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

[Zi Yan <ziy@nvidia.com>]

[-- Attachment #1: Type: text/plain, Size: 3941 bytes --]

On 9 Apr 2024, at 5:31, Baolin Wang wrote:

> On 2024/4/8 22:23, Johannes Weiner wrote:
>> On Mon, Apr 08, 2024 at 09:38:20AM +0200, 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,
>>> 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.
>>> Actually wondering why we're not seeing -next failures already (or maybe I
>>> just haven't noticed).
>>
>> Good catch indeed.
>>
>> Trying to understand why I didn't notice this during testing, and I
>> think it's because I had order-10 pageblocks in my config. There is
>> this in compaction_capture():
>>
>> 	if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
>> 		return false;
>>
>> Most compaction is for order-9 THPs on movable blocks, so I didn't get
>> much capturing in practice in order for that leak to be noticable.
>
> This makes me wonder why not use 'cc->migratetype' for migratetype comparison, so that low-order (like mTHP) compaction can directly get the released pages, which could avoid some compaction scans without mixing the migratetype?
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 2facf844ef84..7a64020f8222 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -622,7 +622,7 @@ compaction_capture(struct capture_control *capc, struct page *page,
>          * and vice-versa but no more than normal fallback logic which can
>          * have trouble finding a high-order free page.
>          */
> -       if (order < pageblock_order && migratetype == MIGRATE_MOVABLE)
> +       if (order < pageblock_order && capc->cc->migratetype != migratetype)
>                 return false;
>
>         capc->page = page;

It is worth trying, since at the original patch time mTHP was not present and
not capturing any MIGRATE_MOVABLE makes sense. But with your change, the capture
will lose the opportunity of letting an unmovable request use a reclaimable
pageblock and vice-versa, like the comment says. Please change the comment
as well and we should monitor potential unmovable and reclaimable regression.


--
Best Regards,
Yan, Zi

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 854 bytes --]