Re: [PATCH 1/2] mm: avoid spurious 'bad pmd' warning messages

[Dave Hansen]

On 05/17/2017 10:16 AM, Ross Zwisler wrote:
> @@ -3061,7 +3061,7 @@ static int pte_alloc_one_map(struct vm_fault *vmf)
>  	 * through an atomic read in C, which is what pmd_trans_unstable()
>  	 * provides.
>  	 */
> -	if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
> +	if (pmd_devmap(*vmf->pmd) || pmd_trans_unstable(vmf->pmd))
>  		return VM_FAULT_NOPAGE;

I'm worried we are very unlikely to get this right in the future.  It's
totally not obvious what the ordering requirement is here.

Could we move pmd_devmap() and pmd_trans_unstable() into a helper that
gets the ordering right and also spells out the ordering requirement?