nettest: a room played over a bad network, measured

scripts/nettest-ns.sh

@@ -0,0 +1,103 @@
+#!/bin/sh
+#
+# Copyright (C) 2004-2026 The thinksynth authors
+#
+# This program is free software; you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by the
+# Free Software Foundation; either version 2 of the License, or (at your
+# option) any later version.
+
+# nettest-ns.sh -- a private network of N hosts for wasm/web/nettest.mjs,
+# with no privilege: a user and network namespace of its own (the hub),
+# and in it one more network namespace per host.
+#
+#   scripts/nettest-ns.sh N command [arg...]
+#
+#        hub: br0 10.77.0.1/24 -- the relay and the harness
+#         |
+#     +---+--------+--- ...
+#    nt1          nt2            veth, hub side: netem here is the host's
+#     |            |               downlink
+#    eth0         eth0           veth, host side: netem here is its uplink
+#  10.77.0.11   10.77.0.12       host i, a namespace held open by a sleep
+#
+# So each host has a link of its own, and a packet between two hosts
+# crosses both: the sender's uplink and the receiver's downlink, as two
+# players at home do. The relay sits on the bridge with no link of its own.
+#
+# The command runs in the hub with, in its environment:
+#
+#   NETTEST_HUB     the hub's address, 10.77.0.1
+#   NETTEST_PIDS    a process in each host's namespace, in host order:
+#                   `nsenter -t PID -n' runs something there, which is how
+#                   a browser is started on a host and how its uplink's
+#                   qdisc is changed
+#
+# The namespaces have no route out, which is the point: WebRTC finds
+# nothing but host candidates, and the mesh has to open on those. Each
+# host has a default route to the hub all the same, because Chromium
+# gathers candidates only on an interface that has one.
+#
+# Everything goes when the command exits, however it ends: this script is
+# the first process of a PID namespace of its own, and the kernel kills
+# everything else in it when it goes, browsers and all, and itself if
+# unshare is killed. A namespace with no process in it is gone.
+
+set -eu
+
+if [ $# -lt 2 ]; then
+    echo "usage: $0 N command [arg...]" >&2
+    exit 2
+fi
+
+PATH="$PATH:/usr/sbin:/sbin"
+
+if [ -z "${NETTEST_IN_HUB:-}" ]; then
+    NETTEST_IN_HUB=1 exec unshare -rnmpf --mount-proc --kill-child "$0" "$@"
+fi
+
+n=$1
+shift
+
+holders=
+
+ip link set lo up
+ip link add br0 type bridge
+ip addr add 10.77.0.1/24 dev br0
+ip link set br0 up
+
+hub=$(readlink /proc/self/ns/net)
+i=1
+while [ "$i" -le "$n" ]; do
+    unshare -n sleep infinity &
+    pid=$!
+    holders="$holders $pid"
+
+    # The holder has its namespace once unshare has made it, which is a
+    # moment after the fork.
+    tries=0
+    while [ "$(readlink "/proc/$pid/ns/net")" = "$hub" ]; do
+        tries=$((tries + 1))
+        if [ "$tries" -gt 200 ]; then
+            echo "$0: host $i got no namespace" >&2
+            exit 1
+        fi
+        sleep 0.01
+    done
+
+    ip link add "nt$i" type veth peer name eth0 netns "$pid"
+    ip link set "nt$i" master br0 up
+    nsenter -t "$pid" -n sh -c "
+        ip link set lo up
+        ip addr add 10.77.0.$((10 + i))/24 dev eth0
+        ip link set eth0 up
+        ip route add default via 10.77.0.1"
+    i=$((i + 1))
+done
+
+export NETTEST_HUB=10.77.0.1
+export NETTEST_PIDS="${holders# }"
+
+# Not exec: the first process of a PID namespace ignores a signal it has
+# no handler for, and the command is to die of a ^C.
+"$@"