Re: [PATCH net-next 2/2] tun: AF_XDP Rx zero-copy support

[Jason Wang]

On Mon, Jan 29, 2024 at 7:10 PM wangyunjian <wangyunjian@xxxxxxxxxx> wrote:
>
> > -----Original Message-----
> > From: Jason Wang [mailto:jasowang@xxxxxxxxxx]
> > Sent: Monday, January 29, 2024 11:05 AM
> > To: wangyunjian <wangyunjian@xxxxxxxxxx>
> > Cc: mst@xxxxxxxxxx; willemdebruijn.kernel@xxxxxxxxx; kuba@xxxxxxxxxx;
> > davem@xxxxxxxxxxxxx; magnus.karlsson@xxxxxxxxx; netdev@xxxxxxxxxxxxxxx;
> > linux-kernel@xxxxxxxxxxxxxxx; kvm@xxxxxxxxxxxxxxx;
> > virtualization@xxxxxxxxxxxxxxx; xudingke <xudingke@xxxxxxxxxx>
> > Subject: Re: [PATCH net-next 2/2] tun: AF_XDP Rx zero-copy support
> >
> > On Sat, Jan 27, 2024 at 5:34 PM wangyunjian <wangyunjian@huaweicom>
> > wrote:
> > >
> > > > > -----Original Message-----
> > > > > From: Jason Wang [mailto:jasowang@xxxxxxxxxx]
> > > > > Sent: Thursday, January 25, 2024 12:49 PM
> > > > > To: wangyunjian <wangyunjian@xxxxxxxxxx>
> > > > > Cc: mst@xxxxxxxxxx; willemdebruijn.kernel@xxxxxxxxx;
> > > > > kuba@xxxxxxxxxx; davem@xxxxxxxxxxxxx; magnus.karlsson@xxxxxxxxx;
> > > > > netdev@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> > > > > kvm@xxxxxxxxxxxxxxx; virtualization@xxxxxxxxxxxxxxx; xudingke
> > > > > <xudingke@xxxxxxxxxx>
> > > > > Subject: Re: [PATCH net-next 2/2] tun: AF_XDP Rx zero-copy support
> > > > >
> > > > > On Wed, Jan 24, 2024 at 5:38 PM Yunjian Wang
> > > > <wangyunjian@xxxxxxxxxx>
> > > > > wrote:
> > > > > >
> > > > > > Now the zero-copy feature of AF_XDP socket is supported by some
> > > > > > drivers, which can reduce CPU utilization on the xdp program.
> > > > > > This patch set allows tun to support AF_XDP Rx zero-copy feature.
> > > > > >
> > > > > > This patch tries to address this by:
> > > > > > - Use peek_len to consume a xsk->desc and get xsk->desc length.
> > > > > > - When the tun support AF_XDP Rx zero-copy, the vq's array maybe
> > empty.
> > > > > > So add a check for empty vq's array in vhost_net_buf_produce().
> > > > > > - add XDP_SETUP_XSK_POOL and ndo_xsk_wakeup callback support
> > > > > > - add tun_put_user_desc function to copy the Rx data to VM
> > > > >
> > > > > Code explains themselves, let's explain why you need to do this.
> > > > >
> > > > > 1) why you want to use peek_len
> > > > > 2) for "vq's array", what does it mean?
> > > > > 3) from the view of TUN/TAP tun_put_user_desc() is the TX path, so
> > > > > I guess you meant TX zerocopy instead of RX (as I don't see codes
> > > > > for
> > > > > RX?)
> > > >
> > > > OK, I agree and use TX zerocopy instead of RX zerocopy. I meant RX
> > > > zerocopy from the view of vhost-net.
> > > >
> > > > >
> > > > > A big question is how could you handle GSO packets from
> > userspace/guests?
> > > >
> > > > Now by disabling VM's TSO and csum feature. XDP does not support GSO
> > > > packets.
> > > > However, this feature can be added once XDP supports it in the future.
> > > >
> > > > >
> > > > > >
> > > > > > Signed-off-by: Yunjian Wang <wangyunjian@xxxxxxxxxx>
> > > > > > ---
> > > > > >  drivers/net/tun.c   | 165
> > > > > +++++++++++++++++++++++++++++++++++++++++++-
> > > > > >  drivers/vhost/net.c |  18 +++--
> > > > > >  2 files changed, 176 insertions(+), 7 deletions(-)
> > >
> > > [...]
> > >
> > > > > >
> > > > > >  static int peek_head_len(struct vhost_net_virtqueue *rvq,
> > > > > > struct sock
> > > > > > *sk)  {
> > > > > > +       struct socket *sock = sk->sk_socket;
> > > > > >         struct sk_buff *head;
> > > > > >         int len = 0;
> > > > > >         unsigned long flags;
> > > > > >
> > > > > > -       if (rvq->rx_ring)
> > > > > > -               return vhost_net_buf_peek(rvq);
> > > > > > +       if (rvq->rx_ring) {
> > > > > > +               len = vhost_net_buf_peek(rvq);
> > > > > > +               if (likely(len))
> > > > > > +                       return len;
> > > > > > +       }
> > > > > > +
> > > > > > +       if (sock->ops->peek_len)
> > > > > > +               return sock->ops->peek_len(sock);
> > > > >
> > > > > What prevents you from reusing the ptr_ring here? Then you don't
> > > > > need the above tricks.
> > > >
> > > > Thank you for your suggestion. I will consider how to reuse the ptr_ring.
> > >
> > > If ptr_ring is used to transfer xdp_descs, there is a problem: After
> > > some xdp_descs are obtained through xsk_tx_peek_desc(), the descs may
> > > fail to be added to ptr_ring. However, no API is available to
> > > implement the rollback function.
> >
> > I don't understand, this issue seems to exist in the physical NIC as well?
> >
> > We get more descriptors than the free slots in the NIC ring.
> >
> > How did other NIC solve this issue?
>
> Currently, physical NICs such as i40e, ice, ixgbe, igc, and mlx5 obtains
> available NIC descriptors and then retrieve the same number of xsk
> descriptors for processing.

Any reason we can't do the same? ptr_ring should be much simpler than
NIC ring anyhow.

Thanks

>
> Thanks
>
> >
> > Thanks
> >
> > >
> > > Thanks
> > >
> > > >
> > > > >
> > > > > Thanks
> > > > >
> > > > >
> > > > > >
> > > > > >         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
> > > > > >         head = skb_peek(&sk->sk_receive_queue);
> > > > > > --
> > > > > > 2.33.0
> > > > > >
> > >
>