Re: vhost changes (batched) in linux-next after 12/13 trigger random crashes in KVM guests after reboot

[Eugenio Pérez]

On Thu, 2020-02-13 at 10:30 +0100, Christian Borntraeger wrote:
> 
> On 12.02.20 17:34, Eugenio PÃrez wrote:
> > On Tue, 2020-02-11 at 14:13 +0100, Christian Borntraeger wrote:
> > > On 11.02.20 14:04, Eugenio PÃrez wrote:
> > > > On Mon, 2020-02-10 at 12:01 +0100, Christian Borntraeger wrote:
> > > > > On 10.02.20 10:47, Eugenio Perez Martin wrote:
> > > > > > Hi Christian.
> > > > > > 
> > > > > > I'm not able to reproduce the failure with eccb852f1fe6bede630e2e4f1a121a81e34354ab commit. Could you add
> > > > > > more
> > > > > > data?
> > > > > > Your configuration (libvirt or qemu line), and host's dmesg output if any?
> > > > > > 
> > > > > > Thanks!
> > > > > 
> > > > > If it was not obvious, this is on s390x, a big endian system.
> > > > > 
> > > > 
> > > > Hi Christian. Thank you very much for your fast responses.
> > > > 
> > > > Could you try this patch on top of eccb852f1fe6bede630e2e4f1a121a81e34354ab?
> > > 
> > > I still get 
> > > [   43.665145] Guest moved used index from 0 to 289
> > > after some reboots.
> > > 
> > > 
> > > > Thanks!
> > > > 
> > > > From 71d0f9108a18aa894cc0c0c1c7efbad39f465a27 Mon Sep 17 00:00:00 2001
> > > > From: =?UTF-8?q?Eugenio=20P=C3=A9rez?= <
> > > > eperezma@xxxxxxxxxx>
> > > > Date: Tue, 11 Feb 2020 13:19:10 +0100
> > > > Subject: [PATCH] vhost: fix return value of vhost_get_vq_desc
> > > > 
> > > > Before of the batch change, it was the chain's head. Need to keep that
> > > > way or we will not be able to free a chain of descriptors.
> > > > 
> > > > Fixes: eccb852f1fe6 ("vhost: batching fetches")
> > > > ---
> > > >  drivers/vhost/vhost.c | 3 +--
> > > >  1 file changed, 1 insertion(+), 2 deletions(-)
> > > > 
> > > > diff --git a/drivers/vhost/vhost.c b/drivers/vhost/vhost.c
> > > > index b5a51b1f2e79..fc422c3e5c08 100644
> > > > --- a/drivers/vhost/vhost.c
> > > > +++ b/drivers/vhost/vhost.c
> > > > @@ -2409,12 +2409,11 @@ int vhost_get_vq_desc(struct vhost_virtqueue *vq,
> > > >  			*out_num += ret;
> > > >  		}
> > > >  
> > > > -		ret = desc->id;
> > > > -
> > > >  		if (!(desc->flags & VRING_DESC_F_NEXT))
> > > >  			break;
> > > >  	}
> > > >  
> > > > +	ret = vq->descs[vq->first_desc].id;
> > > >  	vq->first_desc = i + 1;
> > > >  
> > > >  	return ret;
> > > > 
> > 
> > Sorry, still not able to reproduce the issue.
> > 
> > Could we try to disable all the vhost features?
> > 
> > Thanks!
> > 
> > diff --git a/drivers/vhost/vhost.h b/drivers/vhost/vhost.h
> > index 661088ae6dc7..08f6d2ccb697 100644
> > --- a/drivers/vhost/vhost.h
> > +++ b/drivers/vhost/vhost.h
> > @@ -250,11 +250,11 @@ int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled);
> >         } while (0)
> >  
> >  enum {
> > -       VHOST_FEATURES = (1ULL << VIRTIO_F_NOTIFY_ON_EMPTY) |
> > -                        (1ULL << VIRTIO_RING_F_INDIRECT_DESC) |
> > -                        (1ULL << VIRTIO_RING_F_EVENT_IDX) |
> > -                        (1ULL << VHOST_F_LOG_ALL) |
> > -                        (1ULL << VIRTIO_F_ANY_LAYOUT) |
> > +       VHOST_FEATURES = /* (1ULL << VIRTIO_F_NOTIFY_ON_EMPTY) | */
> > +                        /* (1ULL << VIRTIO_RING_F_INDIRECT_DESC) | */
> > +                        /* (1ULL << VIRTIO_RING_F_EVENT_IDX) | */
> > +                        /* (1ULL << VHOST_F_LOG_ALL) | */
> > +                        /* (1ULL << VIRTIO_F_ANY_LAYOUT) | */
> >                          (1ULL << VIRTIO_F_VERSION_1)
> >  };
> > 
> 
> I still get  guest crashes with this on top of eccb852f1fe6. (The patch did not
> apply, I had to manually comment out these things)
> 

Sorry about that, I C&P transformed tabs to spaces.

Can we try tracing last_avail_idx with the attached patch? Can you enable also line and thread id (dyndbg='+plt')?

Thanks!