test-mptcp-ns.sh

#!/bin/bash -e

# MPTCP network namespace test
# Original script by Matthieu Baerts (matttbe) — issue #24
# Creates a virtual network with 2 paths:
#   Path 0: 20 Mbps, 10ms delay each way
#   Path 1: 10 Mbps, 20ms delay each way
# Transit links use netem shaping with fq_codel underneath to better match
# the fair-queued defaults used by most modern Linux systems.
# Expected aggregate with MPTCP: ~30 Mbps
#
# Usage:
#   sudo ./test-mptcp-ns.sh          # Interactive: runs all tests with human output
#   sudo ./test-mptcp-ns.sh --ci     # CI mode: shorter tests, JSON output, assertions

# set -x

CI=false
if [[ "$1" == "--ci" ]]; then
	CI=true
	shift
fi

XFR="./target/release/xfr"
if [[ ! -x "$XFR" ]]; then
	echo "ERROR: $XFR not found. Run 'cargo build --release' first."
	exit 1
fi

# Check MPTCP support
if ! sysctl -n net.mptcp.enabled >/dev/null 2>&1; then
	echo "SKIP: MPTCP not available in this kernel"
	exit 0
fi
if [[ "$(sysctl -n net.mptcp.enabled)" != "1" ]]; then
	echo "SKIP: MPTCP is disabled (sysctl net.mptcp.enabled=0)"
	exit 0
fi

# Check netem support
if ! modprobe sch_netem 2>/dev/null; then
	echo "SKIP: sch_netem kernel module not available"
	exit 0
fi

export NS=xfr
HOSTS=(cli net0 net1 srv)
NSS=()
FAILED=0

cleanup()
{
	local ns
	for ns in "${NSS[@]}"; do
		ip netns pids "${ns}" 2>/dev/null | xargs -r kill 2>/dev/null
		ip netns del "${ns}" >/dev/null 2>&1
	done
}

trap cleanup EXIT

xfr_srv()
{
	ip netns exec "${NS}_srv" "$XFR" serve &
	sleep .5
}

xfr_cli()
{
	ip netns exec "${NS}_cli" "$XFR" 10.0.0.1 --mptcp --no-tui "$@"
}

xfr_cli_tcp()
{
	ip netns exec "${NS}_cli" "$XFR" 10.0.0.1 --no-tui "$@"
}

# Run a test and extract throughput_mbps from JSON output
# Usage: run_json_test <description> <args...>
# Sets $LAST_MBPS to the throughput value
run_json_test()
{
	local desc="$1"
	shift
	local output
	# JSON goes to stdout, INFO logs go to stderr — capture only stdout
	output=$("$@" --json 2>/dev/null)
	local result
	result=$(echo "$output" | grep -o '"throughput_mbps": *[0-9.]*' | head -1 | sed 's/.*: *//')
	if [[ -z "$result" ]]; then
		echo "FAIL: $desc — no throughput in output"
		echo "  raw output: $output" | head -5
		FAILED=1
		LAST_MBPS=0
		LAST_RTX=0
		LAST_RTT_US=0
		LAST_CWND=0
		return
	fi
	LAST_MBPS="$result"
	# Extract final tcp_info fields from pretty-printed JSON.
	local tcp_info_block
	tcp_info_block=$(echo "$output" | sed -n '/"tcp_info"/,/}/p')
	LAST_RTX=$(echo "$tcp_info_block" | grep '"retransmits"' | head -1 | grep -o '[0-9]*')
	LAST_RTT_US=$(echo "$tcp_info_block" | grep '"rtt_us"' | head -1 | grep -o '[0-9]*')
	LAST_CWND=$(echo "$tcp_info_block" | grep '"cwnd"' | head -1 | grep -o '[0-9]*')
	LAST_RTX="${LAST_RTX:-0}"
	LAST_RTT_US="${LAST_RTT_US:-0}"
	LAST_CWND="${LAST_CWND:-0}"
	if [[ -z "$tcp_info_block" ]]; then
		echo "  $desc: ${LAST_MBPS} Mbps (no tcp_info in output)"
	else
		echo "  $desc: ${LAST_MBPS} Mbps, rtx: ${LAST_RTX}, rtt_us: ${LAST_RTT_US}, cwnd: ${LAST_CWND}"
	fi
}

setup()
{
	local suffix
	for suffix in "${HOSTS[@]}"; do
		local ns="${NS}_${suffix}"
		ip netns add "${ns}"
		NSS+=("${ns}")
		ip -n "${ns}" link set lo up
		ip netns exec "${ns}" sysctl -wq net.ipv4.ip_forward=1
	done

	ip link add "cli" netns "${NS}_net0" type veth peer name "net0" netns "${NS}_cli"
	ip link add "cli" netns "${NS}_net1" type veth peer name "net1" netns "${NS}_cli"

	ip link add "srv" netns "${NS}_net0" type veth peer name "net0" netns "${NS}_srv"
	ip link add "srv" netns "${NS}_net1" type veth peer name "net1" netns "${NS}_srv"

	ip -n "${NS}_cli" link set "net0" up
	ip -n "${NS}_cli" addr add dev "net0" "10.0.10.2/24"
	ip -n "${NS}_cli" route add dev "net0" default via "10.0.10.1" metric "100"
	ip -n "${NS}_cli" mptcp endpoint add "10.0.10.2" dev "net0" subflow
	ip -n "${NS}_cli" mptcp limits set add_addr_accepted 2

	ip -n "${NS}_cli" link set "net1" up
	ip -n "${NS}_cli" addr add dev "net1" "10.0.11.2/24"
	ip -n "${NS}_cli" route add dev "net1" "10.0.1.0/24" via "10.0.11.1"
	ip -n "${NS}_cli" mptcp endpoint add "10.0.11.2" dev "net1" subflow

	ip -n "${NS}_srv" link set "net0" up
	ip -n "${NS}_srv" addr add dev "net0" "10.0.0.1/24"
	ip -n "${NS}_srv" route add dev "net0" default via "10.0.0.2" metric "100"
	ip -n "${NS}_srv" mptcp endpoint add "10.0.0.1" dev "net0" signal
	ip netns exec "${NS}_srv" sysctl -wq net.mptcp.allow_join_initial_addr_port=0

	ip -n "${NS}_srv" link set "net1" up
	ip -n "${NS}_srv" addr add dev "net1" "10.0.1.1/24"
	ip -n "${NS}_srv" route add dev "net1" "10.0.11.0/24" via "10.0.1.2"
	ip -n "${NS}_srv" mptcp endpoint add "10.0.1.1" dev "net1" signal

	ip -n "${NS}_net0" link set "cli" up
	ip -n "${NS}_net0" addr add dev "cli" "10.0.10.1/24"
	tc -n "${NS}_net0" qdisc add dev "cli" root handle 1: netem rate 20mbit delay 10ms limit 100
	tc -n "${NS}_net0" qdisc add dev "cli" parent 1:1 fq_codel
	ip -n "${NS}_net0" link set "srv" up
	ip -n "${NS}_net0" addr add dev "srv" "10.0.0.2/24"
	tc -n "${NS}_net0" qdisc add dev "srv" root handle 1: netem rate 20mbit delay 10ms limit 100
	tc -n "${NS}_net0" qdisc add dev "srv" parent 1:1 fq_codel

	ip -n "${NS}_net1" link set "cli" up
	ip -n "${NS}_net1" addr add dev "cli" "10.0.11.1/24"
	tc -n "${NS}_net1" qdisc add dev "cli" root handle 1: netem rate 10mbit delay 20ms limit 100
	tc -n "${NS}_net1" qdisc add dev "cli" parent 1:1 fq_codel
	ip -n "${NS}_net1" link set "srv" up
	ip -n "${NS}_net1" addr add dev "srv" "10.0.1.2/24"
	tc -n "${NS}_net1" qdisc add dev "srv" root handle 1: netem rate 10mbit delay 20ms limit 100
	tc -n "${NS}_net1" qdisc add dev "srv" parent 1:1 fq_codel
}

# Assert that $1 > $2 (floating point comparison)
assert_gt()
{
	local actual="$1" threshold="$2" label="$3"
	if ! echo "$actual > $threshold" | bc -l | grep -q '^1'; then
		echo "FAIL: $label — expected > ${threshold} Mbps, got ${actual} Mbps"
		FAILED=1
	else
		echo "  PASS: $label (${actual} > ${threshold} Mbps)"
	fi
}

# Assert sender-side TCP_INFO was populated in the final JSON result.
# Retransmits may legitimately be zero on a clean run, so use RTT/cwnd instead.
assert_sender_tcp_info()
{
	local label="$1"
	if [[ "${LAST_RTT_US:-0}" -le 0 || "${LAST_CWND:-0}" -le 0 ]]; then
		echo "FAIL: $label — expected sender-side TCP_INFO (rtt_us > 0, cwnd > 0), got rtt_us=${LAST_RTT_US:-0}, cwnd=${LAST_CWND:-0}"
		FAILED=1
	else
		echo "  PASS: $label (rtt_us=${LAST_RTT_US}, cwnd=${LAST_CWND})"
	fi
}

setup "${@}"
xfr_srv

if [[ "$CI" == "true" ]]; then
	# CI mode: shorter tests, JSON output, throughput assertions
	DURATION="5sec"

	echo "=== MPTCP CI Tests ==="
	echo ""

	echo "--- TCP baseline (single path, expect ~20 Mbps) ---"
	run_json_test "TCP upload" xfr_cli_tcp -t "$DURATION"
	TCP_MBPS="$LAST_MBPS"
	assert_gt "$TCP_MBPS" 10 "TCP throughput > 10 Mbps"
	assert_sender_tcp_info "TCP upload sender-side TCP_INFO present (#26)"

	echo ""
	echo "--- MPTCP upload (both paths, expect ~30 Mbps) ---"
	# Retry once on flake: MPTCP meta socket occasionally reports fresh-
	# connection TCP_INFO (rtt_us=0, cwnd=IW10) at end of test when the path
	# manager switches active subflow. The tests themselves are correct;
	# this guards against a kernel-side timing quirk.
	MPTCP_MAX_ATTEMPTS=2
	for attempt in $(seq 1 $MPTCP_MAX_ATTEMPTS); do
		PREV_FAILED=$FAILED
		FAILED=0
		run_json_test "MPTCP upload" xfr_cli -t "$DURATION"
		MPTCP_MBPS="$LAST_MBPS"
		assert_gt "$MPTCP_MBPS" "$TCP_MBPS" "MPTCP throughput > TCP throughput"
		assert_gt "$MPTCP_MBPS" 20 "MPTCP throughput > 20 Mbps (proves multi-path)"
		assert_sender_tcp_info "MPTCP upload sender-side TCP_INFO present (#26)"
		if [[ "$FAILED" -eq 0 ]]; then
			FAILED=$PREV_FAILED
			break
		fi
		if [[ "$attempt" -lt "$MPTCP_MAX_ATTEMPTS" ]]; then
			echo "  MPTCP upload attempt $attempt failed, retrying..."
			FAILED=$PREV_FAILED
		fi
		# On final failure, leave FAILED=1 to fail the suite.
	done

	echo ""
	echo "--- MPTCP reverse (download, expect ~20 Mbps) ---"
	run_json_test "MPTCP download" xfr_cli -t "$DURATION" -R
	# Download threshold lowered from 20 to 15 Mbps: sender-side byte counter
	# is now clamped to tcpi_bytes_acked (accurate) instead of write()-time
	# counts (~10% overcount due to bufferbloat). The old threshold was
	# calibrated against the overcount behavior. Still well above single-path
	# TCP (10 Mbps), proving multi-path is working.
	assert_gt "$LAST_MBPS" 15 "MPTCP download > 15 Mbps"

	echo ""
	echo "--- MPTCP multi-stream (expect ~30 Mbps) ---"
	run_json_test "MPTCP 4-stream" xfr_cli -t "$DURATION" -P 4
	assert_gt "$LAST_MBPS" 20 "MPTCP multi-stream > 20 Mbps"

	echo ""
	echo "--- MPTCP 10-stream stress (JoinHandle panic regression, #24) ---"
	run_json_test "MPTCP 10-stream" xfr_cli -t "$DURATION" -P 10
	assert_gt "$LAST_MBPS" 20 "MPTCP 10-stream > 20 Mbps"

	echo ""
	if [[ "$FAILED" -eq 0 ]]; then
		echo "=== All MPTCP tests passed ==="
	else
		echo "=== MPTCP tests FAILED ==="
		exit 1
	fi
else
	# Interactive mode: longer tests, human-readable output
	echo ""
	echo "=== Test 1: TCP (single path, expect ~20 Mbps) ==="
	xfr_cli_tcp -t 10sec

	echo ""
	echo "=== Test 2: MPTCP (both paths, expect ~30 Mbps) ==="
	xfr_cli -t 10sec

	echo ""
	echo "=== Test 3: MPTCP reverse (download, expect ~30 Mbps) ==="
	xfr_cli -t 10sec -R

	echo ""
	echo "=== Test 4: MPTCP multi-stream (expect ~30 Mbps) ==="
	xfr_cli -t 10sec -P 4
fi