Re: ERROR: "__umoddi3" [drivers/net/ethernet/mellanox/mlx5/core/mlx5_core.ko] undefined!

[Michal Kubecek]

On Mon, Sep 30, 2019 at 08:40:31PM +0200, Borislav Petkov wrote:
> On Mon, Sep 30, 2019 at 09:55:16AM -0700, Stephen Hemminger wrote:
> > Could also us div_u64_rem here?
> 
> Yah, the below seems to work and the resulting asm looks sensible to me
> but someone should definitely double-check me as I don't know this code
> at all.
> 
> Thx.
> 
> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/steering/dr_icm_pool.c b/drivers/net/ethernet/mellanox/mlx5/core/steering/dr_icm_pool.c
> index 913f1e5aaaf2..b4302658e5f8 100644
> --- a/drivers/net/ethernet/mellanox/mlx5/core/steering/dr_icm_pool.c
> +++ b/drivers/net/ethernet/mellanox/mlx5/core/steering/dr_icm_pool.c
> @@ -137,7 +137,7 @@ dr_icm_pool_mr_create(struct mlx5dr_icm_pool *pool,
>  
>  	icm_mr->icm_start_addr = icm_mr->dm.addr;
>  
> -	align_diff = icm_mr->icm_start_addr % align_base;
> +	div_u64_rem(icm_mr->icm_start_addr, align_base, &align_diff);
>  	if (align_diff)
>  		icm_mr->used_length = align_base - align_diff;
>  
> 

While this fixes 32-bit builds, it breaks 64-bit ones as align_diff is
64-bit and div_u64_rem expects pointer to u32. :-(

I checked that align_base is always a power of two so that we could get
away with

	align_diff = icm_mr->icm_start_addr & (align_base - 1)

I'm not sure, however, if it's safe to assume align_base will always
have to be a power of two or if we should add a check for safety.

(Cc-ing also author of commit 29cf8febd185 ("net/mlx5: DR, ICM pool
memory allocator").)

Michal