Re: [PATCH 6/7] memcg: add per cgroup writeback pages accounting

[Sha Zhengju <handai.szj@gmail.com>]

On 07/09/2012 12:18 PM, Kamezawa Hiroyuki wrote:
> (2012/07/09 13:14), Fengguang Wu wrote:
>> On Mon, Jul 09, 2012 at 11:36:11AM +0800, Sha Zhengju wrote:
>>> On 07/08/2012 10:53 PM, Fengguang Wu wrote:
>>>>> @@ -2245,7 +2252,10 @@ int test_set_page_writeback(struct page *page)
>>>>>   {
>>>>>       struct address_space *mapping = page_mapping(page);
>>>>>       int ret;
>>>>> +    bool locked;
>>>>> +    unsigned long flags;
>>>>>
>>>>> +    mem_cgroup_begin_update_page_stat(page,&locked,&flags);
>>>>>       if (mapping) {
>>>>>           struct backing_dev_info *bdi = mapping->backing_dev_info;
>>>>>           unsigned long flags;
>>>>> @@ -2272,6 +2282,8 @@ int test_set_page_writeback(struct page *page)
>>>>>       }
>>>>>       if (!ret)
>>>>>           account_page_writeback(page);
>>>>> +
>>>>> +    mem_cgroup_end_update_page_stat(page,&locked,&flags);
>>>>>       return ret;
>>>>>
>>>>>   }
>>>> Where is the MEM_CGROUP_STAT_FILE_WRITEBACK increased?
>>>>
>>>
>>> It's in account_page_writeback().
>>>
>>>   void account_page_writeback(struct page *page)
>>>   {
>>> +    mem_cgroup_inc_page_stat(page, MEM_CGROUP_STAT_FILE_WRITEBACK);
>>>       inc_zone_page_state(page, NR_WRITEBACK);
>>>   }
>>
>> I didn't find that chunk, perhaps it's lost due to rebase..
>>
>>> There isn't a unified interface to dec/inc writeback accounting, so
>>> I just follow that.
>>> Maybe we can rework account_page_writeback() to also account
>>> dec in?
>>
>> The current seperate inc/dec paths are fine. It sounds like
>> over-engineering if going any further.
>>
>> I'm a bit worried about some 3rd party kernel module to call
>> account_page_writeback() without 
>> mem_cgroup_begin/end_update_page_stat().
>> Will that lead to serious locking issues, or merely inaccurate
>> accounting?
>>
>
> Ah, Hm. Maybe it's better to add some debug check in
>  mem_cgroup_update_page_stat(). rcu_read_lock_held() or some.
>

This also apply to account_page_dirtied()... But as an "range" lock, I 
think it's common
in current kernel: just as set_page_dirty(), the caller should call it 
under the page lock
(in most cases) and it's his responsibility to guarantee correctness. I 
can add some
comments or debug check as reminding but I think i can only do so...


Thanks,
Sha