Re: [Qemu-devel] [PATCH v2 3/3] net: virtio-net discards TX data after link down

[Michael S. Tsirkin]

On Thu, Nov 10, 2016 at 01:56:05AM +0200, Yuri Benditovich wrote:
> 
> 
> On Wed, Nov 9, 2016 at 10:28 PM, Michael S. Tsirkin <address@hidden> wrote:
> 
>     On Wed, Nov 09, 2016 at 05:22:02PM +0200, address@hidden
>     wrote:
>     > From: Yuri Benditovich <address@hidden>
>     >
>     > https://bugzilla.redhat.com/show_bug.cgi?id=1295637
>     > Upon set_link monitor command or upon netdev deletion
>     > virtio-net sends link down indication to the guest
>     > and stops vhost if one is used.
>     > Guest driver can still submit data for TX until it
>     > recognizes link loss. If these packets not returned by
>     > the host, the Windows guest will never be able to finish
>     > disable/removal/shutdown.
>     > Now each packet sent by guest after NIC indicated link
>     > down will be completed immediately.
>     >
>     > Signed-off-by: Yuri Benditovich <address@hidden>
>     > ---
>     >  hw/net/virtio-net.c | 28 ++++++++++++++++++++++++++++
>     >  1 file changed, 28 insertions(+)
>     >
>     > diff --git a/hw/net/virtio-net.c b/hw/net/virtio-net.c
>     > index 06bfe4b..ab4e18a 100644
>     > --- a/hw/net/virtio-net.c
>     > +++ b/hw/net/virtio-net.c
>     > @@ -218,6 +218,16 @@ static void virtio_net_vnet_endian_status(VirtIONet
>     *n, uint8_t status)
>     >      }
>     >  }
>     >
>     > +static void virtio_net_drop_tx_queue_data(VirtIODevice *vdev, VirtQueue
>     *vq)
>     > +{
>     > +    VirtQueueElement *elem;
>     > +    while ((elem = virtqueue_pop(vq, sizeof(VirtQueueElement)))) {
>     > +        virtqueue_push(vq, elem, 0);
>     > +        virtio_notify(vdev, vq);
>     > +        g_free(elem);
>     > +    }
>     > +}
>     > +
>     >  static void virtio_net_set_status(struct VirtIODevice *vdev, uint8_t
>     status)
>     >  {
>     >      VirtIONet *n = VIRTIO_NET(vdev);
> 
>     I don't like this part. This does too much queue parsing,
>     I would like to just copy head from avail to used ring.
> 
>     For example, people want to support rings >1K in size.
>     Let's add bool virtqueue_drop(vq) and be done with it.
>    
> 
> Please note that this code works only when link is down.
> For me this was too complicated to write simpler procedure
> with the same result.

Yes - it's somewhat problematic and risky that we process
the ring in qemu, but I don't see an easy way around that.
But at least let's limit the processing and assumptions we
make.


> 
>    
>     > @@ -262,6 +272,14 @@ static void virtio_net_set_status(struct
>     VirtIODevice *vdev, uint8_t status)
>     >              } else {
>     >                  qemu_bh_cancel(q->tx_bh);
>     >              }
>     > +            if ((n->status & VIRTIO_NET_S_LINK_UP) == 0 &&
>     > +                (queue_status & VIRTIO_CONFIG_S_DRIVER_OK)) {
>     > +                /* if tx is waiting we are likely have some packets in

... we likely have some ...

>     tx queue
>     > +                 * and disabled notification */

what does this refer to?

>     > +                q->tx_waiting = 0;
>     > +                virtio_queue_set_notification(q->tx_vq, 1);
>     > +                virtio_net_drop_tx_queue_data(vdev, q->tx_vq);
>     > +            }
>     >          }
>     >      }
>     >  }
> 
>     OK but what if guest keeps sending packets? What will drop them?
> 
> 
> This code fixes following problem in original code (example):
> We are in vhost=off and receive kick ->virtio_net_handle_tx_timer
> -> tx_waiting=1, notification disabled, timer set
> Now we receive link loss, cancel the timer and stay with packets in the queue
> and with
> disabled notification. Nobody will return them. (easy to reproduce with timer
> set to 5ms)
> 
> Added code drops packets we already have and ensure we will report them
> as completed to guest. If guest keeps sending packets, they will be dropped 
> in virtio_net_handle_tx_timer and in  virtio_net_handle_tx_bh (in procedures
> just below)
> as we already with link down. 

Yes I get that. I'm just not 100% sure all paths have
us listen on the ioeventfd and handle kicks without races -
this was previously assumed not to matter.


> 
> 
>     > @@ -1319,6 +1337,11 @@ static void virtio_net_handle_tx_timer(
>     VirtIODevice *vdev, VirtQueue *vq)
>     >      VirtIONet *n = VIRTIO_NET(vdev);
>     >      VirtIONetQueue *q = &n->vqs[vq2q(virtio_get_queue_index(vq))];
>     >
>     > +    if (unlikely((n->status & VIRTIO_NET_S_LINK_UP) == 0)) {
>     > +        virtio_net_drop_tx_queue_data(vdev, vq);
>     > +        return;
>     > +    }
>     > +
>     >      /* This happens when device was stopped but VCPU wasn't. */
>     >      if (!vdev->vm_running) {
>     >          q->tx_waiting = 1;
>     > @@ -1345,6 +1368,11 @@ static void virtio_net_handle_tx_bh(VirtIODevice
>     *vdev, VirtQueue *vq)
>     >      VirtIONet *n = VIRTIO_NET(vdev);
>     >      VirtIONetQueue *q = &n->vqs[vq2q(virtio_get_queue_index(vq))];
>     >
>     > +    if (unlikely((n->status & VIRTIO_NET_S_LINK_UP) == 0)) {
>     > +        virtio_net_drop_tx_queue_data(vdev, vq);
>     > +        return;
>     > +    }
>     > +
>     >      if (unlikely(q->tx_waiting)) {
>     >          return;
>     >      }
>     > --
>     > 1.9.1
> 
>