2024-06-12 08:30:04

by Oscar Salvador

Subject: [PATCH] mm/hugetlb: Guard dequeue_hugetlb_folio_nodemask against NUMA_NO_NODE uses

dequeue_hugetlb_folio_nodemask() expects a preferred node where to get the
hugetlb page from.
It does not expect, though, users to pass NUMA_NO_NODE, otherwise we will
get trash when trying to get the zonelist from that node.
All current users are careful enough to not pass NUMA_NO_NODE, but
it opens the door for new users to get this wrong since it is not
documented [0].
Guard against this by getting the local nid if NUMA_NO_NODE was passed.

[0] https://lore.kernel.org/linux-mm/[email protected]/

Closes: https://lore.kernel.org/linux-mm/[email protected]/
Reported-and-tested-by: [email protected]
Signed-off-by: Oscar Salvador <[email protected]>
---
mm/hugetlb.c | 4 ++++
1 file changed, 4 insertions(+)

diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 3518321f6598..6a5ea898e4da 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -1355,6 +1355,10 @@ static struct folio *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_
struct zoneref *z;
int node = NUMA_NO_NODE;

+ /* 'nid' should not be NUMA_NO_NODE. Try to catch any misuse of it and rectifiy. */
+ if (nid == NUMA_NO_NODE)
+ nid = numa_node_id();
+
zonelist = node_zonelist(nid, gfp_mask);

retry_cpuset:
--
2.45.2

2024-06-12 08:54:00

by Muchun Song

Subject: Re: [PATCH] mm/hugetlb: Guard dequeue_hugetlb_folio_nodemask against NUMA_NO_NODE uses

> On Jun 12, 2024, at 16:29, Oscar Salvador <[email protected]> wrote:
>
> dequeue_hugetlb_folio_nodemask() expects a preferred node where to get the
> hugetlb page from.
> It does not expect, though, users to pass NUMA_NO_NODE, otherwise we will
> get trash when trying to get the zonelist from that node.
> All current users are careful enough to not pass NUMA_NO_NODE, but
> it opens the door for new users to get this wrong since it is not
> documented [0].
> Guard against this by getting the local nid if NUMA_NO_NODE was passed.
>
> [0] https://lore.kernel.org/linux-mm/[email protected]/
>
> Closes: https://lore.kernel.org/linux-mm/[email protected]/
> Reported-and-tested-by: [email protected]
> Signed-off-by: Oscar Salvador <[email protected]>

Reviewed-by: Muchun Song <[email protected]>

Thanks.

2024-06-12 17:26:44

by Kasireddy, Vivek

Subject: RE: [PATCH] mm/hugetlb: Guard dequeue_hugetlb_folio_nodemask against NUMA_NO_NODE uses

> dequeue_hugetlb_folio_nodemask() expects a preferred node where to get
> the
> hugetlb page from.
> It does not expect, though, users to pass NUMA_NO_NODE, otherwise we
> will
> get trash when trying to get the zonelist from that node.
> All current users are careful enough to not pass NUMA_NO_NODE, but
> it opens the door for new users to get this wrong since it is not
> documented [0].
> Guard against this by getting the local nid if NUMA_NO_NODE was passed.
>
> [0] https://lore.kernel.org/linux-
> mm/[email protected]/
>
> Closes: https://lore.kernel.org/linux-
> mm/[email protected]/
> Reported-and-tested-by:
> [email protected]
> Signed-off-by: Oscar Salvador <[email protected]>
> ---
> mm/hugetlb.c | 4 ++++
> 1 file changed, 4 insertions(+)
>
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index 3518321f6598..6a5ea898e4da 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -1355,6 +1355,10 @@ static struct folio
> *dequeue_hugetlb_folio_nodemask(struct hstate *h, gfp_t gfp_
> struct zoneref *z;
> int node = NUMA_NO_NODE;
>
> + /* 'nid' should not be NUMA_NO_NODE. Try to catch any misuse of it
> and rectifiy. */
> + if (nid == NUMA_NO_NODE)
> + nid = numa_node_id();
> +
> zonelist = node_zonelist(nid, gfp_mask);
Acked-by: Vivek Kasireddy <[email protected]>

>
> retry_cpuset:
> --
> 2.45.2