On Fri, Oct 05, 2012 at 03:55:36PM +0100, James Bottomley wrote:
> On Fri, 2012-10-05 at 15:48 +0100, Al Viro wrote:
> > On Fri, Oct 05, 2012 at 02:44:24PM +0100, James Bottomley wrote:
> > > On Fri, 2012-10-05 at 12:07 +0100, James Bottomley wrote:
> > > > I tried out the code at
> > > >
> > > > git://git.kernel.org/pub/scm/linux/kernel/git/viro/signal.git
> > > > experimental-kernel_thread
> > > >
> > > > and it gives me this panic on boot.
> > >
> > > OK, found the fix: the idle thread is a kernel thread, but it doesn't
> > > come through kernel_thread(). The fix is to check for it (fortunately
> > > it has the signal usp == 0).
> >
> > Um... I see, but I really wonder if that's the right fix. FWIW, sparc
> > will have the same problem... Hell knows. OTOH, it's a nice way to
> > get of implicit interplay between copy_thread() and idle_regs() - note
> > that SMP architectures doing default idle_regs() need to be damn careful
> > about what they do in their "is that kernel thread" logics; all-zeros
> > pt_regs might give varying results on user_mode(regs) tests, etc.
> > Might be better to go for
> > if (p->flags & PF_KTHREAD) {
> > if (!usp) {
> > we are starting an idle thread
> > } else {
> > we are setting things up for kernel_thread()
> > }
> > } else {
> > we are forking
> > }
> > kind of logics, looking at regs only in the last case. And to hell with
> > (separate and overridable) idle_regs() once everything goes that way...
>
> But there's not a lot of point. forking an idle thread actually doesn't
> care about any of the register execution setup because it never really
> uses it to execute. That's why it was safe for us to use the user
> thread setup ... I suppose the interior of the kernel thread case could
> be conditioned on if (usp).
BTW, speaking of parisc copy_thread()... Why the hell do we bother
with *cregs = *pregs in userland case? It's a part of task_struct,
after all, and we have copied that wholesale in arch_dup_task_struct().
Another thing: why do we bother with
STREG %r30,PT_GR21(%r1)
in fork wrapper? We bloody well know what the offset will be, after all -
right in the beginning of that sucker we'd done
LDREG TI_TASK-THREAD_SZ_ALGN-FRAME_SIZE(%r30), %r1
so we rely on %r30 having been (unsigned long)current_thread_info() +
THREAD_SZ_ALGN + FRAME_SIZE. Then we add FRAME_SIZE again. IOW, the
offset is a known constant. Hell, in child_return you rely on its
value... While we are at it, I'm not sure you need to go through
wrapper_exit on the way out in parent - saving cr27 can be done via
e.g. r28 instead of r3, at which point you can simply branch to
sys_clone() with no work left for wrapper_exit. *Child* obviously
needs to restore these registers, so let it do that in child_return,
but why bother in parent? After all, we are talking about the callee-saved
registers, so sys_clone() is going to revert whatever changes it makes
to them...
BTW, TIF_SYSCALL_TRACE and singlestepping are turned off in child, so I don't
see any need for child_return to know where the parent had come from - it
won't have anything to do in tracesys_exit anyway.
I've folded your fixes and pushed the result; I've added (again, completely
untested) optimizations along the lines of the above on top of those, as
a separate commit. Comments?
--
To unsubscribe from this list: send the line "unsubscribe linux-parisc" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
