The x86_64 implementation of Poly1305 produces the wrong result on some
inputs because poly1305_4block_avx2() incorrectly assumes that when
partially reducing the accumulator, the bits carried from limb 'd4' to
limb 'h0' fit in a 32-bit integer.  This is true for poly1305-generic
which processes only one block at a time.  However, it's not true for
the AVX2 implementation, which processes 4 blocks at a time and
therefore can produce intermediate limbs about 4x larger.

Fix it by making the relevant calculations use 64-bit arithmetic rather
than 32-bit.  Note that most of the carries already used 64-bit
arithmetic, but the d4 -> h0 carry was different for some reason.

To be safe I also made the same change to the corresponding SSE2 code,
though that only operates on 1 or 2 blocks at a time.  I don't think
it's really needed for poly1305_block_sse2(), but it doesn't hurt
because it's already x86_64 code.  It *might* be needed for
poly1305_2block_sse2(), but overflows aren't easy to reproduce there.

This bug was originally detected by my patches that improve testmgr to
fuzz algorithms against their generic implementation.  But also add a
test vector which reproduces it directly (in the AVX2 case).

Fixes: b1ccc8f4 ("crypto: poly1305 - Add a four block AVX2 variant for x86_64")
Fixes: c70f4abe ("crypto: poly1305 - Add a SSE2 SIMD variant for x86_64")
Cc: <stable@vger.kernel.org> # v4.3+
Cc: Martin Willi <martin@strongswan.org>
Cc: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Eric Biggers <ebiggers@google.com>
Reviewed-by: Martin Willi <martin@strongswan.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>