Re: [Qemu-devel] Bug in virtio_net_load

[Dr. David Alan Gilbert]

* Michael S. Tsirkin (address@hidden) wrote:
> On Thu, Jun 30, 2016 at 10:34:51AM +0200, Robin Geuze wrote:
> > Hey,
> > 
> > I work for TransIP and we host a VPS platform based on QEMU/KVM. We are
> > currently running qemu 2.4.0. A few days ago we noticed that live migrations
> > for some of our VM's would fail. Further investigation turned out it was
> > specific to windows server 2012, caused by the fact that the standard virtio
> > driver from RedHat was replaced in windows updates by a driver called
> > "Midfin eFabric" (this driver doesn't really seem to be meant for virtio, we
> > have a case running at MicroSoft about that).  Once we knew how to reproduce
> > we tested this on  QEMU 2.6.0 as well and it also seems to be affected
> > (later we found out that 2.4.0 to 2.6.0 migration does work probably due to
> > pure luck).
> > 
> > We started investigating the problem in QEMU 2.4.0 and noticed it was caused
> > by the fact that virtio_net_device_load requires certain feature flags to be
> > set, specifically to load curr_guest_offloads which is only written and read
> > if the VIRTIO_NET_F_CTRL_GUEST_OFFLOADS flag is set, but those flags are set
> > in virtio_load after the call to virtio_net_device_load. Moving the code
> > setting the feature flags before the call to virtio_net_device_load fixes
> > it, however it introduces another problem. Virtio can have 64-bits feature
> > flags, however the standard save payload for virtio only has space for
> > 32-bits feature flags. This was solved by putting those in a subsection of
> > the vmstate_save_state stuff. Unfortunately this is called (and thus binary
> > offset located) after the virtio_net_device_load code.
> > 
> > There was an attempt to fix this in QEMU 2.6.0. However, this seems to have
> > broken it worse. The write code (virtio_net_save, virtio_save and
> > virtio_net_save_device) still puts the curr_guest_offloads value before the
> > vmstate_save_state data. However the read code expects and tries to read it
> > after the vmstate_save_state data. Should we just also change the
> > virtio_net_save code to have it follow the same order as virtio_net_load? Or
> > will this potentially break more stuff.
> > 
> > Regards,
> > 
> > Robin Geuze
> > 
> > TransIP BV
> 
> After going over it several times, I think the change in 2.6
> was wrong
> 
> 
> 
> commit 1f8828ef573c83365b4a87a776daf8bcef1caa21
> Author: Jason Wang <address@hidden>
> Date:   Fri Sep 11 16:01:56 2015 +0800
> 
>     virtio-net: unbreak self announcement and guest offloads after migration
>     
>     After commit 019a3edbb25f1571e876f8af1ce4c55412939e5d ("virtio: make
>     features 64bit wide"). Device's guest_features was actually set after
>     vdc->load(). This breaks the assumption that device specific load()
>     function can check guest_features. For virtio-net, self announcement
>     and guest offloads won't work after migration.
>     
>     Fixing this by defer them to virtio_net_load() where guest_features
>     were guaranteed to be set. Other virtio devices looks fine.
>     
>     Fixes: 019a3edbb25f1571e876f8af1ce4c55412939e5d
>            ("virtio: make features 64bit wide")
>     Cc: address@hidden
>     Cc: Gerd Hoffmann <address@hidden>
>     Signed-off-by: Jason Wang <address@hidden>
>     Reviewed-by: Michael S. Tsirkin <address@hidden>
>     Signed-off-by: Michael S. Tsirkin <address@hidden>
>     Reviewed-by: Cornelia Huck <address@hidden>
> 
> 
> I'm not sure what was I thinking when I applied this:
> it changes load without changing save - how can this work?
> 
> 
> I am inclined to revert 1f8828ef573c83365b4a87a776daf8bcef1caa21 and
> apply this instead:
> 
> diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c
> index 7ed06ea..18153d5 100644
> --- a/hw/virtio/virtio.c
> +++ b/hw/virtio/virtio.c
> @@ -1499,6 +1499,16 @@ int virtio_load(VirtIODevice *vdev, QEMUFile *f, int 
> version_id)
>      }
>      qemu_get_be32s(f, &features);
>  
> +    /*
> +     * Temporarily set guest_features low bits - needed by
> +     * virtio net load code testing for VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
> +     * VIRTIO_NET_F_GUEST_ANNOUNCE and VIRTIO_NET_F_CTRL_VQ.
> +     *
> +     * Note: devices should always test host features in future - don't 
> create
> +     * new dependencies like this.
> +     */
> +    vdev->guest_features = features;
> +
>      config_len = qemu_get_be32(f);
>  
>      /*
> 
> Could you please confirm whether this help?
> Jason, Cornelia - any comments?
> 
> David, if this goes in I'm afraid your patchset reworking
> save/load will have to be rebased, but I think we want
> the bugfix first and new features/changes second.
> Do you agree?

Yes, bug fixes first.  But actually the merge is trivial, the only
place I think they collide is in virtio_net_load and it's just the difference
of the return virtio_load vs the ret =  virtio_load in the patch you revert.

Dave

> -- 
> MST
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK