Re: [PATCH] x86-32: Allocate irq stacks seperate from percpu area

[Tejun Heo <tj@kernel.org>]

Hello,

On 10/27/2010 03:33 PM, Eric Dumazet wrote:
> Le mercredi 27 octobre 2010 à 11:57 +0200, Peter Zijlstra a écrit :
>> On Wed, 2010-10-27 at 08:07 +0200, Eric Dumazet wrote:
>>>> -     irqctx = &per_cpu(hardirq_stack, cpu);
>>>> +     irqctx = (union irq_ctx *)__get_free_pages(THREAD_FLAGS, THREAD_ORDER);
>>>
>>> Hmm, then we lose NUMA affinity for stacks. 
>>
>> I guess we could use:
>>
>>   alloc_pages_node(cpu_to_node(cpu), THREAD_FLAGS, THREAD_ORDER);
>>
>>
> 
> Anyway, I just discovered per_cpu data on my machine (NUMA capable) all
> sit on a single node, if 32bit kernel used.
> 
> # cat /proc/buddyinfo 
> Node 0, zone      DMA      0      1      0      1      2      1      1      0      1      1      3 
> Node 0, zone   Normal     94    251     81     16      3      2      1      2      1      2    187 
> Node 0, zone  HighMem    113     88     47     36     18      5      4      3      2      0    268 
> Node 1, zone  HighMem    154     97     43     16      9      4      3      2      3      2    482 
...
> 
> I presume node 1 having only HighMem could be the reason ?

What does cpu_to_node() on each cpu say?  Also, do you know why
num_possible_cpus() is 32, not 16?

Thanks.

-- 
tejun