Re: [syzbot] [mm?] WARNING in unmap_page_range (2)

[David Hildenbrand]

> It should be fine, as:
>
> static void make_uffd_wp_pte(struct vm_area_struct *vma,
> 			     unsigned long addr, pte_t *pte)
> {
> 	pte_t ptent = ptep_get(pte);
>
> #ifndef CONFIG_USERFAULTFD_
>
> 	if (pte_present(ptent)) {
> 		pte_t old_pte;
>
> 		old_pte = ptep_modify_prot_start(vma, addr, pte);
> 		ptent = pte_mkuffd_wp(ptent);
> 		ptep_modify_prot_commit(vma, addr, pte, old_pte, ptent);
> 	} else if (is_swap_pte(ptent)) {
> 		ptent = pte_swp_mkuffd_wp(ptent);
> 		set_pte_at(vma->vm_mm, addr, pte, ptent);
> 	} else {                                      <----------------- this must be pte_none() already
> 		set_pte_at(vma->vm_mm, addr, pte,
> 			   make_pte_marker(PTE_MARKER_UFFD_WP));
> 	}
> }

Indeed! Is pte_swp_mkuffd_wp() reasonable for pte markers? I rememebr 
that we don't support multiple markers yet, so it might be good enough.

> > 2) We get the error on arm64, which does *not* support uffd-wp. Do we
> >     maybe end up calling make_uffd_wp_pte() and place a pte marker, even
> >     though we don't have CONFIG_PTE_MARKER_UFFD_WP?
> >
> >
> > static inline bool pte_marker_entry_uffd_wp(swp_entry_t entry)
> > {
> > #ifdef CONFIG_PTE_MARKER_UFFD_WP
> > 	return is_pte_marker_entry(entry) &&
> > 	    (pte_marker_get(entry) & PTE_MARKER_UFFD_WP);
> > #else
> > 	return false;
> > #endif
> > }
> >
> > Will always return false without CONFIG_PTE_MARKER_UFFD_WP.
> >
> > But make_uffd_wp_pte() might just happily place an entry. Hm.
> >
> >
> > The following might fix the problem:

[...]

> I'd like to double check with Muhammad (as I didn't actually follow his
> work in the latest versions.. quite a lot changed), but I _think_
> fundamentally we missed something important in the fast path, and I think
> it applies even to archs that support uffd..
>
> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
> index e91085d79926..3b81baabd22a 100644
> --- a/fs/proc/task_mmu.c
> +++ b/fs/proc/task_mmu.c
> @@ -2171,7 +2171,8 @@ static int pagemap_scan_pmd_entry(pmd_t *pmd, unsigned long start,
>                  return 0;
>          }
>
> -       if (!p->vec_out) {
> +       if (!p->vec_out &&
> +           (p->arg.flags & PM_SCAN_WP_MATCHING))

Ouch, yes. So that's the global fence I was wondering where to find it.

--
Cheers,

David / dhildenb