Re: [PATCH 0/4] Some more lock_page work..

[Linus Torvalds <torvalds@linux-foundation.org>]

On Thu, Oct 15, 2020 at 2:43 AM Kirill A. Shutemov <kirill@shutemov.name> wrote:
>
> Okay, I see what you propose.
>
> But I don't think it addresses race with try_to_unmap():

I don't think it needs to.

Remember: the map_pages() thing is called only for when the page
tables are empty.

So try_to_unmap() will never see the pte entry, and there is nothing
to race with.

So sure, it can "race" with try_to_unmap like you say, but who cares?
The "race" is no different from taking the page lock _after_
try_to_unmap() already ran (and didn't see anything because the page
hadn't been mapped yet).

IOW I don't think try_to_unmap() really matters. The race you outline
can already happen with the "trylock()" - no different from the
trylock just succeeding after the unlock_page().

So you can think of map_pages() as all happening after try_to_unmap()
has already succeeded - and didn't see the new pte that hasn't been
filled in yet.

I do think there is a real race, but it is is with "__remove_mapping()".

That still happens under the page lock, but it doesn't actually
_depend_ on the page lock as far as I can tell. Because I think the
real protection there is that

        if (!page_ref_freeze(page, refcount))
                goto cannot_free;

it that code sees "oh, somebody else has a reference to the page, we
can't remove the mapping".

But it's entirely possible that I don't understand your worry, and I
overlooked something. If so, can you explain using smaller words,
please ;)

             Linus