On Wed, Dec 05, 2012 at 05:11:32PM +0000, Al Viro wrote:

On Wed, Dec 05, 2012 at 04:08:18PM +0000, James Hogan wrote:

quoted

other work not included:
* switch to generic kernel_{thread,execve} (will be posted separately to
  aid review)

quoted

other changes:

...

quoted

* switch to generic sys_execve

... which should've broken your kernel_execve().

BTW, one general note: lose the magical 7th argument.  In the few syscalls
that do need pt_regs, you can bloody well use current_pt_regs() instead;
it's what, (register & constant) + constant?  Passing it as explicit
argument is saving at most two arithmetical insns on those syscalls.
Moreover, for fork/clone/vfork we really don't need it until copy_thread();
one of the changes in -next kills passing pt_regs * to do_fork() and through
all that call chain (furthermore, if you make that
	childregs->ctx.AX[0].U0 = ALIGN(usp, 8);
in copy_thread() conditional on usp != 0, you can go with generic
variants from kernel/fork.c; again, see linux-next for examples of that).
What's left?  sigreturn and sigaltstack?

Seriously, drop that struct pt_regs *; it makes things uglier, (slightly)
hurts the syscall overhead and doesn't help the few syscalls that even
look at that argument.