On Fri, 15 Apr 2022 13:53:54 +0800, Jason Wang <jasowang@xxxxxxxxxx> wrote:
> On Fri, Apr 15, 2022 at 10:23 AM Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx> wrote:
> >
> > On Thu, 14 Apr 2022 17:30:02 +0800, Jason Wang <jasowang@xxxxxxxxxx> wrote:
> > > On Wed, Apr 13, 2022 at 4:47 PM Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx> wrote:
> > > >
> > > > On Wed, 13 Apr 2022 16:00:18 +0800, Jason Wang <jasowang@xxxxxxxxxx> 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@xxxxxxxxxxxxxxxxx>
> > > > > > ---
> > > > > > 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.
> >
> > Yes.
> >
> > > >
> > > > 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.
> >
> > I will keep an eye on this situation.
> >
> > >
> > > >
> > > > >
> > > > >
> > > > > > +
> > > > > > /* 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?
> >
> > I don't think this situation will be encountered, because napi_disable is
> > mutually exclusive, so there will be no unexpected napi enable.
> >
> > Is there something I misunderstood?
>
> So in virtnet_rx_resize() we do:
>
> napi_disable()
> ...
> resize()
> ...
> napi_enalbe()
>
> How can we guarantee that the work is not run after the napi_disable()?
I think you're talking about a situation like this:
virtnet_rx_resize refill work
-----------------------------------------------------------
napi_disable()
... napi_disable()
resize() ...
napi_enable()
...
napi_enalbe()
But in fact:
virtnet_rx_resize refill work
-----------------------------------------------------------
napi_disable()
... napi_disable() <----[0]
resize() |
... |
napi_enalbe() |
napi_disable() <---- [1] here success
napi_enable()
Because virtnet_rx_resize() has already executed napi_disable(), napi_disalbe()
of [0] will wait until [1] to complete.
I'm not sure if my understanding is correct.
Thanks.
>
> Thanks
>
> >
> > Thanks.
> >
> > >
> > > 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
> > > > >
> > > >
> > >
> >
>
