Hi Jérôme,

On 24 Aug 2018, at 15:25, jglisse@redhat.com wrote:

> From: Jérôme Glisse <jglisse@redhat.com>
>
> Before this patch migration pmd entry (!pmd_present()) would have
> been treated as a bad entry (pmd_bad() returns true on migration
> pmd entry). The outcome was that device driver would believe that
> the range covered by the pmd was bad and would either SIGBUS or
> simply kill all the device's threads (each device driver decide
> how to react when the device tries to access poisonnous or invalid
> range of memory).
>
> This patch explicitly handle the case of migration pmd entry which
> are non present pmd entry and either wait for the migration to
> finish or report empty range (when device is just trying to pre-
> fill a range of virtual address and thus do not want to wait or
> trigger page fault).
>
> Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
> Signed-off-by: Jérôme Glisse <jglisse@redhat.com>
> Cc: Ralph Campbell <rcampbell@nvidia.com>
> Cc: John Hubbard <jhubbard@nvidia.com>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> ---
>  mm/hmm.c | 45 +++++++++++++++++++++++++++++++++++++++------
>  1 file changed, 39 insertions(+), 6 deletions(-)
>
> diff --git a/mm/hmm.c b/mm/hmm.c
> index a16678d08127..659efc9aada6 100644
> --- a/mm/hmm.c
> +++ b/mm/hmm.c
> @@ -577,22 +577,47 @@ static int hmm_vma_walk_pmd(pmd_t *pmdp,
>  {
>  	struct hmm_vma_walk *hmm_vma_walk = walk->private;
>  	struct hmm_range *range = hmm_vma_walk->range;
> +	struct vm_area_struct *vma = walk->vma;
>  	uint64_t *pfns = range->pfns;
>  	unsigned long addr = start, i;
>  	pte_t *ptep;
> +	pmd_t pmd;
>
> -	i = (addr - range->start) >> PAGE_SHIFT;
>
>  again:
> -	if (pmd_none(*pmdp))
> +	pmd = READ_ONCE(*pmdp);
> +	if (pmd_none(pmd))
>  		return hmm_vma_walk_hole(start, end, walk);
>
> -	if (pmd_huge(*pmdp) && (range->vma->vm_flags & VM_HUGETLB))
> +	if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB))
>  		return hmm_pfns_bad(start, end, walk);
>
> -	if (pmd_devmap(*pmdp) || pmd_trans_huge(*pmdp)) {
> -		pmd_t pmd;
> +	if (!pmd_present(pmd)) {
> +		swp_entry_t entry = pmd_to_swp_entry(pmd);
> +
> +		if (is_migration_entry(entry)) {

I think you should check thp_migration_supported() here, since PMD migration is only enabled in x86_64 systems.
Other architectures should treat PMD migration entries as bad.

> +			bool fault, write_fault;
> +			unsigned long npages;
> +			uint64_t *pfns;
> +
> +			i = (addr - range->start) >> PAGE_SHIFT;
> +			npages = (end - addr) >> PAGE_SHIFT;
> +			pfns = &range->pfns[i];
> +
> +			hmm_range_need_fault(hmm_vma_walk, pfns, npages,
> +					     0, &fault, &write_fault);
> +			if (fault || write_fault) {
> +				hmm_vma_walk->last = addr;
> +				pmd_migration_entry_wait(vma->vm_mm, pmdp);
> +				return -EAGAIN;
> +			}
> +			return 0;
> +		}
> +
> +		return hmm_pfns_bad(start, end, walk);
> +	}
>

—
Best Regards,
Yan Zi