.1.0 had broken this since LibreSSL is
defining OPENSSL_VERSION_NUMBER in a manner that claims it to be newer
than the current OpenSSL version even though it does not support the
current OpenSSL API.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
Now the these functions cannot be made to fail by forcing the memory
allocation fail since the OpenSSL-internal version is used, add
TEST_FAIL check to allow OOM test cases to be converted to use the
TEST_FAIL mechanism without reducing coverage.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
The EVP_CIPHER_CTX structure will be made opaque in OpenSSL 1.1.0, so
need to use EVP_CIPHER_CTX_new() with it instead of stack memory. The
design here moves the older OpenSSL versions to use that dynamic
allocation design as well to minimize maintenance effort.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
The SSL_CIPHER **cipher argument was marked const in OpenSSL 1.1.0
pre-release 2 similarly to how this is in BoringSSL. Fix build with that
in preparation for supporting OpenSSL 1.1.0.
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
If wpa_supplicant was first configured with PKCS #12 -based client
certificate chain and then used with another network profile that used a
different certificate chain from a X.509 certificate PEM file instead of
PKCS#12, the extra certificate chain was not reconstructed properly with
older versions of OpenSSL that 1.0.2. This could result in the
authentication failing due to the client certificate chain not being
complete or including incorrect certificates.
Fix this by clearing the extra certificate chain when setting up a new
TLS connection with OpenSSL 1.0.1. This allows OpenSSL to build the
chain using the default mechanism in case the new TLS exchange does not
use PKCS#12.
The following hwsim test case sequence was able to find the issue:
ap_wpa2_eap_tls_pkcs12 ap_wpa2_eap_tls_intermediate_ca_ocsp
Signed-off-by: Jouni Malinen <jouni@qca.qualcomm.com>
