Re: [RFC][PATCH 02/24] arc: Provide atomic_{or,xor,and}

[Vineet Gupta]

On Friday 10 July 2015 12:35 PM, Peter Zijlstra wrote:
> On Fri, Jul 10, 2015 at 04:30:46AM +0000, Vineet Gupta wrote:
>> > 
>> > Since we are on the topic, the cmpxchg() loop in arch/arc/kernel/smp.c still
>> > irritates me.
>> > Do we need a new set of primitives to operate atomically on non atomic_t data or
>> > does that mean that the data *not* being atomic_t but requiring such semantics is
>> > the fundamental problem and thus needs to be converted first.
> So if you look at the last patch, there's already a few sites that do
> things like:
>
> +       atomic_or(*mask, (atomic_t *)&flushcache_cpumask);
>
> Which is of course ugly as hell, but does work.
>
> Esp. inside arch code.

Right - I don't have issues with using this API - but this requires atomic_t data
type.  The specific cmpxchg() loop that I'm referring to is not for atomic_t - so
that needs to be converted to atomic_t first ?

>
> Now the 'problem' with cmpxchg/xchg, the instructions working on !atomic
> data is:
>
>   http://lkml.kernel.org/r/alpine.LRH.2.02.1406011342470.20831@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
>   http://lkml.kernel.org/r/20140606175316.GV13930@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
>
> And note that includes some arc.

Correct so we don't mix cmpxchg() with normal load/store.

>
> Adding more such primitives will only make it harder on those already
> 'broken' archs.

Not sure if I follow here - my point was not so much about expanding the
atomic_*() API but whether it makes sense to have "some" API for non atomic_t vs.
converting the non atomic_t to atomic_t and then use the API as that is the
fundamental problem for such cases.

-Vineet
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/