提交 b5638d46 编写于 作者: P Petr Machata 提交者: David S. Miller

selftests: mlxsw: Add a test for UC behavior under MC flood

A so-called "MC-aware" mode has recently been enabled in mlxsw. In
MC-aware mode, BUM traffic is handled in a special way so that when a
switch is flooded with BUM, UC performance isn't unduly impacted.
Without enablement of this mode, a stream of BUM traffic can cause
sustained UC throughput drop in excess of 99 %.

Add a test for this behavior. Compare how much UC throughput degrades as
a stream of broadcast frames floods the switch. A minimal degradation is
tolerated to cover for glitches in traffic injection performance.
Signed-off-by: NPetr Machata <petrm@mellanox.com>
Reviewed-by: NJiri Pirko <jiri@mellanox.com>
Signed-off-by: NIdo Schimmel <idosch@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>
上级 a381ed12
#!/bin/bash
# SPDX-License-Identifier: GPL-2.0
#
# A test for switch behavior under MC overload. An issue in Spectrum chips
# causes throughput of UC traffic to drop severely when a switch is under heavy
# MC load. This issue can be overcome by putting the switch to MC-aware mode.
# This test verifies that UC performance stays intact even as the switch is
# under MC flood, and therefore that the MC-aware mode is enabled and correctly
# configured.
#
# Because mlxsw throttles CPU port, the traffic can't actually reach userspace
# at full speed. That makes it impossible to use iperf3 to simply measure the
# throughput, because many packets (that reach $h3) don't get to the kernel at
# all even in UDP mode (the situation is even worse in TCP mode, where one can't
# hope to see more than a couple Mbps).
#
# So instead we send traffic with mausezahn and use RX ethtool counters at $h3.
# Multicast traffic is untagged, unicast traffic is tagged with PCP 1. Therefore
# each gets a different priority and we can use per-prio ethtool counters to
# measure the throughput. In order to avoid prioritizing unicast traffic, prio
# qdisc is installed on $swp3 and maps all priorities to the same band #7 (and
# thus TC 0).
#
# Mausezahn can't actually saturate the links unless it's using large frames.
# Thus we set MTU to 10K on all involved interfaces. Then both unicast and
# multicast traffic uses 8K frames.
#
# +-----------------------+ +----------------------------------+
# | H1 | | H2 |
# | | | unicast --> + $h2.111 |
# | | | traffic | 192.0.2.129/28 |
# | multicast | | | e-qos-map 0:1 |
# | traffic | | | |
# | $h1 + <----- | | + $h2 |
# +-----|-----------------+ +--------------|-------------------+
# | |
# +-----|-------------------------------------------------|-------------------+
# | + $swp1 + $swp2 |
# | | >1Gbps | >1Gbps |
# | +---|----------------+ +----------|----------------+ |
# | | + $swp1.1 | | + $swp2.111 | |
# | | BR1 | SW | BR111 | |
# | | + $swp3.1 | | + $swp3.111 | |
# | +---|----------------+ +----------|----------------+ |
# | \_________________________________________________/ |
# | | |
# | + $swp3 |
# | | 1Gbps bottleneck |
# | | prio qdisc: {0..7} -> 7 |
# +------------------------------------|--------------------------------------+
# |
# +--|-----------------+
# | + $h3 H3 |
# | | |
# | + $h3.111 |
# | 192.0.2.130/28 |
# +--------------------+
ALL_TESTS="
ping_ipv4
test_mc_aware
"
lib_dir=$(dirname $0)/../../../net/forwarding
NUM_NETIFS=6
source $lib_dir/lib.sh
h1_create()
{
simple_if_init $h1
mtu_set $h1 10000
}
h1_destroy()
{
mtu_restore $h1
simple_if_fini $h1
}
h2_create()
{
simple_if_init $h2
mtu_set $h2 10000
vlan_create $h2 111 v$h2 192.0.2.129/28
ip link set dev $h2.111 type vlan egress-qos-map 0:1
}
h2_destroy()
{
vlan_destroy $h2 111
mtu_restore $h2
simple_if_fini $h2
}
h3_create()
{
simple_if_init $h3
mtu_set $h3 10000
vlan_create $h3 111 v$h3 192.0.2.130/28
}
h3_destroy()
{
vlan_destroy $h3 111
mtu_restore $h3
simple_if_fini $h3
}
switch_create()
{
ip link set dev $swp1 up
mtu_set $swp1 10000
ip link set dev $swp2 up
mtu_set $swp2 10000
ip link set dev $swp3 up
mtu_set $swp3 10000
vlan_create $swp2 111
vlan_create $swp3 111
ethtool -s $swp3 speed 1000 autoneg off
tc qdisc replace dev $swp3 root handle 3: \
prio bands 8 priomap 7 7 7 7 7 7 7 7
ip link add name br1 type bridge vlan_filtering 0
ip link set dev br1 up
ip link set dev $swp1 master br1
ip link set dev $swp3 master br1
ip link add name br111 type bridge vlan_filtering 0
ip link set dev br111 up
ip link set dev $swp2.111 master br111
ip link set dev $swp3.111 master br111
}
switch_destroy()
{
ip link del dev br111
ip link del dev br1
tc qdisc del dev $swp3 root handle 3:
ethtool -s $swp3 autoneg on
vlan_destroy $swp3 111
vlan_destroy $swp2 111
mtu_restore $swp3
ip link set dev $swp3 down
mtu_restore $swp2
ip link set dev $swp2 down
mtu_restore $swp1
ip link set dev $swp1 down
}
setup_prepare()
{
h1=${NETIFS[p1]}
swp1=${NETIFS[p2]}
swp2=${NETIFS[p3]}
h2=${NETIFS[p4]}
swp3=${NETIFS[p5]}
h3=${NETIFS[p6]}
h3mac=$(mac_get $h3)
vrf_prepare
h1_create
h2_create
h3_create
switch_create
}
cleanup()
{
pre_cleanup
switch_destroy
h3_destroy
h2_destroy
h1_destroy
vrf_cleanup
}
ping_ipv4()
{
ping_test $h2 192.0.2.130
}
humanize()
{
local speed=$1; shift
for unit in bps Kbps Mbps Gbps; do
if (($(echo "$speed < 1024" | bc))); then
break
fi
speed=$(echo "scale=1; $speed / 1024" | bc)
done
echo "$speed${unit}"
}
rate()
{
local t0=$1; shift
local t1=$1; shift
local interval=$1; shift
echo $((8 * (t1 - t0) / interval))
}
check_rate()
{
local rate=$1; shift
local min=$1; shift
local what=$1; shift
if ((rate > min)); then
return 0
fi
echo "$what $(humanize $ir) < $(humanize $min_ingress)" > /dev/stderr
return 1
}
measure_uc_rate()
{
local what=$1; shift
local interval=10
local i
local ret=0
# Dips in performance might cause momentary ingress rate to drop below
# 1Gbps. That wouldn't saturate egress and MC would thus get through,
# seemingly winning bandwidth on account of UC. Demand at least 2Gbps
# average ingress rate to somewhat mitigate this.
local min_ingress=2147483648
mausezahn $h2.111 -p 8000 -A 192.0.2.129 -B 192.0.2.130 -c 0 \
-a own -b $h3mac -t udp -q &
sleep 1
for i in {5..0}; do
local t0=$(ethtool_stats_get $h3 rx_octets_prio_1)
local u0=$(ethtool_stats_get $swp2 rx_octets_prio_1)
sleep $interval
local t1=$(ethtool_stats_get $h3 rx_octets_prio_1)
local u1=$(ethtool_stats_get $swp2 rx_octets_prio_1)
local ir=$(rate $u0 $u1 $interval)
local er=$(rate $t0 $t1 $interval)
if check_rate $ir $min_ingress "$what ingress rate"; then
break
fi
# Fail the test if we can't get the throughput.
if ((i == 0)); then
ret=1
fi
done
# Suppress noise from killing mausezahn.
{ kill %% && wait; } 2>/dev/null
echo $ir $er
exit $ret
}
test_mc_aware()
{
RET=0
local -a uc_rate
uc_rate=($(measure_uc_rate "UC-only"))
check_err $? "Could not get high enough UC-only ingress rate"
local ucth1=${uc_rate[1]}
mausezahn $h1 -p 8000 -c 0 -a own -b bc -t udp -q &
local d0=$(date +%s)
local t0=$(ethtool_stats_get $h3 rx_octets_prio_0)
local u0=$(ethtool_stats_get $swp1 rx_octets_prio_0)
local -a uc_rate_2
uc_rate_2=($(measure_uc_rate "UC+MC"))
check_err $? "Could not get high enough UC+MC ingress rate"
local ucth2=${uc_rate_2[1]}
local d1=$(date +%s)
local t1=$(ethtool_stats_get $h3 rx_octets_prio_0)
local u1=$(ethtool_stats_get $swp1 rx_octets_prio_0)
local deg=$(bc <<< "
scale=2
ret = 100 * ($ucth1 - $ucth2) / $ucth1
if (ret > 0) { ret } else { 0 }
")
check_err $(bc <<< "$deg > 10")
local interval=$((d1 - d0))
local mc_ir=$(rate $u0 $u1 $interval)
local mc_er=$(rate $t0 $t1 $interval)
# Suppress noise from killing mausezahn.
{ kill %% && wait; } 2>/dev/null
log_test "UC performace under MC overload"
echo "UC-only throughput $(humanize $ucth1)"
echo "UC+MC throughput $(humanize $ucth2)"
echo "Degradation $deg %"
echo
echo "Full report:"
echo " UC only:"
echo " ingress UC throughput $(humanize ${uc_rate[0]})"
echo " egress UC throughput $(humanize ${uc_rate[1]})"
echo " UC+MC:"
echo " ingress UC throughput $(humanize ${uc_rate_2[0]})"
echo " egress UC throughput $(humanize ${uc_rate_2[1]})"
echo " ingress MC throughput $(humanize $mc_ir)"
echo " egress MC throughput $(humanize $mc_er)"
}
trap cleanup EXIT
setup_prepare
setup_wait
tests_run
exit $EXIT_STATUS
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册