#I LOVE YOU EVERDAY#

#!/bin/bash
# This is the default setting of networking multiqueue and irq affinity
# 1. enable multiqueue if available
# 2. irq affinity optimization
# 3. stop irqbalance service
# 4. adapted to the logic of the new cpu topo
# 5. add cases for the irq and cpu ratio
# 6. tune the smp affinity based on interface
# 7. add cpus topo and bitmap relation

# DEBUG
# set -x

# declare some global variables
version="2.0.2"
ecs_network_log=/var/log/ecs_network_optimization.log
node_dir=/sys/devices/system/node
bit_array=()
cpu_num=0
chunk_size=32
group_size=0

input_para=${1:-new}

# bitmap relation function
function init_bit_array() {
    for ((i = 0; i < group_size; ++i)); do
        bit_array["$i"]=0
    done
}

function set_bit() {
    location=$1
    arr_index=$((location / chunk_size))
    bit_index=$((location % chunk_size))
    bit_array[arr_index]=$((bit_array[arr_index] | (1 << bit_index)))
}

# set and check multiqueue
function set_check_multiqueue() {
    eth=$1
    log_file=$ecs_network_log
    queue_num=$(ethtool -l $eth | grep -ia5 'pre-set' | grep -i combined | awk {'print $2'})
    if [ $queue_num -gt 1 ]; then
        # set multiqueue
        ethtool -L $eth combined $queue_num
        # check multiqueue setting
        cur_q_num=$(ethtool -l $eth | grep -iA5 current | grep -i combined | awk {'print $2'})
        if [ "X$queue_num" != "X$cur_q_num" ]; then
            echo "Failed to set $eth queue size to $queue_num" >>$log_file
            echo "after setting, pre-set queue num: $queue_num , current: $cur_q_num" >>$log_file
            return 1
        else
            echo "OK. set $eth queue size to $queue_num" >>$log_file
            return 0
        fi
    else
        echo "only support $queue_num queue; no need to enable multiqueue on $eth" >>$log_file
        return 1
    fi
}

# check the cpu topology is old or new, to put it simply:
# old topo is like one core with two cpus which are next to each other
# new topo one core with two cpus which are split with the number of cores per socket
function core_cpu_sibings() {
    cpu_topo_thread_sibling=/sys/devices/system/cpu/cpu0/topology/thread_siblings_list
    awk -F, '{ print $2 }' $cpu_topo_thread_sibling
}

# get all virtio_net irqs
function get_irqs() {
    driver="$1"
    awk -v driver="$driver".*put -F: '$0 ~ driver {print $1}' /proc/interrupts | tr -d ' '
}

# get device PCIe device location on NUMA0 or NUMA1
function get_device_numa_location() {
    numa_node_location=-1
    numa_node_file=$(dirname "$(readlink -f /sys/class/net/"$1"/device)")/numa_node
    if [ -f "$numa_node_file" ]; then
        numa_node_location=$(cat "$numa_node_file")
    fi
    echo "$numa_node_location"
}

# print logs of the tuned result
function show_nic_smp_xps_rps_info() {
    ethX=$1
    driver=$2

irqs=($(get_irqs "$driver"))
    echo "$ethX rx queue:" >>$ecs_network_log
    for ((i = 0; i < ${#irqs[@]}; i += 2)); do
        rx_queue=$((i / 2))
        rx_irq=${irqs[i]}
        aff_cpu=$(cat /proc/irq/"$rx_irq"/smp_affinity_list)
        rps_cpu=$(cat /sys/class/net/"$ethX"/queues/rx-"$rx_queue"/rps_cpus)
        rps_flow_cnt=$(cat /sys/class/net/"$ethX"/queues/rx-"$rx_queue"/rps_flow_cnt)
        printf "rx_queue: %-3s irq: %-3s irq_affinity_cpu: %-12s rps_cpu: %s rps_flow_cnt: %s\n" "$rx_queue" "$rx_irq" "$aff_cpu" "$rps_cpu" "$rps_flow_cnt" >>$ecs_network_log
    done

echo "$ethX tx queue:" >>$ecs_network_log
    for ((i = 1; i < ${#irqs[@]}; i += 2)); do
        tx_queue=$((i / 2))
        tx_irq=${irqs[i]}
        aff_cpu=$(cat /proc/irq/"$tx_irq"/smp_affinity_list)
        xps_cpu=$(cat /sys/class/net/"$ethX"/queues/tx-"$tx_queue"/xps_cpus)
        printf "tx_queue: %-3s irq: %-3s irq_affinity_cpu: %-12s xps_cpu: %s\n" "$tx_queue" "$tx_irq" "$aff_cpu" "$xps_cpu" >>$ecs_network_log
    done
}

# get cpus sorted based on NUMA nodes
function get_numa_node_cpus() {
    cpu_list=()
    device_numa_location=$(get_device_numa_location "$1")
    if [ "$device_numa_location" != -1 ]; then
        cpu_pairs=($(cat ${node_dir}/node${device_numa_location}/cpulist | sed 's/,/ /g'))
        for cpu_pair in "${cpu_pairs[@]}"; do
            IFS='-' read -ra range <<<"$cpu_pair"
            for ((j = range[0]; j <= range[1]; j++)); do
                cpu_list+=("$j")
            done
        done
    fi
    for i in $(ls -d $node_dir/node*); do
        i=${i/*node/}
        if [ "$device_numa_location" == "$i" ]; then
            continue
        fi
        cpu_pairs=($(cat ${node_dir}/node${i}/cpulist | sed 's/,/ /g'))
        for cpu_pair in "${cpu_pairs[@]}"; do
            IFS='-' read -ra range <<<"$cpu_pair"
            for ((j = range[0]; j <= range[1]; j++)); do
                cpu_list+=("$j")
            done
        done
    done
    echo "${cpu_list[@]}"
}

# stop irqbalance service
function stop_irqblance() {
    log_file=$ecs_network_log
    ret=0

if [ "X" != "X$(ps -ef | grep irqbalance | grep -v grep)" ]; then
        if which systemctl; then
            systemctl stop irqbalance
            systemctl disable irqbalance
        else
            service irqbalance stop
            chkconfig irqbalance off
        fi
        if [ $? -ne 0 ]; then
            echo "Failed to stop irqbalance" >>$log_file
            ret=1
        fi

else
        echo "OK. irqbalance stoped." >>$log_file
    fi
    return $ret
}

function set_irq_smpaffinity_new() {
    driver="$2"
    irqs=($(get_irqs "$driver"))
    irqs_num=${#irqs[@]}

cpus=($(get_numa_node_cpus "$1"))
    cpu_num=${#cpus[@]}

group_size=$(((cpu_num + chunk_size - 1) / chunk_size))
    ratio=$((cpu_num / (irqs_num / 2)))
    remains=$((ratio * (irqs_num / 2)))

cpu_sibling=$(core_cpu_sibings)

for ((irq = 0; irq < irqs_num; )); do
        init_bit_array

positions=()
        if [ $ratio -eq 1 ]; then
            # tx and tx will fill all the cpu bitmask
            cpu=${cpus[((irq / 2))]}
            positions+=("$cpu")
        elif [ $ratio -eq 2 ] || [ $ratio -eq 4 ]; then
            # if the ratio is 2, then the tx and rx will take one cpu bit in one core siblings
            # if the ratio is 4, mostly there are NUMA nodes, tx and rx take on cpu bit in one
            # core and to fill up the NUMA0 as mush as possible, then to NUMA1
            cpu=${cpus[irq]}
            if [ -n "$cpu_sibling" ]; then
                # new topo
                cpu=$((irq % 2 == 0 ? cpus[irq / 2] : cpus[(irq / 2) + cpu_sibling]))
            fi
            positions+=("$cpu")
        elif [ $ratio -eq 8 ]; then
            # per cpu per tx per rx is far enough, this is moslty AMD instance
            # just filled it to every 4 cpus, based on the experiment
            cpu=${cpus[irq * 4]}
            positions+=("$cpu")
        else
            # other cases like ratio 3 leave it as the default we do.
            # if need to deal with it, just do it based on the cases
            cpu=${cpus[irq]}
            positions+=("$cpu")
        fi

# according to the virtio_net driver the "straggers" are processed too.
        # https://github.com/torvalds/linux/commit/2ca653d607ce59f2729173a7ea56dbfa6330ec88
        if [ $remains != "$cpu_num" ]; then
            positions+=("${cpus[remains]}")
            ((remains += 1))
        fi

for position in "${positions[@]}"; do
            set_bit "$position"
        done

irq_mask=$(print_bit_array)

if [ $ratio -eq 1 ]; then
            # tx and tx will be filled one same bit mask
            echo "$irq_mask" >"/proc/irq/${irqs[irq]}/smp_affinity"
            echo "$irq_mask" >"/proc/irq/${irqs[irq + 1]}/smp_affinity"
            ((irq += 2))
        elif [ $ratio -eq 2 ] || [ $ratio -eq 4 ]; then
            # tx and tx will be filled one same bit mask in one cpu sibling cpus
            echo "$irq_mask" >"/proc/irq/${irqs[irq]}/smp_affinity"
            ((irq += 1))
        elif [ $ratio -eq 8 ]; then
            # tx and rx fill all the cpu maps
            echo "$irq_mask" >"/proc/irq/${irqs[irq]}/smp_affinity"
            ((irq += 1))
        else
            # other cases like ratio 3 leave it as default
            echo "$irq_mask" >"/proc/irq/${irqs[irq]}/smp_affinity"
            ((irq += 1))
        fi
    done
}

function set_irq_smpaffinity_old() {
    node_cpumax=0
    for i in $(ls -d $node_dir/node*); do
        i=${i/*node/}
        node_cpumax=$(cat /sys/devices/system/node/node${i}/cpulist | awk -F- '{print $NF}')
        [[ $node_cpumax -gt 0 ]] && break
    done

echo "max node :$i" >>$ecs_network_log

driver="$2"
    irqs=($(get_irqs "$driver"))
    core=0
    for irq in ${irqs[@]}; do
        VEC=$core
        if [ $VEC -ge 32 ]; then
            let "IDX = $VEC / 32"
            MASK_FILL=""
            MASK_ZERO="00000000"
            for ((i = 1; i <= $IDX; i++)); do
                MASK_FILL="${MASK_FILL},${MASK_ZERO}"
            done
            let "VEC -= 32 * $IDX"
            MASK_TMP=$((1 << $VEC))
            MASK=$(printf "%X%s" $MASK_TMP $MASK_FILL)
        else
            MASK_TMP=$((1 << $VEC))
            MASK=$(printf "%X" $MASK_TMP)
        fi
        echo $MASK >/proc/irq/$irq/smp_affinity
        echo "mask:$MASK, irq:$irq" >>$ecs_network_log
        core=$(((core + 1) % (node_cpumax + 1)))
    done
}

# set irq affinity based on the instance
# reference some logic from virtio_net
function set_irq_smpaffinity() {
    if [ "$input_para" == "new" ]; then
        set_irq_smpaffinity_new "$@"
    else
        set_irq_smpaffinity_old "$@"
    fi
}

# main logic
function main() {
    ret_value=0
    echo "running $0" >$ecs_network_log
    echo "========  ECS network setting starts $(date +'%Y-%m-%d %H:%M:%S') ========" >>$ecs_network_log
    stop_irqblance
    # we assume your NIC interface(s) is/are like eth*
    eth_dirs=$(ls -d /sys/class/net/eth*)
    if [ "X$eth_dirs" = "X" ]; then
        echo "ERROR! can not find any ethX in /sys/class/net/ dir." >>$ecs_network_log
        ret_value=1
    fi
    for i in $eth_dirs; do
        cur_eth=$(basename "$i")
        echo "optimize network performance: current device $cur_eth" >>$ecs_network_log
        # only optimize virtio_net device
        driver=$(basename "$(readlink "$i"/device)")
        if ! echo "$driver" | grep -q virtio; then
            echo "ignore device $cur_eth with driver $driver" >>$ecs_network_log
            continue
        fi
        echo "set and check multiqueue on $cur_eth" >>$ecs_network_log
        set_check_multiqueue "$cur_eth" "$ecs_network_log"
        ret_value=$?
        if [ $ret_value -ne 0 ]; then
            echo "Failed to set multiqueue on $cur_eth or the queue len is 1. ret_value $ret_value" >>$ecs_network_log
            continue
        fi
        echo "Current ========== $cur_eth $driver smp affinity:" >>$ecs_network_log
        show_nic_smp_xps_rps_info "$cur_eth" "$driver"

set_irq_smpaffinity "$cur_eth" "$driver"

echo "After ==========tuned $cur_eth $driver smp affinity:" >>$ecs_network_log
        show_nic_smp_xps_rps_info "$cur_eth" "$driver"
    done

echo "========  ECS network setting END $(date +'%Y-%m-%d %H:%M:%S')  ========" >>$ecs_network_log
    return $ret_value
}

function show_help() {
    echo "Usage: $0 [old|new] [-h] [-v]"
    echo "Options:"
    echo "  -h, --help     Display this help and exit"
    echo "  -v, --version  Display the version info"
}

# program starts here
if [ -n "$1" ]; then
    if [ "$1" == "-v" ] || [ "$1" == "--version" ]; then
        echo $version
        exit 0
    elif [ "$1" != "old" ] && [ "$1" != "new" ]; then
        show_help
        exit 1
    fi
fi
main
exit $?