• Arnd Bergmann's avatar
    crypto: serpent - improve __serpent_setkey with UBSAN · c871c10e
    Arnd Bergmann authored
    When UBSAN is enabled, we get a very large stack frame for
    __serpent_setkey, when the register allocator ends up using more registers
    than it has, and has to spill temporary values to the stack. The code
    was originally optimized for in-order x86-32 CPU implementations using
    older compilers, but it now runs into a highly suboptimal case on all
    CPU architectures, as seen by this warning:
    crypto/serpent_generic.c: In function '__serpent_setkey':
    crypto/serpent_generic.c:436:1: error: the frame size of 2720 bytes is larger than 2048 bytes [-Werror=frame-larger-than=]
    Disabling -fsanitize=alignment would avoid that warning, presumably the
    option turns off a optimization step that is required for getting the
    register allocation right, but there is no easy way to do that on gcc-7
    (gcc-8 introduces a function attribute for this).
    I tried to figure out a way to modify the source code instead, and noticed
    that the two stages of the setkey() function (keyiter and sbox) each are
    fine by themselves, but not when combined into one function. Splitting
    out the entire sbox into a separate function also happens to work fine
    with all compilers I tried (arm, arm64 and x86).
    The setkey function uses a strange way to handle offsets into the key
    array, using both negative and positive index values, as well as adjusting
    the array pointer back and forth. I have checked that this actually
    makes no difference to modern compilers, but I left that untouched
    to make the patch easier to review and to keep the code closer to
    the reference implementation.
    Link: https://patchwork.kernel.org/patch/9189575/Signed-off-by: 's avatarArnd Bergmann <arnd@arndb.de>
    Signed-off-by: 's avatarHerbert Xu <herbert@gondor.apana.org.au>