[Qemu-devel] Symptoms of lost interrupts in virtio?

[Qemu-devel] Symptoms of lost interrupts in virtio?

[Soren Hansen]

Michael,

In Ubuntu we're seeing a bug in qemu-kvm 1.0 where networking stops
working. It's described here:

   https://bugs.launchpad.net/ubuntu/+source/libvirt/+bug/997978

We've not been able to reproduce it ourselves, but the users who are
experiencing it say that it's fixed in 1.1.

I've found this series of commits:

   a821ce5 virtio: order index/descriptor reads
   92045d8 virtio: add missing mb() on enable notification
   a281ebc virtio: add missing mb() on notification

where particularly the last one looks like a likely fix for the problem.

The commit message isn't entirely clear about the symptoms it's meant
to fix, even though the behind-the-covers explanation is very
thorough. It's not clear to me, for instance, if the network would
stop working or if you'd just lose a couple of network packets?

The commit message suggests it only happens on a very specific
processor type. Do you happen to remember which one? If it matches
what our affected users see, that could be clue as to whether this is
the  same problem.

Thanks in advance for your time.

-- 
Soren Hansen
Ubuntu Developer    http://www.ubuntu.com/
OpenStack Developer http://www.openstack.org/

Re: [Qemu-devel] Symptoms of lost interrupts in virtio?

[Michael S. Tsirkin]

On Mon, Sep 03, 2012 at 03:16:00PM +0200, Soren Hansen wrote:
> Michael,
> 
> In Ubuntu we're seeing a bug in qemu-kvm 1.0 where networking stops
> working. It's described here:
> 
>    https://bugs.launchpad.net/ubuntu/+source/libvirt/+bug/997978
> 
> We've not been able to reproduce it ourselves, but the users who are
> experiencing it say that it's fixed in 1.1.
> 
> I've found this series of commits:
> 
>    a821ce5 virtio: order index/descriptor reads
>    92045d8 virtio: add missing mb() on enable notification
>    a281ebc virtio: add missing mb() on notification
> 
> where particularly the last one looks like a likely fix for the problem.
> 
> The commit message isn't entirely clear about the symptoms it's meant
> to fix, even though the behind-the-covers explanation is very
> thorough. It's not clear to me, for instance, if the network would
> stop working or if you'd just lose a couple of network packets?


It can stop working. It's more likely to stop if
event_idx is enabled. Try disabling and see what happens.



> The commit message suggests it only happens on a very specific
> processor type. Do you happen to remember which one? If it matches
> what our affected users see, that could be clue as to whether this is
> the  same problem.

I think it was reported on AMD systems.


> Thanks in advance for your time.
> 
> -- 
> Soren Hansen
> Ubuntu Developer    http://www.ubuntu.com/
> OpenStack Developer http://www.openstack.org/