Re: [PATCH RFC v2 3/4] virtio-net: support transmit timestamp

[Willem de Bruijn <willemdebruijn.kernel@gmail.com>]

On Tue, Feb 9, 2021 at 11:39 AM Michael S. Tsirkin <mst@redhat.com> wrote:
>
> On Tue, Feb 09, 2021 at 01:45:11PM +0800, Jason Wang wrote:
> >
> > On 2021/2/9 上午2:55, Willem de Bruijn wrote:
> > > From: Willem de Bruijn <willemb@google.com>
> > >
> > > Add optional PTP hardware tx timestamp offload for virtio-net.
> > >
> > > Accurate RTT measurement requires timestamps close to the wire.
> > > Introduce virtio feature VIRTIO_NET_F_TX_TSTAMP, the transmit
> > > equivalent to VIRTIO_NET_F_RX_TSTAMP.
> > >
> > > The driver sets VIRTIO_NET_HDR_F_TSTAMP to request a timestamp
> > > returned on completion. If the feature is negotiated, the device
> > > either places the timestamp or clears the feature bit.
> > >
> > > The timestamp straddles (virtual) hardware domains. Like PTP, use
> > > international atomic time (CLOCK_TAI) as global clock base. The driver
> > > must sync with the device, e.g., through kvm-clock.
> > >
> > > Modify can_push to ensure that on tx completion the header, and thus
> > > timestamp, is in a predicatable location at skb_vnet_hdr.
> > >
> > > RFC: this implementation relies on the device writing to the buffer.
> > > That breaks DMA_TO_DEVICE semantics. For now, disable when DMA is on.
> > > The virtio changes should be a separate patch at the least.
> > >
> > > Tested: modified txtimestamp.c to with h/w timestamping:
> > >    -       sock_opt = SOF_TIMESTAMPING_SOFTWARE |
> > >    +       sock_opt = SOF_TIMESTAMPING_RAW_HARDWARE |
> > >    + do_test(family, SOF_TIMESTAMPING_TX_HARDWARE);
> > >
> > > Signed-off-by: Willem de Bruijn <willemb@google.com>
> > > ---
> > >   drivers/net/virtio_net.c        | 61 ++++++++++++++++++++++++++++-----
> > >   drivers/virtio/virtio_ring.c    |  3 +-
> > >   include/linux/virtio.h          |  1 +
> > >   include/uapi/linux/virtio_net.h |  1 +
> > >   4 files changed, 56 insertions(+), 10 deletions(-)
> > >
> > > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> > > index ac44c5efa0bc..fc8ecd3a333a 100644
> > > --- a/drivers/net/virtio_net.c
> > > +++ b/drivers/net/virtio_net.c
> > > @@ -210,6 +210,12 @@ struct virtnet_info {
> > >     /* Device will pass rx timestamp. Requires has_rx_tstamp */
> > >     bool enable_rx_tstamp;
> > > +   /* Device can pass CLOCK_TAI transmit time to the driver */
> > > +   bool has_tx_tstamp;
> > > +
> > > +   /* Device will pass tx timestamp. Requires has_tx_tstamp */
> > > +   bool enable_tx_tstamp;
> > > +
> > >     /* Has control virtqueue */
> > >     bool has_cvq;
> > > @@ -1401,6 +1407,20 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
> > >     return stats.packets;
> > >   }
> > > +static void virtnet_record_tx_tstamp(const struct send_queue *sq,
> > > +                                struct sk_buff *skb)
> > > +{
> > > +   const struct virtio_net_hdr_hash_ts *h = skb_vnet_hdr_ht(skb);
> > > +   const struct virtnet_info *vi = sq->vq->vdev->priv;
> > > +   struct skb_shared_hwtstamps ts;
> > > +
> > > +   if (h->hdr.flags & VIRTIO_NET_HDR_F_TSTAMP &&
> > > +       vi->enable_tx_tstamp) {
> > > +           ts.hwtstamp = ns_to_ktime(le64_to_cpu(h->tstamp));
> > > +           skb_tstamp_tx(skb, &ts);
> >
> >
> > This probably won't work since the buffer is read-only from the device. (See
> > virtqueue_add_outbuf()).
> >
> > Another issue that I vaguely remember that the virtio spec forbids out
> > buffer after in buffer.
>
> Both Driver Requirements: Message Framing and Driver Requirements: Scatter-Gather Support
> have this statement:
>
>         The driver MUST place any device-writable descriptor elements after any device-readable descriptor ele-
>         ments.
>
>
> similarly
>
> Device Requirements: The Virtqueue Descriptor Table
>         A device MUST NOT write to a device-readable buffer, and a device SHOULD NOT read a device-writable
>         buffer.

Thanks. That's clear. So the clean solution would be to add a
device-writable descriptor after the existing device-readable ones.

And the device must be aware that this is to return the tstamp only.
In the example implementation of vhost, it has to exclude this last
descriptor from the msg->msg_iter iovec array with packet data
initialized at get_tx_bufs/init_iov_iter.