Re: [PATCH v11 00/25] mm/gup: track dma-pinned pages: FOLL_PIN

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu 19-12-19 12:30:31, John Hubbard wrote:
> On 12/19/19 5:26 AM, Leon Romanovsky wrote:
> > On Mon, Dec 16, 2019 at 02:25:12PM -0800, John Hubbard wrote:
> > > Hi,
> > > 
> > > This implements an API naming change (put_user_page*() -->
> > > unpin_user_page*()), and also implements tracking of FOLL_PIN pages. It
> > > extends that tracking to a few select subsystems. More subsystems will
> > > be added in follow up work.
> > 
> > Hi John,
> > 
> > The patchset generates kernel panics in our IB testing. In our tests, we
> > allocated single memory block and registered multiple MRs using the single
> > block.
> > 
> > The possible bad flow is:
> >   ib_umem_geti() ->
> >    pin_user_pages_fast(FOLL_WRITE) ->
> >     internal_get_user_pages_fast(FOLL_WRITE) ->
> >      gup_pgd_range() ->
> >       gup_huge_pd() ->
> >        gup_hugepte() ->
> >         try_grab_compound_head() ->
> 
> Hi Leon,
> 
> Thanks very much for the detailed report! So we're overflowing...
> 
> At first look, this seems likely to be hitting a weak point in the
> GUP_PIN_COUNTING_BIAS-based design, one that I believed could be deferred
> (there's a writeup in Documentation/core-api/pin_user_page.rst, lines
> 99-121). Basically it's pretty easy to overflow the page->_refcount
> with huge pages if the pages have a *lot* of subpages.
> 
> We can only do about 7 pins on 1GB huge pages that use 4KB subpages.
> Do you have any idea how many pins (repeated pins on the same page, which
> it sounds like you have) might be involved in your test case,
> and the huge page and system page sizes? That would allow calculating
> if we're likely overflowing for that reason.
> 
> 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).

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
-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR



[Index of Archives]     [Linux Input]     [Video for Linux]     [Gstreamer Embedded]     [Mplayer Users]     [Linux USB Devel]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [Yosemite Backpacking]

  Powered by Linux