Re: [PATCH 1/1] x86: Change _node_to_cpumask_ptr to return const ptr

[Mike Travis <travis@sgi.com>]

Vegard Nossum wrote:
> On Tue, Jul 8, 2008 at 10:51 PM, Mike Travis <travis@sgi.com> wrote:
>> Vegard Nossum wrote:
>>> On Tue, Jul 8, 2008 at 8:05 PM, Mike Travis <travis@sgi.com> wrote:
>>>>>> Note: I did not change node_to_cpumask_ptr() in include/asm-generic/topology.h
>>>>>>      as node_to_cpumask_ptr_next() does change the cpumask value.
>>>>> Hmmm. Does it really?
>>>>>
>>>>> #define node_to_cpumask_ptr_next(v, node)                               \
>>>>>                           _##v = node_to_cpumask(node)
>>>>>
>>>>> This doesn't seem to modify it?
>>>> Well I thought about it.  The pointer (*v) does not change
>>>> but the underlying cpumask variable is updated with the
>>>> cpumask for the (supposedly) new node number.  You can see
>>>> that in this code snippet from kernel/sched.c:
>>>>
>>>>        for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
>>>>                int next_node = find_next_best_node(node, &used_nodes);
>>>>
>>>>                node_to_cpumask_ptr_next(nodemask, next_node);
>>>>                cpus_or(*span, *span, *nodemask);
>>>>        }
>>>>
>>>> In the optimized (x86_64) case, the pointer is simply modified
>>>> to point to the new node_to_cpumask_map[node] entry.  It remains
>>>> a pointer to a const value.
>>>>
>>>> But the non-optimized version replaces the const cpumask value
>>>> with the new cpumask value.  Isn't this breaking the const
>>>> attribute?
>>> No, I think the pointer really should be const. This doesn't guarantee
>>> that the value doesn't change behind our backs, it only prevents us
>>> from modifying it ourselves.
>>>
>>>
>>> Vegard
>>>
>> Is this what you had in mind:
>>
>>
>> --- linux-2.6.tip.orig/include/asm-generic/topology.h
>> +++ linux-2.6.tip/include/asm-generic/topology.h
>> @@ -60,7 +60,7 @@
>>  #ifndef node_to_cpumask_ptr
>>
>>  #define        node_to_cpumask_ptr(v, node)                                    \
>> -               cpumask_t _##v = node_to_cpumask(node), *v = &_##v
>> +               const cpumask_t _##v = node_to_cpumask(node), *v = &_##v
>>
>>  #define node_to_cpumask_ptr_next(v, node)                              \
>>                          _##v = node_to_cpumask(node)
>>
>>
>> (It's taking a while as now I need to do some cross-compile testing.)
> 
> Actually, no.
> 
> We don't want the _##v to be const, do we? What do you think about
> this? (Watch out for whitespace munges)
> 
> diff --git a/include/asm-generic/topology.h b/include/asm-generic/topology.h
> index a6aea79..56957f2 100644
> --- a/include/asm-generic/topology.h
> +++ b/include/asm-generic/topology.h
> @@ -60,7 +60,8 @@
>  #ifndef node_to_cpumask_ptr
> 
>  #define        node_to_cpumask_ptr(v, node)
> -               cpumask_t _##v = node_to_cpumask(node), *v = &_##v
> +               cpumask_t _##v = node_to_cpumask(node);                 \
> +               const cpumask_t *v = &_##v;
> 
>  #define node_to_cpumask_ptr_next(v, node)                              \
>                           _##v = node_to_cpumask(node)
> 
> 
> Vegard
> 

Thanks.  That was my alternative though I was hoping to confirm that
the compiler detected the overwrite by node_to_cpumask_ptr_next().
Unfortunately every non-x86 cross-compile that I have for a machine
that has NUMA is failing in some other way.

I'll resubmit with that change.

Mike