Hi, Andrea,
On Fri, Dec 04, 2020 at 01:12:56PM -0500, Andrea Arcangeli wrote:
> On Thu, Dec 03, 2020 at 11:10:18PM -0500, Andrea Arcangeli wrote:
> > from the pte, one that cannot ever be set in any swp entry today. I
> > assume it can't be _PAGE_SWP_UFFD_WP since that already can be set but
> > you may want to verify it...
>
> I thought more about the above, and I think the already existing
> pte_swp_mkuffd_wp will just be enough without having to reserve an
> extra bitflag if we encode it as a non migration entry.
>
> The check:
>
> if (!pte_present && !pte_none && pte_swp_uffd_wp && not_anonymous_vma && !is_migration_entry)
[1]
>
> should be enough to disambiguate it. When setting it, it'd be enough
> to set the pte to the value _PAGE_SWP_UFFD_WP.
>
> Although if you prefer to check for:
>
> if (!pte_present && !pte_none && swp_type == 1 && swp_offset == 0 && not_anonymous_vma && !is_migration_entry)
[2]
>
> that would do as well.
>
> It's up to you, just my preference is to reuse _PAGE_SWP_UFFD_WP since
> it has already to exist, there are already all the pte_swp_*uffd*
> methods available or uffd-wp cannot work.
Yes, I had the same thought that it would be nice if this special pte can be
still related to _PAGE_SWP_UFFD_WP.
To me, above [2] looks exactly the same as Hugh suggested to check against
swp_type==1 && swp_offset==0, since:
- do_swap_page() basically already means "!pte_present && !pte_none"
- "not_anonymous_vma" seems optional if uffd-wp+shmem will be the first user
of such a swp entry
- "!is_migration_entry" seems optional since if swp_type==1, it will never be
a migration entry
While for above [1] that's the thing I asked besides the current type==1 &
offset=0 proposal. Quotting one of the previous emails:
> So I guess I'll start with type==1 && offset==0.
>
> (PS: I still think "swp_entry(0, _UFFD_SWP_UFFD_WP) && !vma_is_anonymous(vma)"
> could also be a good candidate comparing to "swp_entry(1, 0)" considering
> type==1 here is kind of randomly chosen from all the other numbers except 0;
> but maybe that's not extremely important - the major logic should be the same)
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)
Reasons similar to above.
Thanks!
--
Peter Xu
