Re: [PATCH 10/10] mm: page_alloc: consolidate free page accounting

[Baolin Wang <baolin.wang@linux.alibaba.com>]

On 2024/4/8 15:38, Vlastimil Babka wrote:
> On 4/7/24 12:19 PM, Baolin Wang wrote:
>> On 2024/3/21 02:02, Johannes Weiner wrote:
>>>    
>>> +	account_freepages(page, zone, 1 << order, migratetype);
>>> +
>>>    	while (order < MAX_PAGE_ORDER) {
>>> -		if (compaction_capture(capc, page, order, migratetype)) {
>>> -			__mod_zone_freepage_state(zone, -(1 << order),
>>> -								migratetype);
>>> +		int buddy_mt = migratetype;
>>> +
>>> +		if (compaction_capture(capc, page, order, migratetype))
>>>    			return;
>>> -		}
>>
>> IIUC, if the released page is captured by compaction, then the
>> statistics for free pages should be correspondingly decreased,
>> otherwise, there will be a slight regression for my thpcompact benchmark.
>>
>> thpcompact Percentage Faults Huge
>>                             k6.9-rc2-base        base + patch10 + 2 fixes	
>> Percentage huge-1        78.18 (   0.00%)       71.92 (  -8.01%)
>> Percentage huge-3        86.70 (   0.00%)       86.07 (  -0.73%)
>> Percentage huge-5        90.26 (   0.00%)       78.02 ( -13.57%)
>> Percentage huge-7        92.34 (   0.00%)       78.67 ( -14.81%)
>> Percentage huge-12       91.18 (   0.00%)       81.04 ( -11.12%)
>> Percentage huge-18       89.00 (   0.00%)       79.57 ( -10.60%)
>> Percentage huge-24       90.52 (   0.00%)       80.07 ( -11.54%)
>> Percentage huge-30       94.44 (   0.00%)       96.28 (   1.95%)
>> Percentage huge-32       93.09 (   0.00%)       99.39 (   6.77%)
>>
>> I add below fix based on your fix 2, then the thpcompact Percentage
>> looks good. How do you think for the fix?
> 
> Yeah another well spotted, thanks. "slight regression" is an understatement,

Thanks for helping to confirm.

> this affects not just a "statistics" but very important counter
> NR_FREE_PAGES which IIUC would eventually become larger than reality, make
> the watermark checks false positive and result in depleted reserves etc etc.

Right, agree.

> Actually wondering why we're not seeing -next failures already (or maybe I
> just haven't noticed).
> 
>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>> index 8330c5c2de6b..2facf844ef84 100644
>> --- a/mm/page_alloc.c
>> +++ b/mm/page_alloc.c
>> @@ -805,8 +805,10 @@ static inline void __free_one_page(struct page *page,
>>           while (order < MAX_PAGE_ORDER) {
>>                   int buddy_mt = migratetype;
>>
>> -               if (compaction_capture(capc, page, order, migratetype))
>> +               if (compaction_capture(capc, page, order, migratetype)) {
>> +                       account_freepages(zone, -(1 << order), migratetype);
>>                           return;
>> +               }
>>
>>                   buddy = find_buddy_page_pfn(page, pfn, order, &buddy_pfn);
>>                   if (!buddy)
>>
>> With my fix, the THP percentage looks better:
>>                         k6.9-rc2-base          base + patch10 + 2 fixes	+
>> my fix
>> Percentage huge-1        78.18 (   0.00%)       82.83 (   5.94%)
>> Percentage huge-3        86.70 (   0.00%)       93.47 (   7.81%)
>> Percentage huge-5        90.26 (   0.00%)       94.73 (   4.95%)
>> Percentage huge-7        92.34 (   0.00%)       95.22 (   3.12%)
>> Percentage huge-12       91.18 (   0.00%)       92.40 (   1.34%)
>> Percentage huge-18       89.00 (   0.00%)       85.39 (  -4.06%)
>> Percentage huge-24       90.52 (   0.00%)       94.70 (   4.61%)
>> Percentage huge-30       94.44 (   0.00%)       97.00 (   2.71%)
>> Percentage huge-32       93.09 (   0.00%)       92.87 (  -0.24%)