Re: [PATCH v11 63/65] drm_buddy: fix 64-bit truncation in power-of-2 rounding

[Louis Chauvet]

On Fri, 13 Mar 2026 07:20:28 -0600, Jim Cromie <jim.cromie@xxxxxxxxx> wrote:
> [...]
> drm_test_buddy_alloc_exceeds_max_order() uses the on a u64 value,
> where they silently truncate the 10GB allocation, giving unexpected
> success in DRM-CI.  (see below the snip).
> 
> Fix this by replacing the standard macros with safe 64-bit
> power-of-two calculations using ilog2().

This is a general DRM bug, if you send a new iteration, can you move it at
then start so it can be applied easly?

>
>
> diff --git a/drivers/gpu/buddy.c b/drivers/gpu/buddy.c
> index b27761246d4b..ff158cc1d27e 100644
> --- a/drivers/gpu/buddy.c
> +++ b/drivers/gpu/buddy.c
> @@ -915,7 +915,7 @@ static int __alloc_contig_try_harder(struct gpu_buddy *mm,
>  	u64 modify_size;
>  	int err;
>  
> -	modify_size = rounddown_pow_of_two(size);
> +	modify_size = 1ULL << ilog2(size);

Thanks for catching this issue!

To avoid this kind of issue later / in other parts of the kernel, maybe you
can change the macro itself to properly handle u64? I am thinking about
something similar to ilog2[1]:

( \
	(sizeof(n) <= 4) ?		\
	__rounddown_pow_of_two_u32(n) :	\
	__rounddown_pow_of_two_u64(n)	\
)

[1]:https://elixir.bootlin.com/linux/v6.19.8/source/include/linux/log2.h#L156-L164

-- 
Louis Chauvet <louis.chauvet@xxxxxxxxxxx>