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

[Alex Williamson]

On Fri, 29 Apr 2016 10:06:11 +0300
"Kirill A. Shutemov" <kirill@xxxxxxxxxxxxx> wrote:

> On Thu, Apr 28, 2016 at 08:45:42PM -0600, Alex Williamson wrote:
> > On Fri, 29 Apr 2016 03:51:06 +0300
> > "Kirill A. Shutemov" <kirill@xxxxxxxxxxxxx> wrote:
> >   
> > > On Fri, Apr 29, 2016 at 01:21:27AM +0200, Andrea Arcangeli wrote:  
> > > > Hello Alex and Kirill,
> > > > 
> > > > On Thu, Apr 28, 2016 at 12:58:08PM -0600, Alex Williamson wrote:    
> > > > > > > specific fix to this code is not applicable.  It also still occurs on
> > > > > > > kernels as recent as v4.6-rc5, so the issue hasn't been silently fixed
> > > > > > > yet.  I'm able to reproduce this fairly quickly with the above test,
> > > > > > > but it's not hard to imagine a test w/o any iommu dependencies which
> > > > > > > simply does a user directed get_user_pages_fast() on a set of userspace
> > > > > > > addresses, retains the reference, and at some point later rechecks that
> > > > > > > a new get_user_pages_fast() results in the same page address.  It    
> > > > 
> > > > Can you try to "git revert 1f25fe20a76af0d960172fb104d4b13697cafa84"
> > > > and then apply the below patch on top of the revert?
> > > > 
> > > > Totally untested... if I missed something and it isn't correct, I hope
> > > > this brings us in the right direction faster at least.
> > > > 
> > > > Overall the problem I think is that we need to restore full accuracy
> > > > and we can't deal with false positive COWs (which aren't entirely
> > > > cheap either... reading 512 cachelines should be much faster than
> > > > copying 2MB and using 4MB of CPU cache). 32k vs 4MB. The problem of
> > > > course is when we really need a COW, we'll waste an additional 32k,
> > > > but then it doesn't matter that much as we'd be forced to load 4MB of
> > > > cache anyway in such case. There's room for optimizations but even the
> > > > simple below patch would be ok for now.
> > > > 
> > > > From 09e3d1ff10b49fb9c3ab77f0b96a862848e30067 Mon Sep 17 00:00:00 2001
> > > > From: Andrea Arcangeli <aarcange@xxxxxxxxxx>
> > > > Date: Fri, 29 Apr 2016 01:05:06 +0200
> > > > Subject: [PATCH 1/1] mm: thp: calculate page_mapcount() correctly for THP
> > > >  pages
> > > > 
> > > > This allows to revert commit 1f25fe20a76af0d960172fb104d4b13697cafa84
> > > > and it provides fully accuracy with wrprotect faults so page pinning
> > > > will stop causing false positive copy-on-writes.
> > > > 
> > > > Signed-off-by: Andrea Arcangeli <aarcange@xxxxxxxxxx>
> > > > ---
> > > >  mm/util.c | 5 +++--
> > > >  1 file changed, 3 insertions(+), 2 deletions(-)
> > > > 
> > > > diff --git a/mm/util.c b/mm/util.c
> > > > index 6cc81e7..a0b9f63 100644
> > > > --- a/mm/util.c
> > > > +++ b/mm/util.c
> > > > @@ -383,9 +383,10 @@ struct address_space *page_mapping(struct page *page)
> > > >  /* Slow path of page_mapcount() for compound pages */
> > > >  int __page_mapcount(struct page *page)
> > > >  {
> > > > -	int ret;
> > > > +	int ret = 0, i;
> > > >  
> > > > -	ret = atomic_read(&page->_mapcount) + 1;
> > > > +	for (i = 0; i < HPAGE_PMD_NR; i++)
> > > > +		ret = max(ret, atomic_read(&page->_mapcount) + 1);
> > > >  	page = compound_head(page);
> > > >  	ret += atomic_read(compound_mapcount_ptr(page)) + 1;
> > > >  	if (PageDoubleMap(page))    
> > > 
> > > You are right about the cause. I spend some time on wrong path: I was only
> > > able to trigger the bug with numa balancing enabled, so I assumed
> > > something is wrong in that code...
> > > 
> > > I would like to preserve current page_mapcount() behaviouts.
> > > I think this fix is better:  
> > 
> > This also seems to work in my testing, but assuming all else being
> > equal, there is a performance difference between the two for this test
> > case in favor of Andrea's solution.  Modifying the test to exit after
> > the first set of iterations, my system takes on average 107s to complete
> > with the solution below or 103.5s with the other approach.  Please note
> > that I have every mm debugging option I could find enabled and THP
> > scanning full speed on the system, so I don't know how this would play
> > out in a more tuned configuration.
> > 
> > The only reason I noticed is that I added a side test to sleep a random
> > number of seconds and kill the test program because sometimes killing
> > the test triggers errors.  I didn't see any errors with either of these
> > solutions, but suspected the first solution was completing more
> > iterations for similar intervals.  Modifying the test to exit seems to
> > prove that true.
> > 
> > I can't speak to which is the more architecturally correct solution,
> > but there may be a measurable performance difference to consider.  
> 
> 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.

I can't explain it either, I won't claim to understand either solution,
but with all the kernel hacking vm debug/sanity options disabled, there
still appears to be a very slight advantage to Andrea's proposal.  I
expect the test program should show this even if you're having a
difficult time using it to reproduce the bug.  I ran Andrea's patch
overnight, no issues reported.  Running your patch for an extended test
now.  Thanks,

Alex