After merging with 2.4.6 kernel, ret_from_sys_call (and
o32_ret_from_sys_call) does not check whether it returns to kernel
mode or not.
syscall may happen in kernel mode, so we should check KU_USER bits (as
done in past code). Is this right?
At least, currently DO_FAULT() jumps to ret_from_sys_call and it may
cause problems. If page fault happened in kernel code when any
signals pending, do_signal() is called before returning to kernel and
it fails to setup sigcontext.