[lkml]   [2010]   [Apr]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [BUGFIX][mm][PATCH] fix migration race in rmap_walk
On Sat, Apr 24, 2010 at 11:02:00AM +0900, KAMEZAWA Hiroyuki wrote:
> On Fri, 23 Apr 2010 16:58:01 +0100
> Mel Gorman <> wrote:
> > > I had considered this idea as well as it is vaguely similar to how zones get
> > > resized with a seqlock. I was hoping that the existing locking on anon_vma
> > > would be usable by backing off until uncontended but maybe not so lets
> > > check out this approach.
> > >
> >
> > A possible combination of the two approaches is as follows. It uses the
> > anon_vma lock mostly except where the anon_vma differs between the page
> > and the VMAs being walked in which case it uses the seq counter. I've
> > had it running a few hours now without problems but I'll leave it
> > running at least 24 hours.
> >
> ok, I'll try this, too.
> > ==== CUT HERE ====
> > mm,migration: Prevent rmap_walk_[anon|ksm] seeing the wrong VMA information by protecting against vma_adjust with a combination of locks and seq counter
> >
> > vma_adjust() is updating anon VMA information without any locks taken.
> > In constract, file-backed mappings use the i_mmap_lock. This lack of
> > locking can result in races with page migration. During rmap_walk(),
> > vma_address() can return -EFAULT for an address that will soon be valid.
> > This leaves a dangling migration PTE behind which can later cause a
> > BUG_ON to trigger when the page is faulted in.
> >
> > With the recent anon_vma changes, there is no single anon_vma->lock that
> > can be taken that is safe for rmap_walk() to guard against changes by
> > vma_adjust(). Instead, a lock can be taken on one VMA while changes
> > happen to another.
> >
> > What this patch does is protect against updates with a combination of
> > locks and seq counters. First, the vma->anon_vma lock is taken by
> > vma_adjust() and the sequence counter starts. The lock is released and
> > the sequence ended when the VMA updates are complete.
> >
> > The lock serialses rmap_walk_anon when the page and VMA share the same
> > anon_vma. Where the anon_vmas do not match, the seq counter is checked.
> > If a change is noticed, rmap_walk_anon drops its locks and starts again
> > from scratch as the VMA list may have changed. The dangling migration
> > PTE bug was not triggered after several hours of stress testing with
> > this patch applied.
> >
> > [ Use of a seq counter]
> > Signed-off-by: Mel Gorman <>
> I think this patch is nice!

It looks nice but it still broke after 28 hours of running. The
seq-counter is still insufficient to catch all changes that are made to
the list. I'm beginning to wonder if a) this really can be fully safely
locked with the anon_vma changes and b) if it has to be a spinlock to
catch the majority of cases but still a lazy cleanup if there happens to
be a race. It's unsatisfactory and I'm expecting I'll either have some
insight to the new anon_vma changes that allow it to be locked or Rik
knows how to restore the original behaviour which as Andrea pointed out
was safe.

> Acked-by: KAMEZAWA Hiroyuki <>

Mel Gorman
Part-time Phd Student Linux Technology Center
University of Limerick IBM Dublin Software Lab

 \ /
  Last update: 2010-04-24 12:47    [W:0.098 / U:9.692 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site