Re: [PATCH v0] mm/page_alloc: Cleanup for __del_page_from_free_list()

[Zi Yan <ziy@nvidia.com>]

On 1 Oct 2025, at 0:38, jinji zhong wrote:

>> On 30 Sep 2025, at 9:55, Vlastimil Babka wrote:
>
>>> On 9/25/25 10:50, zhongjinji wrote:
>>>> It is unnecessary to set page->private in __del_page_from_free_list().
>>>>
>>>> If the page is about to be allocated, page->private will be cleared by
>>>> post_alloc_hook() before the page is handed out. If the page is expanded
>>>> or merged, page->private will be reset by set_buddy_order, and no one
>>>> will retrieve the page's buddy_order without the PageBuddy flag being set.
>>>> If the page is isolated, it will also reset page->private when it
>>>> succeeds.
>>>
>>> Seems correct.
>
>> This means high order free pages will have head[2N].private set to a non-zero
>> value, where head[N*2].private is 1, head[N*(2^2)].private is 2, ...
>> head[N*(2^M)].private is M and head[0].private is the actual free page order.
>> If such a high order free page is used as high order folio, it should be fine.
>> But if user allocates a non-compound high order page and uses split_page()
>> to get a list of order-0 pages from this high order page, some pages will
>> have non zero private. I wonder if these users are prepared for that.
>
> Having non-empty page->private in tail pages of non-compound high-order
> pages is not an issue, as pages from the pcp lists never guarantee their
> initial state. If ensuring empty page->private for tail pages is required,

Sure. But is it because all page allocation users return used pages with
->private set back to 0? And can all page allocation users handle non-zero
->private? Otherwise, it can cause subtle bugs.

> we should handle this in prep_new_page(), similar to the approach taken in
> prep_compound_page().
>
>> For example, kernel/events/ring_buffer.c does it. In its comment, it says
>> “set its first page's private to this order; !PagePrivate(page) means it's
>> just a normal page.”
>> (see https://elixir.bootlin.com/linux/v6.17/source/kernel/events/ring_buffer.c#L634)
>
> PagePrivate is a flag in page->flags that indicates page->private is
> already in use. While PageBuddy serves a similar purpose, it additionally
> signifies that the page is part of the buddy system.

OK. You mean ->private will never be used if PagePrivate is not set
in ring buffer code?

If you are confident about it is OK to make some pages’ ->private not being
zero at allocation, I am not going to block the patch.

>
>> I wonder if non zero page->private would cause any issue there.
>
>> Maybe split_page() should set all page->private to 0.
>
>> Let me know if I get anything wrong.
>
>>>
>>>> Since __del_page_from_free_list() is a hot path in the kernel, it would be
>>>> better to remove the unnecessary set_page_private().
>>>>
>>>> Signed-off-by: zhongjinji <zhongjinji@honor.com>
>>>
>>> Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
>>>
>>>> ---
>>>>  mm/page_alloc.c | 1 -
>>>>  1 file changed, 1 deletion(-)
>>>>
>>>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>>>> index d1d037f97c5f..1999eb7e7c14 100644
>>>> --- a/mm/page_alloc.c
>>>> +++ b/mm/page_alloc.c
>>>> @@ -868,7 +868,6 @@ static inline void __del_page_from_free_list(struct page *page, struct zone *zon
>>>>
>>>>  	list_del(&page->buddy_list);
>>>>  	__ClearPageBuddy(page);
>>>> -	set_page_private(page, 0);
>>>>  	zone->free_area[order].nr_free--;
>>>>
>>>>  	if (order >= pageblock_order && !is_migrate_isolate(migratetype))
>
>
>> Best Regards,
>> Yan, Zi


Best Regards,
Yan, Zi