Emulation of the tlbwr instruction, which writes a TLB entry to a random
index in the TLB, currently uses get_random_bytes() to generate a 4 byte
random number which we then mask to form the index. This is overkill in
a couple of ways:

  - We don't need 4 bytes here since we mask the value to form a 6 bit
    number anyway, so we waste /dev/random entropy generating 3 random
    bytes that are unused.

  - We don't need crypto-grade randomness here - the architecture spec
    allows implementations to use any algorithm & merely encourages that
    some pseudo-randomness be used rather than a simple counter. The
    fast prandom_u32() function fits that criteria well.

So rather than using get_random_bytes() & consuming /dev/random entropy,
switch to using the faster prandom_u32_max() which provides what we need
here whilst also performing the masking/modulo for us.
Signed-off-by: Paul Burton <paul.burton@mips.com>
Reported-by: George Spelvin <lkml@sdf.org>
Cc: James Hogan <jhogan@kernel.org>
Cc: linux-mips@vger.kernel.org