iopsys-feed/qosmngr/files/linux/lib/qos/qos.sh

#!/bin/sh
. /lib/functions.sh
#set -x

# include common code
. /lib/qos/ip_rule.sh
. /lib/qos/iptables.sh
. /lib/qos/ebtables.sh
. /lib/qos/classify.sh

MAJOR=""

POLICER_COUNT=0
Q_COUNT=0
SP_Q_PRIO=7
SOQ_wgt=0

get_max_burst_size_per_queue() {
	# maximum burst size for tc class
	echo "4289999999"
}

handle_q_weight() {
	local qid="$1" #queue section ID
	config_get ifname "$qid" "ifname"
	# If ifname is empty that is good enough to break
	if [ -z "$ifname" ];then
		return
	fi

	config_get sc_alg "$qid" "scheduling"
	config_get weight "$qid" "weight"

	if [ "$ifname" == "$2" ] && [ "$sc_alg" == "WRR" ]; then
		SOQ_wgt=`expr $SOQ_wgt + $weight`
	fi
}

# Function to handle a queue order and
# update total number of queues
handle_q_order() {
	local qid="$1" #queue section ID

	config_get is_enable "$qid" "enable" 1

	# No need to configure disabled queues
	if [ $is_enable == '0' ]; then
		return
	fi

	config_get ifname "$qid" "ifname"
	# If ifname is empty that is good enough to break
	if [ -z "$ifname" ];then
		return
	fi

	# Create precedence file containing queue order per
	# interface.
	local precedence_file="/tmp/qos/$ifname/q_order"
	local q_no=$(cat /tmp/qos/$ifname/q_idx)

	config_get precedence "$qid" "precedence"
	value=${precedence}_q${q_no}
	echo $value >> $precedence_file

	# Update the number of queues per interface.
	q_no=$((q_no + 1))
	echo $q_no > /tmp/qos/$ifname/q_idx
}

# Sort queue, lower value in uci means higher precedence, so this
# function sorts the precedence in decending order
sort_q_by_precedence() {
	ifname="$1"

	local order_file="/tmp/qos/$ifname/q_order"
	local tmp_order_file="/tmp/qos/$ifname/q_order.tmp"

	sort -n -k1 $order_file > $tmp_order_file
	cp $tmp_order_file $order_file
	rm -f $tmp_order_file
}

sort_by_precedence() {
	for interf in $(jsonfilter -i /etc/board.json -e @.network.lan.ports[*] -e @.network.lan.device -e @.network.wan.device | xargs); do
		sort_q_by_precedence $interf
	done
}

# function to handle a queue section
handle_queue() {
	local qid="$1" #queue section ID
	local port="$2"
	local port_bw="$3"
	local root="$4"
	local port_bs="$5"
	local par_class="$6"
	local priority=7
	local bs=0

	config_get is_enable "$qid" "enable"

	# no need to configure disabled queues
	if [ "$is_enable" == "0" ]; then
		return
	fi

	config_get ifname "$qid" "ifname"
	# if ifname is empty that is good enough to break
	if [ -z "$ifname" ];then
		return
	fi

	# This is to get the qid per interface.
	if [ "$port" != "$ifname" ]; then
		return
	fi

	local precedence_file="/tmp/qos/$ifname/q_order"

	local temp_order=0
	while read -r line; do
		line_qid=${line: -1}
		if [ "$line_qid" == "$Q_COUNT" ]; then
			break
		fi
		temp_order=$((temp_order + 1))
	done < "$precedence_file"

	# precedence_file so the order is calculated accordingly.
	local order=`expr $SP_Q_PRIO - $temp_order`

	config_get sc_alg "$qid" "scheduling"
	config_get wgt "$qid" "weight" 1
	config_get rate "$qid" "rate"
	config_get bs "$qid" "burst_size" 1500
	config_get qsize "$qid" "queue_size" 1024

	[ "$rate" == "0" ] && rate="$port_bw"
	[ "$bs" == "0" ] && bs=1500

	bs=$(( bs * 8  ))
	rate_bs=$(( rate * 25 ))
	if [ $rate_bs -gt $bs ]; then
		bs=$rate_bs
	fi
	# burst size in tc class is given in bytes
	bs=$(( bs * 1000 ))
	local max_bs=$(get_max_burst_size_per_queue )
	if [ $bs -gt $max_bs ]; then
		bs=$max_bs
	fi

	local salg=1

	case "$sc_alg" in
		"SP") salg=1
		;;
		"WRR") salg=2
		;;
		"WDRR") salg=3
		;;
		"WFQ") salg=4
		;;
	esac

	[ -f /etc/board.json ] || return 0
	if [ $salg -ne 2 ]; then
		priority=`expr 7 - $order`
	fi

	local ceil_rate=$port_bw
	if [ "$rate" -lt "$port_bw" ]; then
		ceil_rate=$rate
	fi

	local cls_priority=`expr $order + 1`
	cls_priority=${cls_priority}0

	if [ $salg -eq 2 ]; then
		# maximum quantum value 60000
		# quantum value calculate based on the configured weight value for the queue
		quantum=`expr $wgt \* 60000 / $SOQ_wgt`
		tc class add dev $port parent ${root}:${par_class} classid ${root}:${cls_priority} htb rate ${rate}kbit prio ${priority} ceil ${ceil_rate}kbit quantum ${quantum} burst $bs cburst $port_bs
		tc qdisc add dev $port parent ${root}:${cls_priority} handle ${cls_priority}: sfq perturb 10
	else
		tc class add dev $port parent ${root}:${par_class} classid ${root}:${cls_priority} htb rate ${rate}kbit prio ${priority} ceil ${ceil_rate}kbit burst $bs cburst $port_bs
	fi
	tc filter add dev $port parent ${root}:0 prio $order u32 match mark $order 0xf flowid ${root}:${cls_priority}

	Q_COUNT=$((Q_COUNT + 1))
}

iptables_set_traffic_class() {
	IP_RULE="$IP_RULE -j MARK --set-mark $1"
}

# marking value be decimal for linux target as it uses set-mark whereas other
# targets uses set-xmark, hence this function can't make it common
ip_rule_get_converted_tos() {
	echo $1
}

ebt_match_ipv6_dscp() {
	#when ethertype is not configured by user then both proto rules of ipv4
	#and ipv6 to be installed so update BR6_RULE string as well otherwise
	#update BR_RULE only for installation of ipv6 proto rule only.
	if [ ! -z "$BR6_RULE" ]; then
		BR6_RULE="$BR6_RULE --ip6-tclass $1"
	else
		BR_RULE="$BR_RULE --ip6-tclass $1"
	fi
}

broute_filter_on_dscp() {
	BR_RULE="$BR_RULE --ip-tos $1"
}

broute_ipv4_rule_options()
{
	local cid=$1
	config_get protocol "$cid" "proto"
	config_get dscp_filter "$cid" "dscp_filter"

	set_ip_addr $cid ebt_match_src_ip ebt_match_dst_ip

	if [ ! -z $dscp_filter ]; then
		local tos_val
		local tos_hex

		tos_val=$((dscp_filter<<2))
		tos_hex=$(printf "%x" $tos_val)
		broute_filter_on_dscp "$tos_hex"
	fi

	if [ ! -z $protocol ]; then
		local proto_num=$(protocol_string_to_num "$protocol")
		ebt_match_ip_protocol "$proto_num"

		#port installation for protol tcp/udp/sctp
		if [ $proto_num = "6" ] || [ $proto_num = "17" ] || [ $proto_num = "132" ] ; then
			set_ports "$cid" ebt_match_ip_src_port ebt_match_ip_dst_port
		fi
	fi
}

broute_ipv6_rule_options()
{
	local cid=$1

	config_get protocol "$cid" "proto"
	config_get dscp_filter "$cid" "dscp_filter"

	set_ip_addr $cid ebt_match_ipv6_src_ip ebt_match_ipv6_dst_ip

	if [ ! -z $dscp_filter ]; then
		local tos_val
		local tos_hex

		tos_val=$(($dscp_filter<<2))
		tos_hex=$(printf "%x" $tos_val)
		ebt_match_ipv6_dscp "$tos_hex"
	fi

	if [ ! -z $protocol ]; then
		local proto_num=$(protocol_string_to_num "$protocol")
		ebt_match_ipv6_protocol "$proto_num"

		#port installation for protol tcp/udp/sctp
		if [ $proto_num = "6" ] || [ $proto_num = "17" ] || [ $proto_num = "132" ]; then
			set_ports "$cid" ebt_match_ipv6_src_port ebt_match_ipv6_dst_port
		fi
	fi
}

broute_rule_set_traffic_class() {
	#when ethertype is not configured by user then both proto rules of ipv4
	#and ipv6 to be installed so update BR6_RULE string as well otherwise
	#update BR_RULE only for installation of ipv6 proto rule only.
	BR_RULE="$BR_RULE -j mark --mark-or 0x$1 --mark-target ACCEPT"
	if [ ! -z "$BR6_RULE" ]; then
		BR6_RULE="$BR6_RULE -j mark --mark-or 0x$1 --mark-target ACCEPT"
	fi
}

#function to handle a policer section
handle_policer() {
	local p_sec="$1" # policer section ID
	local dir=1 # default direction, upstream

	config_get is_enable "$p_sec" "enable"

	#no need to configure disabled policer
	if [ $is_enable == '0' ]; then
		return
	fi

	POLICER_COUNT=$((POLICER_COUNT + 1))
}

handle_policer_rules() {
	local c_sec=$1
	local policer_name
	local ifname
	local pname
	local pindex=-1
	local ingress_rate=0
	local in_burst_size=0

	config_get policer_name "$c_sec" "policer"
	if [ -z "$policer_name" ];then
		# no need to apply policer if policer not present in this
		# classification rule
		return
	fi

	config_get ifname "$c_sec" "ifname"
	if [ -z "$ifname" ]; then
		# cannot associate policer as interface is not mentioned
		return
	fi

	local i=0
	local max_policer_inst=$(cat /tmp/qos/max_policer_inst)
	while :
	do
		if [ $i -eq $max_policer_inst ]; then
			break
		fi

		pname="$(uci -q get qos.@policer[$i].name)"
		if [ "$policer_name" == "$pname" ]; then
			pindex=$i
			ingress_rate=$(uci -q get qos.@policer[$i].committed_rate)
			in_burst_rate=$(uci -q get qos.@policer[$i].committed_burst_size)
			break
		fi
		i=$((i + 1))
	done

	if [ $pindex -lt 0 ]; then
		# policer not found, no need to proceed further
		return
	fi

	config_ingress_rate_limit $ifname $ingress_rate $in_burst_size $pindex

}

config_ingress_rate_limit() {
	local ifname="$1"
	local ingress_rate=$2
	local in_burst_size=$3
	local pindex="$4"
	local wanport="$(jsonfilter -i /etc/board.json -e @.network.wan.device)"

	# Unit in uci file is in bps while that accepted by ethswctl is kbits
	if [ $ingress_rate -lt 1000 ]; then
		return
	fi

	ingress_rate=$((ingress_rate / 1000))

	local bs=0
	local avg_bs=$(( ingress_rate * 25 ))
	config_get b_size "$sid" "burst_size"
	if [ "$in_burst_size" > "0" ]; then
		bs=`expr $in_burst_size / 1000`
	fi
	bs=$(( bs * 8 ))
	if [ $avg_bs -gt $bs ]; then
		bs="$avg_bs"
	fi

	tc qdisc add dev $ifname ingress

	tc filter add dev $ifname parent ffff: protocol ip prio $pindex u32 match ip src 0.0.0.0/0 police rate ${ingress_rate}kbit burst $bs drop flowid :$pindex

}

pre_configure_queue() {
	# Delete queues
	for intf in $(jsonfilter -i /etc/board.json -e @.network.lan.ports[*] -e @.network.lan.device -e @.network.wan.device | xargs); do
		rm -rf /tmp/qos/$intf

		mkdir -p /tmp/qos/$intf
		touch /tmp/qos/$intf/q_order
		touch /tmp/qos/$intf/q_idx
		echo 0 > /tmp/qos/$intf/q_idx

		tc qdisc del dev $intf root
		tc qdisc del dev $intf ingress 2> /dev/null
	done
}

get_link_rate() {
	local ifname="$1"
	local phycap="$(ethtool $ifname | grep -A 10 "Supported link modes" | grep 00 | tail -n 1 | awk '{print$NF}')"
	local speed=1000

	# Get the max capability of this port
	case "$phycap" in
		10000*) speed=10000 ;;
		5000*) speed=5000 ;;
		2500*) speed=2500 ;;
		1000*) speed=1000 ;;
		100*) speed=100 ;;
		10*) speed=10 ;;
		*) speed=1000 ;;
	esac

	echo "$speed"
}

configure_queue() {
	qdisc_idx=0
	local bs=1500
	local rate=0
	# Load UCI file
	config_load qos
	config_foreach handle_q_order queue
	sort_by_precedence

	get_intf_shaper_config() {
		local b_size
		sid="$1" #shaper section ID

		config_get is_enable "$sid" "enable"
		# no need to configure disabled queues
		if [ "$is_enable" == "0" ]; then
			return
		fi


		config_get ifname "$sid" "ifname"
		# if ifname is empty that is good enough to break
		if [ -z "$ifname" ] || ! [ "$ifname" == "$2" ];then
			return
		fi

		config_get rate "$sid" "rate"
		# Convert the rate from bps to kbps.
		if [ $rate -lt 1000 ];then
			return
		fi

		rate=$(( rate / 1000 ))

		#burst size is either bitrate*0.00025 or max_packat_size in bit whichever is greater
		local avg_bs=$(( rate * 25 ))
		config_get b_size "$sid" "burst_size"
		if [ "$b_size" > "0" ]; then
			bs="$b_size"
		fi
		bs=$(( bs * 8 ))
		if [ $avg_bs -gt $bs ]; then
			bs="$avg_bs"
		fi
		bs=$(( bs / 1000 ))
	}
	local wanport="$(jsonfilter -i /etc/board.json -e @.network.wan.device)"
	for interf in $(jsonfilter -i /etc/board.json -e @.network.lan.ports[*] -e @.network.lan.device -e @.network.wan.device | xargs); do
		Q_COUNT=0
		rate=0
		# sp queue have max priority value = no. of queue configured on the port
		# hence read and update SP_Q_PRIO here
		local q_no=$(cat /tmp/qos/$interf/q_idx)
		SP_Q_PRIO=`expr $q_no - 1`
		qdisc_idx=`expr $qdisc_idx + 1`
		# link_rate is in mbps and rate is in kbp
		link_rate=$(get_link_rate "$interf")
		# Read the shaper configuration for interface
		config_foreach get_intf_shaper_config shaper $interf
		if [ "$rate" == "0" ]; then
			rate=$(( link_rate * 1000 ))
		fi

		p_qdisc_idx=`expr $qdisc_idx + 11`
		# port shaper needs to add at root tbf qdisc, child class gets token based on rate & burst_size
		tc qdisc add dev $interf root handle ${p_qdisc_idx}: tbf rate ${rate}kbit burst ${bs}k limit 1M
		tc qdisc add dev $interf parent ${p_qdisc_idx}: handle ${qdisc_idx}: htb default 10

		local par_class=1
		tc class add dev $interf parent ${qdisc_idx}: classid ${qdisc_idx}:$par_class htb rate ${rate}kbit prio 8 ceil ${rate}kbit
		tc filter add dev $interf parent ${qdisc_idx}:0 matchall flowid ${qdisc_idx}:$par_class

		# if qdisc_idx is the index corresponds to WAN port, then
		# it would be the MAJOR portion of the destination class ID.
		# under the same qdisc. We are programming for WAN port only.
		if [ "$interf" == "$wanport" ]; then
			MAJOR="$qdisc_idx"
		fi

		#calculate sum of queue weight, to derive the quantum value
		SOQ_wgt=0
		config_foreach handle_q_weight queue $interf

		config_foreach handle_queue queue $interf $rate $qdisc_idx $bs $par_class
	done
}

configure_policer() {

        # Delete policer
        local i=0
        local max_p_inst=0
        if [ -f "/tmp/qos/max_policer_inst" ]; then
                max_p_inst=$(cat /tmp/qos/max_policer_inst)
        fi

        # reset the policer counter
        echo 0 > /tmp/qos/max_policer_inst
        # Load UCI file
        config_load qos
        config_foreach handle_policer policer
        echo $POLICER_COUNT > /tmp/qos/max_policer_inst
}

configure_qos() {
	pre_configure_queue
	configure_queue
	configure_policer
	configure_classify
	if [ -f "/tmp/qos/classify.ebtables" ]; then
		sh /tmp/qos/classify.ebtables
	fi
}

reload_qos() {
	local service_name="$1"

	if [ -z "$service_name" ]; then
		configure_qos
	elif [ "$service_name" == "queue" ]; then
		pre_configure_queue
		configure_queue
	elif [ "$service_name" == "classify" ]; then
		configure_classify
		if [ -f "/tmp/qos/classify.ebtables" ]; then
			sh /tmp/qos/classify.ebtables
		fi
	elif [ "$service_name" == "policer" ]; then
		configure_policer
	fi
}

get_cfg_added_deleted()
{
	# return true if there is difference in number of queue configuration
	# in /etc/config/qos and /tmp/qos/qos, false is returned if both file
	# has same count of queue configuration.
	local queue=0
	local classify=0
	local shaper=0
	local policer=0
	local old_cfg="false"

	config_cb() {
		# invoked on the just previous config_load, get the count of
		# queue configuration in /etc/config/qos and /tmp/qos/qos.
		cfg_type="$1"
		cfg_name="$2"
		if [ -z $cfg_name ] || [ -z $cfg_type ]; then
			return
		fi

		if [ "$old_cfg" == "false" ]; then
			eval $cfg_type=$(($cfg_type + 1))
		else
			eval $cfg_type=$(($cfg_type - 1))
		fi
		option_cb() {
			local option="$1"
			local value="$2"
			if [ -z "$option" ] || [ -z "$value" ]; then
				return
			fi
			if [ "$old_cfg" == "false" ]; then
				eval $cfg_type=$(($cfg_type + 1))
			else
				eval $cfg_type=$(($cfg_type - 1))
			fi
		}
	}
	# config_load will trigger call for config_cb that is getting count
	# of qos configuration, respective config counts will come.
	config_load qos
	# config_load will trigger call for config_cb that is decreasing count
	# of qos configuration.
	old_cfg="true"
	UCI_CONFIG_DIR="/tmp/qos"
	config_load qos
	UCI_CONFIG_DIR="/etc/config"
	reset_cb

	if [ $classify -ne 0 ]; then
		modified_config="classify"
	fi
	if [ $shaper -ne 0 ]; then
		modified_config="$modified_config shaper"
	fi
	if [ $policer -ne  0 ]; then
		modified_config="$modified_config policer"
	fi
	if [ $queue -eq 0 ]; then
		echo "$modified_config"
		return
	else
		echo "queue"
		return
	fi
}

# reload_qos_service is invoked on qos service reload.
reload_qos_service() {
	q_cfg_restart="false"
	policer="false"
	classify="false"
	shaper="false"

	setup_qos

	if ! [[ -f "/etc/config/qos"  && -f "/tmp/qos/qos" ]]; then
		configure_qos
		cp /etc/config/qos /tmp/qos/qos
	fi

	config_cb() {
		# this is invoked when config_load is called.
		cfg_type="$1"
		cfg_name="$2"
		if [ -z $cfg_name ] || [ -z $cfg_type ]; then
			return
		fi
		option_cb() {
			# checking for if any parameters value is modified in queue cfg
			# comparsion is done between /etc/config/qos and /tmp/qos/qos
			local option="$1"
			local value="$2"
			local old_value=""
			if [ -z "$option" ] || [ -z "$value" ]; then
				return
			fi
			old_value=$(uci -q -c "/tmp/qos" get qos.$cfg_name.$option)
			if ! [ "$old_value" == "$value" ]; then
				if [ "$cfg_type" == "queue" ]; then
					q_cfg_restart="true"
				else
					eval $cfg_type="true"
				fi
			fi
		}
	}
	# if there is new addition/deletion of queue configuration
	# then return is queue to trigger restart of qos.
	# Otehrwise shaper policer classify
	# respective operation config is invoked that does not change queue statistics
	cfg_added_deleted=$(get_cfg_added_deleted)
	if [ "$cfg_added_deleted" == "queue" ]; then
		configure_qos
	else
		q_cfg_restart="false"
		# config_load will trigger call for config_cb that is checking
		# for modification in config value of queue config.
		# if change of value of queue config is there then q_cfg_restart
		# is set as true, else other qos config flag is set as true.
		config_load qos
		reset_cb
		if [ "$q_cfg_restart" == "true" ]; then
			configure_qos
		else
			for config in $cfg_added_deleted
			do
				eval $config="true"
			done
			if [ "$shaper" == "true" ]; then
				# for tc, shaper is implemented as part of q setup
				# hence, if shaper is modified, reload queue
				reload_qos "queue"
			fi
			if [ "$policer" == "true" ]; then
				reload_qos "policer"
				# change in policer config may need reconfiguration
				# of classifier
				reload_qos "classify"
			fi
			if [ "$classify" == "true" ]; then
				reload_qos "classify"
			fi
		fi
	fi
	cp /etc/config/qos /tmp/qos/qos
}