On 2022/6/12 23:44, Muchun Song wrote:
> On Sat, Jun 11, 2022 at 10:13:52AM +0800, Miaohe Lin wrote:
>> Since commit 5232c63f46fd ("mm: Make compound_pincount always available"),
>> compound_pincount_ptr is stored at first tail page now. So we should call
>> prep_compound_head() after the first tail page is initialized to take
>> advantage of the likelihood of that tail struct page being cached given
>> that we will read them right after in prep_compound_head().
>>
>> Signed-off-by: Miaohe Lin <linmiaohe@xxxxxxxxxx>
>> Cc: Joao Martins <joao.m.martins@xxxxxxxxxx>
>> ---
>> v2:
>> Don't move prep_compound_head() outside loop per Joao.
>> ---
>> mm/page_alloc.c | 17 +++++++++++------
>> 1 file changed, 11 insertions(+), 6 deletions(-)
>>
>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>> index 4c7d99ee58b4..048df5d78add 100644
>> --- a/mm/page_alloc.c
>> +++ b/mm/page_alloc.c
>> @@ -6771,13 +6771,18 @@ static void __ref memmap_init_compound(struct page *head,
>> set_page_count(page, 0);
>>
>> /*
>> - * The first tail page stores compound_mapcount_ptr() and
>> - * compound_order() and the second tail page stores
>> - * compound_pincount_ptr(). Call prep_compound_head() after
>> - * the first and second tail pages have been initialized to
>> - * not have the data overwritten.
>> + * The first tail page stores compound_mapcount_ptr(),
>> + * compound_order() and compound_pincount_ptr(). Call
>> + * prep_compound_head() after the first tail page have
>> + * been initialized to not have the data overwritten.
>> + *
>> + * Note the idea to make this right after we initialize
>> + * the offending tail pages is trying to take advantage
>> + * of the likelihood of those tail struct pages being
>> + * cached given that we will read them right after in
>> + * prep_compound_head().
>> */
>> - if (pfn == head_pfn + 2)
>> + if (unlikely(pfn == head_pfn + 1))
>> prep_compound_head(head, order);
>
> For me it is weird not to put this out of the loop. I saw the reason
> is because of the caching suggested by Joao. But I think this is not
> a hot path and putting it out of the loop may be more intuitive at least
> for me. Maybe this optimization is unnecessary (maybe I am wrong).
> And it will be consistent with prep_compound_page() (at least it does
> not do the similar optimization) if we drop this optimization.
This is also what I thought in the first version. :)
>
> Hi Joao,
>
> I am wondering is it a significant optimization for zone device memory?
> I found this code existed from the 1st version you introduced. So
> I suspect maybe you have some numbers, would you like to share with us?
Those numbers would be really helpful.
>
> Thanks.
Thanks!
>
> .
>
