[RFC net-next 0/3] Add RSS and LRO support

[Frank Wunderlich]

From: Frank Wunderlich <frank-w at public-files.de>

This series is currently only for discussion to get the upported SDK driver
changes in a good shape.
To use it some other parts are still missing like USXGMII PCS support on
mt7988.

patches are upported from mtk SDK:
- https://git01.mediatek.com/plugins/gitiles/openwrt/feeds/mtk-openwrt-feeds/+/refs/heads/master/master/files/target/linux/mediatek/patches-6.12/999-eth-08-mtk_eth_soc-add-register-definitions-for-rss-lro-reg.patch
- https://git01.mediatek.com/plugins/gitiles/openwrt/feeds/mtk-openwrt-feeds/+/refs/heads/master/master/files/target/linux/mediatek/patches-6.12/999-eth-09-mtk_eth_soc-add-rss-support.patch
- https://git01.mediatek.com/plugins/gitiles/openwrt/feeds/mtk-openwrt-feeds/+/refs/heads/master/master/files/target/linux/mediatek/patches-6.12/999-eth-10-mtk_eth_soc-add-hw-lro-support.patch

RSS / LRO

HW-acceleration for ending traffic. For routed traffic PPE is needed and
hw offloading in nftables.
Bridged traffic may need additional changes (openwrt use bridger utility).

RSS (Receive Side Scaling)

using 4 additional IRQ for spreading load

cat /proc/interrupts | grep ethernet

echo 1 > /proc/irq/105/smp_affinity
echo 2 > /proc/irq/106/smp_affinity
echo 4 > /proc/irq/107/smp_affinity
echo 8 > /proc/irq/108/smp_affinity

moving tx frame-engine irq to different cpu (here 3rd)
echo 4 > /proc/irq/103/smp_affinity

disable RPS (Receive Packet Steering) for all macs:

echo 0 > /sys/devices/platform/soc/15100000.ethernet/net/eth0/queues/rx-0/rps_cpus

pay attention on iperf-version (iperf 3.17 is ok, 3.12 is not)

traffic must be created using multiple streams so that it can be splitted, so use
multithreaded iperf3

on R4: bin/iperf3 -s
on the other side: iperf3 -c 192.168.1.1 -i 1 -P 4

you should reach ~9.3 GBit/s

and see spreading load over CPU cores

root at bpi-r4-phy-8G:~# cat /proc/interrupts | grep eth
103: 20 198366 0 0 GICv3 229 Level 15100000.ethernet
105: 3611 0 0 0 GICv3 221 Level 15100000.ethernet, 15100000.ethernet
106: 2 6842 0 0 GICv3 222 Level 15100000.ethernet, 15100000.ethernet
107: 4 0 27643 0 GICv3 223 Level 15100000.ethernet, 15100000.ethernet
108: 3 0 0 27925 GICv3 224 Level 15100000.ethernet, 15100000.ethernet

using the iperf3 from debian bookworm (3.12) results in only 6.7GBit/s, so
newer version is needed (not tested yet in trixie).

LRO(Large Receive Offload)

Add HW LRO RX rule:

ethtool -N [interface] flow-type tcp4 dst-ip [IP] action 0 loc [0/1]

Delete HW LRO RX rule:

ethtool -N [interface] delete [0/1]

Enable/Disable HW LRO rule:

ethtool -K [interface] lro [on | off]

Show the current offload features:

ethtool -k [interface]

example:

ethtool -N eth2 flow-type tcp4 dst-ip 192.168.1.1 action 0 loc 0
ethtool -K eth2 lro on ethtool -k eth2

using iperf(2) instead of iperf3 to reach full traffic!

verify with propritary debugfs (not part of this series)

Enable HW LRO rings
echo 4 1 > /proc/mtketh/hw_lro_auto_tlb
Enable HW LRO statistics
echo 5 1 > /proc/mtketh/hw_lro_auto_tlb

cat /proc/mtketh/hw_lro_stats

Mason Chang (3):
  net: ethernet: mtk_eth_soc: Add register definitions for RSS and LRO
  net: ethernet: mtk_eth_soc: Add RSS support
  net: ethernet: mtk_eth_soc: Add LRO support

 drivers/net/ethernet/mediatek/mtk_eth_soc.c | 770 ++++++++++++++++----
 drivers/net/ethernet/mediatek/mtk_eth_soc.h | 171 +++--
 2 files changed, 757 insertions(+), 184 deletions(-)

-- 
2.43.0