158. Transmit Segmentation Offload (TSO) Tests¶

158.1. Description¶

This document provides the plan for testing the TSO (Transmit Segmentation Offload, also called Large Send offload - LSO) feature of Intel Ethernet Controller, including Intel 82599 10GbE Ethernet Controller and Intel® Ethernet Converged Network Adapter XL710-QDA2. TSO enables the TCP/IP stack to pass to the network device a larger ULP datagram than the Maximum Transmit Unit Size (MTU). NIC divides the large ULP datagram to multiple segments according to the MTU size.

158.2. Prerequisites¶

Hardware:: Intel® Ethernet 700 Series, Intel® Ethernet 800 Series and 82599/500 Series

The DUT must take one of the Ethernet controller ports connected to a port on another device that is controlled by the Scapy packet generator.

The Ethernet interface identifier of the port that Scapy will use must be known. On tester, all offload feature should be disabled on tx port, and start rx port capture:

ifconfig <tx port> mtu 9000
ethtool -K <tx port> rx off tx off tso off gso off gro off lro off
ip l set <tx port> up
tcpdump -n -e -i <rx port> -s 0 -w /tmp/cap

On DUT, run pmd with parameter “–enable-rx-cksum”. Then enable TSO on tx port and checksum on rx port. The test commands is below:

#enable hw checksum on rx port
csum set ip hw 0
csum set udp hw 0
csum set tcp hw 0
csum set sctp hw 0
set fwd csum

# enable TSO on tx port
*tso set 800 1

158.3. Test case: csum fwd engine, use TSO¶

This test uses Scapy to send out one large TCP package. The dut forwards package with TSO enable on tx port while rx port turns checksum on. After package send out by TSO on tx port, the tester receives multiple small TCP package.

Turn off tx port by ethtool on tester:

ethtool -K <tx port> rx off tx off tso off gso off gro off lro off
ip l set <tx port> up

capture package rx port on tester:

tcpdump -n -e -i <rx port> -s 0 -w /tmp/cap

Launch the userland testpmd application on DUT as follows:

./x86_64-native-linuxapp-gcc/app/dpdk-testpmd -c 0xffffffff -n 2 -- -i --rxd=512 --txd=512
--burst=32 --rxfreet=64 --mbcache=128 --portmask=0x3 --txpt=36 --txht=0 --txwt=0
--txfreet=32 --txrst=32 --enable-rx-cksum
  testpmd> set verbose 1
  # should stop ports before set csum and start ports after the settings
  testpmd> port stop all
# enable hw checksum on rx port
testpmd> csum set ip hw 0
testpmd> csum set udp hw 0
testpmd> csum set tcp hw 0
testpmd> csum set sctp hw 0
testpmd> csum set outer-ip hw 0
testpmd> csum parse-tunnel on 0
# enable TSO on tx port
testpmd> tso set 800 1
# set fwd engine and start
testpmd> set fwd csum
testpmd> port start all
testpmd> set promisc all off
testpmd> start

Test IPv4() in scapy:

sendp([Ether(dst="%s", src="52:00:00:00:00:00")/IP(src="192.168.1.1",dst="192.168.1.2")/UDP(sport=1021,dport=1021)/Raw(load="\x50"*%s)], iface="%s")

Test IPv6() in scapy:

sendp([Ether(dst="%s", src="52:00:00:00:00:00")/IPv6(src="FE80:0:0:0:200:1FF:FE00:200", dst="3555:5555:6666:6666:7777:7777:8888:8888")/UDP(sport=1021,dport=1021)/Raw(load="\x50"*%s)], iface="%s"

158.4. Test case: csum fwd engine, use TSO tunneling¶

not support nic: IXGBE_10G-82599_SFP, IGC-I225_LM, IGC-I226_LM.

This test uses Scapy to send out one large TCP package. The dut forwards package with TSO enable on tx port while rx port turns checksum on. After package send out by TSO on tx port, the tester receives multiple small TCP package.

Turn off tx port by ethtool on tester:

ethtool -K <tx port> rx off tx off tso off gso off gro off lro off
ip l set <tx port> up

capture package rx port on tester:

tcpdump -n -e -i <rx port> -s 0 -w /tmp/cap

Launch the userland testpmd application on DUT as follows:

./x86_64-native-linuxapp-gcc/app/dpdk-testpmd -c 0xffffffff -n 2 -- -i --rxd=512 --txd=512
--burst=32 --rxfreet=64 --mbcache=128 --portmask=0x3 --txpt=36 --txht=0 --txwt=0
--txfreet=32 --txrst=32 --enable-rx-cksum
  testpmd> set verbose 1

  testpmd> port stop all
# enable hw checksum on rx port
testpmd> csum set ip hw 0
testpmd> csum set udp hw 0
testpmd> csum set tcp hw 0
testpmd> csum set sctp hw 0
testpmd> csum set outer-ip hw 0
#Intel® Ethernet 700 Series not support outer udp
testpmd> csum set outer-udp hw 0
testpmd> csum parse-tunnel on 0

# enable hw checksum on tx port
testpmd> csum set ip hw 1
testpmd> csum set udp hw 1
testpmd> csum set tcp hw 1
testpmd> csum set sctp hw 1
#csum set outer-ip must be set to hw if outer L3 is IPv4
testpmd> csum set outer-ip hw 1
#csum parse-tunnel must be set so that tunneled packets are recognized
testpmd> csum parse-tunnel on 1
#Intel® Ethernet 700 Series not support outer udp
testpmd> csum set outer-udp hw 1

# enable TSO on tx port
testpmd> tunnel_tso set 800 1
# enable VXLAN protocol on ports
testpmd> rx_vxlan_port add 4789 0
# set fwd engine and start
testpmd> set fwd csum
testpmd> port start all
testpmd> set promisc all off
testpmd> start

Test vxlan() in scapy:

sendp([Ether(dst="%s",src="52:00:00:00:00:00")/IP(src="192.168.1.1",dst="192.168.1.2")/UDP(sport=1021,dport=4789)/VXLAN(vni=1234)/Ether(dst=%s,src="52:00:00:00:00:00")/IP(src="192.168.1.1",dst="192.168.1.2")/UDP(sport=1021,dport=1021)/Raw(load="\x50"*%s)], iface="%s"

Test nvgre() in scapy:

sendp([Ether(dst="%s",src="52:00:00:00:00:00")/IP(src="192.168.1.1",dst="192.168.1.2",proto=47)/GRE(key_present=1,proto=0x6558,key=0x00001000)/Ether(dst="%s",src="52:00:00:00:00:00")/IP(src="192.168.1.1",dst="192.168.1.2")/TCP(sport=1021,dport=1021)/("X"*%s)], iface="%s")

158.5. Test case: TSO performance¶

Set the packet stream to be sent out from packet generator before testing as below.

Frame Size	1S/1C/1T	1S/1C/1T	1S/2C/1T	1S/2C/2T	1S/2C/2T
64
65
128
256
512
1024
1280
1518

Then run the test application as below:

./x86_64-native-linuxapp-gcc/app/dpdk-testpmd -c 0xffffffff -n 2 -- -i --rxd=512 --txd=512
--burst=32 --rxfreet=64 --mbcache=128 --portmask=0x3 --txpt=36 --txht=0 --txwt=0
--txfreet=32 --txrst=32 --enable-rx-cksum

The -n command is used to select the number of memory channels. It should match the number of memory channels on that setup.