test/adq: a functional perf test for tcp with ADQ enabled

We apply ADQ on target, then run perf, so that we can verify
SPDK implementation interacting with ADQ.
Although the functional test cases here are relatively simplified,
and in the CI, we have to add -debug compilation option,
so the perf performance with ADQ may not be ideal. We will give
another performance test reference cases of ADQ later.

Change-Id: I5341a7fcd61334ef78084302a4ae70f8ec9b9e46
Signed-off-by: wanghailiangx <hailiangx.e.wang@intel.com>
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/7476
Community-CI: Broadcom CI <spdk-ci.pdl@broadcom.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Reviewed-by: Karol Latecki <karol.latecki@intel.com>
Reviewed-by: Kamil Godzwon <kamilx.godzwon@intel.com>
Reviewed-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Xiaodong Liu <xiaodong.liu@intel.com>

test/nvmf/nvmf.sh

@@ -58,6 +58,13 @@ run_test "nvmf_nmic" test/nvmf/target/nmic.sh "${TEST_ARGS[@]}"
 run_test "nvmf_fio_target" test/nvmf/target/fio.sh "${TEST_ARGS[@]}"
 run_test "nvmf_bdevio" test/nvmf/target/bdevio.sh "${TEST_ARGS[@]}"
 if [[ $NET_TYPE == phy ]]; then
+	if [ "$SPDK_TEST_NVMF_TRANSPORT" = "tcp" ]; then
+		gather_supported_nvmf_pci_devs
+		TCP_INTERFACE_LIST=("${net_devs[@]}")
+		if ((${#TCP_INTERFACE_LIST[@]} > 0)); then
+			run_test "nvmf_perf_adq" test/nvmf/target/perf_adq.sh "${TEST_ARGS[@]}"
+		fi
+	fi
 	run_test "nvmf_shutdown" test/nvmf/target/shutdown.sh "${TEST_ARGS[@]}"
 	#TODO: disabled due to intermittent failures. Need to triage.
 	# run_test "nvmf_srq_overwhelm" test/nvmf/target/srq_overwhelm.sh $TEST_ARGS
@@ -71,6 +78,7 @@ timing_enter host
 
 run_test "nvmf_identify" test/nvmf/host/identify.sh "${TEST_ARGS[@]}"
 run_test "nvmf_perf" test/nvmf/host/perf.sh "${TEST_ARGS[@]}"
+
 run_test "nvmf_failover" test/nvmf/host/failover.sh "${TEST_ARGS[@]}"
 if [[ $SPDK_TEST_USDT -eq 1 ]]; then
 	run_test "nvmf_multipath" test/nvmf/host/multipath.sh "${TEST_ARGS[@]}"

test/nvmf/target/perf_adq.sh

@@ -0,0 +1,173 @@
+#!/usr/bin/env bash
+
+testdir=$(readlink -f $(dirname $0))
+rootdir=$(readlink -f $testdir/../../..)
+source $rootdir/test/common/autotest_common.sh
+source $rootdir/test/nvmf/common.sh
+
+gather_supported_nvmf_pci_devs
+TCP_INTERFACE_LIST=("${net_devs[@]}")
+if ((${#TCP_INTERFACE_LIST[@]} == 0)); then
+	echo "ERROR: Physical TCP interfaces are not ready"
+	exit 1
+fi
+
+rpc_py="$rootdir/scripts/rpc.py"
+perf="$SPDK_EXAMPLE_DIR/perf"
+enable_adq=0
+
+function pre_conf_for_adq() {
+	# Enable adding flows to hardware
+	"${NVMF_TARGET_NS_CMD[@]}" ethtool --offload $NVMF_TARGET_INTERFACE hw-tc-offload on
+	# ADQ driver turns on this switch by default, we need to turn it off for SPDK testing
+	"${NVMF_TARGET_NS_CMD[@]}" ethtool --set-priv-flags $NVMF_TARGET_INTERFACE channel-pkt-inspect-optimize off
+	# Since sockets are non-blocking, a non-zero value of net.core.busy_read is sufficient
+	sysctl -w net.core.busy_poll=1
+	sysctl -w net.core.busy_read=1
+}
+
+# Configuring traffic classes
+function traffic_classes() {
+	tc=/usr/sbin/tc
+	# Create 2 traffic classes and 2 tc1 queues
+	"${NVMF_TARGET_NS_CMD[@]}" $tc qdisc add dev $NVMF_TARGET_INTERFACE root \
+		mqprio num_tc 2 map 0 1 queues 2@0 2@2 hw 1 mode channel
+	"${NVMF_TARGET_NS_CMD[@]}" $tc qdisc add dev $NVMF_TARGET_INTERFACE ingress
+	# TC filter is configured using target address (traddr) and port number (trsvcid) to steer packets
+	"${NVMF_TARGET_NS_CMD[@]}" $tc filter add dev $NVMF_TARGET_INTERFACE protocol \
+		ip parent ffff: prio 1 flower dst_ip $NVMF_FIRST_TARGET_IP/32 ip_proto tcp dst_port $NVMF_PORT skip_sw hw_tc 1
+	# Setup mechanism for Tx queue selection based on Rx queue(s) map
+	"${NVMF_TARGET_NS_CMD[@]}" $rootdir/scripts/perf/nvmf/set_xps_rxqs $NVMF_TARGET_INTERFACE
+}
+
+function start_nvmf_target() {
+	nvmfappstart -m $1 --wait-for-rpc
+	trap 'process_shm --id $NVMF_APP_SHM_ID; clean_ints_files; nvmftestfini; exit 1' SIGINT SIGTERM EXIT
+	$rpc_py sock_impl_set_options --enable-placement-id $enable_adq --enable-zerocopy-send-server -i posix
+	$rpc_py framework_start_init
+	$rpc_py nvmf_create_transport $NVMF_TRANSPORT_OPTS --io-unit-size 8192 --sock-priority $enable_adq
+	$rpc_py bdev_malloc_create 64 512 -b Malloc1
+	$rpc_py nvmf_create_subsystem nqn.2016-06.io.spdk:cnode1 -a -s SPDK00000000000001
+	$rpc_py nvmf_subsystem_add_ns nqn.2016-06.io.spdk:cnode1 Malloc1
+	$rpc_py nvmf_subsystem_add_listener nqn.2016-06.io.spdk:cnode1 -t $TEST_TRANSPORT -a $NVMF_FIRST_TARGET_IP -s $NVMF_PORT
+}
+
+function reload_driver() {
+	rmmod ice
+	modprobe ice
+	sleep 5
+}
+
+function clean_ints_files() {
+	rm -f temp_ints1.log
+	rm -f temp_ints2.log
+}
+
+function get_nvmf_poll_groups() {
+	"$rpc_py" thread_get_pollers | jq -r '.threads[] | .active_pollers[] |
+	select(.name == "nvmf_poll_group_poll").busy_count'
+}
+
+function num_busy_count() {
+	get_pollers_busy_count0=($(get_nvmf_poll_groups))
+	sleep 2
+	get_pollers_busy_count1=($(get_nvmf_poll_groups))
+	local num=0
+	for i in "${!get_pollers_busy_count0[@]}"; do
+		increment=$((get_pollers_busy_count1[i] - get_pollers_busy_count0[i]))
+		if ((increment > 0)); then
+			((++num))
+		fi
+	done
+	echo $num
+}
+
+function compare_ints() {
+	grep $2 < /proc/interrupts | awk '{
+		for (i = 1; i <= NF; i++) {
+				val[i]=$i;
+				if (i>1)
+					printf "%d\n", $i;
+		}
+	}' > temp_ints1.log
+	sleep $1
+	grep $2 < /proc/interrupts | awk '{
+		for (i = 1; i <= NF; i++) {
+				val[i]=$i;
+				if (i>1)
+					printf "%d\n", $i;
+		}
+	}' > temp_ints2.log
+
+	if diff temp_ints1.log temp_ints2.log > /dev/null; then
+		return 0
+	fi
+	return 1
+}
+
+function check_ints_result() {
+	# We only test check_ints three times here, as long as there is no interruption once,
+	# we consider it is pass. Of course, ideally, one time is enough.
+	for ((i = 1; i <= 3; i++)); do
+		if compare_ints 2 $NVMF_TARGET_INTERFACE; then
+			return 0
+		fi
+	done
+	return 1
+}
+
+# Clear the previous configuration that may have an impact.
+# At present, ADQ configuration is only applicable to the ice driver.
+reload_driver
+
+# Testcase 1 and Testcase 2 show the SPDK interacting with ADQ.
+# The number of continuously increasing nvmf_poll_group_poll's busy_count, we define it as "num_busy_count".
+# When ADQ enabled, num_busy_count will be equal to the number of tc1 queues of traffic classes.
+# When ADQ disabled, num_busy_count will be equal to the smaller value of initiator connections and target cores.
+# Testcase 1: Testing 2 traffic classes and 2 tc1 queues without ADQ
+nvmftestinit
+start_nvmf_target 0xF
+sleep 2
+$perf -q 64 -o 4096 -w randread -t 10 -c 0x70 \
+	-r "trtype:${TEST_TRANSPORT} adrfam:IPv4 traddr:${NVMF_FIRST_TARGET_IP} trsvcid:${NVMF_PORT} \
+subnqn:nqn.2016-06.io.spdk:cnode1" &
+perfpid=$!
+sleep 2
+if [[ $(num_busy_count) -ne 3 ]]; then
+	echo "ERROR: num_busy_count != cores of initiators! Testcase 1 failed."
+	exit 1
+fi
+wait $perfpid
+clean_ints_files
+nvmftestfini
+
+reload_driver
+
+# Testcase 2: Testing 2 traffic classes and 2 tc1 queues with ADQ
+enable_adq=1
+nvmftestinit
+sleep 2
+pre_conf_for_adq
+traffic_classes
+start_nvmf_target 0xF
+sleep 2
+# The number of I/O connections from initiator is the core count * qpairs per ns, so here its 6.
+# ADQ on target side will work if 6 connections are matched to two out of four cores on the target.
+$perf -q 64 -o 4096 -w randread -t 15 -P 2 -c 0x70 \
+	-r "trtype:${TEST_TRANSPORT} adrfam:IPv4 traddr:${NVMF_FIRST_TARGET_IP} trsvcid:${NVMF_PORT} \
+subnqn:nqn.2016-06.io.spdk:cnode1" &
+perfpid=$!
+sleep 3
+if ! check_ints_result; then
+	echo "ERROR: check_ints failed! There is interruption in perf, this is not what we expected."
+	exit 1
+fi
+if [[ $(num_busy_count) -ne 2 ]]; then
+	echo "ERROR: num_busy_count != tc1 queues of traffic classes! Testcase 2 failed."
+	exit 1
+fi
+wait $perfpid
+clean_ints_files
+nvmftestfini
+
+trap - SIGINT SIGTERM EXIT