Re: [PATCH 04/14] mm,migration: Allow the migration of PageSwapCache pages

[Andrea Arcangeli <aarcange@redhat.com>]

On Fri, Apr 23, 2010 at 09:39:48PM +0200, Andrea Arcangeli wrote:
> On Fri, Apr 23, 2010 at 08:23:12PM +0100, Mel Gorman wrote:
> > On Fri, Apr 23, 2010 at 08:31:35PM +0200, Andrea Arcangeli wrote:
> > > Hi Mel,
> > > 
> > > On Thu, Apr 22, 2010 at 04:44:43PM +0100, Mel Gorman wrote:
> > > > heh, I thought of a similar approach at the same time as you but missed
> > > > this mail until later. However, with this approach I suspect there is a
> > > > possibility that two walkers of the same anon_vma list could livelock if
> > > > two locks on the list are held at the same time. Am still thinking of
> > > > how it could be resolved without introducing new locking.
> > > 
> > > Trying to understand this issue and I've some questions. This
> > > vma_adjust and lock inversion troubles with the anon-vma lock in
> > > rmap_walk are a new issue introduced by the recent anon-vma changes,
> > > right?
> > > 
> > 
> > In a manner of speaking. There was no locking going on but prior to the
> > anon_vma changes, there would have been only one anon_vma lock and the
> > fix would be easier - just take the lock on anon_vma->lock while the
> > VMAs are being updated.
> 
> So it was very much a bug before too and we could miss to find some
> pte mapping the page if vm_start was adjusted?

Well I looked deeper into it myself as I wanted to have this bit (and
other bits) sorted out in aa.git, and definitely this is a bug
introduced by the newest anon-vma changes in 2.6.34-rc so aa.git
cannot be affected as it's using the 2.6.33 anon-vma (and prev) code.

vma_adjust already takes the anon_vma->lock and of course I also
further verified that trying to apply your snippet to vma_adjust
results in immediately deadlock as the very same lock is already taken
in my tree as it's the same anon-vma (simpler). So aa.git will be
immune from these bugs for now.