Re: [PATCH] mm/page_alloc: Fix try_alloc_pages

[Vlastimil Babka]

On 4/1/25 05:23, Alexei Starovoitov wrote:
> From: Alexei Starovoitov <ast@xxxxxxxxxx>
> 
> Fix an obvious bug. try_alloc_pages() should set_page_refcounted.
> 
> Fixes: 97769a53f117 ("mm, bpf: Introduce try_alloc_pages() for opportunistic page allocation")
> Signed-off-by: Alexei Starovoitov <ast@xxxxxxxxxx>

Acked-by: Vlastimil BAbka <vbabka@xxxxxxx>

> ---
> 
> As soon as I fast forwarded and rerun the tests the bug was
> seen immediately.
> I'm completely baffled how I managed to lose this hunk.

I think the earlier versions were done on older base than v6.14-rc1 which
acquired efabfe1420f5 ("mm/page_alloc: move set_page_refcounted() to callers
of get_page_from_freelist()")

> I'm pretty sure I manually tested various code paths of
> trylock logic with CONFIG_DEBUG_VM=y.
> Pure incompetence :(
> Shame.
> ---
>  mm/page_alloc.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index ffbb5678bc2f..c0bcfe9d0dd9 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -7248,6 +7248,9 @@ struct page *try_alloc_pages_noprof(int nid, unsigned int order)
>  
>  	/* Unlike regular alloc_pages() there is no __alloc_pages_slowpath(). */
>  
> +	if (page)
> +		set_page_refcounted(page);

Note for the later try-kmalloc integration, slab uses frozen pages now, so
we'll need to split out a frozen variant of this API.

But this is ok as a bugfix for now.

> +
>  	if (memcg_kmem_online() && page &&
>  	    unlikely(__memcg_kmem_charge_page(page, alloc_gfp, order) != 0)) {
>  		free_pages_nolock(page, order);