test/vhost_perf: refactor test scripts to use disk map

Use provided configuration file via --disk-map option
instead of creating bdevs and VMs in ordered sequence
(e.g. 0 to 10, etc.).

This allows to:
- specify which NVMe device we want to use
  (PCI BDF identifier)
- how to name it in  SPDK nvme bdev configuration
- how many splits or lvol bdevs to create on this device
- which VMs should use created bdevs

With CPU mask configuration file this allows to better
control resources when running the test (especially in
case of NUMA optimization where using sequential for/while
loops is not a good approach).

vm_count and max_disks parameters removed. These are not
needed anymore as they're controlled by config file.

Example of config file contents:
(BDF,Spdk NvmeBdev name,Split count,VM list)
0000:1b:00.0,Nvme1,2,2 3
0000:89:00.0,Nvme3,4,4 5 6 7

Change-Id: I9fc73458825d8072537aa04880765a048e034ce4
Signed-off-by: Karol Latecki <karol.latecki@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/464565
Reviewed-by: Shuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>

test/vhost/perf_bench/vhost_perf.sh

@@ -5,7 +5,7 @@ rootdir=$(readlink -f $testdir/../../..)
 source $rootdir/test/common/autotest_common.sh
 source $rootdir/test/vhost/common.sh
 
-vm_count=1
+vhost_num="0"
 vm_memory=2048
 vm_sar_enable=false
 host_sar_enable=false
@@ -13,19 +13,26 @@ sar_delay="0"
 sar_interval="1"
 sar_count="10"
 vm_throttle=""
-max_disks=""
 ctrl_type="spdk_vhost_scsi"
 use_split=false
 kernel_cpus=""
 run_precondition=false
 lvol_stores=()
 lvol_bdevs=()
+split_bdevs=()
 used_vms=""
 wwpn_prefix="naa.5001405bc6498"
 
 fio_bin="--fio-bin=/home/sys_sgsw/fio_ubuntu"
 fio_iterations=1
 precond_fio_bin="/usr/src/fio/fio"
+disk_map=""
+
+disk_cfg_bdfs=()
+disk_cfg_spdk_names=()
+disk_cfg_splits=()
+disk_cfg_vms=()
+disk_cfg_kernel_names=()
 
 function usage()
 {
@@ -39,10 +46,6 @@ function usage()
 	echo "                            of binary is recommended."
 	echo "    --fio-job=PATH          Fio config to use for test."
 	echo "    --fio-iterations=INT    Number of times to run specified workload."
-	echo "    --vm-count=INT          Total number of virtual machines to launch in this test;"
-	echo "                            Each VM will get one bdev (lvol or split vbdev)"
-	echo "                            to run FIO test."
-	echo "                            Default: 1"
 	echo "    --vm-memory=INT         Amount of RAM memory (in MB) to pass to a single VM."
 	echo "                            Default: 2048 MB"
 	echo "    --vm-image=PATH         OS image to use for running the VMs."
@@ -53,8 +56,6 @@ function usage()
 	echo "    --sar-interval=INT      Interval (seconds) argument for SAR. Default: 1s."
 	echo "    --sar-count=INT         Count argument for SAR. Default: 10."
 	echo "    --vm-throttle-iops=INT  I/Os throttle rate in IOPS for each device on the VMs."
-	echo "    --max-disks=INT         Maximum number of NVMe drives to use in test."
-	echo "                            Default: will use all available NVMes."
 	echo "    --ctrl-type=TYPE        Controller type to use for test:"
 	echo "                            spdk_vhost_scsi - use spdk vhost scsi"
 	echo "                            spdk_vhost_blk - use spdk vhost block"
@@ -66,6 +67,12 @@ function usage()
 	echo "    --precond-fio-bin       FIO binary used for SPDK fio plugin precondition. Default: /usr/src/fio/fio."
 	echo "    --custom-cpu-cfg=PATH   Custom CPU config for test."
 	echo "                            Default: spdk/test/vhost/common/autotest.config"
+	echo "    --disk-map              Disk map for given test. Specify which disks to use, their SPDK name,"
+	echo "                            how many times to split them and which VMs should be attached to created bdevs."
+	echo "                            Example:"
+	echo "                            NVME PCI BDF,Spdk Bdev Name,Split Count,VM List"
+	echo "                            0000:1a:00.0,Nvme0,2,0 1"
+	echo "                            0000:1b:00.0,Nvme1,2,2 3"
 	echo "-x                          set -x for script debug"
 	exit 0
 }
@@ -88,17 +95,16 @@ function cleanup_lvol_cfg()
 function cleanup_split_cfg()
 {
 	notice "Removing split vbdevs"
-	for (( i=0; i<max_disks; i++ ));do
+	for disk in "${disk_cfg_spdk_names[@]}"; do
-		$rpc_py bdev_split_delete Nvme${i}n1
+		$rpc_py bdev_split_delete ${disk}n1
 	done
 }
 
 function cleanup_parted_config()
 {
-	local disks
+	notice "Removing parted disk configuration"
-	disks=$(find /dev/ -maxdepth 1 -name 'nvme*n1' | sort --version-sort)
+	for disk in "${disk_cfg_kernel_names[@]}"; do
-	for disk in $disks; do
+		parted -s /dev/${disk}n1 rm 1
-		parted -s $disk rm 1
 	done
 }
 
@@ -109,6 +115,36 @@ function cleanup_kernel_vhost()
 	cleanup_parted_config
 }
 
+function create_vm() {
+	vm_num=$1
+	setup_cmd="vm_setup --disk-type=$ctrl_type --force=$vm_num --memory=$vm_memory --os=$VM_IMAGE"
+	if [[ "$ctrl_type" == "kernel_vhost" ]]; then
+		x=$(printf %03d $vm_num)
+		setup_cmd+=" --disks=${wwpn_prefix}${x}"
+	else
+		setup_cmd+=" --disks=0"
+	fi
+	$setup_cmd
+	used_vms+=" $vm_num"
+	echo "Added to used vms"
+	echo $used_vms
+}
+
+function create_spdk_controller() {
+	vm_num=$1
+	bdev=$2
+
+	if [[ "$ctrl_type" == "spdk_vhost_scsi" ]]; then
+		$rpc_py vhost_create_scsi_controller naa.0.$vm_num
+		notice "Created vhost scsi controller naa.0.$vm_num"
+		$rpc_py vhost_scsi_controller_add_target naa.0.$vm_num 0 $bdev
+		notice "Added LUN 0/$bdev to controller naa.0.$vm_num"
+	elif [[ "$ctrl_type" == "spdk_vhost_blk" ]]; then
+		$rpc_py vhost_create_blk_controller naa.0.$vm_num $bdev
+		notice "Created vhost blk controller naa.0.$vm_num $bdev"
+	fi
+}
+
 while getopts 'xh-:' optchar; do
 	case "$optchar" in
 		-)
@@ -117,7 +153,6 @@ while getopts 'xh-:' optchar; do
 			fio-bin=*) fio_bin="--fio-bin=${OPTARG#*=}" ;;
 			fio-job=*) fio_job="${OPTARG#*=}" ;;
 			fio-iterations=*) fio_iterations="${OPTARG#*=}" ;;
-			vm-count=*) vm_count="${OPTARG#*=}" ;;
 			vm-memory=*) vm_memory="${OPTARG#*=}" ;;
 			vm-image=*) VM_IMAGE="${OPTARG#*=}" ;;
 			vm-sar-enable) vm_sar_enable=true ;;
@@ -126,13 +161,13 @@ while getopts 'xh-:' optchar; do
 			sar-interval=*) sar_interval="${OPTARG#*=}" ;;
 			sar-count=*) sar_count="${OPTARG#*=}" ;;
 			vm-throttle-iops=*) vm_throttle="${OPTARG#*=}" ;;
-			max-disks=*) max_disks="${OPTARG#*=}" ;;
 			ctrl-type=*) ctrl_type="${OPTARG#*=}" ;;
 			use-split) use_split=true ;;
 			run-precondition) run_precondition=true ;;
 			precond-fio-bin=*) precond_fio_bin="${OPTARG#*=}" ;;
 			limit-kernel-vhost=*) kernel_cpus="${OPTARG#*=}" ;;
 			custom-cpu-cfg=*) custom_cpu_cfg="${OPTARG#*=}" ;;
+			disk-map=*) disk_map="${OPTARG#*=}" ;;
 			*) usage $0 "Invalid argument '$OPTARG'" ;;
 		esac
 		;;
@@ -147,6 +182,10 @@ rpc_py="$rootdir/scripts/rpc.py -s $(get_vhost_dir 0)/rpc.sock"
 
 if [[ -n $custom_cpu_cfg ]]; then
 	source $custom_cpu_cfg
+	vhost_reactor_mask="vhost_${vhost_num}_reactor_mask"
+	vhost_reactor_mask="${!vhost_reactor_mask}"
+	vhost_master_core="vhost_${vhost_num}_master_core"
+	vhost_master_core="${!vhost_master_core}"
 fi
 
 if [[ -z $fio_job ]]; then
@@ -155,40 +194,11 @@ if [[ -z $fio_job ]]; then
 fi
 
 trap 'error_exit "${FUNCNAME}" "${LINENO}"' INT ERR
-notice "Get NVMe disks:"
-nvmes=($(iter_pci_class_code 01 08 02))
 
-if [[ -z $max_disks ]]; then
+if [[ -z $disk_map ]]; then
-	max_disks=${#nvmes[@]}
+	fail "No disk map provided for test. Exiting."
 fi
 
-if [[ ${#nvmes[@]} -lt max_disks ]]; then
-	fail "Number of NVMe drives (${#nvmes[@]}) is lower than number of requested disks for test ($max_disks)"
-fi
-
-
-# Calculate number of needed splits per NVMe
-# so that each VM gets it's own bdev during test.
-splits=()
-
-if [[ $vm_count -le $max_disks ]]; then
-	for i in $(seq 0 $((max_disks - 1))); do
-		splits+=("1")
-	done
-else
-	#Calculate least minimum number of splits on each disks
-	for i in $(seq 0 $((max_disks - 1))); do
-		splits+=( $((vm_count / max_disks)) )
-	done
-	# Split up the remainder
-	for i in $(seq 0 $((vm_count % max_disks - 1))); do
-		(( splits[i]++ ))
-	done
-fi
-notice "Preparing NVMe setup..."
-notice "Using $max_disks physical NVMe drives"
-notice "Nvme split list: ${splits[*]}"
-
 # ===== Precondition NVMes if specified =====
 if [[ $run_precondition == true ]]; then
 	# Using the same precondition routine possible for lvols thanks
@@ -204,102 +214,136 @@ if [[ $run_precondition == true ]]; then
     --iodepth=32 --filename="${fio_filename}" || true
 fi
 
-# ===== Prepare NVMe splits & run vhost process =====
+set +x
+readarray disk_cfg < $disk_map
+for line in "${disk_cfg[@]}"; do
+	echo $line
+	IFS=","
+	s=($line)
+	disk_cfg_bdfs+=(${s[0]})
+	disk_cfg_spdk_names+=(${s[1]})
+	disk_cfg_splits+=(${s[2]})
+	disk_cfg_vms+=("${s[3]}")
+
+	# Find kernel nvme names
 	if [[ "$ctrl_type" == "kernel_vhost" ]]; then
+		tmp=$(find /sys/devices/pci* -name ${s[0]} -print0 | xargs sh -c 'ls $0/nvme')
+		disk_cfg_kernel_names+=($tmp)
+		IFS=" "
+	fi
+done
+unset IFS
+set -x
+
+
+if [[ "$ctrl_type" == "kernel_vhost" ]]; then
+	notice "Configuring kernel vhost..."
 	trap 'vm_kill_all; sleep 1; cleanup_kernel_vhost; error_exit "${FUNCNAME}" "${LINENO}"' INT ERR
+
 	# Split disks using parted for kernel vhost
 	newline=$'\n'
-	for (( i=0; i<max_disks; i++ ));do
+	backstores=()
-		parted -s /dev/nvme${i}n1 mklabel msdos
+	for (( i=0; i<${#disk_cfg_kernel_names[@]}; i++ )); do
-		parted -s /dev/nvme${i}n1 mkpart extended 2048s 100%
+		nvme=${disk_cfg_kernel_names[$i]}
-		part_size=$((100/${splits[$i]})) # Split 100% of disk into roughly even parts
+		splits=${disk_cfg_splits[$i]}
-		echo "  Creating ${splits[$i]} partitions of relative disk size ${part_size}"
+		notice "  Creating extended partition on disk /dev/${nvme}n1"
+		parted -s /dev/${nvme}n1 mklabel msdos
+		parted -s /dev/${nvme}n1 mkpart extended 2048s 100%
 
-		for p in $(seq 0 $((${splits[$i]} - 1))); do
+		part_size=$((100/${disk_cfg_splits[$i]})) # Split 100% of disk into roughly even parts
+		echo "  Creating  ${splits} partitions of relative disk size ${part_size}"
+		for p in $(seq 0 $((splits - 1))); do
 			p_start=$((p*part_size))
 			p_end=$((p_start+part_size))
-			parted -s /dev/nvme${i}n1 mkpart logical ${p_start}% ${p_end}%
+			parted -s /dev/${nvme}n1 mkpart logical ${p_start}% ${p_end}%
+			sleep 3
 		done
-	done
-	sleep 1
 
 		# Prepare kernel vhost configuration
 		# Below grep: match only NVMe partitions which are not "Extended" type.
 		# For example: will match nvme0n1p15 but not nvme0n1p1
-	partitions=$(find /dev/ -maxdepth 1 -name 'nvme*' | sort --version-sort | grep -P 'p(?!1$)\d+')
+		partitions=$(find /dev -name "${nvme}n1*" | sort --version-sort | grep -P 'p(?!1$)\d+')
-	backstores=()
-
 		# Create block backstores for vhost kernel process
 		for p in $partitions; do
 			backstore_name=$(basename $p)
 			backstores+=("$backstore_name")
 			targetcli backstores/block create $backstore_name $p
 		done
+		partitions=($partitions)
 
 		# Create kernel vhost controllers and add LUNs
-	for ((i=0; i<${#backstores[*]}; i++)); do
+		# Setup VM configurations
+		vms_to_run=(${disk_cfg_vms[i]})
+		for (( j=0; j<${#vms_to_run[@]}; j++ )); do
 			# WWPN prefix misses 3 characters. Need to complete it
 			# using block backstore number
-		x=$(printf %03d $i)
+			x=$(printf %03d ${vms_to_run[$j]})
 			wwpn="${wwpn_prefix}${x}"
 			targetcli vhost/ create $wwpn
-		targetcli vhost/$wwpn/tpg1/luns create /backstores/block/${backstores[$i]}
+			targetcli vhost/$wwpn/tpg1/luns create /backstores/block/$(basename ${partitions[$j]})
+			create_vm ${vms_to_run[j]}
+			sleep 1
 		done
+	done
+	targetcli ls
 else
-	# Run vhost process and prepare split vbdevs or lvol bdevs
+	notice "Configuring SPDK vhost..."
-	notice "running SPDK vhost"
+	vhost_run "${vhost_num}" "--no-gen-nvme" "-p ${vhost_master_core}" "-m ${vhost_reactor_mask}"
-	vhost_run 0
 	notice "..."
 
 	if [[ $use_split == true ]]; then
-		notice "Using split vbdevs"
+		notice "Configuring split bdevs configuration..."
 		trap 'cleanup_split_cfg; error_exit "${FUNCNAME}" "${LINENO}"' INT ERR
-		split_bdevs=()
+		for (( i=0; i<${#disk_cfg_bdfs[@]}; i++ )); do
-		for (( i=0; i<max_disks; i++ ));do
+			nvme_bdev=$($rpc_py bdev_nvme_attach_controller -b ${disk_cfg_spdk_names[$i]} -t pcie -a ${disk_cfg_bdfs[$i]})
-			out=$($rpc_py bdev_split_create Nvme${i}n1 ${splits[$i]})
+			notice "Created NVMe Bdev: $nvme_bdev with BDF ${disk_cfg_bdfs[$i]}"
-			for s in $(seq 0 $((${splits[$i]}-1))); do
+
-				split_bdevs+=("Nvme${i}n1p${s}")
+			splits=$($rpc_py bdev_split_create $nvme_bdev ${disk_cfg_splits[$i]})
+			splits=($splits)
+			notice "Created splits: ${splits[*]} on Bdev ${nvme_bdev}"
+			for s in "${splits[@]}"; do
+				split_bdevs+=($s)
 			done
+
+			vms_to_run=(${disk_cfg_vms[i]})
+			for (( j=0; j<${#vms_to_run[@]}; j++ )); do
+				notice "Setting up VM ${vms_to_run[j]}"
+				create_spdk_controller "${vms_to_run[j]}" ${splits[j]}
+				create_vm ${vms_to_run[j]}
+			done
+			echo " "
 		done
 		bdevs=("${split_bdevs[@]}")
 	else
-		notice "Using logical volumes"
+		notice "Configuring LVOLs..."
 		trap 'cleanup_lvol_cfg; error_exit "${FUNCNAME}" "${LINENO}"' INT ERR
-		for (( i=0; i<max_disks; i++ ));do
+		for (( i=0; i<${#disk_cfg_bdfs[@]}; i++ )); do
-			ls_guid=$($rpc_py bdev_lvol_create_lvstore Nvme${i}n1 lvs_$i --clear-method none)
+			nvme_bdev=$($rpc_py bdev_nvme_attach_controller -b ${disk_cfg_spdk_names[$i]} -t pcie -a ${disk_cfg_bdfs[$i]})
+			notice "Created NVMe Bdev: $nvme_bdev with BDF ${disk_cfg_bdfs[$i]}"
+
+			ls_guid=$($rpc_py bdev_lvol_create_lvstore $nvme_bdev lvs_$i --clear-method none)
 			lvol_stores+=("$ls_guid")
-			for (( j=0; j<${splits[$i]}; j++)); do
+			notice "Created Lvol Store: $ls_guid on Bdev $nvme_bdev"
+
+			vms_to_run=(${disk_cfg_vms[i]})
+			for (( j=0; j<${disk_cfg_splits[$i]}; j++)); do
 				free_mb=$(get_lvs_free_mb "$ls_guid")
-				size=$((free_mb / (${splits[$i]}-j) ))
+				size=$((free_mb / ((${disk_cfg_splits[$i]}-j)) ))
 				lb_name=$($rpc_py bdev_lvol_create -u $ls_guid lbd_$j $size --clear-method none)
 				lvol_bdevs+=("$lb_name")
-			done
+				notice "Created LVOL Bdev $lb_name on Lvol Store $ls_guid on Bdev $nvme_bdev"
-		done
-		bdevs=("${lvol_bdevs[@]}")
-	fi
-fi
 
-# Prepare VMs and controllers
+				notice "Setting up VM ${vms_to_run[j]}"
-for (( i=0; i<vm_count; i++)); do
+				create_spdk_controller "${vms_to_run[j]}" ${lb_name}
-	vm="vm_$i"
+				create_vm ${vms_to_run[j]}
-
-	setup_cmd="vm_setup --disk-type=$ctrl_type --force=$i --memory=$vm_memory"
-	setup_cmd+=" --os=$VM_IMAGE"
-
-	if [[ "$ctrl_type" == "spdk_vhost_scsi" ]]; then
-		$rpc_py vhost_create_scsi_controller naa.0.$i
-		$rpc_py vhost_scsi_controller_add_target naa.0.$i 0 ${bdevs[$i]}
-		setup_cmd+=" --disks=0"
-	elif [[ "$ctrl_type" == "spdk_vhost_blk" ]]; then
-		$rpc_py vhost_create_blk_controller naa.$i.$i ${bdevs[$i]}
-		setup_cmd+=" --disks=$i"
-	elif [[ "$ctrl_type" == "kernel_vhost" ]]; then
-		x=$(printf %03d $i)
-		setup_cmd+=" --disks=${wwpn_prefix}${x}"
-	fi
-	$setup_cmd
-	used_vms+=" $i"
 			done
+			echo " "
+		done
+		$rpc_py bdev_lvol_get_lvstores
+	fi
+	$rpc_py bdev_get_bdevs
+	$rpc_py vhost_get_controllers
+fi
 
 # Start VMs
 # Run VMs
@@ -383,7 +427,6 @@ for i in $(seq 1 $fio_iterations); do
 		done
 	fi
 
-
 	wait $fio_pid
 	mv $VHOST_DIR/fio_results/$fio_log_fname $VHOST_DIR/fio_results/$fio_log_fname.$i
 	sleep 1
@@ -401,7 +444,7 @@ else
 	else
 		cleanup_lvol_cfg
 	fi
-	vhost_kill 0
+	vhost_kill "${vhost_num}"
 fi
 
 if [[ -n "$kernel_cpus" ]]; then