Re: [Bug #11342] Linux 2.6.27-rc3: kernel BUG at mm/vmalloc.c - bisected

[Mike Travis <travis@sgi.com>]

Ingo Molnar wrote:
> * Christoph Lameter <cl@linux-foundation.org> wrote:
> 
>> Alan D. Brunelle wrote:
>>
>>> I think you're right: the kernel as a whole may not be ready for 4,096
>>> CPUs apparently...
>> Mike has been working diligently on getting all these cpumasks off the 
>> stack for the last months and has created an infrastructure to do 
>> this. So I think we are close. It might just be a matter of merging 
>> some more patches that are still left in Ingo's tree.
> 
> hm, there are no such patches left that i know of - the only bits in 
> -tip are the zero-based percpu, which was found to be a bit fragile in 
> testing:

Yes, it's just a case of new changes abusing the stack.
> 
>  earth4:~/tip> git-log-line --author=Travis linus..
>  d379497: Zero based percpu: infrastructure to rebase the per cpu area to zero
>  b3a0cb4: x86: extend percpu ops to 64 bit
> 
> [and it has no relevance to stack footprint.]
> 
> So i dont think the current cpumask_t approach will work. We simply 
> should not get into an endless fight against the windmills that 
> introduce on-stack cpumask_t again and again. We should just take the 
> plunge once and do a clean alloc/free cpumask model. Most of the hotpath 
> cpumasks are constant or pre-constructed, so they are not a real issue.

It would have been nice to know this 9 months ago... ;-)
> 
> Plus, on the general question of stack footprint problems and the 
> difficulty of debugging them, the worst-case stack footprint tracer i 
> wrote for -rt some time ago should be dusted off as well and put into 
> ftrace. David has something quite close to that for Sparc64 already.
> 
> 	Ingo

I'll start experimenting with globally changing cpumask_t to be a pointer,
and see what falls out.

Thanks,
Mike