Re: [PATCH v9 31/32] virtio_net: support rx/tx queue resize

[Jason Wang <jasowang@redhat.com>]

On Wed, Apr 13, 2022 at 4:47 PM Xuan Zhuo <xuanzhuo@linux.alibaba.com> wrote:
>
> On Wed, 13 Apr 2022 16:00:18 +0800, Jason Wang <jasowang@redhat.com> wrote:
> >
> > 在 2022/4/6 上午11:43, Xuan Zhuo 写道:
> > > This patch implements the resize function of the rx, tx queues.
> > > Based on this function, it is possible to modify the ring num of the
> > > queue.
> > >
> > > There may be an exception during the resize process, the resize may
> > > fail, or the vq can no longer be used. Either way, we must execute
> > > napi_enable(). Because napi_disable is similar to a lock, napi_enable
> > > must be called after calling napi_disable.
> > >
> > > Signed-off-by: Xuan Zhuo <xuanzhuo@linux.alibaba.com>
> > > ---
> > >   drivers/net/virtio_net.c | 81 ++++++++++++++++++++++++++++++++++++++++
> > >   1 file changed, 81 insertions(+)
> > >
> > > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> > > index b8bf00525177..ba6859f305f7 100644
> > > --- a/drivers/net/virtio_net.c
> > > +++ b/drivers/net/virtio_net.c
> > > @@ -251,6 +251,9 @@ struct padded_vnet_hdr {
> > >     char padding[4];
> > >   };
> > >
> > > +static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf);
> > > +static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf);
> > > +
> > >   static bool is_xdp_frame(void *ptr)
> > >   {
> > >     return (unsigned long)ptr & VIRTIO_XDP_FLAG;
> > > @@ -1369,6 +1372,15 @@ static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
> > >   {
> > >     napi_enable(napi);
> > >
> > > +   /* Check if vq is in reset state. The normal reset/resize process will
> > > +    * be protected by napi. However, the protection of napi is only enabled
> > > +    * during the operation, and the protection of napi will end after the
> > > +    * operation is completed. If re-enable fails during the process, vq
> > > +    * will remain unavailable with reset state.
> > > +    */
> > > +   if (vq->reset)
> > > +           return;
> >
> >
> > I don't get when could we hit this condition.
>
>
> In patch 23, the code to implement re-enable vq is as follows:
>
> +static int vp_modern_enable_reset_vq(struct virtqueue *vq)
> +{
> +       struct virtio_pci_device *vp_dev = to_vp_device(vq->vdev);
> +       struct virtio_pci_modern_device *mdev = &vp_dev->mdev;
> +       struct virtio_pci_vq_info *info;
> +       unsigned long flags, index;
> +       int err;
> +
> +       if (!vq->reset)
> +               return -EBUSY;
> +
> +       index = vq->index;
> +       info = vp_dev->vqs[index];
> +
> +       /* check queue reset status */
> +       if (vp_modern_get_queue_reset(mdev, index) != 1)
> +               return -EBUSY;
> +
> +       err = vp_active_vq(vq, info->msix_vector);
> +       if (err)
> +               return err;
> +
> +       if (vq->callback) {
> +               spin_lock_irqsave(&vp_dev->lock, flags);
> +               list_add(&info->node, &vp_dev->virtqueues);
> +               spin_unlock_irqrestore(&vp_dev->lock, flags);
> +       } else {
> +               INIT_LIST_HEAD(&info->node);
> +       }
> +
> +       vp_modern_set_queue_enable(&vp_dev->mdev, index, true);
> +
> +       if (vp_dev->per_vq_vectors && info->msix_vector != VIRTIO_MSI_NO_VECTOR)
> +               enable_irq(pci_irq_vector(vp_dev->pci_dev, info->msix_vector));
> +
> +       vq->reset = false;
> +
> +       return 0;
> +}
>
>
> There are three situations where an error will be returned. These are the
> situations I want to handle.

Right, but it looks harmless if we just schedule the NAPI without the check.

>
> But I'm rethinking the question, and I feel like you're right, although the
> hardware setup may fail. We can no longer sync with the hardware. But using it
> as a normal vq doesn't have any problems.

Note that we should make sure the buggy(malicous) device won't crash
the codes by changing the queue_reset value at its will.

>
> >
> >
> > > +
> > >     /* If all buffers were filled by other side before we napi_enabled, we
> > >      * won't get another interrupt, so process any outstanding packets now.
> > >      * Call local_bh_enable after to trigger softIRQ processing.
> > > @@ -1413,6 +1425,15 @@ static void refill_work(struct work_struct *work)
> > >             struct receive_queue *rq = &vi->rq[i];
> > >
> > >             napi_disable(&rq->napi);
> > > +
> > > +           /* Check if vq is in reset state. See more in
> > > +            * virtnet_napi_enable()
> > > +            */
> > > +           if (rq->vq->reset) {
> > > +                   virtnet_napi_enable(rq->vq, &rq->napi);
> > > +                   continue;
> > > +           }
> >
> >
> > Can we do something similar in virtnet_close() by canceling the work?
>
> I think there is no need to cancel the work here, because napi_disable will wait
> for the napi_enable of the resize. So if the re-enable failed vq is used as a normal
> vq, this logic can be removed.

Actually I meant the part of virtnet_rx_resize().

If we don't synchronize with the refill work, it might enable NAPI unexpectedly?

Thanks

>
>
> >
> >
> > > +
> > >             still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
> > >             virtnet_napi_enable(rq->vq, &rq->napi);
> > >
> > > @@ -1523,6 +1544,10 @@ static void virtnet_poll_cleantx(struct receive_queue *rq)
> > >     if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
> > >             return;
> > >
> > > +   /* Check if vq is in reset state. See more in virtnet_napi_enable() */
> > > +   if (sq->vq->reset)
> > > +           return;
> >
> >
> > We've disabled TX napi, any chance we can still hit this?
>
> Same as above.
>
> >
> >
> > > +
> > >     if (__netif_tx_trylock(txq)) {
> > >             do {
> > >                     virtqueue_disable_cb(sq->vq);
> > > @@ -1769,6 +1794,62 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
> > >     return NETDEV_TX_OK;
> > >   }
> > >
> > > +static int virtnet_rx_resize(struct virtnet_info *vi,
> > > +                        struct receive_queue *rq, u32 ring_num)
> > > +{
> > > +   int err;
> > > +
> > > +   napi_disable(&rq->napi);
> > > +
> > > +   err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_free_unused_buf);
> > > +   if (err)
> > > +           goto err;
> > > +
> > > +   if (!try_fill_recv(vi, rq, GFP_KERNEL))
> > > +           schedule_delayed_work(&vi->refill, 0);
> > > +
> > > +   virtnet_napi_enable(rq->vq, &rq->napi);
> > > +   return 0;
> > > +
> > > +err:
> > > +   netdev_err(vi->dev,
> > > +              "reset rx reset vq fail: rx queue index: %td err: %d\n",
> > > +              rq - vi->rq, err);
> > > +   virtnet_napi_enable(rq->vq, &rq->napi);
> > > +   return err;
> > > +}
> > > +
> > > +static int virtnet_tx_resize(struct virtnet_info *vi,
> > > +                        struct send_queue *sq, u32 ring_num)
> > > +{
> > > +   struct netdev_queue *txq;
> > > +   int err, qindex;
> > > +
> > > +   qindex = sq - vi->sq;
> > > +
> > > +   virtnet_napi_tx_disable(&sq->napi);
> > > +
> > > +   txq = netdev_get_tx_queue(vi->dev, qindex);
> > > +   __netif_tx_lock_bh(txq);
> > > +   netif_stop_subqueue(vi->dev, qindex);
> > > +   __netif_tx_unlock_bh(txq);
> > > +
> > > +   err = virtqueue_resize(sq->vq, ring_num, virtnet_sq_free_unused_buf);
> > > +   if (err)
> > > +           goto err;
> > > +
> > > +   netif_start_subqueue(vi->dev, qindex);
> > > +   virtnet_napi_tx_enable(vi, sq->vq, &sq->napi);
> > > +   return 0;
> > > +
> > > +err:
> >
> >
> > I guess we can still start the queue in this case? (Since we don't
> > change the queue if resize fails).
>
> Yes, you are right.
>
> Thanks.
>
> >
> >
> > > +   netdev_err(vi->dev,
> > > +              "reset tx reset vq fail: tx queue index: %td err: %d\n",
> > > +              sq - vi->sq, err);
> > > +   virtnet_napi_tx_enable(vi, sq->vq, &sq->napi);
> > > +   return err;
> > > +}
> > > +
> > >   /*
> > >    * Send command via the control virtqueue and check status.  Commands
> > >    * supported by the hypervisor, as indicated by feature bits, should
> >
>