Re: [PATCH 2/2] mm/rmap: Fixup copying of soft dirty and uffd ptes

On Tuesday, 25 August 2020 1:43:59 AM AEST Peter Xu wrote:

...

--- a/mm/migrate.c +++ b/mm/migrate.c @@ -2427,9 +2427,11 @@ static int migrate_vma_collect_pmd(pmd_t *pmdp,

	entry = make_migration_entry(page, mpfn &
	
				     MIGRATE_PFN_WRITE);
	
	swp_pte = swp_entry_to_pte(entry);

• 	if (pte_soft_dirty(pte))
  

• 	if ((is_swap_pte(pte) && pte_swp_soft_dirty(pte))
  

• 		|| (!is_swap_pte(pte) && pte_soft_dirty(pte)))
  
  	swp_pte = pte_swp_mksoft_dirty(swp_pte);
  
  

• 	if (pte_uffd_wp(pte))
  

• 	if ((is_swap_pte(pte) && pte_swp_uffd_wp(pte))
  

• 		|| (!is_swap_pte(pte) && pte_uffd_wp(pte)))
  
  	swp_pte = pte_swp_mkuffd_wp(swp_pte);
  

  set_pte_at(mm, addr, ptep, swp_pte);

The worst case is we'll call is_swap_pte() four times for each entry. Also considering we know it's not a pte_none() when reach here, how about:

if (pte_present(pte)) { // pte handling of both soft dirty and uffd-wp } else { // swap handling of both soft dirty and uffd-wp }

?

Works for me, I'd missed we knew it was not a pte_none() so will respin. Thanks!

- Alistair

Thanks,