trickleshaper

#! /usr/bin/env bash
#
# This shaping script was developed to support host-fair bandwidth
# allocation in networks with very low aggregate backhaul bandwidth
# (<500k) and on hosts where fq_cake is not yet available. It is
# loosely adapted from Wondershaper's
# (https://github.com/magnific0/wondershaper) simple CLI and with much
# technical inspiration from the Bufferbloat Project's
# (https://www.bufferbloat.net/projects/) excellent work. It has not
# been optimized for high throughput networks. If you find yourself
# operating at throughputs consistently higher than 1Mbps, you should
# strongly consdier fq_codel or fq_cake, which are much more general,
# flexible, and optimized.
#
# This script as written assumes a local area /24 network, and creates
# 256 host specific queues keyed on the last octet of the IP
# address. If you're reading this and running a different size
# network, you should probably modify to meet your network specific
# conditions.
#
# This code has no warranty to the greatest extent of the law, implied
# or otherwise. Use at your own risk.
#
# Copyright 2019 Matthew Johnson <matt9j@cs.washington.edu>
# Licensed under the GPLv3 https://www.gnu.org/licenses/gpl-3.0.en.html

VERSION=1.0.0

usage()
{
cat << EOF
USAGE: $0 [-hcs] [-a <adapter>] [-d <rate>] [-u <rate>]

Limit the bandwidth of an adapter

OPTIONS:
   -h           Show this message
   -a <adapter> Set the adapter
   -d <rate>    Set maximum download rate (in Kbps) and/or
   -u <rate>    Set maximum upload rate (in Kbps)
   -p           Use presets in /etc/conf.d/wondershaper.conf
   -c           Clear the limits from adapter
   -s           Show the current status of adapter
   -v           Show the current version

MODES:
   trickleshaper -a <adapter> -d <rate> -u <rate>
   trickleshaper -c -a <adapter>
   trickleshaper -s -a <adapter>

EXAMPLES:
   trickleshaper -a eth0 -d 1024 -u 512
   trickleshaper -a eth0 -u 512
   trickleshaper -c -a eth0

EOF
}

IFACE=
IFB=ifb0

while getopts hvd:u:a:csz: o
do case "$o" in
       h)      usage
               exit 1;;
       v)      echo "Version $VERSION"
               exit 0;;
       d)      DSPEED=$OPTARG;;
       u)      USPEED=$OPTARG;;
       a)      IFACE=$OPTARG;;
       c)      MODE="clear";;
       s)      MODE="status";;
       z)      ADMIN_IP=$OPTARG;;
       [?])    usage
               exit 1;;
esac
done

clearUplink()
{
    echo "Clearing existing uplink"
    tc qdisc del dev $IFACE root
}

clearDownlink()
{
    echo "Clearing existing downlink"
    tc qdisc del dev $IFB ingress
    tc qdisc del dev $IFB root
    tc qdisc del dev $IFACE ingress
}


if [[ ! -z $MODE ]] && [[ -z $IFACE ]]
then
    echo "Please supply the adapter name for the mode."
    echo ""
    usage
    exit 1
fi

if [ "$MODE" = "status" ]
then
    echo "=================================="
    echo "=============Interface============"
    echo "=================================="
    tc -g -s -p qdisc show dev $IFACE
    tc -g -s -p class show dev $IFACE
    tc -g -s -p filter show dev $IFACE

    echo "=================================="
    echo "==========Ingress Buffer=========="
    echo "=================================="
    tc -g -s -p qdisc show dev $IFB
    tc -g -s -p class show dev $IFB
    tc -g -s -p filter show dev $IFB
    exit
fi

if [ "$MODE" = "clear" ]
then
    clearUplink
    clearDownlink
    exit
fi

if ( [[ -z $DSPEED ]] && [[ -z $USPEED ]] ) || [[ -z $IFACE ]]
then
    usage
    exit 1
fi

echo "---Start Egress---"
clearUplink

echo "    ---Add root qdiscs"
tc qdisc add dev $IFACE root handle 1: tbf rate ${USPEED}kbit burst 30kbit latency 10ms
# The latency parameter does not matter since we overwrite the TBF's
# internal qdisc with QFQ, but it is required. See `$man tc-tbf` for
# more information.

tc qdisc add dev $IFACE parent 1:1 handle 2: qfq

echo "    ---Add app classes"
# 2:1 for EF and voice traffic
tc class add dev $IFACE parent 2: classid 2:1 qfq weight 20
# 2:2 for small packets (dangerous since not flow aware, could cause reordering)
# TODO(matt9j) Use CAKE to target small *flows*
tc class add dev $IFACE parent 2: classid 2:2 qfq weight 20
# 2:3 normal best effort traffic
tc class add dev $IFACE parent 2: classid 2:3 qfq weight 10
# Bulk
tc class add dev $IFACE parent 2: classid 2:4 qfq weight 5

for appIndex in `seq 1 4`
do
    echo "    ===============Class "$appIndex"==========="
    echo "        ---Use QFQ in each class"
    qdiscHandle=2$appIndex
    tc qdisc add dev $IFACE parent 2:$appIndex handle $qdiscHandle: qfq

    echo "        ---Add shapers per ip"
    for i in `seq 1 256`
    do
        classid=$qdiscHandle:$(printf %x $i)
        tc class add dev $IFACE parent $qdiscHandle: classid $classid qfq weight 10
        tc qdisc add dev $IFACE parent $classid sfq\
           perturb 30 headdrop probability 0.5 redflowlimit 20000 ecn harddrop
    done

    echo "        ---Add ip source filter"
    tc filter add dev $IFACE parent $qdiscHandle: protocol all prio 1\
       handle 0xff$qdiscHandle\
       flow map key nfct-src and 0xff divisor 256 baseclass $qdiscHandle:1
done

echo "    ---Add app class filters"
# TODO(matt9j) Identify heavy flows and send to bottom band

# Prioritize DSCP EF
tc filter add dev $IFACE parent 2: protocol ip prio 10 u32 \
   match ip dsfield 0xb8 0xfc flowid 2:1

tc filter add dev $IFACE parent 2: protocol ip prio 10 u32 \
   match ip6 priority 0xb8 0xfc flowid 2:1

# Prioritize slightly larger UDP packets to support realtime audio streaming
# This catches WhatsApp and Signal voice at <256 bytes/packet
tc filter add dev $IFACE parent 2: protocol ip prio 11 u32 \
   match u16 0x0000 0xff00 at 2 \
   match ip protocol 0x11 0xff flowid 2:2

tc filter add dev $IFACE parent 2: protocol ip prio 11 u32 \
   match u16 0x0000 0xff00 at 4 \
   match ip6 protocol 0x11 0xff flowid 2:2

# Prioritize all small packets (<128 bytes payload)
tc filter add dev $IFACE parent 2: protocol ip prio 12 u32 \
   match u16 0x0000 0xff80 at 2 flowid 2:2

tc filter add dev $IFACE parent 2: protocol ip prio 12 u32 \
   match u16 0x0000 0xff80 at 4 flowid 2:2

# By default send all traffic to the middle band.
# Large prio ("low priority") ensures other filters take precedence
# over this default.
tc filter add dev $IFACE parent 2: prio 1000 matchall flowid 2:3

echo ""
echo "---Start Ingress--"
echo "    ---Add IFB"
# Add the IFB interface
modprobe ifb numifbs=1
ip link set dev $IFB up

echo "    ---Clear existing ingress"
clearDownlink

echo "    ---Redirect to the ifb"
# Redirect ingress (incoming) to egress ifb0
tc qdisc add dev $IFACE handle ffff: ingress
tc filter add dev $IFACE parent ffff: matchall action mirred egress redirect dev $IFB

echo "    ---Add root qdiscs"
tc qdisc add dev $IFB root handle 1: tbf rate ${DSPEED}kbit burst 30kbit latency 10ms
# The latency parameter does not matter since we overwrite the TBF's
# internal qdisc with QFQ, but it is required. See `$man tc-tbf` for
# more information.
tc qdisc add dev $IFB parent 1: handle 2: qfq

echo "    ---Add app classes"
# 2:1 for EF and voice traffic
tc class add dev $IFB parent 2: classid 2:1 qfq weight 20
# 2:2 for small packets (dangerous since not flow aware, could cause reordering)
# TODO(matt9j) Use CAKE to target small *flows*
tc class add dev $IFB parent 2: classid 2:2 qfq weight 20
# 2:3 normal best effort traffic
tc class add dev $IFB parent 2: classid 2:3 qfq weight 10
# Bulk
tc class add dev $IFB parent 2: classid 2:4 qfq weight 5

if [[ ! -z $ADMIN_IP ]]
then
    echo "    ---Add admin shortcut"

    tc class add dev $IFB parent 2: classid 2:5 qfq weight 30
    tc qdisc add dev $IFB parent 2:5 sfq\
       perturb 30 headdrop probability 0.5 redflowlimit 20000 ecn harddrop

    # Add exception filter
    tc filter add dev $IFB parent 2: protocol ip prio 1 handle 0x1337 \
       u32 match ip dst $ADMIN_IP flowid 2:5
fi

for appIndex in `seq 1 4`
do
    echo "    ===============Class "$appIndex"==========="
    echo "        ---Use QFQ in each class"
    qdiscHandle=2$appIndex
    tc qdisc add dev $IFB parent 2:$appIndex handle $qdiscHandle: qfq

    echo "        ---Add shapers per ip"
    for i in `seq 1 256`
    do
        classid=$qdiscHandle:$(printf %x $i)
        tc class add dev $IFB parent $qdiscHandle: classid $classid qfq weight 10
        tc qdisc add dev $IFB parent $classid sfq\
           perturb 30 headdrop probability 0.5 redflowlimit 20000 ecn harddrop
    done

    echo "        ---Add ip dest filter"
    tc filter add dev $IFB parent $qdiscHandle: protocol all prio 1\
       handle 0xff$qdiscHandle\
       flow map key nfct-dst and 0xff divisor 256 baseclass $qdiscHandle:1
done

echo "    ---Add app class filters"
# TODO(matt9j) Identify heavy flows and send to bottom band

# Prioritize DSCP EF
tc filter add dev $IFB parent 2: protocol ip prio 10 u32 \
   match ip dsfield 0xb8 0xfc flowid 2:1

tc filter add dev $IFB parent 2: protocol ip prio 10 u32 \
   match ip6 priority 0xb8 0xfc flowid 2:1

# Prioritize slightly larger UDP packets to support realtime audio streaming
# This catches WhatsApp and Signal voice at <256 bytes/packet
tc filter add dev $IFB parent 2: protocol ip prio 11 u32 \
   match u16 0x0000 0xff00 at 2 \
   match ip protocol 0x11 0xff flowid 2:2

tc filter add dev $IFB parent 2: protocol ip prio 11 u32 \
   match u16 0x0000 0xff00 at 4 \
   match ip6 protocol 0x11 0xff flowid 2:2

# Prioritize Small packets (<128 bytes payload)
tc filter add dev $IFB parent 2: protocol ip prio 12 u32 \
   match u16 0x0000 0xff80 at 2 flowid 2:2

tc filter add dev $IFB parent 2: protocol ip prio 12 u32 \
   match u16 0x0000 0xff80 at 4 flowid 2:2

# By default send all traffic to the middle band.
# Large prio ("low priority") ensures other filters take precedence
# over this default.
tc filter add dev $IFB parent 2: prio 1000 matchall flowid 2:3