[PATCH] mm/vmscan: restore allowed mask in alloc_demote_folio()

[PATCH] mm/vmscan: restore allowed mask in alloc_demote_folio()

[Bing Jiao]

In alloc_demote_folio(), mtc->nmask is set to NULL for the first
allocation. If that succeeds, it returns without restoring mtc->nmask
to allowed_mask. For subsequent allocations from the migrate_pages()
batch, mtc->nmask will be NULL. If the target node then becomes full,
the fallback allocation will use nmask = NULL, allocating from any
node allowed by the task cpuset, which for kswapd is all nodes.

To address this issue, restore the mtc->nmask to its original allowed
nodemask after the first allocation.

Signed-off-by: Bing Jiao <bingjiao@google.com>
---
 mm/vmscan.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/vmscan.c b/mm/vmscan.c
index cbffc0a27824..b42abd17aee7 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -985,11 +985,11 @@ static struct folio *alloc_demote_folio(struct folio *src,
 	mtc->nmask = NULL;
 	mtc->gfp_mask |= __GFP_THISNODE;
 	dst = alloc_migration_target(src, (unsigned long)mtc);
+	mtc->nmask = allowed_mask;
 	if (dst)
 		return dst;

 	mtc->gfp_mask &= ~__GFP_THISNODE;
-	mtc->nmask = allowed_mask;

 	return alloc_migration_target(src, (unsigned long)mtc);
 }
--
2.53.0.473.g4a7958ca14-goog

Re: [PATCH] mm/vmscan: restore allowed mask in alloc_demote_folio()

[David Hildenbrand (Arm)]

On 3/2/26 08:03, Bing Jiao wrote:
> In alloc_demote_folio(), mtc->nmask is set to NULL for the first
> allocation. If that succeeds, it returns without restoring mtc->nmask
> to allowed_mask. For subsequent allocations from the migrate_pages()
> batch, mtc->nmask will be NULL. If the target node then becomes full,
> the fallback allocation will use nmask = NULL, allocating from any
> node allowed by the task cpuset, which for kswapd is all nodes.
> 
> To address this issue, restore the mtc->nmask to its original allowed
> nodemask after the first allocation.
> 

That would be

Fixes: 320080272892 ("mm/demotion: demote pages according to allocation fallback order")

?

> Signed-off-by: Bing Jiao <bingjiao@google.com>
> ---
>  mm/vmscan.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index cbffc0a27824..b42abd17aee7 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -985,11 +985,11 @@ static struct folio *alloc_demote_folio(struct folio *src,
>  	mtc->nmask = NULL;
>  	mtc->gfp_mask |= __GFP_THISNODE;
>  	dst = alloc_migration_target(src, (unsigned long)mtc);
> +	mtc->nmask = allowed_mask;
>  	if (dst)
>  		return dst;
> 
>  	mtc->gfp_mask &= ~__GFP_THISNODE;
> -	mtc->nmask = allowed_mask;
> 
>  	return alloc_migration_target(src, (unsigned long)mtc);
>  }
> --
> 2.53.0.473.g4a7958ca14-goog
> 

Maybe we should just not touch the original mtc?

diff --git a/mm/vmscan.c b/mm/vmscan.c
index de62225b381a..f07716e5389e 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -985,9 +985,9 @@ static void folio_check_dirty_writeback(struct folio *folio,
 static struct folio *alloc_demote_folio(struct folio *src,
                unsigned long private)
 {
+       struct migration_target_control *mtc, target_nid_mtc;
        struct folio *dst;
        nodemask_t *allowed_mask;
-       struct migration_target_control *mtc;
 
        mtc = (struct migration_target_control *)private;
 
@@ -1001,15 +1001,12 @@ static struct folio *alloc_demote_folio(struct folio *src,
         * a demotion of cold pages from the target memtier. This can result
         * in the kernel placing hot pages in slower(lower) memory tiers.
         */
-       mtc->nmask = NULL;
-       mtc->gfp_mask |= __GFP_THISNODE;
-       dst = alloc_migration_target(src, (unsigned long)mtc);
+       target_nid_mtc = *mtc;
+       target_nid_mtc.nmask = NULL;
+       target_nid_mtc.gfp_mask |= __GFP_THISNODE;
+       dst = alloc_migration_target(src, (unsigned long)&target_nid_mtc);
        if (dst)
                return dst;
-
-       mtc->gfp_mask &= ~__GFP_THISNODE;
-       mtc->nmask = allowed_mask;
-
        return alloc_migration_target(src, (unsigned long)mtc);
 }
 


-- 
Cheers,

David

Re: [PATCH] mm/vmscan: restore allowed mask in alloc_demote_folio()

[Bing Jiao]

On Mon, Mar 02, 2026 at 09:00:07AM +0100, David Hildenbrand (Arm) wrote:
> On 3/2/26 08:03, Bing Jiao wrote:
> > In alloc_demote_folio(), mtc->nmask is set to NULL for the first
> > allocation. If that succeeds, it returns without restoring mtc->nmask
> > to allowed_mask. For subsequent allocations from the migrate_pages()
> > batch, mtc->nmask will be NULL. If the target node then becomes full,
> > the fallback allocation will use nmask = NULL, allocating from any
> > node allowed by the task cpuset, which for kswapd is all nodes.
> >
> > To address this issue, restore the mtc->nmask to its original allowed
> > nodemask after the first allocation.
> >
>
> That would be
>
> Fixes: 320080272892 ("mm/demotion: demote pages according to allocation fallback order")
>
> ?

Thanks for pointing it out. Will add it in the new patch.

> > Signed-off-by: Bing Jiao <bingjiao@google.com>
> > ---
> >  mm/vmscan.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > index cbffc0a27824..b42abd17aee7 100644
> > --- a/mm/vmscan.c
> > +++ b/mm/vmscan.c
> > @@ -985,11 +985,11 @@ static struct folio *alloc_demote_folio(struct folio *src,
> >  	mtc->nmask = NULL;
> >  	mtc->gfp_mask |= __GFP_THISNODE;
> >  	dst = alloc_migration_target(src, (unsigned long)mtc);
> > +	mtc->nmask = allowed_mask;
> >  	if (dst)
> >  		return dst;
> >
> >  	mtc->gfp_mask &= ~__GFP_THISNODE;
> > -	mtc->nmask = allowed_mask;
> >
> >  	return alloc_migration_target(src, (unsigned long)mtc);
> >  }
> > --
> > 2.53.0.473.g4a7958ca14-goog
> >
>
> Maybe we should just not touch the original mtc?
>
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index de62225b381a..f07716e5389e 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -985,9 +985,9 @@ static void folio_check_dirty_writeback(struct folio *folio,
>  static struct folio *alloc_demote_folio(struct folio *src,
>                 unsigned long private)
>  {
> +       struct migration_target_control *mtc, target_nid_mtc;
>         struct folio *dst;
>         nodemask_t *allowed_mask;
> -       struct migration_target_control *mtc;
>
>         mtc = (struct migration_target_control *)private;
>
> @@ -1001,15 +1001,12 @@ static struct folio *alloc_demote_folio(struct folio *src,
>          * a demotion of cold pages from the target memtier. This can result
>          * in the kernel placing hot pages in slower(lower) memory tiers.
>          */
> -       mtc->nmask = NULL;
> -       mtc->gfp_mask |= __GFP_THISNODE;
> -       dst = alloc_migration_target(src, (unsigned long)mtc);
> +       target_nid_mtc = *mtc;
> +       target_nid_mtc.nmask = NULL;
> +       target_nid_mtc.gfp_mask |= __GFP_THISNODE;
> +       dst = alloc_migration_target(src, (unsigned long)&target_nid_mtc);
>         if (dst)
>                 return dst;
> -
> -       mtc->gfp_mask &= ~__GFP_THISNODE;
> -       mtc->nmask = allowed_mask;
> -
>         return alloc_migration_target(src, (unsigned long)mtc);
>  }
>
>
>
> --
> Cheers,
>
> David

Thank you for the suggestion, David.

I agree that not touching the original mtc is a better. It makes
the distinction between the two allocation attempts much clearer
and avoids the side-effect bug. Will update it then.

Best,
Bing