Bad page state splats on arm64, v4.11-rc{3,4}

[Catalin Marinas]

On Mon, Apr 03, 2017 at 12:37:51PM +0100, Will Deacon wrote:
> On Mon, Apr 03, 2017 at 11:56:29AM +0100, Mark Rutland wrote:
> > On Fri, Mar 31, 2017 at 06:58:45PM +0100, Mark Rutland wrote:
> > > Hi,
> > > 
> > > I'm seeing intermittent bad page state splats on arm64 with 4.11-rc3 and
> > > v4.11-rc4. I have not tested earlier kernels, or other architectures.
> > > 
> > > So far, it looks like the flags are always bad in the same
> > > way:
> > > 
> > > 	bad because of flags: 0x80(waiters)
> > > 
> > > ... though I don't know if that's definitely the case for splat 4, the
> > > BUG at mm/page_alloc.c:800.
> > > 
> > > I see this in QEMU VMs launched by Syzkaller, triggering once every few
> > > hours. So far, I have not been able to reproduce the issue in any other
> > > way (including using syz-repro).
> > 
> > It looks like this may be an issue with the arm64 HUGETLB code.
> > 
> > I wasn't able to trigger the issue over the weekend on a kernel with
> > HUGETLBFS disabled. There are known issues with our handling of
> > contiguous entries, and this might be an artefact of that.
> 
> After chatting with Punit, it looks like this might be because the GUP
> code doesn't handle huge ptes (which we create using the contiguous hint),
> so follow_page_pte ends up with one of those and goes wrong. In particular,
> the migration code will certainly do the wrong thing.
> 
> I'll probably revert the contiguous support (again) if testing indicates
> that it makes this issue disappear.

It might be worth checking with Punit's patches as well:

https://marc.info/?l=linux-arm-kernel&m=149089199018167&w=2

-- 
Catalin