Re: [PATCH v2] mm: Don't fault around userfaultfd-registered regions on reads

[Andrea Arcangeli]

On Fri, Dec 04, 2020 at 02:23:29PM -0500, Peter Xu wrote:
> If we see [1]:
> 
>   if (!pte_present && !pte_none && pte_swp_uffd_wp && not_anonymous_vma && !is_migration_entry)
> 
> Then it's fundamentally the same as:
> 
>   swp_entry(0, _UFFD_SWP_UFFD_WP) && !vma_is_anonymous(vma)

Yes conceptually it's the same, but in practice it's different value
in the raw pte if you set  swp_offset  to _UFFD_SWP_UFFD_WP.

Setting swp_offset to _UFFD_SWP_UFFD_WP is just confusing, it's better
magic type 1 offset 0 then to even touch _UFFD_SWP_UFFD_WP if you
reserve the magic value in the swap entry.

pte_swp_uffd_wp or _UFFD_SWP_UFFD_WP are the raw pte value.

swp_entry(0, _UFFD_SWP_UFFD_WP) is not the raw pte value, and it has
to be first converted to the raw pte value with swp_entry_to_pte, so
the _UFFD_SWP_UFFD_WP gets shifted left.

If we use _UFFD_SWP_UFFD_WP it looks much cleaner to keep it in the
pte, not in the swp entry, since then you can use the already existing
methods that only can take in input the pte_t (not the swp_entry_t).

Thanks,
Andrea