The gf128mul library has different variants with different
memory/performance tradeoffs, where the faster ones use 4k or 64k lookup
tables precomputed at runtime, which are based on one of the
multiplication factors, which is commonly the key for keyed hash
algorithms such as GHASH.

The slowest variant is gf128_mul_lle() [and its bbe/ble counterparts],
which does not use precomputed lookup tables, but it still relies on a
single u16[256] lookup table which is input independent. The use of such
a table may cause the execution time of gf128_mul_lle() to correlate
with the value of the inputs, which is generally something that must be
avoided for cryptographic algorithms. On top of that, the function uses
a sequence of if () statements that conditionally invoke be128_xor()
based on which bits are set in the second argument of the function,
which is usually a pointer to the multiplication factor that represents
the key.

In order to remove the correlation between the execution time of
gf128_mul_lle() and the value of its inputs, let's address the
identified shortcomings:
- add a time invariant version of gf128mul_x8_lle() that replaces the
  table lookup with the expression that is used at compile time to
  populate the lookup table;
- make the invocations of be128_xor() unconditional, but pass a zero
  vector as the third argument if the associated bit in the key is
  cleared.

The resulting code is likely to be significantly slower. However, given
that this is the slowest version already, making it even slower in order
to make it more secure is assumed to be justified.

The bbe and ble counterparts could receive the same treatment, but the
former is never used anywhere in the kernel, and the latter is only
used in the driver for a asynchronous crypto h/w accelerator (Chelsio),
where timing variances are unlikely to matter.
Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>