Re: [PATCH] x86, percpu: revert commit fe8e0c25

[Brian Gerst]

On Tue, Oct 26, 2010 at 9:34 AM, Tejun Heo <tj@xxxxxxxxxx> wrote:
> Hello,
>
> On 10/26/2010 03:25 PM, Ingo Molnar wrote:
>>
>> * Tejun Heo <tj@xxxxxxxxxx> wrote:
>>
>>> Commit fe8e0c25 (x86, 32-bit: Align percpu area and irq stacks to THREAD_SIZE)
>>> aligned PERCPU section to THREAD_SIZE which can be larger than PAGE_SIZE,
>>> introduced DEFINE_PER_CPU_MULTIPAGE_ALIGNED() and used it to make irq stacks
>>> aligned to THREAD_SIZE on x86_32.
>>>
>>> This won't work. ÂThe PERCPU output section is used as the template to prepare the
>>> percpu area and the actual percpu area is _alwasy_ aligned to PAGE_SIZE whether
>>> the source area is aligned to larger size or not.
>>
>> The problem is, this will reintroduce a nasty boot crash which commit fe8e0c25
>> fixed. Do you say that fe8e0c25 didnt have the alignment effect?
>
> AFAICS, not in a way which is correct. ÂThe patch probably made the
> following two differences.
>
> * The stack in the template area is THREAD_SIZE aligned. ÂIf something
> Âwas dereferencing it before percpu init, this could have helped.
> ÂIIRC, x86 early init code does use the template area.
>
> * The percpu address would be THREAD_SIZE aligned while the translated
> Âkernel address for each cpu wouldn't be. ÂFor masking stack pointer
> Âto find out task struct, I don't think aligning the percpu address
> Âwould have been helpful.
>
> It's simply broken and needs to be reverted. ÂIf the patch somehow
> fixed boot crash, yeah, we probably want to put a fix for it first
> tho.
>
> Thanks.

Probably the best fix is to go back to allocating the stacks with
get_free_pages(), and only keep the pointers in percpu memory.

--
Brian Gerst
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/