[PATCH] mm: migrate: Simplify the refcount validation when migrating hugetlb mapping

[PATCH] mm: migrate: Simplify the refcount validation when migrating hugetlb mapping

[Baolin Wang]

There is no need to validate the hugetlb page's refcount before trying
to freeze the hugetlb page's expected refcount, instead we can just
rely on the page_ref_freeze() to simplify the validation.

Moreover we are always under the page lock when migrating the hugetlb
page mapping, which means nowhere else can remove it from the page cache,
so we can remove the xas_load() validation under the i_pages lock.

Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Cc: Matthew Wilcox <willy@infradead.org>
---
 mm/migrate.c | 5 -----
 1 file changed, 5 deletions(-)

diff --git a/mm/migrate.c b/mm/migrate.c
index a3d8c2b..b267827 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -477,11 +477,6 @@ int migrate_huge_page_move_mapping(struct address_space *mapping,
 
 	xas_lock_irq(&xas);
 	expected_count = 2 + page_has_private(page);
-	if (page_count(page) != expected_count || xas_load(&xas) != page) {
-		xas_unlock_irq(&xas);
-		return -EAGAIN;
-	}
-
 	if (!page_ref_freeze(page, expected_count)) {
 		xas_unlock_irq(&xas);
 		return -EAGAIN;
-- 
1.8.3.1

Re: [PATCH] mm: migrate: Simplify the refcount validation when migrating hugetlb mapping

[Mike Kravetz]

On 4/11/22 04:34, Baolin Wang wrote:
> There is no need to validate the hugetlb page's refcount before trying
> to freeze the hugetlb page's expected refcount, instead we can just
> rely on the page_ref_freeze() to simplify the validation.
> 
> Moreover we are always under the page lock when migrating the hugetlb
> page mapping, which means nowhere else can remove it from the page cache,
> so we can remove the xas_load() validation under the i_pages lock.

I agree with the reasoning here.

> 
> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
> Cc: Matthew Wilcox <willy@infradead.org>
> ---
>  mm/migrate.c | 5 -----
>  1 file changed, 5 deletions(-)

Matthew added that extra xas_load() and check both here and in
migrate_page_move_mapping.  However, migrate_page_move_mapping has been
converted to folio and does not appear to have the same check.

Ideally we will convert hugetlbfs to folio someday and rewrite
migrate_huge_page_move_mapping as well.

Would be good to get Matthew's opinion.

Acked-by: Mike Kravetz <mike.kravetz@oracle.com>
-- 
Mike Kravetz

> 
> diff --git a/mm/migrate.c b/mm/migrate.c
> index a3d8c2b..b267827 100644
> --- a/mm/migrate.c
> +++ b/mm/migrate.c
> @@ -477,11 +477,6 @@ int migrate_huge_page_move_mapping(struct address_space *mapping,
>  
>  	xas_lock_irq(&xas);
>  	expected_count = 2 + page_has_private(page);
> -	if (page_count(page) != expected_count || xas_load(&xas) != page) {
> -		xas_unlock_irq(&xas);
> -		return -EAGAIN;
> -	}
> -
>  	if (!page_ref_freeze(page, expected_count)) {
>  		xas_unlock_irq(&xas);
>  		return -EAGAIN;