On Tue, Dec 18, 2018 at 6:45 AM Björn Töpel <bjorn.to...@gmail.com> wrote:
>
> Den mån 17 dec. 2018 kl 20:40 skrev William Tu <u9012...@gmail.com>:
> >
> > The patch adds support for AF_XDP async xmit. Users can use
> > AF_XDP on both side of the veth and get better performance, with
> > the cost of ksoftirqd doing the xmit. The veth_xsk_async_xmit
> > simply kicks the napi function, veth_poll, to run, and the transmit
> > logic is implemented there.
> >
> > Tested using two namespaces, one runs xdpsock and the other runs
> > xdp_rxq_info. A simple script comparing the performance with/without
> > AF_XDP shows improvement from 724Kpps to 1.1Mpps.
> >
> > ip netns add at_ns0
> > ip link add p0 type veth peer name p1
> > ip link set p0 netns at_ns0
> > ip link set dev p1 up
> > ip netns exec at_ns0 ip link set dev p0 up
> >
> > # receiver
> > ip netns exec at_ns0 xdp_rxq_info --dev p0 --action XDP_DROP
> >
> > # sender
> > xdpsock -i p1 -t -N -z
> > or
> > xdpsock -i p1 -t -S
> >
> > Signed-off-by: William Tu <u9012...@gmail.com>
> > ---
> > drivers/net/veth.c | 247
> > ++++++++++++++++++++++++++++++++++++++++++++++++++++-
> > 1 file changed, 245 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/net/veth.c b/drivers/net/veth.c
> > index f412ea1cef18..5171ddad5973 100644
> > --- a/drivers/net/veth.c
> > +++ b/drivers/net/veth.c
> > @@ -25,6 +25,10 @@
> > #include <linux/ptr_ring.h>
> > #include <linux/bpf_trace.h>
> > #include <linux/net_tstamp.h>
> > +#include <net/xdp_sock.h>
> > +#include <linux/mm.h>
> > +#include <linux/slab.h>
> > +#include <net/page_pool.h>
> >
> > #define DRV_NAME "veth"
> > #define DRV_VERSION "1.0"
> > @@ -53,6 +57,7 @@ struct veth_rq {
> > bool rx_notify_masked;
> > struct ptr_ring xdp_ring;
> > struct xdp_rxq_info xdp_rxq;
> > + struct xdp_umem *xsk_umem;
> > };
> >
> > struct veth_priv {
> > @@ -61,6 +66,11 @@ struct veth_priv {
> > struct bpf_prog *_xdp_prog;
> > struct veth_rq *rq;
> > unsigned int requested_headroom;
> > +
> > + /* AF_XDP zero-copy */
> > + struct xdp_umem **xsk_umems;
> > + u16 num_xsk_umems_used;
> > + u16 num_xsk_umems;
> > };
> >
>
> The umems are since commit 661b8d1b0e3a ("net: add umem reference in
> netdev{_rx}_queue") part of the netdev structure, so instead of
> storing the umem in the priv area, you can pull it from the netdev
> device. Doing so means that you can remove all the *_umem() functions
> below.
>
Thanks, will use it in next version.
> Jan (added as Cc:) just did that work on the i40e driver (patches
> posted on intel-wired-lan list).
>
> > /*
> > @@ -742,10 +752,87 @@ static int veth_poll(struct napi_struct *napi, int
> > budget)
> > struct veth_rq *rq =
> > container_of(napi, struct veth_rq, xdp_napi);
> > unsigned int xdp_xmit = 0;
> > - int done;
> > + int tx_budget = budget;
> > + int done = 0;
> > +
> > + /* tx: use netif_tx_napi_add or here? */
> > + while (rq->xsk_umem && tx_budget--) {
> > + struct veth_priv *priv, *peer_priv;
> > + struct net_device *dev, *peer_dev;
> > + unsigned int inner_xdp_xmit = 0;
> > + unsigned int metasize = 0;
> > + struct veth_rq *peer_rq;
> > + struct xdp_frame *xdpf;
> > + bool dropped = false;
> > + struct sk_buff *skb;
> > + struct page *page;
> > + char *vaddr;
> > + void *addr;
> > + u32 len;
> > +
> > + if (!xsk_umem_consume_tx_virtual(rq->xsk_umem, &vaddr,
> > &len))
> > + break;
> > +
> > + page = dev_alloc_page();
> > + if (!page)
> > + return -ENOMEM;
> > +
> > + addr = page_to_virt(page);
> > + xdpf = addr;
> > + memset(xdpf, 0, sizeof(*xdpf));
> > +
> > + addr += sizeof(*xdpf);
> > + memcpy(addr, vaddr, len);
> > +
> > + xdpf->data = addr + metasize;
> > + xdpf->len = len;
> > + xdpf->headroom = 0;
> > + xdpf->metasize = metasize;
> > + xdpf->mem.type = MEM_TYPE_PAGE_SHARED;
> > +
> > + /* Invoke peer rq to rcv */
> > + dev = rq->dev;
> > + priv = netdev_priv(dev);
> > + peer_dev = priv->peer;
> > + peer_priv = netdev_priv(peer_dev);
> > + peer_rq = peer_priv->rq;
There is a bug here, where in this case we'll always use peer_rq[0]
Instead, we should do peer_priv->rq[qid];
> > +
> > + /* put into peer rq */
> > + skb = veth_xdp_rcv_one(peer_rq, xdpf, &inner_xdp_xmit);
> > + if (!skb) {
> > + /* Peer side has XDP program attached */
> > + if (inner_xdp_xmit & VETH_XDP_TX) {
> > + /* Not supported */
> > + xsk_umem_complete_tx(rq->xsk_umem, 1);
> > + xsk_umem_consume_tx_done(rq->xsk_umem);
> > + xdp_return_frame(xdpf);
> > + goto skip_tx;
> > + } else if (inner_xdp_xmit & VETH_XDP_REDIR) {
> > + xdp_do_flush_map();
> > + } else {
> > + dropped = true;
> > + }
> > + } else {
> > + /* Peer side has no XDP attached */
> > + napi_gro_receive(&peer_rq->xdp_napi, skb);
> > + }
> > + xsk_umem_complete_tx(rq->xsk_umem, 1);
> > + xsk_umem_consume_tx_done(rq->xsk_umem);
> > +
> > + /* update peer stats */
> > + u64_stats_update_begin(&peer_rq->stats.syncp);
> > + peer_rq->stats.xdp_packets++;
> > + peer_rq->stats.xdp_bytes += len;
> > + if (dropped)
> > + rq->stats.xdp_drops++;
> > + u64_stats_update_end(&peer_rq->stats.syncp);
> > + done++;
> > + }
>
> Refactor to a function?
Sure, will do it.
Regards,
William
