This makes the 64-bit kernel use 64-bit signed integers for the counter
(effectively supporting 32-bit of active count in the semaphore), thus
avoiding things like overflow of the mmap_sem if you use a really crazy
number of threads

Note: Ideally the type in the structure should be atomic_long_t rather
than "long". However, there's some nasty issues with that. It needs to
be initialized statically -and- lib/rwsem.c does things like

        sem->count = RWSEM_UNLOCKED_VALUE;

Now, if you mix in the fact that atomic_* types are actually structures
with one member and note typedefs of a scalar, it makes its really nasty.

So I stuck to what we did before using a long and casts for now.
Signed-off-by: NBenjamin Herrenschmidt <benh@kernel.crashing.org>