path: root/net/test/run_net_test.sh

#!/bin/bash

# Builds mysteriously fail if stdout is non-blocking.
fixup_ptys() {
  python3 << 'EOF'
import fcntl, os, sys
fd = sys.stdout.fileno()
flags = fcntl.fcntl(fd, fcntl.F_GETFL)
flags &= ~(fcntl.FASYNC | os.O_NONBLOCK | os.O_APPEND)
fcntl.fcntl(fd, fcntl.F_SETFL, flags)
EOF
}

# Common kernel options
OPTIONS=" ANDROID GKI_NET_XFRM_HACKS"
OPTIONS="$OPTIONS DEBUG_SPINLOCK DEBUG_ATOMIC_SLEEP DEBUG_MUTEXES DEBUG_RT_MUTEXES"
OPTIONS="$OPTIONS WARN_ALL_UNSEEDED_RANDOM IKCONFIG IKCONFIG_PROC"
OPTIONS="$OPTIONS DEVTMPFS DEVTMPFS_MOUNT FHANDLE"
OPTIONS="$OPTIONS IPV6 IPV6_ROUTER_PREF IPV6_MULTIPLE_TABLES IPV6_ROUTE_INFO"
OPTIONS="$OPTIONS TUN SYN_COOKIES IP_ADVANCED_ROUTER IP_MULTIPLE_TABLES"
OPTIONS="$OPTIONS NETFILTER NETFILTER_ADVANCED NETFILTER_XTABLES"
OPTIONS="$OPTIONS NETFILTER_XT_MARK NETFILTER_XT_TARGET_MARK"
OPTIONS="$OPTIONS IP_NF_IPTABLES IP_NF_MANGLE IP_NF_FILTER"
OPTIONS="$OPTIONS IP6_NF_IPTABLES IP6_NF_MANGLE IP6_NF_FILTER INET6_IPCOMP"
OPTIONS="$OPTIONS IPV6_OPTIMISTIC_DAD"
OPTIONS="$OPTIONS IPV6_ROUTE_INFO IPV6_ROUTER_PREF"
OPTIONS="$OPTIONS NETFILTER_XT_TARGET_IDLETIMER"
OPTIONS="$OPTIONS NETFILTER_XT_TARGET_NFLOG"
OPTIONS="$OPTIONS NETFILTER_XT_MATCH_POLICY"
OPTIONS="$OPTIONS NETFILTER_XT_MATCH_QUOTA"
OPTIONS="$OPTIONS NETFILTER_XT_MATCH_QUOTA2"
OPTIONS="$OPTIONS NETFILTER_XT_MATCH_QUOTA2_LOG"
OPTIONS="$OPTIONS NETFILTER_XT_MATCH_SOCKET"
OPTIONS="$OPTIONS NETFILTER_XT_MATCH_QTAGUID"
OPTIONS="$OPTIONS INET_DIAG INET_UDP_DIAG INET_DIAG_DESTROY"
OPTIONS="$OPTIONS IP_SCTP"
OPTIONS="$OPTIONS IP_NF_TARGET_REJECT IP_NF_TARGET_REJECT_SKERR"
OPTIONS="$OPTIONS IP6_NF_TARGET_REJECT IP6_NF_TARGET_REJECT_SKERR"
OPTIONS="$OPTIONS NET_KEY XFRM_USER XFRM_STATISTICS CRYPTO_CBC"
OPTIONS="$OPTIONS CRYPTO_CTR CRYPTO_HMAC CRYPTO_AES CRYPTO_SHA1"
OPTIONS="$OPTIONS CRYPTO_XCBC CRYPTO_CHACHA20POLY1305"
OPTIONS="$OPTIONS CRYPTO_USER INET_ESP INET_XFRM_MODE_TRANSPORT"
OPTIONS="$OPTIONS INET_XFRM_MODE_TUNNEL INET6_ESP"
OPTIONS="$OPTIONS INET6_XFRM_MODE_TRANSPORT INET6_XFRM_MODE_TUNNEL"
OPTIONS="$OPTIONS CRYPTO_SHA256 CRYPTO_SHA512 CRYPTO_AES_X86_64 CRYPTO_NULL"
OPTIONS="$OPTIONS CRYPTO_GCM CRYPTO_ECHAINIV NET_IPVTI"
OPTIONS="$OPTIONS DUMMY"

# Kernel version specific options
OPTIONS="$OPTIONS XFRM_INTERFACE"                # Various device kernels
OPTIONS="$OPTIONS XFRM_MIGRATE"                  # Added in 5.10
OPTIONS="$OPTIONS CGROUP_BPF"                    # Added in android-4.9
OPTIONS="$OPTIONS NF_SOCKET_IPV4 NF_SOCKET_IPV6" # Added in 4.9
OPTIONS="$OPTIONS INET_SCTP_DIAG"                # Added in 4.7
OPTIONS="$OPTIONS SOCK_CGROUP_DATA"              # Added in 4.5
OPTIONS="$OPTIONS CRYPTO_ECHAINIV"               # Added in 4.1
OPTIONS="$OPTIONS BPF_SYSCALL"                   # Added in 3.18
OPTIONS="$OPTIONS IPV6_VTI"                      # Added in 3.13
OPTIONS="$OPTIONS IPV6_PRIVACY"                  # Removed in 3.12
OPTIONS="$OPTIONS NETFILTER_TPROXY"              # Removed in 3.11

# UML specific options
OPTIONS="$OPTIONS BLK_DEV_UBD HOSTFS"

# QEMU specific options
OPTIONS="$OPTIONS PCI VIRTIO VIRTIO_PCI VIRTIO_BLK NET_9P NET_9P_VIRTIO 9P_FS"
OPTIONS="$OPTIONS CRYPTO_DEV_VIRTIO SERIAL_8250 SERIAL_8250_PCI"
OPTIONS="$OPTIONS SERIAL_8250_CONSOLE PCI_HOST_GENERIC SERIAL_AMBA_PL011"
OPTIONS="$OPTIONS SERIAL_AMBA_PL011_CONSOLE"

# Obsolete options present at some time in Android kernels
OPTIONS="$OPTIONS IP_NF_TARGET_REJECT_SKERR IP6_NF_TARGET_REJECT_SKERR"

# b/262323440 - UML *sometimes* seems to have issues with:
#   UPSTREAM: hardening: Clarify Kconfig text for auto-var-init
# which is in 4.14.~299/4.19.~266 LTS and which does:
#   prompt "Initialize kernel stack variables at function entry"
#   default GCC_PLUGIN_STRUCTLEAK_BYREF_ALL if COMPILE_TEST && GCC_PLUGINS
#   default INIT_STACK_ALL_PATTERN if COMPILE_TEST && CC_HAS_AUTO_VAR_INIT_PATTERN
# + default INIT_STACK_ALL_ZERO if CC_HAS_AUTO_VAR_INIT_PATTERN
#   default INIT_STACK_NONE
# and thus presumably switches from INIT_STACK_NONE to INIT_STACK_ALL_ZERO
#
# My guess it that this is triggering some sort of UML and/or compiler bug...
# Let's just turn it off... we don't care that much.
OPTIONS="$OPTIONS INIT_STACK_NONE"

# These two break the flo kernel due to differences in -Werror on recent GCC.
DISABLE_OPTIONS=" REISERFS_FS ANDROID_PMEM"

# How many TAP interfaces to create to provide the VM with real network access
# via the host. This requires privileges (e.g., root access) on the host.
#
# This is not needed to run the tests, but can be used, for example, to allow
# the VM to update system packages, or to write tests that need access to a
# real network. The VM does not set up networking by default, but it contains a
# DHCP client and has the ability to use IPv6 autoconfiguration. This script
# does not perform any host-level setup beyond configuring tap interfaces;
# configuring IPv4 NAT and/or IPv6 router advertisements or ND proxying must
# be done separately.
NUMTAPINTERFACES=0

# The root filesystem disk image we'll use.
ROOTFS=${ROOTFS:-net_test.rootfs.20221014}
COMPRESSED_ROOTFS=$ROOTFS.xz
URL=https://dl.google.com/dl/android/$COMPRESSED_ROOTFS

# Parse arguments and figure out which test to run.
ARCH=${ARCH:-um}
J=${J:-$(nproc)}
MAKE="make"
OUT_DIR=$(readlink -f ${OUT_DIR:-.})
KERNEL_DIR=$(readlink -f ${KERNEL_DIR:-.})
if [ "$OUT_DIR" != "$KERNEL_DIR" ]; then
    MAKE="$MAKE O=$OUT_DIR"
fi
SCRIPT_DIR=$(dirname $(readlink -f $0))
CONFIG_SCRIPT=${KERNEL_DIR}/scripts/config
CONFIG_FILE=${OUT_DIR}/.config
consolemode=
netconfig=
testmode=
cmdline=
nowrite=1
nobuild=0
norun=0

KVER_MAJOR="$(sed -rn 's@^ *VERSION *= *([0-9]+)$@\1@p'    < "${KERNEL_DIR}/Makefile")"
KVER_MINOR="$(sed -rn 's@^ *PATCHLEVEL *= *([0-9]+)$@\1@p' < "${KERNEL_DIR}/Makefile")"
KVER_LEVEL="$(sed -rn 's@^ *SUBLEVEL *= *([0-9]+)$@\1@p'   < "${KERNEL_DIR}/Makefile")"
KVER="${KVER_MAJOR}.${KVER_MINOR}.${KVER_LEVEL}"
echo "Detected kernel version ${KVER}"

if [[ -z "${DEFCONFIG:-}" ]]; then
  case "${ARCH}" in
    um)
      export DEFCONFIG=defconfig
      ;;
    arm64)
      if [[ -e arch/arm64/configs/gki_defconfig ]]; then
        export DEFCONFIG=gki_defconfig
      elif [[ -e arch/arm64/configs/cuttlefish_defconfig ]]; then
        export DEFCONFIG=cuttlefish_defconfig
      fi
      ;;
    x86_64)
      if [[ -e arch/x86/configs/gki_defconfig ]]; then
        export DEFCONFIG=gki_defconfig
      elif [[ -e arch/x86/configs/x86_64_cuttlefish_defconfig ]]; then
        export DEFCONFIG=x86_64_cuttlefish_defconfig
      fi
  esac
fi

if tty >/dev/null; then
  verbose=
else
  verbose=1
fi

test=all_tests.sh
while [[ -n "$1" ]]; do
  if [[ "$1" == "--builder" || "$1" == "-b" ]]; then
    consolemode="con=null,fd:1"
    testmode=builder
    shift
  elif [[ "$1" == "--readwrite" || "$1" == "--rw" ]]; then
    nowrite=0
    shift
  elif [[ "$1" == "--readonly" ||  "$1" == "--ro" ]]; then
    nowrite=1
    shift
  elif [[ "$1" == "--nobuild" ]]; then
    nobuild=1
    shift
  elif [[ "$1" == "--norun" ]]; then
    norun=1
    shift
  elif [[ "$1" == "--verbose" ]]; then
    verbose=1
    shift
  elif [[ "$1" == "--noverbose" ]]; then
    verbose=
    shift
  else
    test=$1
    break  # Arguments after the test file are passed to the test itself.
  fi
done

# Check that test file exists and is readable
test_file=$SCRIPT_DIR/$test
if [[ ! -e $test_file ]]; then
  echo "test file '${test_file}' does not exist"
  exit 1
fi

if [[ ! -x $test_file ]]; then
  echo "test file '${test_file}' is not executable"
  exit 1
fi

# Collect trailing arguments to pass to $test
test_args=${@:2}

function isRunningTest() {
  ! (( norun ))
}

function isBuildOnly() {
  (( norun )) && ! (( nobuild ))
}

if ! isRunningTest && ! isBuildOnly; then
  echo "Usage:" >&2
  echo "  $0 [--builder] [--readonly|--ro|--readwrite|--rw] [--nobuild] [--verbose] [<test>]" >&2
  echo "      - if [<test>] is not specified, run all_tests.sh" >&2
  echo "  $0 --norun" >&2
  exit 1
fi

cd $OUT_DIR
echo Running tests from: `pwd`

set -e

# Check if we need to uncompress the disk image.
# We use xz because it compresses better: to 42M vs 72M (gzip) / 62M (bzip2).
cd $SCRIPT_DIR
if [ ! -f $ROOTFS ]; then
  echo "Deleting $COMPRESSED_ROOTFS" >&2
  rm -f $COMPRESSED_ROOTFS
  echo "Downloading $URL" >&2
  wget -nv $URL
  echo "Uncompressing $COMPRESSED_ROOTFS" >&2
  unxz $COMPRESSED_ROOTFS
fi
if ! [[ "${ROOTFS}" =~ ^/ ]]; then
  ROOTFS="${SCRIPT_DIR}/${ROOTFS}"
fi
echo "Using $ROOTFS"
cd -

# If network access was requested, create NUMTAPINTERFACES tap interfaces on
# the host, and prepare UML command line params to use them. The interfaces are
# called <user>TAP0, <user>TAP1, on the host, and eth0, eth1, ..., in the VM.
if (( $NUMTAPINTERFACES > 0 )); then
  user=${USER:0:10}
  tapinterfaces=
  for id in $(seq 0 $(( NUMTAPINTERFACES - 1 )) ); do
    tap=${user}TAP$id
    tapinterfaces="$tapinterfaces $tap"
    mac=$(printf fe:fd:00:00:00:%02x $id)
    if [ "$ARCH" == "um" ]; then
      netconfig="$netconfig eth$id=tuntap,$tap,$mac"
    else
      netconfig="$netconfig -netdev tap,id=hostnet$id,ifname=$tap,script=no,downscript=no"
      netconfig="$netconfig -device virtio-net-pci,netdev=hostnet$id,id=net$id,mac=$mac"
    fi
  done

  for tap in $tapinterfaces; do
    if ! ip link list $tap > /dev/null; then
      echo "Creating tap interface $tap" >&2
      sudo tunctl -u $USER -t $tap
      sudo ip link set $tap up
    fi
  done
fi

if [[ -n "${KERNEL_BINARY:-}" ]]; then
  nobuild=1
else
  # Set default KERNEL_BINARY location if it was not provided.
  if [ "$ARCH" == "um" ]; then
    KERNEL_BINARY=./linux
  elif [ "$ARCH" == "i386" -o "$ARCH" == "x86_64" -o "$ARCH" == "x86" ]; then
    KERNEL_BINARY=./arch/x86/boot/bzImage
  elif [ "$ARCH" == "arm64" ]; then
    KERNEL_BINARY=./arch/arm64/boot/Image.gz
  fi
fi

if ((nobuild == 0)); then
  make_flags=
  if [ "$ARCH" == "um" ]; then
    # Exporting ARCH=um SUBARCH=x86_64 doesn't seem to work, as it
    # "sometimes" (?) results in a 32-bit kernel.
    make_flags="$make_flags ARCH=$ARCH SUBARCH=${SUBARCH:-x86_64} CROSS_COMPILE= "
  fi
  if [[ -n "${CC:-}" ]]; then
    # The CC flag is *not* inherited from the environment, so it must be
    # passed in on the command line.
    make_flags="$make_flags CC=$CC"
  fi

  # If there's no kernel config at all, create one or UML won't work.
  [ -f $CONFIG_FILE ] || (cd $KERNEL_DIR && $MAKE $make_flags $DEFCONFIG)

  # Enable the kernel config options listed in $OPTIONS.
  $CONFIG_SCRIPT --file $CONFIG_FILE ${OPTIONS// / -e }

  # Increase acceptable frame size.
  $CONFIG_SCRIPT --file $CONFIG_FILE --set-val FRAME_WARN 3172

  # Disable the kernel config options listed in $DISABLE_OPTIONS.
  $CONFIG_SCRIPT --file $CONFIG_FILE ${DISABLE_OPTIONS// / -d }

  echo "Running: $MAKE $make_flags olddefconfig"
  $MAKE $make_flags olddefconfig

  # Compile the kernel.
  if [ "$ARCH" == "um" ]; then
    echo "Running: $MAKE -j$J $make_flags linux"
    $MAKE -j$J $make_flags linux
  else
    echo "Running: $MAKE -j$J $make_flags"
    $MAKE -j$J $make_flags
  fi
fi

if (( norun == 1 )); then
  exit 0
fi

if (( nowrite == 1 )); then
  cmdline="ro"
fi

if (( verbose == 1 )); then
  cmdline="$cmdline verbose=1"
fi

cmdline="$cmdline panic=1 init=/sbin/net_test.sh"
cmdline="$cmdline net_test_args=\"$test_args\" net_test_mode=$testmode"

# Experience shows that we need at least 128 bits of entropy for the
# kernel's crng init to complete (before it fully initializes stuff behaves
# *weirdly* and there's plenty of kernel warnings and some tests even fail),
# hence net_test.sh needs at least 32 hex chars (which is the amount of hex
# in a single random UUID) provided to it on the kernel cmdline.
#
# Just to be safe, we'll pass in 384 bits, and we'll do this as a random
# 64 character base64 seed (because this is shorter than base16).
# We do this by getting *three* random UUIDs and concatenating their hex
# digits into an *even* length hex encoded string, which we then convert
# into base64.
entropy="$(cat /proc/sys/kernel/random{/,/,/}uuid | tr -d '\n-')"
entropy="$(xxd -r -p <<< "${entropy}" | base64 -w 0)"
cmdline="${cmdline} random.trust_cpu=on entropy=${entropy}"

if [ "$ARCH" == "um" ]; then
  # Get the absolute path to the test file that's being run.
  cmdline="$cmdline net_test=/host$SCRIPT_DIR/$test"

  # We'd use UML's /proc/exitcode feature to communicate errors on test failure,
  # if not for UML having a tendency to crash during shutdown,
  # so instead use an extra serial line we'll redirect to an open fd...
  cmdline="$cmdline exitcode=/dev/ttyS3"

  # Map the --readonly flag to UML block device names
  if ((nowrite == 0)); then
    blockdevice=ubda
  else
    blockdevice=ubdar
  fi

  # Create a temp file for 'serial line 3' for return code.
  SSL3="$(mktemp)"

  exitcode=0
  $KERNEL_BINARY >&2 3>"${SSL3}" umid=net_test mem=512M \
    $blockdevice=$ROOTFS $netconfig $consolemode ssl3=null,fd:3 $cmdline \
  || exitcode=$?

  # Return to beginning of line (via carriage return) after the above newline moved us down.
  echo -en '\r'
  # re-enable: 'postprocess output' and 'translate newline to carriage return-newline'
  stty opost onlcr || :

  if [[ "${exitcode}" == 134 && -s "${SSL3}" && "$(tr -d '\r' < "${SSL3}")" == 0 ]]; then
    # Sometimes the tests all pass, but UML crashes during the shutdown process itself.
    # As such we can't actually rely on the /proc/exitcode returned value.
    echo "Warning: UML appears to have crashed after successfully executing the tests." 1>&2
  elif [[ "${exitcode}" != 0 ]]; then
    echo "Warning: UML exited with ${exitcode} instead of zero." 1>&2
  fi

  if [[ -s "${SSL3}" ]]; then
    exitcode="$(tr -d '\r' < "${SSL3}")"
    echo "Info: retrieved exit code ${exitcode}." 1>&2
  fi

  rm -f "${SSL3}"
  unset SSL3

  # UML is kind of crazy in how guest syscalls work.  It requires host kernel
  # to not be in vsyscall=none mode.
  if [[ "${exitcode}" != '0' ]]; then
    {
      # Hopefully one of these exists
      cat /proc/config || :
      zcat /proc/config.gz || :
      cat "/boot/config-$(uname -r)" || :
      zcat "/boot/config-$(uname -r).gz" || :
    } 2>/dev/null \
    | egrep -q '^CONFIG_LEGACY_VSYSCALL_NONE=y' \
    && ! egrep -q '(^| )vsyscall=(native|emulate|xonly)( |$)' /proc/cmdline \
    && {
      echo -e "\r"
      echo -e "-----=====-----\r"
      echo -e "If above you saw a 'net_test.sh[1]: segfault at ...' followed by\r"
      echo -e "'Kernel panic - not syncing: Attempted to kill init!' then please\r"
      echo -e "set 'vsyscall=emulate' on *host* kernel command line.\r"
      echo -e "On Linux 5.2+ you can instead use the slightly safer 'vsyscall=xonly'.\r"
      echo -e "(for example via GRUB_CMDLINE_LINUX in /etc/default/grub)\r"
      echo -e "-----=====-----\r"
    }
  fi
else
  # We boot into the filesystem image directly in all cases
  cmdline="$cmdline root=/dev/vda"

  # The path is stripped by the 9p export; we don't need SCRIPT_DIR
  cmdline="$cmdline net_test=/host/$test"

  # Map the --readonly flag to a QEMU block device flag
  if ((nowrite > 0)); then
    blockdevice=",readonly=on"
  else
    blockdevice=
  fi
  blockdevice="-drive file=$ROOTFS,format=raw,if=none,id=drive-virtio-disk0$blockdevice"
  blockdevice="$blockdevice -device virtio-blk-pci,drive=drive-virtio-disk0"

  # Pass through our current console/screen size to inner shell session
  read rows cols < <(stty size 2>/dev/null)
  [[ -z "${rows}" ]] || cmdline="${cmdline} console_rows=${rows}"
  [[ -z "${cols}" ]] || cmdline="${cmdline} console_cols=${cols}"
  unset rows cols

  # QEMU has no way to modify its exitcode; simulate it with a serial port.
  #
  # Choose to do it this way over writing a file to /host, because QEMU will
  # initialize the 'exitcode' file for us, it avoids unnecessary writes to the
  # host filesystem (which is normally not written to) and it allows us to
  # communicate an exit code back in cases we do not have /host mounted.
  #
  if [ "$ARCH" == "i386" -o "$ARCH" == "x86_64" -o "$ARCH" == "x86" ]; then
    # Assume we have hardware-accelerated virtualization support for amd64
    qemu="qemu-system-x86_64 -machine pc,accel=kvm -cpu host"

    # We know 'ttyS0' will be our serial port on x86 from the hard-coded
    # '-serial mon:stdio' flag below
    cmdline="$cmdline console=ttyS0"

    # The assignment of 'ttyS1' here is magical; we know ttyS0 was used up
    # by '-serial mon:stdio', and so this second serial port will be 'ttyS1'
    cmdline="$cmdline exitcode=/dev/ttyS1"
  elif [ "$ARCH" == "arm64" ]; then
    # This uses a software model CPU, based on cortex-a57
    qemu="qemu-system-aarch64 -machine virt -cpu cortex-a57"

    # We know 'ttyAMA0' will be our serial port on arm64 from the hard-coded
    # '-serial mon:stdio' flag below
    cmdline="$cmdline console=ttyAMA0"

    # The kernel will print messages via a virtual ARM serial port (ttyAMA0),
    # but for command line consistency with x86, we put the exitcode serial
    # port on the PCI bus, and it will be the only one.
    cmdline="$cmdline exitcode=/dev/ttyS0"
  fi

  $qemu >&2 -name net_test -m 512 \
    -kernel $KERNEL_BINARY \
    -no-user-config -nodefaults -no-reboot \
    -display none -nographic -serial mon:stdio -parallel none \
    -smp 4,sockets=4,cores=1,threads=1 \
    -device virtio-rng-pci \
    -chardev file,id=exitcode,path=exitcode \
    -device pci-serial,chardev=exitcode \
    -fsdev local,security_model=mapped-xattr,id=fsdev0,fmode=0644,dmode=0755,path=$SCRIPT_DIR \
    -device virtio-9p-pci,id=fs0,fsdev=fsdev0,mount_tag=host \
    $blockdevice $netconfig -append "$cmdline"
  [[ -s exitcode ]] && exitcode=`cat exitcode | tr -d '\r'` || exitcode=1
  rm -f exitcode
fi

# UML reliably screws up the ptys, QEMU probably can as well...
fixup_ptys
stty sane || :
tput smam || :

echo "Returning exit code ${exitcode}." 1>&2
exit "${exitcode}"