Re: [PATCH] Revert "mm/page_alloc: protect pcp->batch accesses with ACCESS_ONCE"

[Zefan Li]

On 2016/3/31 9:14, Hekuang wrote:
> Hi
> 
> å 2016/3/30 19:10, Michal Hocko åé:
>> On Wed 30-03-16 18:51:12, Hekuang wrote:
>>> hi
>>>
>>> å 2016/3/30 18:38, Mel Gorman åé:
>>>> On Wed, Mar 30, 2016 at 10:22:07AM +0000, He Kuang wrote:
>>>>> This reverts commit 998d39cb236fe464af86a3492a24d2f67ee1efc2.
>>>>>
>>>>> When local irq is disabled, a percpu variable does not change, so we can
>>>>> remove the access macros and let the compiler optimize the code safely.
>>>>>
>>>> batch can be changed from other contexts. Why is this safe?
>>>>
>>> I've mistakenly thought that per_cpu variable can only be accessed by that
>>> cpu.
>> git blame would point you to 998d39cb236f ("mm/page_alloc: protect
>> pcp->batch accesses with ACCESS_ONCE"). I haven't looked into the code
>> deeply to confirm this is still the case but it would be a good lead
>> that this is not that simple. ACCESS_ONCE resp. {READ,WRITE}_ONCE are
>> usually quite subtle so I would encourage you or anybody else who try to
>> remove them to study the code and the history deeper before removing
>> them.
>>
> Thank you for responding, I've read that commit and related articles and not sending
> mail casually, though you may think it's a stupid patch. I'm a beginner and I think
> sending mails to maillist is a effective way to learn kernel, And, sure i'll be more careful and
> well prepared next time :)
> 

pcp->batch can be changed in a different cpu. You may read percpu_pagelist_fraction_sysctl_handler()
to see how that can happen.