Re: [git pull] cpus4096 tree, part 3

[Mike Travis <travis@sgi.com>]

Rusty Russell wrote:
> On Sunday 04 January 2009 07:26:03 Linus Torvalds wrote:
>> On Sat, 3 Jan 2009, Ingo Molnar wrote:
>>>> Has anybody looked at what the stack size is with MAXSMP set with an 
>>>> allyesconfig? And what areas are still problematic, if any? Are we going 
>>>> to have some code-paths that still essentially have 1kB+ of stack space 
>>>> just because they haven't been converted and still have the cpu mask on 
>>>> stack?
>>> ok, indeed testing of that is in order now.
>> Well, since I can compile a allyesconfig pretty quickly, I did the static 
>> part. It looks better than it used to, and I think most of the huge stacks 
>> are totally unrealted to cpu masks. But not all.
>>
>> But it looks like we have a few:
>>
>>  - flush_tlb_current_task:
>> 	cpumask_t cpu_mask;
>>  - flush_tlb_mm:
>> 	cpumask_t cpu_mask;
> ...
>>  - acpi_cpufreq_target:
>> 	cpumask_t online_policy_cpus
> 
> Mike?  These are x86-specific...

I've been testing the heck out of it... ;-)

> 
>>  - local_cpus_show:
>> 	cpumask_t mask;
>>  - local_cpulist_show:
>> 	cpumask_t mask;

Yes, these are in my "real soon now" patchset.  Trivial.
> 
> Yes, this removal is still in my queue.  I'll double-check that all the
> archs have the new "cpumask_of_pcibus". (cpumask:replace-and-remove-pcibus_to_cpumask.patch "cpumask: remove the now-obsoleted pcibus_to_cpumask()").
> 
>> and then we have a number of things that have "struct cpufreq_policy" on 
>> the stack, and those things have two cpumask_t's in each.
> 
> Yep, we have the conversion for that too.  Mike, it's cpumask:convert-drivers_acpi.patch "cpumask: convert struct cpufreq_policy to cpumask_var_t."
> 
That's part of what I'm testing above.

>> The rest of the high-stack-usage cases - from a _very_ quick look - seem 
>> to be unrelated to CPU masks, but in the "more than 1kB of stack" group 
>> about a third (wild handwaving eyeballing) of them do seem to be related 
>> to cpumask.
> 
> Mike was tracking this; I think he has a script to set NR_CPUS small then
> large and dump the changes.

It's looking pretty good, only 11 > 1k and 19 more > 512.
 
Thanks,
Mike