Re: 5.0-rc1 KVM inspired "BUG: Bad page state in process" spew

[Sean Christopherson]

On Wed, Jan 09, 2019 at 04:03:16PM +0100, Mike Galbraith wrote:
> On Wed, 2019-01-09 at 15:42 +0100, Adam Borowski wrote:
> > On Wed, Jan 09, 2019 at 06:38:58AM +0100, Mike Galbraith wrote:
> > > KVM seems to be busted in master ATM.  All I have to do to have host
> > > start screaming and maybe exploding (if the guest doesn't do so first)
> > > is to try to install a (obese in this case) kernel over nfs mount of
> > > the host in a guest.
> > > 
> > > Kernel producing the spew below is 3bd6e94, config attached.
> > 
> > I get same, except that the BUGs were preceded by a bunch of warnings,
> 
> Yeah, I was in too much of a rush...
> 
> > > homer: # grep BUG: /netconsole.log
> > > [ 1531.909703] BUG: Bad page state in process X  pfn:100491
> > > [ 1531.958141] BUG: Bad page state in process systemd-journal  pfn:100412
> > > [ 1532.662359] BUG: Bad page state in process X  pfn:10043f
> > > [ 1532.664033] BUG: Bad page state in process X  pfn:10044d
> > > [ 1532.686433] BUG: Bad page state in process systemd-journal  pfn:1027b0
> > 
> > the first one being:
> > 
> > Jan  9 00:41:22 umbar kernel: [74122.790461] WARNING: CPU: 2 PID: 26769 at arch/x86/kvm/mmu.c:830 mmu_spte_clear_track_bits+0x7e/0x100
> 
> ...I also get oodles of those.

There's also a bugzilla bug that's probably the same thing:
https://bugzilla.kernel.org/show_bug.cgi?id=202189.

I'm 99.9% confident this is only manifests when reclaiming from the guest,
i.e. host is swapping out a VM's memory.  The WARNING is complaining that
KVM is trying to reclaim a page before it has been removed from KVM's MMU,
i.e. a use-after-free scenario is imminent, and the stack trace shows the
kernel is reclaming.  The bug also listed host swapping as a possible
requirement, and last but not least, I was only able to reproduce this by
forcing reclaim.

I'll try to bisect.