On Thu, Jun 9, 2016 at 9:47 AM, Stefan Hajnoczi <stefa...@gmail.com> wrote:
> On Thu, Jun 9, 2016 at 9:25 AM, Paolo Bonzini <pbonz...@redhat.com> wrote:
>> On 09/06/2016 09:35, Stefan Hajnoczi wrote:
>>>
>>> if (s->dataplane) {
>>> virtio_blk_data_plane_stop(s->dataplane);
>>> }
>>>
>>> virtio_save(vdev, f);
>>> }
>>>
>>> We reach the situation I described where BDS AioContext changes but
>>> mirror_run() is still in the IOThread AioContext.
>>
>> Why is the virtio_blk_data_plane_stop necessary? It used to sync
>> between vring and virtio state, but that's not required anymore---and we
>> know that the virtqueues are quiescent at this point...
>
> I agree. It doesn't seem necessary anymore unless I'm missing a race
> condition with ioeventfd.
>
> It was introduced so that the savevm command (without live migration!) works:
>
> commit 10a06fd65f667a972848ebbbcac11bdba931b544
> Author: Pavel Butsykin <pbutsy...@virtuozzo.com>
> Date: Mon Oct 26 14:42:57 2015 +0300
>
> virtio: sync the dataplane vring state to the virtqueue before virtio_save
>
> We can remove it now.
BTW this doesn't mean that the blockjobs AioContext following problem
is solved. AioContexts can still change if the guest resets the
virtio device, for example.
I suspect we'll still hit undefined behavior when that happens.
Stefan
