The bulk allocation path in alloc_pages_bulk_noprof() currently misses
trace_mm_page_alloc() and kmsan_alloc_page() calls, leaving bulk-allocated
pages invisible to ftrace/BPF/perf and leaving KMSAN shadow memory stale.

Add both calls after set_page_refcounted() in the bulk loop to match
the standard allocation path. The gfp mask passed to kmsan_alloc_page()
is stripped of __GFP_RECLAIM because the bulk loop runs under the PCP
spinlock, and KMSAN's stack depot allocation must not sleep.

Both are no-ops when their respective features are disabled, so there
is no overhead in production kernels.

Suggested-by: Gregory Price <gourry@gourry.net>
Signed-off-by: Qiqi Liu <liuqiqi@kylinos.cn>
---
 mm/page_alloc.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 12fac9084c48..73c73499a051 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -5280,6 +5280,8 @@ unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
 
 		prep_new_page(page, 0, gfp, ALLOC_DEFAULT);
 		set_page_refcounted(page);
+		trace_mm_page_alloc(page, 0, gfp, ac.migratetype);
+		kmsan_alloc_page(page, 0, gfp & ~__GFP_RECLAIM);
 		page_array[nr_populated++] = page;
 	}
 
-- 
2.25.1

On 9/8/26 08:50, Qiqi Liu wrote:
> The bulk allocation path in alloc_pages_bulk_noprof() currently misses
> trace_mm_page_alloc() and kmsan_alloc_page() calls, leaving bulk-allocated
> pages invisible to ftrace/BPF/perf and leaving KMSAN shadow memory stale.
> 
> Add both calls after set_page_refcounted() in the bulk loop to match
> the standard allocation path. The gfp mask passed to kmsan_alloc_page()
> is stripped of __GFP_RECLAIM because the bulk loop runs under the PCP
> spinlock, and KMSAN's stack depot allocation must not sleep.
> 
> Both are no-ops when their respective features are disabled, so there
> is no overhead in production kernels.
> 
> Suggested-by: Gregory Price <gourry@gourry.net>
> Signed-off-by: Qiqi Liu <liuqiqi@kylinos.cn>
> ---
>  mm/page_alloc.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 12fac9084c48..73c73499a051 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -5280,6 +5280,8 @@ unsigned long alloc_pages_bulk_noprof(gfp_t gfp, int preferred_nid,
>  
>  		prep_new_page(page, 0, gfp, ALLOC_DEFAULT);
>  		set_page_refcounted(page);
> +		trace_mm_page_alloc(page, 0, gfp, ac.migratetype);
> +		kmsan_alloc_page(page, 0, gfp & ~__GFP_RECLAIM);

With normal page allocation the set_page_refcounted() happens after
trace+kmsan. While it currently shouldn't matter, it could be more future
proof to keep the same order.

>  		page_array[nr_populated++] = page;
>  	}
>

On Tue, Sep 08, 2026 at 10:38:46AM +0200, Vlastimil Babka wrote:
> With normal page allocation the set_page_refcounted() happens after
> trace+kmsan. While it currently shouldn't matter, it could be more future
> proof to keep the same order.

Hi Vlastimil,

Thanks for the review. You're right — I'll swap the order in v4 so that
trace_mm_page_alloc() and kmsan_alloc_page() come before set_page_refcounted()
to match the normal allocation path.

Best regards,
Qiqi Liu