| Message ID | 20190216083524.860-1-rdong.ge@gmail.com |
|---|---|
| State | Accepted |
| Delegated to: | David Miller |
| Headers | show |
| Series | [v2] net: dsa: Implement flow_dissect callback for tag_dsa. | expand |
Hi Andrew I have tested the L3 forwarding throughput performance of my box (with an intel 4-core processor and each core’s frequency is 2.20GHz). In my test scenario, I generated 200 UDP flows (frame size is 64 bytes) with different src/dst pairs from eth1 to eth0. Eth1 is a slave DSA device from mv88e6190x, eth0 is an intel NIC. Then get the max fps with no frame losing. The max fps with RPS working is 400235fps, frames are hashed to four cores’ backlog. And the max fps without RPS is 199686fps. Rundong Rundong Ge <rdong.ge@gmail.com> 于2019年2月16日周六 下午4:35写道： > > RPS not work for DSA devices since the 'skb_get_hash' > will always get the invalid hash for dsa tagged packets. > > "[PATCH] tag_mtk: add flow_dissect callback to the ops struct" > introduced the flow_dissect callback to get the right hash for > MTK tagged packet. Tag_dsa and tag_edsa also need to implement > the callback. > > Signed-off-by: Rundong Ge <rdong.ge@gmail.com> > --- > net/dsa/tag_dsa.c | 9 +++++++++ > net/dsa/tag_edsa.c | 9 +++++++++ > 2 files changed, 18 insertions(+) > > diff --git a/net/dsa/tag_dsa.c b/net/dsa/tag_dsa.c > index 8b2f92e..67ff3fa 100644 > --- a/net/dsa/tag_dsa.c > +++ b/net/dsa/tag_dsa.c > @@ -146,8 +146,17 @@ static struct sk_buff *dsa_rcv(struct sk_buff *skb, struct net_device *dev, > return skb; > } > > +static int dsa_tag_flow_dissect(const struct sk_buff *skb, __be16 *proto, > + int *offset) > +{ > + *offset = 4; > + *proto = ((__be16 *)skb->data)[1]; > + return 0; > +} > + > const struct dsa_device_ops dsa_netdev_ops = { > .xmit = dsa_xmit, > .rcv = dsa_rcv, > + .flow_dissect = dsa_tag_flow_dissect, > .overhead = DSA_HLEN, > }; > diff --git a/net/dsa/tag_edsa.c b/net/dsa/tag_edsa.c > index f5b87ee..234585e 100644 > --- a/net/dsa/tag_edsa.c > +++ b/net/dsa/tag_edsa.c > @@ -165,8 +165,17 @@ static struct sk_buff *edsa_rcv(struct sk_buff *skb, struct net_device *dev, > return skb; > } > > +static int edsa_tag_flow_dissect(const struct sk_buff *skb, __be16 *proto, > + int *offset) > +{ > + *offset = 8; > + *proto = ((__be16 *)skb->data)[3]; > + return 0; > +} > + > const struct dsa_device_ops edsa_netdev_ops = { > .xmit = edsa_xmit, > .rcv = edsa_rcv, > + .flow_dissect = edsa_tag_flow_dissect, > .overhead = EDSA_HLEN, > }; > -- > 1.8.3.1 >
On Sat, Feb 16, 2019 at 04:38:25PM +0800, Rundong Ge wrote: > Hi Andrew > > I have tested the L3 forwarding throughput performance of my box (with > an intel 4-core processor and each core’s frequency is 2.20GHz). > In my test scenario, I generated 200 UDP flows (frame size is 64 > bytes) with different src/dst pairs from eth1 to eth0. Eth1 is a slave > DSA device from mv88e6190x, eth0 is an intel NIC. Then get the max fps > with no frame losing. > > The max fps with RPS working is 400235fps, frames are hashed to four > cores’ backlog. > And the max fps without RPS is 199686fps. Nice numbers, thanks. Andrew
On Sat, Feb 16, 2019 at 08:35:24AM +0000, Rundong Ge wrote: > RPS not work for DSA devices since the 'skb_get_hash' > will always get the invalid hash for dsa tagged packets. > > "[PATCH] tag_mtk: add flow_dissect callback to the ops struct" > introduced the flow_dissect callback to get the right hash for > MTK tagged packet. Tag_dsa and tag_edsa also need to implement > the callback. > > Signed-off-by: Rundong Ge <rdong.ge@gmail.com> Reviewed-by: Andrew Lunn <andrew@lunn.ch> Andrew
From: Rundong Ge <rdong.ge@gmail.com> Date: Sat, 16 Feb 2019 08:35:24 +0000 > RPS not work for DSA devices since the 'skb_get_hash' > will always get the invalid hash for dsa tagged packets. > > "[PATCH] tag_mtk: add flow_dissect callback to the ops struct" > introduced the flow_dissect callback to get the right hash for > MTK tagged packet. Tag_dsa and tag_edsa also need to implement > the callback. > > Signed-off-by: Rundong Ge <rdong.ge@gmail.com> Applied to net-next.
diff --git a/net/dsa/tag_dsa.c b/net/dsa/tag_dsa.c index 8b2f92e..67ff3fa 100644 --- a/net/dsa/tag_dsa.c +++ b/net/dsa/tag_dsa.c @@ -146,8 +146,17 @@ static struct sk_buff *dsa_rcv(struct sk_buff *skb, struct net_device *dev, return skb; } +static int dsa_tag_flow_dissect(const struct sk_buff *skb, __be16 *proto, + int *offset) +{ + *offset = 4; + *proto = ((__be16 *)skb->data)[1]; + return 0; +} + const struct dsa_device_ops dsa_netdev_ops = { .xmit = dsa_xmit, .rcv = dsa_rcv, + .flow_dissect = dsa_tag_flow_dissect, .overhead = DSA_HLEN, }; diff --git a/net/dsa/tag_edsa.c b/net/dsa/tag_edsa.c index f5b87ee..234585e 100644 --- a/net/dsa/tag_edsa.c +++ b/net/dsa/tag_edsa.c @@ -165,8 +165,17 @@ static struct sk_buff *edsa_rcv(struct sk_buff *skb, struct net_device *dev, return skb; } +static int edsa_tag_flow_dissect(const struct sk_buff *skb, __be16 *proto, + int *offset) +{ + *offset = 8; + *proto = ((__be16 *)skb->data)[3]; + return 0; +} + const struct dsa_device_ops edsa_netdev_ops = { .xmit = edsa_xmit, .rcv = edsa_rcv, + .flow_dissect = edsa_tag_flow_dissect, .overhead = EDSA_HLEN, };
RPS not work for DSA devices since the 'skb_get_hash' will always get the invalid hash for dsa tagged packets. "[PATCH] tag_mtk: add flow_dissect callback to the ops struct" introduced the flow_dissect callback to get the right hash for MTK tagged packet. Tag_dsa and tag_edsa also need to implement the callback. Signed-off-by: Rundong Ge <rdong.ge@gmail.com> --- net/dsa/tag_dsa.c | 9 +++++++++ net/dsa/tag_edsa.c | 9 +++++++++ 2 files changed, 18 insertions(+)