psa_cipher_encrypt() and psa_cipher_decrypt() sometimes add a zero offset to
a null pointer when the cipher does not use an IV. This is undefined
behavior, although it works as naively expected on most platforms. This
can cause a crash with modern Clang+ASan (depending on compiler optimizations).
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
Now when operation holds pointer to dynamically allocated buffer for password key we can't do copy of the operation object in test instead we need to re-initialize operation object after error.
Signed-off-by: Przemek Stekiel <przemyslaw.stekiel@mobica.com>
Instead of corrupting the public key part of the message,
corrupt the proof part. A proof is conceptually similar to a signature,
and changing anything in it should make it invalid with
a high probability.
Also, instead of shifting data, perform a bitflip.
Signed-off-by: Andrzej Kurek <andrzej.kurek@arm.com>
We're not strictly required to abort, but at least to leave the context
is an invalid state. For "late" functions like input() and output(),
calling abort() is the easiest way to do that. Do it systematically for
input() and output() by using a wrapper. psa_pake_get_implicit_key() was
already doing it. For "early" function, we can just leave the operation
in its current state which is already invalid.
Restore previous tests about that. Not adding systematic tests, though,
just test the two functions that are the most important, and more likely
to return errors.
Since we now abort in more cases, we need to make sure we don't
invalidate the operation that's going to be re-used later in the test.
For that reason, use a copy of the operation for calls to input() and
output() that are expected to return errors.
Signed-off-by: Manuel Pégourié-Gonnard <manuel.pegourie-gonnard@arm.com>
Also, the error code changed from INSUFFICIENT_MEMORY to INVALID_DATA.
Temporarily remove a test about aborting the operation on error.
Auto-abort will be re-introduced in the next commit in a more systematic
way.
Signed-off-by: Manuel Pégourié-Gonnard <manuel.pegourie-gonnard@arm.com>
The format used by the mbedtls_ecjpake_xxx() APIs and that defined by
the PSA Crypto PAKE extension are quite different; the former is
tailored to the needs of TLS while the later is quite generic and plain.
Previously we only addressed some part of this impedance mismatch: the
different number of I/O rounds, but failed to address the part where the
legacy API adds some extras (length bytes, ECParameters) that shouldn't
be present in the PSA Crypto version. See comments in the code.
Add some length testing as well; would have caught the issue.
Signed-off-by: Manuel Pégourié-Gonnard <manuel.pegourie-gonnard@arm.com>
Systematically replace "TEST_ASSERT( $x <= $y )" by "TEST_LE_U( $x, $y )" in
test_suite_psa_crypto. In this file, all occurrences of this pattern are
size_t so unsigned.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
We want to check:
1. actual output <= PSA_RAW_KEY_AGREEMENT_OUTPUT_SIZE (the output fits
if the caller uses the key-specific buffer size macro)
2. actual output <= PSA_RAW_KEY_AGREEMENT_OUTPUT_MAX_SIZE (the output fits
if the caller uses the generic buffer size macro)
3. PSA_RAW_KEY_AGREEMENT_OUTPUT_SIZE <= PSA_RAW_KEY_AGREEMENT_OUTPUT_MAX_SIZE
(consistency in the calculation)
We were only testing (1) and (2). Test (3) as well. (1) and (3) together
imply (2) so there's no need to test (2).
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
For multipart encrpytion, call psa_cipher_finish(). This is not actually
necessary for non-pathological implementations of ECB (the only currently
supported IV-less cipher algorithm) because it requires the input to be a
whole number of blocks and non-pathological implementations emit the output
block from update() as soon as an input block is available. But in principle
a driver could delay output and thus require a call to finish().
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
