path: root/init

#!/bin/bash

# Init script for build2 Build OS.
#
# Loosely based on the one that comes in Debian initrd.img (since we are
# using its kernel image).
#
trap "exit 1" ERR
set -o errtrace # Trap in functions.

# Note: diagnostics goes to stdout.
#
function info () { echo "$*"; }
function error ()
{
  if [ "$#" -gt 0 ]; then
    info "$*";
  fi

  # The setsid voodoo (taken from Debian init's panic()) is to enable job
  # control.
  #
  info "type Ctrl-D to exit shell and reboot"
  setsid /bin/bash -c "exec /bin/bash -i <>/dev/tty1 1>&0 2>&1"
  reboot -f
}

# Some pre-systemd utilities (like reboot) come from klibc-utils.
#
export PATH=/sbin:/usr/sbin:/bin:/usr/bin:/usr/lib/klibc/bin/

# One would expect rootflags=size=Xg to work but it doesn't (perhaps init
# is expected to interpret it)?
#
mount -o remount,size=2G /

mkdir -p /sys /proc
mount -t sysfs -o nodev,noexec,nosuid sysfs /sys
mount -t proc -o nodev,noexec,nosuid proc /proc

info "init starting up..."

mount -t devtmpfs -o nosuid,mode=0755 udev /dev

# Prepare the /dev directory.
#
ln -s /proc/self/fd /dev/fd
ln -s /proc/self/fd/0 /dev/stdin
ln -s /proc/self/fd/1 /dev/stdout
ln -s /proc/self/fd/2 /dev/stderr

mkdir -p /dev/pts
mount -t devpts -o noexec,nosuid,gid=5,mode=0620 devpts /dev/pts || true

# If tmpfs size is unspecified, then the default is 50%. Note that for /tmp in
# particular we may (temporarily) utilize several gigabytes while building the
# toolchain.
#
mkdir -p /run
mount -t tmpfs -o nodev,noexec,nosuid,size=512M,mode=0755 tmpfs /run

mkdir -p /tmp
mount -t tmpfs -o nodev,nosuid,size=2G,mode=1777 tmpfs /tmp

# Start udev.
#
# Based on Debian initrd's init-{top,bottom}/udev
#
info "starting udev..."

if [ -w /sys/kernel/uevent_helper ]; then
  echo >/sys/kernel/uevent_helper
fi

SYSTEMD_LOG_LEVEL=info /lib/systemd/systemd-udevd --daemon --resolve-names=never

udevadm trigger --type=subsystems --action=add
udevadm trigger --type=devices --action=add
udevadm settle || true

# On 6-series kernels we seem to be executed a lot earlier (or a lot faster)
# with many devices (Ethernet, USB storage) not being discovered yet (and
# devices that require firmware generally taking a while). So let's wait a
# bit for things to settle down.
#
for s in 5 4 3 2 1; do
  info "waiting for devices ${s}s..."
  sleep 1
done

# Detect hardware sensors.
#
sensors-detect --auto

# Initialize KVM.
#
#if ! (/sbin/modprobe kvm_intel || /sbin/modprobe kvm_amd); then
#  error "no virtualization support available (is it disabled in BIOS?)"
#fi

# Parse the kernel command line. This is complicated by the fact that the
# values can be quoted, for example:
#
# foo='foo fox'
# bar="bar 'box'"
#
# Or (as rewritten by GRUB):
#
# 'foo=foo fox'
# "bar=bar 'box'"
#
# First we separete quoted variables and arguments with newlines (giving
# priority to assignments). Then we replace whitespaces with newline on
# lines that don't contain quotes. Finally, clean up by removing blank
# lines.
#
# Note: the same code as in buildos.
#
readarray -t cmdline < <(cat /proc/cmdline | \
  sed -r -e "s/([^ ]+=)?('[^']*'|\"[^\"]*\")/\n\1\2\n/g" | \
  sed -r -e "/['\"]/!s/ /\n/g" |
  sed -r -e '/^\s*$/d')

# Enter all buildos variables as bash variables.
#
info "command line:"
for v in "${cmdline[@]}"; do

  # Rewrite "x=y" as x="y" (as well as the single-quote variant).
  #
  v1="$(sed -n -re "s/^\"([^= ]+)=(.*)\"\$/\1=\"\2\"/p" <<<"$v")"
  if [ -n "$v1" ]; then
    v="$v1"
  else
    v1="$(sed -n -re "s/^'([^= ]+)=(.*)'\$/\1='\2'/p" <<<"$v")"
    if [ -n "$v1" ]; then
      v="$v1"
    fi
  fi

  var="$(sed -n -re 's/^buildos\.([^= ]+)=.*$/\1/p' <<<"$v")" # Extract name.

  if [ -n "$var" ]; then
    val="$(sed -re 's/^[^= ]+=(.*)$/\1/' <<<"$v")"            # Extract value.
    val="$(sed -re "s/^('(.*)'|\"(.*)\")\$/\2\3/" <<<"$val")" # Strip quoted.
    info "  $var=$val"

    # If the variable contains a dot, then it is a toolchain variable and we
    # don't care about those in init.
    #
    if [[ "$var" != *.* ]]; then
      declare "$var=$val"
    fi
  fi
done

# Figure out network configuration and generate the corresponding
# /etc/network/interfaces.
#
info "starting network..."

# We are using udev's predictable interface names. The two character prefixes
# based on the type of interface:
#
#  en -- ethernet
#  sl -- serial line IP (slip)
#  wl -- wlan
#  ww -- wwan
#
eth_all="$(cd /sys/class/net && ls -d en?*)"

if [ -z "$eth_all" ]; then
  info "no ethernet interfaces found among:"
  ip link show
  error
fi

eth=
eth_up=
for s in 1 2 4 8; do

  # Try to bring them all up and find the one that has carrier.
  #
  for i in $eth_all; do
    ip link set "$i" up || true
  done

  sleep "$s"

  for i in $eth_all; do
    if [ "$(cat "/sys/class/net/$i/carrier")" -eq "1" ]; then
      info "detected carrier on $i"
      eth_up+=" $i"
    fi
  done

  # Bring them all down.
  #
  for i in $eth_all; do
    ip link set "$i" down || true
  done

  # If we didn't find anything, try to wait for carrier longer.
  #
  if [ -z "$eth_up" ]; then
    continue
  fi

  # If we end up with several interfaces we simply unleash dhcp on all of
  # them and use the first that gets configured.
  #
  # Note also that it's possible the interface that we want is not yet ready
  # in which case we will try to wait for carrier a bit longer.
  #
  for i in $eth_up; do
    if dhclient -v "$i"; then
      eth="$i"
      break
    fi
  done

  if [ -n "$eth" ]; then
    break
  fi
done

if [ -z "$eth_up" ]; then
  info "no ethernet interfaces with carrier among:"
  ip link show
  error
fi

if [ -z "$eth" ]; then
  info "no ethernet interfaces with DHCP among:"
  ip link show
  error
fi

# Global and local MAC addresses (used below for br0 and br1, respectively).
# Derive the local address from the global by fixing the first octet to 02
# (locally-assigned).
#
gmac="$(cat "/sys/class/net/$eth/address")"
lmac="$(sed -re 's/..:(.+)/02:\1/g' <<<"$gmac")"

info "configured $eth ($gmac)"

# Machine id.
#
mid="$(sed -re 's/://g' <<<"$gmac")"

# Set the hostname.
#
hname="$(hostname)"
if [ "$hname" = "(none)" ]; then
  hname="build-$mid"
  hostname "$hname"
fi
echo "$hname" >/etc/hostname

info "hostname $hname"

# Stop DHCP client without releasing the lease and deconfigure the interface.
# The plan is to generate a bridge-based /etc/network/interfaces configuration
# based on what we have discovered and then let the systemd networking bringup
# to configure everything (at which point we will hopefully reuse the lease).
#
dhclient -x 2>/dev/null

# @@ Needs to be made configurable. Something like 172.23.0.0/16.
#
priv_network="172.23.0.0"
priv_netmask="255.255.0.0"
priv_netbase="$(sed -e 's/^\(.*\)\.0\.0$/\1/' <<<"$priv_network")"

# Note that if we don't assign the bridge MAC address, then it will keep
# changing every time an interface with a greater address (e.g., a tap)
# joins the bridge. Needless to say, constantly changing MAC will wreck
# all kinds of networking havoc.
#
cat <<EOF >/etc/network/interfaces
auto lo
iface lo inet loopback

# Public bridge.
#
auto br0
iface br0 inet dhcp
    bridge_ports   $eth
    bridge_stp     off
    bridge_maxwait 0
    bridge_fd      0
    bridge_hw      $gmac
    post-up ip link set $eth txqueuelen 4000
    post-up ip link set br0 txqueuelen 4000

# Private bridge with NAT to br0.
#
auto br1
iface br1 inet static
    address ${priv_netbase}.0.1
    netmask $priv_netmask
    bridge_ports   none
    bridge_stp     off
    bridge_maxwait 0
    bridge_fd      0
    bridge_hw      $lmac
    #post-up ip link set br1 address $lmac
    post-up ip link set br1 txqueuelen 4000
    post-up iptables -t nat -A POSTROUTING -o br0 -j MASQUERADE
    post-up iptables -A FORWARD -i br0 -o br1 -m state --state RELATED,ESTABLISHED -j ACCEPT
    post-up iptables -A FORWARD -i br1 -o br0 -j ACCEPT
EOF

cat <<EOF >/etc/dnsmasq.d/br1-dhcp
interface=br1
bind-interfaces
dhcp-range=${priv_netbase}.1.1,${priv_netbase}.255.255,$priv_netmask,2h
EOF

# Figure out disk configuration and generate the corresponding /etc/fstab.
#
fstab=/etc/fstab
#fstab=/dev/stdout

echo -n '' >$fstab

l=
state=
machines=
while read l || [ -n "$l" ]; do
  d="$(sed -re 's/.*NAME=\"([^\"]+)\".*/\1/'   <<<"$l")"
  t="$(sed -re 's/.*FSTYPE=\"([^\"]*)\".*/\1/' <<<"$l")"
  l="$(sed -re 's/.*LABEL=\"([^\"]*)\".*/\1/'  <<<"$l")"

  # Strip the buildos prefix from the label. If the result is empty then this
  # disk/patition hasn't been labeled for use by us.
  #
  l="$(sed -n -re 's/^buildos\.([^ ]+)$/\1/p' <<<"$l")"

  if [ -z "$l" ]; then
    continue
  fi

  # Handle buildos.state.
  #
  if [ "$l" == "state" ]; then

    if [ -n "$state" ]; then
      error "multiple disks labeled with buildos.state"
    fi

    if [ -z "$t" ]; then
      error "no filesystem on $d labeled with buildos.state"
    fi

    info "mounting $d (buildos.state) on /state as $t"

    # Check it.
    #
    if ! fsck -n -t "$t" "$d"; then
      info "$d (buildos.state) has errors; run fsck -t $type $d"
      error
    fi

    o="defaults,noatime,nodiratime"
    echo "$d /state $t $o 0 0" >>$fstab

    # Mount it now since we need it below.
    #
    mkdir -p "/state"
    mount -t "$t" -o "$o" "$d" /state

    state="true"
    continue
  fi

  # Handle buildos.machines and buildos.machines.* mounts.
  #
  if [[ "$l" == "machines" ]] || [[ "$l" =~ "machines.".+ ]]; then

    if [ "$t" != "btrfs" ]; then
      error "non-btrfs filesystem on $d labeled with buildos.machines"
    fi

    if [ "$l" = "machines" ]; then
      # Single mount.
      #
      if [ -n "$machines" ]; then
	error "multiple disks labeled with buildos.machines/machines.*"
      fi

      m=/build/machines/default
      machines="single"
    else
      # Multiple mounts.
      #
      if [ "$machines" = "single" ]; then
	error "multiple disks labeled with buildos.machines/machines.*"
      fi

      n="$(sed -n -re 's/^machines\.([^ ]+)$/\1/p' <<<"$l")"
      m="/build/machines/$n"
      machines="multiple"
    fi

    info "mounting $d (buildos.$l) on $m"

    # Check it.
    #
    if ! btrfs check -p "$d"; then
      info "$d (buildos.$l) has errors; run btrfs check -p --repair $d"
      error
    fi

    o="defaults,noatime,nodiratime,user_subvol_rm_allowed"
    echo "$d $m btrfs $o 0 0" >>$fstab

    # Mount it and change the owner of the filesystem root.
    #
    mkdir -p "$m"
    mount -t btrfs -o "$o" "$d" "$m"
    chown build:build "$m"

    continue
  fi
done < <(lsblk --pairs --paths --output NAME,FSTYPE,LABEL)
#done <<EOF
#NAME="/dev/sda" FSTYPE="btrfs" LABEL="buildos.machines.vol1"
#NAME="/dev/sdb" FSTYPE="btrfs" LABEL="buildos.machines.vol2"
#EOF

if [ -z "$state" ]; then
  info "no disks labaled with buildos.state among:"
  lsblk --paths --output NAME,TYPE,FSTYPE,SIZE,LABEL,UUID
  info "consider formatting and/or labelling a suitable disk"
  error
fi

if [ -z "$machines" ]; then
  info "no disks labaled with buildos.machines* among:"
  lsblk --paths --output NAME,TYPE,FSTYPE,SIZE,LABEL,UUID
  info "consider formatting and/or labelling a suitable disk"
  error
fi

# Create /build/tftp. We make it a size-limited tmpfs since potentially-
# compromized VMs will be able to upload to.
#
mkdir -p /build/tftp
o="nodev,noexec,nosuid,size=512M"
o+=",mode=0755,uid=$(id -u build),gid=$(id -g build)"
echo "tmpfs /build/tftp tmpfs $o 0 0" >>$fstab

# Generate a host key. Used, for example, by the bbot agent to authenticate
# with a controller.
#
mkdir -p /state/etc
if [ ! -e /state/etc/host-key.pem ]; then
  openssl genrsa 4096 >/state/etc/host-key.pem
fi

# Configure Postfix.
#
cat <<<"$hname" >/etc/mailname

sed -r -i \
 -e "s%^(myhostname).*%\1 = $hname%" \
 -e 's%^(mydestination).*%\1 = $myhostname, localhost.localdomain, localhost%' \
 -e 's%^(mynetworks).*%\1 = 127.0.0.0/8 [::ffff:127.0.0.0]/104 [::1]/128%' \
 -e "s%^(relayhost).*%\1 = $smtp_relay%" \
 /etc/postfix/main.cf

# Make admin alias for buildos.admin_email, alias root as admin.
#
cat <<EOF >>/etc/aliases
admin: $admin_email
root: admin
EOF
newaliases

# Configure OpenSSH server. Things that we do:
#
# - Change host key locations to (persistent) /state/etc/ssh/ and remove
#   existing keys. If no corresponding key exists in /state, generate it.
#
# - Disable root login.
#
# - Disable password authentication.
#
sed -r -i \
  -e "s%^#?HostKey +(.+)%HostKey /state\1%" \
  -e "s%^#?PermitRootLogin.*%PermitRootLogin no%" \
  -e "s%^#?PasswordAuthentication.*%PasswordAuthentication no%" \
 /etc/ssh/sshd_config

# Generate missing keys.
#
mkdir -p /state/etc/ssh
for k in $(echo /etc/ssh/ssh_host_*_key | \
	      sed -re 's%/etc/ssh/ssh_host_([^_]+)_key%\1%g'); do
  if [ ! -e "/state/etc/ssh/ssh_host_${k}_key" ]; then
    ssh-keygen -N "" -t "$k" -f "/state/etc/ssh/ssh_host_${k}_key"
  fi
done
rm -f /etc/ssh/ssh_host_*_key*

# Add buildos.ssh_key to build's authorized_keys.
#
if [ -n "$ssh_key" ]; then
  info "adding buildos.ssh_key to ~build/.ssh/authorized_keys"
  mkdir -p /build/.ssh
  echo "$ssh_key" >>/build/.ssh/authorized_keys

  chown build:build /build/.ssh /build/.ssh/authorized_keys
  chmod 700 /build/.ssh
  chmod 600 /build/.ssh/authorized_keys
fi

# Configure the TFTP server (tftpd-hpa).
#
# We could have ran it as user 'build' but since it is read-only, the default
# user 'tftp' works just as well. Note that by default it is running chroot'ed
# (--secure) so no symlinks pointing outside /build/tftp will work.
#
sed -r -i \
  -e "s%^(TFTP_DIRECTORY).*%\1=\"/build/tftp\"%" \
 /etc/default/tftpd-hpa

# Hand off to systemd. But first arrange to keep console output (which
# becomes tty1).
#
mkdir -p /etc/systemd/system/getty@tty1.service.d
cat <<EOF >/etc/systemd/system/getty@tty1.service.d/noclear.conf
[Service]
TTYVTDisallocate=no
EOF

# Stop udev daemon (systemd will start its own).
#
udevadm control --exit

# Get rid of klibc tools.
#
export PATH=/sbin:/usr/sbin:/bin:/usr/bin

exec /lib/systemd/systemd \
  --show-status=1 \
  --machine-id="00000000000000000000$mid" \
  </dev/console >/dev/console 2>&1