Re: [外部邮件] Re: [PATCH] virtio_ring: reset IN_ORDER state in virtqueue_init()

[Eugenio Perez Martin]

On Tue, Nov 3, 2026 at 10:41 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
>
> On Tue, Nov 03, 2026 at 10:36:26AM +0100, Eugenio Perez Martin wrote:
> > On Fri, Sep 11, 2026 at 8:05 AM Li,Rongqing <lirongqing@xxxxxxxxx> wrote:
> > >
> > > > >
> > > > > The IN_ORDER feature uses additional software state to track the next
> > > > > available descriptor and in-flight descriptor batches.
> > > > >
> > > > > virtqueue_init() resets the common virtqueue state, but free_head and
> > > > > batch_last.id are initialized only when the virtqueue is created. When
> > > > > an IN_ORDER virtqueue is reset, these fields can retain state from the
> > > > > previous queue instance.
> > > > >
> > > > > Reset free_head and invalidate batch_last when initializing an
> > > > > IN_ORDER virtqueue. Keep this in virtqueue_init() so the state is
> > > > > initialized consistently for both newly created and reset virtqueues.
> > > > >
> > > > > Fixes: f6a15d854986 ("virtio_ring: add in order support")
> > > >
> > > > I think this is needed for -stable.
> > > >
> > > > > Signed-off-by: Li RongQing <lirongqing@xxxxxxxxx>
> > > > > ---
> > > > >  drivers/virtio/virtio_ring.c | 5 +++++
> > > > >  1 file changed, 5 insertions(+)
> > > > >
> > > > > diff --git a/drivers/virtio/virtio_ring.c
> > > > > b/drivers/virtio/virtio_ring.c index db678f5..d0dc464 100644
> > > > > --- a/drivers/virtio/virtio_ring.c
> > > > > +++ b/drivers/virtio/virtio_ring.c
> > > > > @@ -559,6 +559,11 @@ static void virtqueue_init(struct vring_virtqueue *vq,
> > > > u32 num)
> > > > >         vq->in_use = false;
> > > > >         vq->last_add_time_valid = false;  #endif
> > > > > +
> > > > > +       if (virtqueue_is_in_order(vq)) {
> > > > > +               vq->free_head = 0;
> > > > > +               vq->batch_last.id = UINT_MAX;
> > > > > +       }
> > > >
> > > > I would try to reuse or refactor virtqueue_vring_attach_split() to reduce code
> > > > duplication.
> > > >
> > > > Btw have you checked Qemu shadow virtqueue as it might suffer from the same
> > > > issue.
> > > >
> > >
> > > I checked QEMU's shadow virtqueue implementation. batch_last is reinitialized in vhost_svq_start(),
> > > but free_head is not. Since the SVQ object can be reused across stop/start, free_head may retain stale
> > > state. So a corresponding fix should be needed
> > >
> > > void vhost_svq_start(VhostShadowVirtqueue *svq, VirtIODevice *vdev,
> > >                       VirtQueue *vq, VhostIOVATree *iova_tree)
> > >  {
> > >      ...
> > >      svq->shadow_avail_idx = 0;
> > >      svq->shadow_used_idx = 0;
> > > +    svq->free_head = 0;
> > >      memset(&svq->batch_last, 0, sizeof(svq->batch_last));
> > >      svq->last_used = 0;
> > >      svq->last_used_idx = 0;
> > >
> >
> > I'll ack this patch if it is sent to the QEMU mailing list, but the
> > guest is not able to reset a virtqueue without deallocating it
>
> what would "deallocating" mean here? who's doing the deallocating?
>

I meant that QEMU free the shadow virtqueue members on vring stop, but
I was wrong about free_head. What QEMU actually free is each
descriptor state (and the queue itself), as it could be resized on
each reset. free_head member should be set to 0 too. Li RongQing,
would you mind sending a patch to the QEMU mailing list with that?

Thanks!