On 12/20/19 1:21 AM, Jan Kara wrote:
...

quoted

So, ideas and next steps:

1. Assuming that you *are* hitting this, I think I may have to fall back to
implementing the "deferred" part of this design, as part of this series, after
all. That means:

  For the pin/unpin calls at least, stop treating all pages as if they are
  a cluster of PAGE_SIZE pages; instead, retrieve a huge page as one page.
  That's not how it works now, and the need to hand back a huge array of
  subpages is part of the problem. This affects the callers too, so it's not
  a super quick change to make. (I was really hoping not to have to do this
  yet.)

Does that mean that you would need to make all GUP users huge page aware?
Otherwise I don't see how what you suggest would work... And I don't think
making all GUP users huge page aware is realistic (effort-wise) or even
wanted (maintenance overhead in all those places).

Well, pretty much yes. It's really just the pin_user_pages*() callers, but
the internals, follow_page() and such, are so interconnected right now that
it would probably blow up into a huge effort, as you point out.

I believe there might be also a different solution for this: For
transparent huge pages, we could find a space in 'struct page' of the
second page in the huge page for proper pin counter and just account pins
there so we'd have full width of 32-bits for it.

								Honza

OK, let me pursue that. Given that I shouldn't need to handle pages
splitting, it should be not *too* bad.

I am starting to think that I should just post the first 9 or so 
prerequisite patches (first 9 patches, plus the v4l2 fix that arguably should 
have been earlier in the sequence I guess), as 5.6 candidates, while I go
back to the drawing board here. 

thanks,
-- 
John Hubbard
NVIDIA