Message-ID: <1697597798.3310452-1-xuanzhuo@linux.alibaba.com>
Date: Wed, 18 Oct 2023 10:56:38 +0800
From: Xuan Zhuo <xuanzhuo@...ux.alibaba.com>
To: Jason Wang <jasowang@...hat.com>
Cc: netdev@...r.kernel.org,
 "David S. Miller" <davem@...emloft.net>,
 Eric Dumazet <edumazet@...gle.com>,
 Jakub Kicinski <kuba@...nel.org>,
 Paolo Abeni <pabeni@...hat.com>,
 "Michael S. Tsirkin" <mst@...hat.com>,
 Alexei Starovoitov <ast@...nel.org>,
 Daniel Borkmann <daniel@...earbox.net>,
 Jesper Dangaard Brouer <hawk@...nel.org>,
 John Fastabend <john.fastabend@...il.com>,
 virtualization@...ts.linux-foundation.org,
 bpf@...r.kernel.org
Subject: Re: [PATCH net-next v1 00/19] virtio-net: support AF_XDP zero copy

On Wed, 18 Oct 2023 10:46:38 +0800, Jason Wang <jasowang@...hat.com> wrote:
> On Tue, Oct 17, 2023 at 7:28 PM Xuan Zhuo <xuanzhuo@...ux.alibaba.com> wrote:
> >
> > On Tue, 17 Oct 2023 14:43:33 +0800, Xuan Zhuo <xuanzhuo@...ux.alibaba.com> wrote:
> > > On Tue, 17 Oct 2023 14:26:01 +0800, Jason Wang <jasowang@...hat.com> wrote:
> > > > On Tue, Oct 17, 2023 at 2:17 PM Xuan Zhuo <xuanzhuo@...ux.alibaba.com> wrote:
> > > > >
> > > > > On Tue, 17 Oct 2023 13:27:47 +0800, Jason Wang <jasowang@...hat.com> wrote:
> > > > > > On Tue, Oct 17, 2023 at 11:28 AM Jason Wang <jasowang@...hat.com> wrote:
> > > > > > >
> > > > > > > On Tue, Oct 17, 2023 at 11:26 AM Xuan Zhuo <xuanzhuo@...ux.alibaba.com> wrote:
> > > > > > > >
> > > > > > > > On Tue, 17 Oct 2023 11:20:41 +0800, Jason Wang <jasowang@...hat.com> wrote:
> > > > > > > > > On Tue, Oct 17, 2023 at 11:11 AM Xuan Zhuo <xuanzhuo@...ux.alibaba.com> wrote:
> > > > > > > > > >
> > > > > > > > > > On Tue, 17 Oct 2023 10:53:44 +0800, Jason Wang <jasowang@...hat.com> wrote:
> > > > > > > > > > > On Mon, Oct 16, 2023 at 8:00 PM Xuan Zhuo <xuanzhuo@...ux.alibaba.com> wrote:
> > > > > > > > > > > >
> > > > > > > > > > > > ## AF_XDP
> > > > > > > > > > > >
> > > > > > > > > > > > XDP socket(AF_XDP) is an excellent bypass kernel network framework. The zero
> > > > > > > > > > > > copy feature of xsk (XDP socket) needs to be supported by the driver. The
> > > > > > > > > > > > performance of zero copy is very good. mlx5 and intel ixgbe already support
> > > > > > > > > > > > this feature, This patch set allows virtio-net to support xsk's zerocopy xmit
> > > > > > > > > > > > feature.
> > > > > > > > > > > >
> > > > > > > > > > > > At present, we have completed some preparation:
> > > > > > > > > > > >
> > > > > > > > > > > > 1. vq-reset (virtio spec and kernel code)
> > > > > > > > > > > > 2. virtio-core premapped dma
> > > > > > > > > > > > 3. virtio-net xdp refactor
> > > > > > > > > > > >
> > > > > > > > > > > > So it is time for Virtio-Net to complete the support for the XDP Socket
> > > > > > > > > > > > Zerocopy.
> > > > > > > > > > > >
> > > > > > > > > > > > Virtio-net can not increase the queue num at will, so xsk shares the queue with
> > > > > > > > > > > > kernel.
> > > > > > > > > > > >
> > > > > > > > > > > > On the other hand, Virtio-Net does not support generate interrupt from driver
> > > > > > > > > > > > manually, so when we wakeup tx xmit, we used some tips. If the CPU run by TX
> > > > > > > > > > > > NAPI last time is other CPUs, use IPI to wake up NAPI on the remote CPU. If it
> > > > > > > > > > > > is also the local CPU, then we wake up napi directly.
> > > > > > > > > > > >
> > > > > > > > > > > > This patch set includes some refactor to the virtio-net to let that to support
> > > > > > > > > > > > AF_XDP.
> > > > > > > > > > > >
> > > > > > > > > > > > ## performance
> > > > > > > > > > > >
> > > > > > > > > > > > ENV: Qemu with vhost-user(polling mode).
> > > > > > > > > > > >
> > > > > > > > > > > > Sockperf: https://github.com/Mellanox/sockperf
> > > > > > > > > > > > I use this tool to send udp packet by kernel syscall.
> > > > > > > > > > > >
> > > > > > > > > > > > xmit command: sockperf tp -i 10.0.3.1 -t 1000
> > > > > > > > > > > >
> > > > > > > > > > > > I write a tool that sends udp packets or recvs udp packets by AF_XDP.
> > > > > > > > > > > >
> > > > > > > > > > > >                   | Guest APP CPU |Guest Softirq CPU | UDP PPS
> > > > > > > > > > > > ------------------|---------------|------------------|------------
> > > > > > > > > > > > xmit by syscall   |   100%        |                  |   676,915
> > > > > > > > > > > > xmit by xsk       |   59.1%       |   100%           | 5,447,168
> > > > > > > > > > > > recv by syscall   |   60%         |   100%           |   932,288
> > > > > > > > > > > > recv by xsk       |   35.7%       |   100%           | 3,343,168
> > > > > > > > > > >
> > > > > > > > > > > Any chance we can get a testpmd result (which I guess should be better
> > > > > > > > > > > than PPS above)?
> > > > > > > > > >
> > > > > > > > > > Do you mean testpmd + DPDK + AF_XDP?
> > > > > > > > >
> > > > > > > > > Yes.
> > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > Yes. This is probably better because my tool does more work. That is not a
> > > > > > > > > > complete testing tool used by our business.
> > > > > > > > >
> > > > > > > > > Probably, but it would be appealing for others. Especially considering
> > > > > > > > > DPDK supports AF_XDP PMD now.
> > > > > > > >
> > > > > > > > OK.
> > > > > > > >
> > > > > > > > Let me try.
> > > > > > > >
> > > > > > > > But could you start to review firstly?
> > > > > > >
> > > > > > > Yes, it's in my todo list.
> > > > > >
> > > > > > Speaking too fast, I think if it doesn't take too long time, I would
> > > > > > wait for the result first as netdim series. One reason is that I
> > > > > > remember claims to be only 10% to 20% loss comparing to wire speed, so
> > > > > > I'd expect it should be much faster. I vaguely remember, even a vhost
> > > > > > can gives us more than 3M PPS if we disable SMAP, so the numbers here
> > > > > > are not as impressive as expected.
> > > > >
> > > > >
> > > > > What is SMAP? Cloud you give me more info?
> > > >
> > > > Supervisor Mode Access Prevention
> > > >
> > > > Vhost suffers from this.
> > > >
> > > > >
> > > > > So if we think the 3M as the wire speed, you expect the result
> > > > > can reach 2.8M pps/core, right?
> > > >
> > > > It's AF_XDP that claims to be 80% if my memory is correct. So a
> > > > correct AF_XDP implementation should not sit behind this too much.
> > > >
> > > > > Now the recv result is 2.5M(2463646) pps/core.
> > > > > Do you think there is a huge gap?
> > > >
> > > > You never describe your testing environment in details. For example,
> > > > is this a virtual environment? What's the CPU model and frequency etc.
> > > >
> > > > Because I never see a NIC whose wire speed is 3M.
> > > >
> > > > >
> > > > > My tool makes udp packet and lookup route, so it take more much cpu.
> > > >
> > > > That's why I suggest you to test raw PPS.
> > >
> > > OK. Let's align some info.
> > >
> > > 1. My test env is vhost-user. Qemu + vhost-user(polling mode).
> > >    I do not use the DPDK, because that there is some trouble for me.
> > >    I use the VAPP (https://github.com/fengidri/vapp) as the vhost-user device.
> > >    That has two threads all are busy mode for tx and rx.
> > >    tx thread consumes the tx ring and drop the packet.
> > >    rx thread put the packet to the rx ring.
> > >
> > > 2. My Host CPU: Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHz
> > >
> > > 3. From this http://fast.dpdk.org/doc/perf/DPDK_23_03_Intel_virtio_performance_report.pdf
> > >    I think we can align that the vhost max speed is 8.5 MPPS.
> > >    Is that ok?
> > >    And the expected AF_XDP pps is about 6 MPPS.
> > >
> > > 4. About the raw PPS, I agree that. I will test with testpmd.
> > >
> >
> > ## testpmd command
> >
> > ./build/app/dpdk-testpmd -l 1-2 --no-pci --main-lcore=2 \
> >         --vdev net_af_xdp0,iface=ens5,queue_count=1,busy_budget=0 \
> >         --log-level=pmd.net.af_xdp:8 \
> >         -- -i -a --nb-cores=1 --rxq=1 --txq=1 --forward-mode=macswap
> >
> > ## work without the follow patch[0]
> >
> > testpmd> show port stats all
> >
> >   ######################## NIC statistics for port 0  ########################
> >   RX-packets: 3615824336 RX-missed: 0          RX-bytes:  202486162816
> >   RX-errors: 0
> >   RX-nombuf:  0
> >   TX-packets: 3615795592 TX-errors: 20738      TX-bytes:  202484553152
> >
> >   Throughput (since last show)
> >   Rx-pps:      3790446          Rx-bps:   1698120056
> >   Tx-pps:      3790446          Tx-bps:   1698120056
> >   ############################################################################
> >
> >
> > ## work with the follow patch[0]
> >
> > testpmd> show port stats all
> >
> >   ######################## NIC statistics for port 0  ########################
> >   RX-packets: 68152727   RX-missed: 0          RX-bytes:  3816552712
> >   RX-errors: 0
> >   RX-nombuf:  0
> >   TX-packets: 68114967   TX-errors: 33216      TX-bytes:  3814438152
> >
> >   Throughput (since last show)
> >   Rx-pps:      6333196          Rx-bps:   2837272088
> >   Tx-pps:      6333227          Tx-bps:   2837285936
> >   ############################################################################
> >
> > I search the dpdk code that the dpdk virtio driver has the similar code.
> >
> > virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
> > {
> >         [...]
> >
> >         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
> >
> >                 [...]
> >
> >                 /* Enqueue Packet buffers */
> >                 virtqueue_enqueue_xmit(txvq, txm, slots, use_indirect,
> >                         can_push, 0);
> >         }
> >
> >         [...]
> >
> >         if (likely(nb_tx)) {
> > -->             vq_update_avail_idx(vq);
> >
> >                 if (unlikely(virtqueue_kick_prepare(vq))) {
> >                         virtqueue_notify(vq);
> >                         PMD_TX_LOG(DEBUG, "Notified backend after xmit");
> >                 }
> >         }
> > }
> >
> > ## patch[0]
> >
> > diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> > index 51d8f3299c10..cfe556b5d88f 100644
> > --- a/drivers/virtio/virtio_ring.c
> > +++ b/drivers/virtio/virtio_ring.c
> > @@ -687,12 +687,7 @@ static inline int virtqueue_add_split(struct virtqueue *_vq,
> >         avail = vq->split.avail_idx_shadow & (vq->split.vring.num - 1);
> >         vq->split.vring.avail->ring[avail] = cpu_to_virtio16(_vq->vdev, head);
> >
> > -       /* Descriptors and available array need to be set before we expose the
> > -        * new available array entries. */
> > -       virtio_wmb(vq->weak_barriers);
> >         vq->split.avail_idx_shadow++;
> > -       vq->split.vring.avail->idx = cpu_to_virtio16(_vq->vdev,
> > -                                               vq->split.avail_idx_shadow);
> >         vq->num_added++;
> >
> >         pr_debug("Added buffer head %i to %p\n", head, vq);
> > @@ -700,8 +695,12 @@ static inline int virtqueue_add_split(struct virtqueue *_vq,
> >
> >         /* This is very unlikely, but theoretically possible.  Kick
> >          * just in case. */
> > -       if (unlikely(vq->num_added == (1 << 16) - 1))
> > +       if (unlikely(vq->num_added == (1 << 16) - 1)) {
> > +               virtio_wmb(vq->weak_barriers);
> > +               vq->split.vring.avail->idx = cpu_to_virtio16(_vq->vdev,
> > +                                                            vq->split.avail_idx_shadow);
> >                 virtqueue_kick(_vq);
> > +       }
> >
> >         return 0;
> >
> > @@ -742,6 +741,9 @@ static bool virtqueue_kick_prepare_split(struct virtqueue *_vq)
> >          * event. */
> >         virtio_mb(vq->weak_barriers);
> >
> > +       vq->split.vring.avail->idx = cpu_to_virtio16(_vq->vdev,
> > +                                               vq->split.avail_idx_shadow);
> > +
>
> Looks like an interesting optimization.
>
> Would you mind posting this with numbers separately?

I will post this later.


>
> Btw, does the current API require virtqueue_kick_prepare() to be done
> before a virtqueue_notify(). If not, we need do something similar in
> virtqueue_notify()?

As I know, prepare is done before a notify.

I will check this doubly.

Thanks.


>
> Thanks
>
> >         old = vq->split.avail_idx_shadow - vq->num_added;
> >         new = vq->split.avail_idx_shadow;
> >         vq->num_added = 0;
> >
> > ---------------
> >
> > Thanks.
> >
> >
> > >
> > > Thanks.
> > >
> > >
> > > >
> > > > Thanks
> > > >
> > > > >
> > > > > I am confused.
> > > > >
> > > > >
> > > > > What is SMAP? Could you give me more information?
> > > > >
> > > > > So if we use 3M as the wire speed, you would expect the result to be 2.8M
> > > > > pps/core, right?
> > > > >
> > > > > Now the recv result is 2.5M (2463646 = 3,343,168/1.357) pps/core. Do you think
> > > > > the difference is big?
> > > > >
> > > > > My tool makes udp packets and looks up routes, so it requires more CPU.
> > > > >
> > > > > I'm confused. Is there something I misunderstood?
> > > > >
> > > > > Thanks.
> > > > >
> > > > > >
> > > > > > Thanks
> > > > > >
> > > > > > >
> > > > > > > >
> > > > > > > >
> > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > What I noticed is that the hotspot is the driver writing virtio desc. Because
> > > > > > > > > > the device is in busy mode. So there is race between driver and device.
> > > > > > > > > > So I modified the virtio core and lazily updated avail idx. Then pps can reach
> > > > > > > > > > 10,000,000.
> > > > > > > > >
> > > > > > > > > Care to post a draft for this?
> > > > > > > >
> > > > > > > > YES, I is thinking for this.
> > > > > > > > But maybe that is just work for split. The packed mode has some troubles.
> > > > > > >
> > > > > > > Ok.
> > > > > > >
> > > > > > > Thanks
> > > > > > >
> > > > > > > >
> > > > > > > > Thanks.
> > > > > > > >
> > > > > > > > >
> > > > > > > > > Thanks
> > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > Thanks.
> > > > > > > > > >
> > > > > > > > > > >
> > > > > > > > > > > Thanks
> > > > > > > > > > >
> > > > > > > > > > > >
> > > > > > > > > > > > ## maintain
> > > > > > > > > > > >
> > > > > > > > > > > > I am currently a reviewer for virtio-net. I commit to maintain AF_XDP support in
> > > > > > > > > > > > virtio-net.
> > > > > > > > > > > >
> > > > > > > > > > > > Please review.
> > > > > > > > > > > >
> > > > > > > > > > > > Thanks.
> > > > > > > > > > > >
> > > > > > > > > > > > v1:
> > > > > > > > > > > >     1. remove two virtio commits. Push this patchset to net-next
> > > > > > > > > > > >     2. squash "virtio_net: virtnet_poll_tx support rescheduled" to xsk: support tx
> > > > > > > > > > > >     3. fix some warnings
> > > > > > > > > > > >
> > > > > > > > > > > > Xuan Zhuo (19):
> > > > > > > > > > > >   virtio_net: rename free_old_xmit_skbs to free_old_xmit
> > > > > > > > > > > >   virtio_net: unify the code for recycling the xmit ptr
> > > > > > > > > > > >   virtio_net: independent directory
> > > > > > > > > > > >   virtio_net: move to virtio_net.h
> > > > > > > > > > > >   virtio_net: add prefix virtnet to all struct/api inside virtio_net.h
> > > > > > > > > > > >   virtio_net: separate virtnet_rx_resize()
> > > > > > > > > > > >   virtio_net: separate virtnet_tx_resize()
> > > > > > > > > > > >   virtio_net: sq support premapped mode
> > > > > > > > > > > >   virtio_net: xsk: bind/unbind xsk
> > > > > > > > > > > >   virtio_net: xsk: prevent disable tx napi
> > > > > > > > > > > >   virtio_net: xsk: tx: support tx
> > > > > > > > > > > >   virtio_net: xsk: tx: support wakeup
> > > > > > > > > > > >   virtio_net: xsk: tx: virtnet_free_old_xmit() distinguishes xsk buffer
> > > > > > > > > > > >   virtio_net: xsk: tx: virtnet_sq_free_unused_buf() check xsk buffer
> > > > > > > > > > > >   virtio_net: xsk: rx: introduce add_recvbuf_xsk()
> > > > > > > > > > > >   virtio_net: xsk: rx: introduce receive_xsk() to recv xsk buffer
> > > > > > > > > > > >   virtio_net: xsk: rx: virtnet_rq_free_unused_buf() check xsk buffer
> > > > > > > > > > > >   virtio_net: update tx timeout record
> > > > > > > > > > > >   virtio_net: xdp_features add NETDEV_XDP_ACT_XSK_ZEROCOPY
> > > > > > > > > > > >
> > > > > > > > > > > >  MAINTAINERS                                 |   2 +-
> > > > > > > > > > > >  drivers/net/Kconfig                         |   8 +-
> > > > > > > > > > > >  drivers/net/Makefile                        |   2 +-
> > > > > > > > > > > >  drivers/net/virtio/Kconfig                  |  13 +
> > > > > > > > > > > >  drivers/net/virtio/Makefile                 |   8 +
> > > > > > > > > > > >  drivers/net/{virtio_net.c => virtio/main.c} | 652 +++++++++-----------
> > > > > > > > > > > >  drivers/net/virtio/virtio_net.h             | 359 +++++++++++
> > > > > > > > > > > >  drivers/net/virtio/xsk.c                    | 545 ++++++++++++++++
> > > > > > > > > > > >  drivers/net/virtio/xsk.h                    |  32 +
> > > > > > > > > > > >  9 files changed, 1247 insertions(+), 374 deletions(-)
> > > > > > > > > > > >  create mode 100644 drivers/net/virtio/Kconfig
> > > > > > > > > > > >  create mode 100644 drivers/net/virtio/Makefile
> > > > > > > > > > > >  rename drivers/net/{virtio_net.c => virtio/main.c} (91%)
> > > > > > > > > > > >  create mode 100644 drivers/net/virtio/virtio_net.h
> > > > > > > > > > > >  create mode 100644 drivers/net/virtio/xsk.c
> > > > > > > > > > > >  create mode 100644 drivers/net/virtio/xsk.h
> > > > > > > > > > > >
> > > > > > > > > > > > --
> > > > > > > > > > > > 2.32.0.3.g01195cf9f
> > > > > > > > > > > >
> > > > > > > > > > >
> > > > > > > > > >
> > > > > > > > >
> > > > > > > >
> > > > > >
> > > > >
> > > >
> > >
> >
>

Powered by blists - more mailing lists