Re: [PATCH] mm: reuse the unshared swapcache page in do_wp_page

[David Hildenbrand <david@redhat.com>]

On 20.01.22 18:48, Nadav Amit wrote:
> 
>> On Jan 20, 2022, at 6:15 AM, David Hildenbrand <david@redhat.com> wrote:
>>
>> On 17.01.22 14:31, zhangliang (AG) wrote:
>>> Sure, I will do that :)
>>
>> I'm polishing up / testing the patches and might send something out for discussion shortly.
>> Just a note that on my branch was a version with a wrong condition that should have been fixed now.
>>
> 
> Sorry for being late for the discussion.
> 
> David, does any of it regards the lru_cache_add() reference issue that I
> mentioned? [1]

No, unfortunately not in that part of my work. *Maybe* we could also try
to handle that reference similarly to the swapcache, but the question is
if we can't wait for PageAnonExclusive.

Right now I have the following in mind to get most parts working as
exptected:

1. Optimize reuse logic for the swapcache as it seems to be easy
2. Streamline COW logic and remove reuse_swap_page() -- fix the CVE for
   THP.
3. Introduce PageAnonExclusive and allow FOLL_PIN only on
   PageAnonExclusive pages.
4. Convert O_DIRECT to FOLL_PIN

We will never ever have to copy a page PageAnonExclusive page in the COW
handler and can immediately reuse it without even locking the page. The
existing reuse logic is essentially then used to reset PageAnonExclusive
on a page (thus it makes sense to work on it) where the flag is not set
anymore -- or on a fresh page if we have to copy.

That implies that all these additional references won't care if your app
doesn't fork() or KSM isn't active. Consequently, anything that
read-protects anonymous pages will work as expected and should be as
fast as it gets.

Sounds good? At least to me. If only swap/migration entries wouldn't be
harder to handle than I'd wish, that's why it's taking a little and will
take a little longer.

> 
> Seems to me that any solution should also regard this problem, or am I
> missing something?

The bigger picture has to handle it, yes!

-- 
Thanks,

David / dhildenb