Re: [PATCH] coredump: Replace opencoded set_mask_bits()

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Aug 07, 2015 at 09:28:05PM +0530, Vineet Gupta wrote:
> On Friday 07 August 2015 09:15 PM, Peter Zijlstra wrote:
> > On Fri, Aug 07, 2015 at 09:05:06PM +0530, Vineet Gupta wrote:
> >> On Friday 07 August 2015 08:27 PM, Peter Zijlstra wrote:
> >>> On Fri, Aug 07, 2015 at 08:14:03PM +0530, Vineet Gupta wrote:
> >>>
> >>>>> See, I have such a cmpxchg loop in ARC code - originally from Peter :-)
> >>>>> arch/arc/kernel/smp.c. @ipi_data_ptr is NOT atomic_t
> >>>>>
> >>>>> 	do {
> >>>>> 		new = old = ACCESS_ONCE(*ipi_data_ptr);
> >>>>> 		new |= 1U << msg;
> >>>>> 	} while (cmpxchg(ipi_data_ptr, old, new) != old);
> >>>>>
> >>> Well, you'll have atomic_or() real soon now.
> >>
> >> Doesn't help my cause - ipi_data_ptr is not atomic_t - hence my prev question in
> >> this thread
> > 
> > A cast will work :-)
> > 
> 
> How ? We have
> 
> typedef struct {
>  int counter;
> } atomic_t;

ARC is 32bit, right? So int and unsigned long are of the same size.
Therefore:

	atomic_or(1 << msg, (atomic_t *)ipi_data_ptr);

Ugly, yes, but it should DTRT.

> > But yes, ideally everything will be type safe because of those archs
> > that cannot have atomic RmW ops like !ARC_HAS_LLSC.
> 
> Type safe - how / what ?

All atomic stuff restricted to atomic*t and bitmap functions (and
ideally we'd also have bitmap_t to avoid passing random unsigned long *
into bitmap functions and praying it all works, we do, and it doesn't,
well mostly :-).

> > Mixing cmpxchg()/xchg() with regular stores is broken on those.
> 
> Right, but how does that relate to this discussion - perhaps I shd stop talking -
> long friday already :-)

:-)

Well, its a very good argument for why we should not use cmpxchg/xchg on
!atomic*t types, and therefore why the function at hand (set_mask_bit)
should really be on an atomic_t.

That said, it will probably make the fs code fugly for having to use
atomic_t and all its accessors all over the place.