The stack allocation in crypt() is bogus as whether tmp_src/tmp_dst
is used is determined by factors unrelated to nbytes and
src->length/dst->length.

Since the condition for whether tmp_src/tmp_dst are used is very
complex, let's allocate them always instead of guessing.

This fixes a number of weird crashes including those AES crashes
that people have been seeing with the 2.4 backport + ipt_conntrack.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: James Morris <jmorris@redhat.com>
Signed-off-by: David S. Miller <davem@redhat.com>