Received-SPF: pass (google.com: domain of linux-kernel+bounces-36295-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:4601:e00::3 as permitted sender) client-ip=2604:1380:4601:e00::3;
Precedence: bulk
MIME-Version: 1.0
References: <20240122102256.261374-1-liangchen.linux@gmail.com> <4ada64aa-87ec-4340-9892-be52c1f27a97@linux.alibaba.com>
In-Reply-To: <4ada64aa-87ec-4340-9892-be52c1f27a97@linux.alibaba.com>
From: Liang Chen <liangchen.linux@gmail.com>
Date: Wed, 24 Jan 2024 10:04:51 +0800
Message-ID: <CAKhg4t+ayD-MPK8FwZFURZyX32d8gk26GdytwRiRjJBwBbaR6w@mail.gmail.com>
Subject: Re: [PATCH] virtio_net: Support RX hash XDP hint
To: Heng Qi <hengqi@linux.alibaba.com>
Cc: mst@redhat.com, jasowang@redhat.com, 
	virtualization@lists.linux-foundation.org, linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Mon, Jan 22, 2024 at 7:10=E2=80=AFPM Heng Qi <hengqi@linux.alibaba.com> =
wrote:
>
> Hi Liang Chen,
>
> =E5=9C=A8 2024/1/22 =E4=B8=8B=E5=8D=886:22, Liang Chen =E5=86=99=E9=81=93=
:
> > The RSS hash report is a feature that's part of the virtio specificatio=
n.
> > Currently, virtio backends like qemu, vdpa (mlx5), and potentially vhos=
t
> > (still a work in progress as per [1]) support this feature. While the
> > capability to obtain the RSS hash has been enabled in the normal path,
> > it's currently missing in the XDP path. Therefore, we are introducing X=
DP
> > hints through kfuncs to allow XDP programs to access the RSS hash.
> >
> > Signed-off-by: Liang Chen <liangchen.linux@gmail.com>
> > ---
> >   drivers/net/virtio_net.c | 56 +++++++++++++++++++++++++++++++++++++++=
+
> >   1 file changed, 56 insertions(+)
> >
> > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> > index d7ce4a1011ea..1463a4709e3c 100644
> > --- a/drivers/net/virtio_net.c
> > +++ b/drivers/net/virtio_net.c
> > @@ -4579,6 +4579,60 @@ static void virtnet_set_big_packets(struct virtn=
et_info *vi, const int mtu)
> >       }
> >   }
> >
> > +static int virtnet_xdp_rx_hash(const struct xdp_md *_ctx, u32 *hash,
> > +                        enum xdp_rss_hash_type *rss_type)
> > +{
> > +     const struct xdp_buff *xdp =3D (void *)_ctx;
> > +     struct virtio_net_hdr_v1_hash *hdr_hash;
> > +     struct virtnet_info *vi;
> > +
> > +     if (!(xdp->rxq->dev->features & NETIF_F_RXHASH))
>
> I think 'vi->has_rss_hash_report' should be used here.
> NETIF_F_RXHASH cannot guarantee that the hash report feature is negotiate=
d,
> and accessing hash_report and hash_value is unsafe at this time.
>

My understanding is that rxhash in dev->features is turned on only if
the corresponding hw feature is set. We will double check on this.

> > +             return -ENODATA;
> > +
> > +     vi =3D netdev_priv(xdp->rxq->dev);
> > +     hdr_hash =3D (struct virtio_net_hdr_v1_hash *)(xdp->data - vi->hd=
r_len);
>
> If the virtio-net-hdr is overrided by the XDP prog, how can this be done
> correctly and as expected?
>

Yeah, thanks for bringing up this concern. We are actively working on
a fix of this issue(possibly with a wrapper structure of xdp_buff).

Thanks,
Liang

> Thanks,
> Heng
>
> > +
> > +     switch (__le16_to_cpu(hdr_hash->hash_report)) {
> > +             case VIRTIO_NET_HASH_REPORT_TCPv4:
> > +                     *rss_type =3D XDP_RSS_TYPE_L4_IPV4_TCP;
> > +                     break;
> > +             case VIRTIO_NET_HASH_REPORT_UDPv4:
> > +                     *rss_type =3D XDP_RSS_TYPE_L4_IPV4_UDP;
> > +                     break;
> > +             case VIRTIO_NET_HASH_REPORT_TCPv6:
> > +                     *rss_type =3D XDP_RSS_TYPE_L4_IPV6_TCP;
> > +                     break;
> > +             case VIRTIO_NET_HASH_REPORT_UDPv6:
> > +                     *rss_type =3D XDP_RSS_TYPE_L4_IPV6_UDP;
> > +                     break;
> > +             case VIRTIO_NET_HASH_REPORT_TCPv6_EX:
> > +                     *rss_type =3D XDP_RSS_TYPE_L4_IPV6_TCP_EX;
> > +                     break;
> > +             case VIRTIO_NET_HASH_REPORT_UDPv6_EX:
> > +                     *rss_type =3D XDP_RSS_TYPE_L4_IPV6_UDP_EX;
> > +                     break;
> > +             case VIRTIO_NET_HASH_REPORT_IPv4:
> > +                     *rss_type =3D XDP_RSS_TYPE_L3_IPV4;
> > +                     break;
> > +             case VIRTIO_NET_HASH_REPORT_IPv6:
> > +                     *rss_type =3D XDP_RSS_TYPE_L3_IPV6;
> > +                     break;
> > +             case VIRTIO_NET_HASH_REPORT_IPv6_EX:
> > +                     *rss_type =3D XDP_RSS_TYPE_L3_IPV6_EX;
> > +                     break;
> > +             case VIRTIO_NET_HASH_REPORT_NONE:
> > +             default:
> > +                     *rss_type =3D XDP_RSS_TYPE_NONE;
> > +     }
> > +
> > +     *hash =3D __le32_to_cpu(hdr_hash->hash_value);
> > +     return 0;
> > +}
> > +
> > +static const struct xdp_metadata_ops virtnet_xdp_metadata_ops =3D {
> > +     .xmo_rx_hash                    =3D virtnet_xdp_rx_hash,
> > +};
> > +
> >   static int virtnet_probe(struct virtio_device *vdev)
> >   {
> >       int i, err =3D -ENOMEM;
> > @@ -4613,6 +4667,8 @@ static int virtnet_probe(struct virtio_device *vd=
ev)
> >       dev->ethtool_ops =3D &virtnet_ethtool_ops;
> >       SET_NETDEV_DEV(dev, &vdev->dev);
> >
> > +     dev->xdp_metadata_ops =3D &virtnet_xdp_metadata_ops;
> > +
> >       /* Do we support "hardware" checksums? */
> >       if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
> >               /* This opens up the world of extra features. */
>