The batman-adv unicast packets contain a full layer 2 frame in encapsulated form. The flow dissector must therefore be able to parse the batman-adv unicast header to reach the layer 2+3 information to allow RPS to schedule the flow to different cores.
Signed-off-by: Sven Eckelmann sven.eckelmann@openmesh.com ---
I've played around with the RPS+XPS to better distribute traffic to different cores on modern SoCs. It was observed that the batman-adv unicast traffic stayed at one core and this limited the overall throughput.
The test was done between a modern PC which was running iperf over batman-adv. Another PC was also running iperf was just transmitting the traffic over ethernet. A QCA Dakota (IPQ4018) based device was in the middle and received the ethernet traffic on one port, encapsulated it as batman-adv unicast and transferred the resulting packet via ethernet to the modern PC.
$ iperf -c 192.168.23.2 -t 30 -i 1 -P16
It was observed that only the receive processing was affected. Adding this relative simple patch allowed me to increase the ethernet throughput from 689 Mbits/sec (with one core doing the interrupt processing and 1 core completely busy with the other processing) to 910 Mbits/sec (one core doing the interrupt processing and the other 3 cores were sharing the load). In theory, the device with the patch could have handled a lot more traffic than 910 Mbit/sec - at least the RPS cores were not even near 100% load.
This should bring a lot more when testing with good 802.11ac WiFi links because the wifi stack/driver are usually stressing the CPU a lot more than normal ethernet drivers.
The change itself is not really invasive but could be a lot cleaner when the packet header information would be available under include/... . And I don't know yet whether it would make sense to add a new flow dissector key with the batman-adv destination address. This is also the reason why I only post this to the b.a.t.m.a.n@lists.open-mesh.org mailing list.
Anyway, maybe some other people ran into the same problem and want to test it. --- net/core/flow_dissector.c | 33 +++++++++++++++++++++++++++++++++ 1 file changed, 33 insertions(+)
diff --git a/net/core/flow_dissector.c b/net/core/flow_dissector.c index 0a977373d003..bbf894fff3fc 100644 --- a/net/core/flow_dissector.c +++ b/net/core/flow_dissector.c @@ -595,6 +595,39 @@ bool __skb_flow_dissect(const struct sk_buff *skb,
break; } + case htons(ETH_P_BATMAN): { + /* TODO provide struct in include/uapi/... header */ + struct batadv_unicast_packet { + u8 packet_type; + u8 version; + u8 ttl; + u8 ttvn; + u8 dest[ETH_ALEN]; + }; + + struct { + struct batadv_unicast_packet batadv_unicast; + struct ethhdr eth; + } *hdr, _hdr; + + hdr = __skb_header_pointer(skb, nhoff, sizeof(_hdr), data, hlen, + &_hdr); + if (!hdr) + goto out_bad; + + if (hdr->batadv_unicast.version != 15) + break; + + if (hdr->batadv_unicast.packet_type != 0x40) + break; + + /* TODO set batman-adv dest address as flow key? */ + + proto = hdr->eth.h_proto; + nhoff += sizeof(*hdr); + + goto again; + } case htons(ETH_P_8021AD): case htons(ETH_P_8021Q): { const struct vlan_hdr *vlan;
On Montag, 27. November 2017 13:33:41 CET Sven Eckelmann wrote:
The batman-adv unicast packets contain a full layer 2 frame in encapsulated form. The flow dissector must therefore be able to parse the batman-adv unicast header to reach the layer 2+3 information to allow RPS to schedule the flow to different cores.
Signed-off-by: Sven Eckelmann sven.eckelmann@openmesh.com
[...]
Arghh, this should have been an RFC. I will mark it this way in patchwork but please don't be confused on the mailing list about the subject.
Kind regards, Sven
b.a.t.m.a.n@lists.open-mesh.org