Don't use an all.sh component because there isn't one that does what we
want (modern Clang with ASan, and test everything).
* We need to set CC explicitly or tweak PATH, because clang in $PATH on
Travis focal instances is Clang 7 which is too old (we want Clang 10).
* Travis lacks the array of versions of openssl and gnutls that we normally
use for testing, so we need to exclude some tests (or build our
own multiple versions of openssl and gnutls).
The SSL test exclusions are ad hoc and based on what currently works.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
OpenSSL may be configured to support features such as cipher suites or
protocol versions that are disabled by default. Enable them all: we're
testing, we don't care about enabling insecure stuff. This is not needed
with the builds of OpenSSL that we're currently using on the Jenkins CI, but
it's needed with more recent versions such as typically found on developer
machines, and with future CI additions.
The syntax to do that was only introduced in OpenSSL 1.1.0; fortunately we
don't need to do anything special with earlier versions.
With OpenSSL 1.1.1f on Ubuntu 20.04, this allows SHA-1 in certificates,
which is still needed for a few test cases in ssl-opt.sh. Curiously, this is
also needed for the cipher suite TLS-DHE-PSK-WITH-ARIA-128-GCM-SHA256 (and
no other, including other DHE-PSK or ARIA cipher suites).
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
Having multiple definitions was cumbersome, and meant we might forget the
definition when adding an inline definition to a file that didn't have one
before (as I did when I added an inline definition in common.h).
Resolves#6649.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
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>
- avoid long unstructured paragraph with long messy sentences
- de-emphasize "no longer depends on MD" and emphasize "can work in
some driver-only builds" instead - that's what users are interested in
(building without MD is just the current way to accomplish that, but
that will change in the future)
Signed-off-by: Manuel Pégourié-Gonnard <manuel.pegourie-gonnard@arm.com>
The loop ends when there are no more bits to process, with one twist: when
that happens, we need to clear the window one last time. Since the window
does not start empty (E_limbs==0 is not supported), the loop always starts
with a non-empty window and some bits to process. So it's correct to move
the window clearing logic to the end of the loop. This lets us exit the loop
when the end of the exponent is reached.
It would be clearer not to do the final window clearing inside the loop, so
we wouldn't need to repeat the loop termination condition (end of exponent
reached) inside the loop. However, this requires duplicating the code to
clear the window. Empirically, this causes a significant code size increase,
even if the window clearing code is placed into a function.
Signed-off-by: Gilles Peskine <Gilles.Peskine@arm.com>
