Re: [BUG] vfio device assignment regression with THP ref counting redesign

[Andrea Arcangeli]

On Fri, Apr 29, 2016 at 10:06:11AM +0300, Kirill A. Shutemov wrote:
> Hm. I just woke up and haven't got any coffee yet, but I don't why my
> approach would be worse for performance. Both have the same algorithmic
> complexity.

Even before looking at the overall performance, I'm not sure your
patch is really fixing it all: you didn't touch reuse_swap_page which
is used by do_wp_page to know if it can call do_wp_page_reuse. Your
patch would still trigger a COW instead of calling do_wp_page_reuse,
but it would only happen if the page was pinned after the pmd split,
which is probably not what the testcase is triggering. My patch
instead fixed that too.

total_mapcount returns the wrong value for reuse_swap_page, which is
probably why you didn't try to use it there.

The main issue of my patch is that it has a performance downside that
is page_mapcount becomes expensive for all other usages, which is
better than breaking vfio but I couldn't use total_mapcount again
because it counts things wrong in reuse_swap_page.

Like I said there's room for optimizations so today I tried to
optimize more stuff...