Re: [PATCH 2/2] x86: modify show_shared_cpu_map in intel_cacheinfo

[Mike Travis <travis@sgi.com>]

Bert Wesarg wrote:
> On Fri, Mar 28, 2008 at 7:19 PM, Mike Travis <travis@sgi.com> wrote:
>>  > Aren't the most cpumaps (like cpu/cpu*/topology/*_siblings or
>>  > node/node*/cpumap) bitmasks?
>>
>>  I did an informal survey and you are right, the majority of references do use
>>  cpumask_scnprintf instead of cpulist_scnprintf.  Maybe the later function was
>>  added later?
>>
>>  To me though, it would seem that:
>>
>>  240-255
>>
>>  is more readable than:
>>
>>  00000000,00000000,00000000,00000000,00000000,00000000,00000000,0000ffff
>>
>>  And as I mentioned, bitmask_parselist() [libbitmask(3)] does parse the output.
> But libbitmask has a bitmask_parsehex() too. (but thanks for the
> pointer to this code).
> 
> Anyway, your above example is wrong, the most significant bits comes first:
> 
> ffff0000,00000000,00000000,00000000,00000000,00000000,00000000,00000000
> 
> This makes it not more readable, but I think readability isn't in this
> case of that much importance.

The original problem was how to avoid allocating a large stack space to display
cpu ids.  By using cpulist_scnprintf, it accomplishes this without, what I think
is too much pain.  If it's really that much of a problem, I will rework this patch.
But the length of the line with 4096 cpus will be 1152 bytes  Is this really
better?

> 
> I further think, this problem could be easily solved, if NR_CPUS and
> possibly your nr_cpus_ids is somehow exported to user space.
> 
> With this information, the user is not surprised to see more that 1024
> bits (=CPU_SETSIZE, which is currently the glibc constant for the
> sched_{set,get}affinity() API). Also the glibc has the new variable
> cpu_set_t size API (since 2.7).

Yes, thanks.  That is being dealt with in another task.

Thanks,
Mike