Re: [PATCH v1 1/1] mm/gup: skip pinnable check for refs==1

[John Hubbard]

On 1/31/22 23:43, John Hubbard wrote:
> > > pin_user_pages
> > >    __get_user_pages_locked
> > >      follow_page_mask
> > >        follow_page_pte
> > >          try_grab_page
> > >            !is_pinnable_page(page)
> > >              return NULL;
> > >          return ERR_PTR(-ENOMEM);
> > >       return -ENOMEM without faultin_page
> >
> > Yes, that's all clear.
> > ...oh, but I guess you're pointing out that it's always going to be
> page-at-a-time this deep in the pin_user_pages() call path. Which is true.
>
> I hadn't worked through how to fix this yet, my initial reaction was
> that allowing single refs to go through, while prohibiting multiple refs,
> was clearly *not* the way to go.

OK, so after looking at this some more, I think that the real problem
with commit 54d516b1d62f ("mm/gup: small refactoring: simplify
try_grab_page()") is that it funnels fast and slow gup through the same
routine (try_grab_compound_head()). And the problem with *that*, is that
try_grab_compound_head() is coded up with that assumption that it is
being called *only* by fast-gup:

		/*
		 * Can't do FOLL_LONGTERM + FOLL_PIN gup fast path if not in a
		 * right zone, so fail and let the caller fall back to the slow
		 * path.
		 */
		if (unlikely((flags & FOLL_LONGTERM) &&
			     !is_pinnable_page(page)))
			return NULL;

Now, to fix that, I'd really rather not conflate "refs == 1" with "on
the slow path", because that's a conceptual mismatch.

So, other ideas:

a) Remove the above check, and fail fast gup at a different point for
the (FOLL_LONGTERM && !is_pinnable) case. Haven't looked closely at this
yet.

b) Pass in FOLL_FAST_ONLY from all call sites *except* try_grag_page().
Skip the above check in slow-gup (!FOLL_FAST_ONLY) cases.

This makes the code ugly, though, and I'm just listing it here for
completeness.

c) Just call the entire refactoring idea a mistake, and roll it back
either entirely, or enough to keep fast and slow gup separate.

Unless a better idea shows up, (c) is probably the way to go, I think...


thanks,
--
John Hubbard
NVIDIA