ceph/src/spdk/scripts/setup.sh

   1 #!/usr/bin/env bash
   2
   3 set -e
   4
   5 rootdir=$(readlink -f $(dirname $0))/..
   6
   7 function linux_iter_pci {
   8         # Argument is the class code
   9         # TODO: More specifically match against only class codes in the grep
  10         # step.
  11         lspci -mm -n | grep $1 | tr -d '"' | awk -F " " '{print "0000:"$1}'
  12 }
  13
  14 function linux_bind_driver() {
  15         bdf="$1"
  16         driver_name="$2"
  17         old_driver_name="no driver"
  18         ven_dev_id=$(lspci -n -s $bdf | cut -d' ' -f3 | sed 's/:/ /')
  19
  20         if [ -e "/sys/bus/pci/devices/$bdf/driver" ]; then
  21                 old_driver_name=$(basename $(readlink /sys/bus/pci/devices/$bdf/driver))
  22
  23                 if [ "$driver_name" = "$old_driver_name" ]; then
  24                         return 0
  25                 fi
  26
  27                 echo "$ven_dev_id" > "/sys/bus/pci/devices/$bdf/driver/remove_id" 2> /dev/null || true
  28                 echo "$bdf" > "/sys/bus/pci/devices/$bdf/driver/unbind"
  29         fi
  30
  31         echo "$bdf ($ven_dev_id): $old_driver_name -> $driver_name"
  32
  33         echo "$ven_dev_id" > "/sys/bus/pci/drivers/$driver_name/new_id" 2> /dev/null || true
  34         echo "$bdf" > "/sys/bus/pci/drivers/$driver_name/bind" 2> /dev/null || true
  35
  36         iommu_group=$(basename $(readlink -f /sys/bus/pci/devices/$bdf/iommu_group))
  37         if [ -e "/dev/vfio/$iommu_group" ]; then
  38                 if [ "$username" != "" ]; then
  39                         chown "$username" "/dev/vfio/$iommu_group"
  40                 fi
  41         fi
  42 }
  43
  44 function linux_hugetlbfs_mount() {
  45         mount | grep '^hugetlbfs ' | awk '{ print $3 }'
  46 }
  47
  48 function configure_linux {
  49         driver_name=vfio-pci
  50         if [ -z "$(ls /sys/kernel/iommu_groups)" ]; then
  51                 # No IOMMU. Use uio.
  52                 driver_name=uio_pci_generic
  53         fi
  54
  55         # NVMe
  56         modprobe $driver_name || true
  57         for bdf in $(linux_iter_pci 0108); do
  58                 linux_bind_driver "$bdf" "$driver_name"
  59         done
  60
  61
  62         # IOAT
  63         TMP=`mktemp`
  64         #collect all the device_id info of ioat devices.
  65         grep "PCI_DEVICE_ID_INTEL_IOAT" $rootdir/include/spdk/pci_ids.h \
  66         | awk -F"x" '{print $2}' > $TMP
  67
  68         for dev_id in `cat $TMP`; do
  69                 # Abuse linux_iter_pci by giving it a device ID instead of a class code
  70                 for bdf in $(linux_iter_pci $dev_id); do
  71                         linux_bind_driver "$bdf" "$driver_name"
  72                 done
  73         done
  74         rm $TMP
  75
  76         echo "1" > "/sys/bus/pci/rescan"
  77
  78         hugetlbfs_mount=$(linux_hugetlbfs_mount)
  79
  80         if [ -z "$hugetlbfs_mount" ]; then
  81                 hugetlbfs_mount=/mnt/huge
  82                 echo "Mounting hugetlbfs at $hugetlbfs_mount"
  83                 mkdir -p "$hugetlbfs_mount"
  84                 mount -t hugetlbfs nodev "$hugetlbfs_mount"
  85         fi
  86         echo "$NRHUGE" > /proc/sys/vm/nr_hugepages
  87
  88         if [ "$driver_name" = "vfio-pci" ]; then
  89                 if [ "$username" != "" ]; then
  90                         chown "$username" "$hugetlbfs_mount"
  91                 fi
  92
  93                 MEMLOCK_AMNT=`ulimit -l`
  94                 if [ "$MEMLOCK_AMNT" != "unlimited" ] ; then
  95                         MEMLOCK_MB=$(( $MEMLOCK_AMNT / 1024 ))
  96                         echo ""
  97                         echo "Current user memlock limit: ${MEMLOCK_MB} MB"
  98                         echo ""
  99                         echo "This is the maximum amount of memory you will be"
 100                         echo "able to use with DPDK and VFIO if run as current user."
 101                         echo -n "To change this, please adjust limits.conf memlock "
 102                         echo "limit for current user."
 103
 104                         if [ $MEMLOCK_AMNT -lt 65536 ] ; then
 105                                 echo ""
 106                                 echo "## WARNING: memlock limit is less than 64MB"
 107                                 echo -n "## DPDK with VFIO may not be able to initialize "
 108                                 echo "if run as current user."
 109                         fi
 110                 fi
 111         fi
 112 }
 113
 114 function reset_linux {
 115         # NVMe
 116         modprobe nvme || true
 117         for bdf in $(linux_iter_pci 0108); do
 118                 linux_bind_driver "$bdf" nvme
 119         done
 120
 121
 122         # IOAT
 123         TMP=`mktemp`
 124         #collect all the device_id info of ioat devices.
 125         grep "PCI_DEVICE_ID_INTEL_IOAT" $rootdir/include/spdk/pci_ids.h \
 126         | awk -F"x" '{print $2}' > $TMP
 127
 128         modprobe ioatdma || true
 129         for dev_id in `cat $TMP`; do
 130                 # Abuse linux_iter_pci by giving it a device ID instead of a class code
 131                 for bdf in $(linux_iter_pci $dev_id); do
 132                         linux_bind_driver "$bdf" ioatdma
 133                 done
 134         done
 135         rm $TMP
 136
 137         echo "1" > "/sys/bus/pci/rescan"
 138
 139         hugetlbfs_mount=$(linux_hugetlbfs_mount)
 140         rm -f "$hugetlbfs_mount"/spdk*map_*
 141 }
 142
 143 function status_linux {
 144         echo "NVMe devices"
 145
 146         echo -e "BDF\t\tNuma Node\tDriver name\t\tDevice name"
 147         for bdf in $(linux_iter_pci 0108); do
 148                 driver=`grep DRIVER /sys/bus/pci/devices/$bdf/uevent |awk -F"=" '{print $2}'`
 149                 node=`cat /sys/bus/pci/devices/$bdf/numa_node`;
 150                 if [ "$driver" = "nvme" ]; then
 151                         name="\t"`ls /sys/bus/pci/devices/$bdf/nvme`;
 152                 else
 153                         name="-";
 154                 fi
 155                 echo -e "$bdf\t$node\t\t$driver\t\t$name";
 156         done
 157
 158         echo "I/OAT DMA"
 159
 160         #collect all the device_id info of ioat devices.
 161         TMP=`grep "PCI_DEVICE_ID_INTEL_IOAT" $rootdir/include/spdk/pci_ids.h \
 162         | awk -F"x" '{print $2}'`
 163         echo -e "BDF\t\tNuma Node\tDriver Name"
 164         for dev_id in $TMP; do
 165                 # Abuse linux_iter_pci by giving it a device ID instead of a class code
 166                 for bdf in $(linux_iter_pci $dev_id); do
 167                         driver=`grep DRIVER /sys/bus/pci/devices/$bdf/uevent |awk -F"=" '{print $2}'`
 168                         node=`cat /sys/bus/pci/devices/$bdf/numa_node`;
 169                         echo -e "$bdf\t$node\t\t$driver"
 170                 done
 171         done
 172 }
 173
 174 function configure_freebsd {
 175         TMP=`mktemp`
 176
 177         # NVMe
 178         GREP_STR="class=0x010802"
 179
 180         # IOAT
 181         grep "PCI_DEVICE_ID_INTEL_IOAT" $rootdir/include/spdk/pci_ids.h \
 182         | awk -F"x" '{print $2}' > $TMP
 183         for dev_id in `cat $TMP`; do
 184                 GREP_STR="${GREP_STR}\|chip=0x${dev_id}8086"
 185         done
 186
 187         AWK_PROG="{if (count > 0) printf \",\"; printf \"%s:%s:%s\",\$2,\$3,\$4; count++}"
 188         echo $AWK_PROG > $TMP
 189
 190         BDFS=`pciconf -l | grep "${GREP_STR}" | awk -F: -f $TMP`
 191
 192         kldunload nic_uio.ko || true
 193         kenv hw.nic_uio.bdfs=$BDFS
 194         kldload nic_uio.ko
 195         rm $TMP
 196
 197         kldunload contigmem.ko || true
 198         kenv hw.contigmem.num_buffers=$((NRHUGE * 2 / 256))
 199         kenv hw.contigmem.buffer_size=$((256 * 1024 * 1024))
 200         kldload contigmem.ko
 201 }
 202
 203 function reset_freebsd {
 204         kldunload contigmem.ko || true
 205         kldunload nic_uio.ko || true
 206 }
 207
 208 : ${NRHUGE:=1024}
 209
 210 username=$1
 211 mode=$2
 212
 213 if [ "$username" = "reset" -o "$username" = "config" -o "$username" = "status" ]; then
 214         mode="$username"
 215         username=""
 216 fi
 217
 218 if [ "$mode" == "" ]; then
 219         mode="config"
 220 fi
 221
 222 if [ "$username" = "" ]; then
 223         username="$SUDO_USER"
 224         if [ "$username" = "" ]; then
 225                 username=`logname 2>/dev/null` || true
 226         fi
 227 fi
 228
 229 if [ `uname` = Linux ]; then
 230         if [ "$mode" == "config" ]; then
 231                 configure_linux
 232         elif [ "$mode" == "reset" ]; then
 233                 reset_linux
 234         elif [ "$mode" == "status" ]; then
 235                 status_linux
 236         fi
 237 else
 238         if [ "$mode" == "config" ]; then
 239                 configure_freebsd
 240         elif [ "$mode" == "reset" ]; then
 241                 reset_freebsd
 242         fi
 243 fi