| Message ID | 152775721817.24817.11576562399044807823.stgit@firesoul |
|---|---|
| State | Accepted, archived |
| Delegated to: | BPF Maintainers |
| Headers | show |
| Series | bpf/xdp: add flags argument to ndo_xdp_xmit and flag flush operation | expand |
On Thu, May 31, 2018 at 2:00 AM, Jesper Dangaard Brouer <brouer@redhat.com> wrote: > This is the first real user of the XDP_XMIT_FLUSH flag. > > As pointed out many times, XDP_REDIRECT without using BPF maps is > significant slower than the map variant. This is primary due to the > lack of bulking, as the ndo_xdp_flush operation is required after each > frame (to avoid frames hanging on the egress device). > > It is still possible to optimize this case. Instead of invoking two > NDO indirect calls, which are very expensive with CONFIG_RETPOLINE, > instead instruct ndo_xdp_xmit to flush via XDP_XMIT_FLUSH flag. > > Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> Acked-by: Song Liu <songliubraving@fb.com> > --- > net/core/filter.c | 3 +-- > 1 file changed, 1 insertion(+), 2 deletions(-) > > diff --git a/net/core/filter.c b/net/core/filter.c > index 6a21dbcad350..6981b4608979 100644 > --- a/net/core/filter.c > +++ b/net/core/filter.c > @@ -3056,10 +3056,9 @@ static int __bpf_tx_xdp(struct net_device *dev, > if (unlikely(!xdpf)) > return -EOVERFLOW; > > - sent = dev->netdev_ops->ndo_xdp_xmit(dev, 1, &xdpf, 0); > + sent = dev->netdev_ops->ndo_xdp_xmit(dev, 1, &xdpf, XDP_XMIT_FLUSH); > if (sent <= 0) > return sent; > - dev->netdev_ops->ndo_xdp_flush(dev); > return 0; > } > >
diff --git a/net/core/filter.c b/net/core/filter.c index 6a21dbcad350..6981b4608979 100644 --- a/net/core/filter.c +++ b/net/core/filter.c @@ -3056,10 +3056,9 @@ static int __bpf_tx_xdp(struct net_device *dev, if (unlikely(!xdpf)) return -EOVERFLOW; - sent = dev->netdev_ops->ndo_xdp_xmit(dev, 1, &xdpf, 0); + sent = dev->netdev_ops->ndo_xdp_xmit(dev, 1, &xdpf, XDP_XMIT_FLUSH); if (sent <= 0) return sent; - dev->netdev_ops->ndo_xdp_flush(dev); return 0; }
This is the first real user of the XDP_XMIT_FLUSH flag. As pointed out many times, XDP_REDIRECT without using BPF maps is significant slower than the map variant. This is primary due to the lack of bulking, as the ndo_xdp_flush operation is required after each frame (to avoid frames hanging on the egress device). It is still possible to optimize this case. Instead of invoking two NDO indirect calls, which are very expensive with CONFIG_RETPOLINE, instead instruct ndo_xdp_xmit to flush via XDP_XMIT_FLUSH flag. Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com> --- net/core/filter.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)