#!/bin/bash

# IHK SMP-x86 example boot script.
# author: Balazs Gerofi <bgerofi@riken.jp>
#      Copyright (C) 2014  RIKEN AICS
# 
# This is an example script for loading IHK, configuring a partition and 
# booting McKernel on it. 
# The script reserves half of the CPU cores and 512MB of RAM from NUMA node 0 
# when IHK is loaded for the first time, otherwise it destroys the current 
# McKernel instance and reboots it using the same set of resources as it used
# previously. 
# Note that the script does not output anything unless an error occurs.

prefix="@prefix@"
BINDIR="${prefix}/bin"
SBINDIR="${prefix}/sbin"
ETCDIR=@ETCDIR@
KMODDIR="${prefix}/kmod"
KERNDIR="${prefix}/@TARGET@/kernel"
ENABLE_MCOVERLAYFS="@ENABLE_MCOVERLAYFS@"

mem="512M@0"
cpus=""

INTERVAL=1
LOGMODE=0
facility="LOG_LOCAL6"
chown_option=`logname 2> /dev/null`

if [ "`systemctl status irqbalance_mck.service 2> /dev/null |grep -E 'Active: active'`" != "" -o "`systemctl status irqbalance.service 2> /dev/null |grep -E 'Active: active'`" != "" ]; then
irqbalance_used="yes"
else
irqbalance_used="no"
fi

while getopts :i:k:c:m:o:f: OPT
do
	case ${OPT} in
	f)	facility=${OPTARG}
		;;
	o)	chown_option=${OPTARG}
		;;
	i)	INTERVAL=${OPTARG}
		expr "${INTERVAL}" + 1 > /dev/null 2>&1
		if [ $? -ge 2 ]
		then
			echo "invalid -i value" >&2
			exit 1
		fi
		if [ ${INTERVAL} -le 0 ]
		then
			echo "invalid -i value" >&2
			exit 1
		fi
		;;
	k)	LOGMODE=${OPTARG}
		expr "${LOGMODE}" + 1 > /dev/null 2>&1
		if [ $? -ge 2 ]
		then
			echo "invalid -k value" >&2
			exit 1
		fi
		if [ ${LOGMODE} -lt 0 -o ${LOGMODE} -gt 2 ]
		then
			echo "invalid -k value" >&2
			exit 1
		fi
		;;
	c) cpus=${OPTARG}
		;;
	m) mem=${OPTARG}
		;;
	*)  echo "invalid option -${OPT}" >&2
		exit 1
	esac
done

ihk_ikc_irq_core=0

release=`uname -r`
major=`echo ${release} | sed -e 's/^\([0-9]*\).*/\1/'`
minor=`echo ${release} | sed -e 's/^[0-9]*.\([0-9]*\).*/\1/'`
patch=`echo ${release} | sed -e 's/^[0-9]*.[0-9]*.\([0-9]*\).*/\1/'`
linux_version_code=`expr \( ${major} \* 65536 \) + \( ${minor} \* 256 \) + ${patch}`
rhel_release=`echo ${release} | sed -e 's/^[0-9]*.[0-9]*.[0-9]*-\([0-9]*\).*/\1/'`
if [ "${release}" == "${rhel_release}" ]; then rhel_release=""; fi
enable_mcoverlay="no"
if [ "${ENABLE_MCOVERLAYFS}" == "yes" ]; then
	if [ "${rhel_release}" == "" ]; then
		if [ ${linux_version_code} -ge 262144 -a ${linux_version_code} -lt 262400 ]; then
			enable_mcoverlay="yes"
		fi
	else
		if [ ${linux_version_code} -eq 199168 -a ${rhel_release} -ge 327 ]; then
			enable_mcoverlay="yes"
		fi
	fi
fi

if [ "$cpus" == "" ]; then
	# Get the number of CPUs on NUMA node 0
	nr_cpus=`lscpu --parse | awk -F"," '{if ($4 == 0) print $4}' | wc -l`

	# Use the second half of the cores
	let nr_cpus="$nr_cpus / 2"
	cpus=`lscpu --parse | awk -F"," '{if ($4 == 0) print $1}' | tail -n $nr_cpus | xargs echo -n | sed 's/ /,/g'`
	if [ "$cpus" == "" ]; then echo "error: no available CPUs on NUMA node 0?" >&2; exit 1; fi
fi

# Remove mcoverlay if loaded
if [ "$enable_mcoverlay" == "yes" ]; then
	if [ "`lsmod | grep mcoverlay`" != "" ]; then
		if [ "`cat /proc/mounts | grep /tmp/mcos/mcos0_sys`" != "" ]; then umount -l /tmp/mcos/mcos0_sys; fi
		if [ "`cat /proc/mounts | grep /tmp/mcos/mcos0_proc`" != "" ]; then umount -l /tmp/mcos/mcos0_proc; fi
		if [ "`cat /proc/mounts | grep /tmp/mcos/linux_proc`" != "" ]; then umount -l /tmp/mcos/linux_proc; fi
		if [ "`cat /proc/mounts | grep /tmp/mcos`" != "" ]; then umount -l /tmp/mcos; fi
		if [ -e /tmp/mcos ]; then rm -rf /tmp/mcos; fi
		if ! rmmod mcoverlay; then echo "error: removing mcoverlay" >&2; exit 1; fi
	fi
fi

# Stop irqbalance
if [ "${irqbalance_used}" == "yes" ]; then
    systemctl stop irqbalance_mck.service 2>/dev/null
    if ! systemctl stop irqbalance.service 2>/dev/null ; then echo "error: stopping irqbalance" >&2; exit 1; fi;
fi

# Load IHK if not loaded
if [ "`lsmod | grep ihk`" == "" ]; then
        if ! insmod ${KMODDIR}/ihk.ko; then echo "error: loading ihk" >&2; exit 1; fi;
fi

# Load IHK-SMP if not loaded and reserve CPUs and memory
if [ "`lsmod | grep ihk_smp_x86`" == "" ]; then
        ihk_irq=""
        for i in `seq 64 255`; do 
                if [ ! -d /proc/irq/$i ] && [ "`cat /proc/interrupts | grep ":" | awk '{print $1}' | grep -o '[0-9]*' | grep -e '^$i$'`" == "" ]; then
                        ihk_irq=$i 
                        break
                fi
        done
        if [ "$ihk_irq" == "" ]; then echo "error: no IRQ available" >&2; exit 1; fi
        if ! insmod ${KMODDIR}/ihk-smp-x86.ko ihk_start_irq=$ihk_irq ihk_ikc_irq_core=$ihk_ikc_irq_core; then echo "error: loading ihk-smp-x86" >&2; exit 1; fi;
        if ! ${SBINDIR}/ihkconfig 0 reserve cpu ${cpus}; then echo "error: reserving CPUs" >&2; exit 1; fi
        if ! ${SBINDIR}/ihkconfig 0 reserve mem ${mem}; then echo "error: reserving memory" >&2; exit 1; fi
# If loaded, but no resources allocated, get CPUs and memory
else
        if ! ${SBINDIR}/ihkconfig 0 query cpu > /dev/null; then echo "error: querying cpus" >&2; exit 1; fi
        cpus_allocated=`${SBINDIR}/ihkosctl 0 query cpu`
		if [ "$cpus_allocated" == "" ]; then
			if ! ${SBINDIR}/ihkconfig 0 reserve cpu ${cpus}; then echo "error: reserving CPUs" >&2; exit 1; fi
		fi

        if ! ${SBINDIR}/ihkosctl 0 query mem > /dev/null; then echo "error: querying memory" >&2; exit 1; fi
        mem_allocated=`${SBINDIR}/ihkosctl 0 query mem`
		if [ "$mem_allocated" == "" ]; then
			if ! ${SBINDIR}/ihkconfig 0 reserve mem ${mem}; then echo "error: reserving memory" >&2; exit 1; fi
		fi
fi

# Load mcctrl if not loaded
if [ "`lsmod | grep mcctrl`" == "" ]; then
	if ! insmod ${KMODDIR}/mcctrl.ko; then echo "error: inserting mcctrl.ko" >&2; exit 1; fi
fi

# Check for existing OS instance and destroy
if [ -c /dev/mcos0 ]; then
        # Query CPU cores and memory of OS instance so that the same values are used as previously
        if ! ${SBINDIR}/ihkosctl 0 query cpu > /dev/null; then echo "error: querying cpus" >&2; exit 1; fi
        cpus=`${SBINDIR}/ihkosctl 0 query cpu`
        if ! ${SBINDIR}/ihkosctl 0 query mem > /dev/null; then echo "error: querying memory" >&2; exit 1; fi
        mem=`${SBINDIR}/ihkosctl 0 query mem`

        if ! ${SBINDIR}/ihkconfig 0 destroy 0; then echo "warning: destroy failed" >&2; fi	
else
        # Otherwise query IHK-SMP for resources
        if ! ${SBINDIR}/ihkconfig 0 query cpu > /dev/null; then echo "error: querying cpus" >&2; exit 1; fi
        cpus=`${SBINDIR}/ihkconfig 0 query cpu`
        if ! ${SBINDIR}/ihkconfig 0 query mem > /dev/null; then echo "error: querying memory" >&2; exit 1; fi
        mem=`${SBINDIR}/ihkconfig 0 query mem`	
fi

if ! ${SBINDIR}/ihkconfig 0 create; then echo "error: create" >&2; exit; fi
if ! ${SBINDIR}/ihkosctl 0 assign cpu ${cpus}; then echo "error: assign CPUs" >&2; exit 1; fi
if ! ${SBINDIR}/ihkosctl 0 assign mem ${mem}; then echo "error: assign memory" >&2; exit 1; fi
if ! ${SBINDIR}/ihkosctl 0 load ${KERNDIR}/mckernel.img; then echo "error: loading kernel image" >&2; exit 1; fi
if ! ${SBINDIR}/ihkosctl 0 kargs "hidos ksyslogd=${LOGMODE}"; then echo "error: setting kernel arguments" >&2; exit 1; fi
if ! ${SBINDIR}/ihkosctl 0 boot; then echo "error: booting" >&2; exit 1; fi
if ! chown ${chown_option} /dev/mcd* /dev/mcos*; then echo "error: chowning device files" >&2; exit 1; fi

if [ "$enable_mcoverlay" == "yes" ]; then
	if [ ! -e /tmp/mcos ]; then mkdir -p /tmp/mcos; fi
	if ! mount -t tmpfs tmpfs /tmp/mcos; then echo "error: mount /tmp/mcos" >&2; exit 1; fi
	if [ ! -e /tmp/mcos/linux_proc ]; then mkdir -p /tmp/mcos/linux_proc; fi
	if ! mount --bind /proc /tmp/mcos/linux_proc; then echo "error: mount /tmp/mcos/linux_proc" >&2; exit 1; fi
	if ! insmod ${KMODDIR}/mcoverlay.ko; then echo "error: inserting mcoverlay.ko" >&2; exit 1; fi
	while [ ! -e /proc/mcos0 ]
	do
		sleep 1
	done
	if [ ! -e /tmp/mcos/mcos0_proc ]; then mkdir -p /tmp/mcos/mcos0_proc; fi
	if [ ! -e /tmp/mcos/mcos0_proc_upper ]; then mkdir -p /tmp/mcos/mcos0_proc_upper; fi
	if [ ! -e /tmp/mcos/mcos0_proc_work ]; then mkdir -p /tmp/mcos/mcos0_proc_work; fi
	if ! mount -t mcoverlay mcoverlay -o lowerdir=/proc/mcos0:/proc,upperdir=/tmp/mcos/mcos0_proc_upper,workdir=/tmp/mcos/mcos0_proc_work,nocopyupw,nofscheck /tmp/mcos/mcos0_proc; then echo "error: mount /tmp/mcos/mcos0_proc" >&2; exit 1; fi
	mount --make-rprivate /proc

	while [ ! -e /sys/devices/virtual/mcos/mcos0/sys/devices/system/node ]
	do
		sleep 1
	done
	if [ ! -e /tmp/mcos/mcos0_sys ]; then mkdir -p /tmp/mcos/mcos0_sys; fi
	if [ ! -e /tmp/mcos/mcos0_sys_upper ]; then mkdir -p /tmp/mcos/mcos0_sys_upper; fi
	if [ ! -e /tmp/mcos/mcos0_sys_work ]; then mkdir -p /tmp/mcos/mcos0_sys_work; fi
	if ! mount -t mcoverlay mcoverlay -o lowerdir=/sys/devices/virtual/mcos/mcos0/sys:/sys,upperdir=/tmp/mcos/mcos0_sys_upper,workdir=/tmp/mcos/mcos0_sys_work,nocopyupw,nofscheck /tmp/mcos/mcos0_sys; then echo "error: mount /tmp/mcos/mcos0_sys" >&2; exit 1; fi
	mount --make-rprivate /sys
	for cpuid in `find /sys/devices/system/cpu/* -maxdepth 0 -name "cpu[0123456789]*" -printf "%f "`; do
		if [ ! -e "/sys/devices/virtual/mcos/mcos0/sys/devices/system/cpu/$cpuid" ]; then
			rm -rf /tmp/mcos/mcos0_sys/devices/system/cpu/$cpuid
		fi
	done
	for nodeid in `find /sys/devices/system/node/* -maxdepth 0 -name "node[0123456789]*" -printf "%f "`; do
		if [ ! -e "/sys/devices/virtual/mcos/mcos0/sys/devices/system/node/$nodeid" ]; then
			rm -rf /tmp/mcos/mcos0_sys/devices/system/node/$nodeid
		else
			# Delete non-existent symlinks
			for cpuid in `find /sys/devices/system/node/$nodeid/* -maxdepth 0 -name "cpu[0123456789]*" -printf "%f "`; do
				if [ ! -e "/sys/devices/virtual/mcos/mcos0/sys/devices/system/node/$nodeid/$cpuid" ]; then
					rm -f /tmp/mcos/mcos0_sys/devices/system/node/$nodeid/$cpuid
				fi
			done

			rm -f /tmp/mcos/mcos0_sys/devices/system/node/$nodeid/memory*
		fi
	done
	for cpuid in `find /sys/bus/cpu/devices/* -maxdepth 0 -name "cpu[0123456789]*" -printf "%f "`; do
		if [ ! -e "/sys/devices/virtual/mcos/mcos0/sys/bus/cpu/devices/$cpuid" ]; then
			rm -rf /tmp/mcos/mcos0_sys/bus/cpu/devices/$cpuid
		fi
	done
fi
if [ ${LOGMODE} -ne 0 ]
then
# mcklogd survives when McKernel isn't shut down by mcstop+release.sh
	pkill mcklogd    
	SBINDIR=${SBINDIR} ${SBINDIR}/mcklogd -i ${INTERVAL} -f ${facility}
fi

# Start irqbalance with CPUs and IRQ for McKernel banned
if [ "${irqbalance_used}" == "yes" ]; then
    if ! etcdir=@ETCDIR@ perl -e 'use File::Copy qw(copy); $etcdir=$ENV{'etcdir'}; @files = grep { -f } glob "/proc/irq/*/smp_affinity"; foreach $file (@files) { $rel = substr($file, 1); $dir=substr($rel, 0, length($rel)-length("/smp_affinity")); if(0) { print "cp $file $etcdir/$rel\n";} if(system("mkdir -p $etcdir/$dir")){ exit 1;} if(!copy($file,"$etcdir/$rel")){ exit 1;} }' ; then echo "error: saving /proc/irq/*/smp_affinity" >&2; exit 1; fi;

    ncpus=`lscpu | grep -E '^CPU\(s\):' | awk '{print $2}'`
    smp_affinity_mask=`echo $cpus | ncpus=$ncpus perl -e 'while(<>){@tokens = split /,/;foreach $token (@tokens) {@nums = split /-/,$token; for($num = $nums[0]; $num <= $nums[$#nums]; $num++) {$ndx=int($num/32); $mask[$ndx] |= (1<<($num % 32))}}} $nint32s = int(($ENV{'ncpus'}+31)/32); for($j = $nint32s - 1; $j >= 0; $j--) { if($j != $nint32s - 1){print ",";} $nblks = $j == $nint32s - 1 ? int(($ENV{'ncpus'} % 32)/4) : 8; for($i = $nblks - 1;$i >= 0;$i--){ printf("%01x",($mask[$j] >> ($i*4)) & 0xf);}}'`

    if ! ncpus=$ncpus smp_affinity_mask=$smp_affinity_mask perl -e '@dirs = grep { -d } glob "/proc/irq/*"; foreach $dir (@dirs) { $hit = 0; $affinity_str = `cat $dir/smp_affinity`; chomp $affinity_str; @int32strs = split /,/, $affinity_str; @int32strs_mask=split /,/, $ENV{'smp_affinity_mask'}; for($i=0;$i <= $#int32strs_mask; $i++) { $int32strs_inv[$i] = sprintf("%08x",hex($int32strs_mask[$i])^0xffffffff); if($i == 0) { $len = int((($ENV{'ncpus'}%32)+3)/4); $int32strs_inv[$i] = substr($int32strs_inv[$i], -$len, $len); } } $inv = join(",", @int32strs_inv); $nint32s = int(($ENV{'ncpus'}+31)/32); for($j = $nint32s - 1; $j >= 0; $j--) { if(hex($int32strs[$nint32s - 1 - $j]) & hex($int32strs_mask[$nint32s - 1 - $j])) { $hit = 1; }} if($hit == 1) { $cmd = "echo $inv > $dir/smp_affinity 2>/dev/null"; system $cmd;}}'; then echo "error: modifying /proc/irq/*/smp_affinity" >&2; exit 1; fi;

    banirq=`cat /proc/interrupts| perl -e 'while(<>) { if(/^\s*(\d+).*IHK\-SMP\s*$/) {print $1;}}'`

    sed "s/%mask%/$smp_affinity_mask/g" $ETCDIR/irqbalance_mck.in | sed "s/%banirq%/$banirq/g" > $ETCDIR/irqbalance_mck
    if ! systemctl link $ETCDIR/irqbalance_mck.service >/dev/null 2>/dev/null; then echo "error: linking irqbalance_mck" >&2; exit 1; fi;
    if ! systemctl start irqbalance_mck.service 2>/dev/null ; then echo "error: starting irqbalance_mck" >&2; exit 1; fi;
#    echo cpus=$cpus mask=$smp_affinity_mask banirq=$banirq
fi