path: root/src/initrd/btrfs-create.sh

#!/bin/sh

. loop-layer.sh
. mdadm-dup.sh

losetup() { /sbin/losetup "$@"; }

luks_secret()
{
  local parms=$-; # this junk keeps set -x from being too annoying
  set +x
  [ -n "$luks_secret" ] || luks_secret="$(head -c256 /dev/urandom)"
  printf %s "$luks_secret"
  case $parms in *x*) set -x; set -x ;; esac
}

floor4()
{
  # Negatives round up, but aren't used.
  echo $(($1 / 4 * 4))
}

ceil4()
{
  local x="$1"
  [ $((x % 4)) -eq 0 ] || x=$((x + 4 - x % 4))
  printf '%d\n' "$x"
}

kernel_commandline_has()
{
    local v="$1" c
    read c < /proc/cmdline
    for c in $c
    do
        case "$c" in
            "$v"|"$v"=*) true; return;;
        esac
    done
    false
}

netbooted()
{
    kernel_commandline_has BOOTIF
}

cdrom_has_rootfs()
{
    if netbooted
    then
        false
    else
        bootwait samizdat-cdrom
        [ -d /cdrom/rootfs ]
    fi
}

losetup_layers()
{
    if cdrom_has_rootfs
    then
        # TODO: This is some kind of shortcut or short circuit to find these
        # files, that ought to be found through the grok-block system (i.e.,
        # event-driven rather than polling).
        local fs fs_rw
        for fs in /cdrom/rootfs/*.btrfs; do
            fs_rw=/"${fs##*/}".rw
            dd if=/dev/zero of="$fs_rw" bs=1M count=10
            losetup_snapshot "$fs" "$fs_rw" || return
        done
    elif [ -e /dev/disk/by-partlabel/samizdat-rootfs ]
    then
        local dev
        for dev in samizdat-rootfs
        do
            umount /dev/disk/by-partlabel/$dev
            dd if=/dev/zero of=/$dev.rw bs=1M count=10
            dm_snapshot /dev/disk/by-partlabel/$dev /$dev.rw
        done
    else
        bootwait samizdat-nbd-dev
        local dev
        for dev in nbd0 nbd1
        do
            dd if=/dev/zero of=/$dev.rw bs=1M count=10
            dm_snapshot /dev/$dev /$dev.rw
        done
    fi
}

init_samizdat()
{
  local blockdev="$1" imgfile="$2" uuid

  losetup_layers                                   || return
  modprobe btrfs                                   || return
  btrfs device scan                                || return

  uuid=$(choose_uuid)                              || return
  [ "$uuid" ]                                      || return

  mount -t btrfs UUID="$uuid" /root || return

  btrfs device add "$blockdev" /root               || return
  mount -o rw,remount /root                        || return
  samizdat_movemounts "$imgfile"                   || return

  initialize_root_filesystem                       || return
}

movemounts()
{
    # Move mounted filesystems to the future root filesystem
    local dev mp rest target
    while read dev mp rest; do
        case "$mp" in
            /root/*|/root|/|/proc|/dev|/dev/pts|/sys|/run) continue ;;
            *) target=/root/$mp ;;
        esac
        mkdir -p "$target"
        mount -n -o move "$mp" "$target"
    done </proc/mounts
}

samizdat_movemounts()
{
    local imgfile="$1" mountpoint

    if [ "$imgfile" ]; then
        mountpoint=$(mountpoint_of "$imgfile") || return
        mkdir /root/outerfs
        mount -o move "$mountpoint" /root/outerfs
    fi
    mkdir /root/cdrom
    mount -o move /cdrom /root/cdrom
    umount /plaintext
    movemounts
    true
}

mountpoint_of()
{
    local f="$1"
    while ! mountpoint -q "$f"; do
        f=$(dirname "$f")
        [ "$f" != '.' ] || return 1
    done
    printf '%s\n' "$f"
}

copy_execs()
{
    local d b
    d=$1
    shift

    for b in "$@"; do
        [ -x "/root/$d/$b" ] && continue
        [ -x "/root/usr/$d/$b" ] && continue
        [ "$(which "$b")" ] || continue
        cp "$(which "$b")" "/root/$d/$b"
    done
}

is_subvolume()
{
    btrfs subvolume show "$1" >/dev/null 2>&1
}

make_subvolume_idem()
{
    is_subvolume "$1" && return
    if [ -d "$1" ]; then
        rmdir "$1" || mv "$1" "$1"~
        btrfs subvolume create "$1"
        mv "$1"~/* "$1"~/.* "$1"/ 2>&1
        rmdir "$1"~
    else
        if [ -e "$1" ]; then
            rm -f "$1"
        fi
        btrfs subvolume create "$1"
    fi
}

initialize_root_filesystem()
{
    local uhome=/home/u
    local uhome_subs="${uhome} ${uhome}/.cache ${uhome}/.stack ${uhome}/.rustup
        ${uhome}/src/fsmgr/_build ${uhome}/Downloads ${uhome}/.mozilla"
    for d in /root /srv /var/cache/apt/archives /home $uhome_subs
    do
        make_subvolume_idem /root/${d}
    done
    chroot /root chown -R u:u ${uhome}

    mv /root/root/.gnupg /root/root/.gnupg~
    mv /gpg/gnupghome /root/root/.gnupg                 || return

    copy_execs sbin  mdadm dmsetup cryptsetup fsck.hfsplus
    copy_execs bin   btrfs rsync gpg gpg2 gpg-agent

    # Copy these over unconditionally, because they ought to remain in sync with
    # the initrd.
    cp /bin/mdadm-dup.sh /root/sbin/
    cp /bin/samizdat-eject.sh /root/sbin/

    sed -i -e 's/^root:x:/root::/' /root/etc/passwd
    cp /patchroot/* /root/root/

    true
}

# Get the uuid of the filesystem with the most devices,
# excluding filesystems that don't incorporate loop devices.
# This is used to choose the latest seed -- which should have
# the most layers.
choose_uuid()
{
    local seen_loop= seen_uuid= seen_devs=
    btrfs filesystem show |
        while read line; do
            case "$line" in
                Label*)
                    seen_uuid=${line##*uuid: }
                    seen_devs=
                    seen_loop=
                    ;;
                *Total\ devices*)
                    seen_devs=${line#*Total devices }
                    seen_devs=${seen_devs%% *}
                    ;;
                *path\ /dev/mapper/*)
                    seen_loop=t;;
            esac
            [ "$seen_loop" ] && echo "$seen_devs $seen_uuid"
        done |
        uniq | sort -nr | head -n1 | (read _ x; echo $x)
}

filesystem_incomplete()
{
    local n
    n=$(btrfs filesystem show "$1" | sed -ne 's/.*Total devices \([^ ]*\) .*/\1/p')
    [ "$n" != 1 ]
}

partition_new_hard_drive_DESTROYING_EVERYTHING()
{
    # TODO: get the actual size of the btrfs master and actually triple it. The empty
    # space can be left available to allow the local machine to boot a
    # locally-regenerated rootfs seed.
    sz=3
    u=GiB
    actual_sz=$sz$u
    tripled_sz=$((sz * 3))$u
    pct=100% # TODO: use 50%

    local target="$1" sz=6GiB
    # [ "$(parted -sm "$target" print | grep -c :)"  = 1 ] || return
    parted "$target" -sm                                                     \
           unit B                                                            \
           mklabel gpt                                                       \
           mkpart samizdat-grub-incomplete 32KiB 8MiB                        \
           set 1 bios_grub on                                                \
           mkpart samizdat-plaintext-incomplete btrfs 64MiB $actual_sz       \
           mkpart samizdat-luks-encrypted-incomplete $tripled_sz $pct        \
        &&
        udevadm settle
}

mark_partitions_as_complete()
{
    local dev="$1"
    # TODO: Verify existing names
    parted "$dev" -sm \
           name 1 samizdat-grub \
           name 2 samizdat-plaintext \
           name 3 samizdat-luks-encrypted
}

open_samizdat()
{
  local blockdev=/dev/mapper/samizdatcrypt fs

  if filesystem_incomplete "$blockdev"; then
     losetup_layers
  fi
  modprobe btrfs || return
  btrfs device scan || return
  mount -t btrfs "$blockdev" /root || return
  samizdat_movemounts "$imgfile"
  LoSetup -D
}

init_samizdat_lodev()
{
  local imgfile="$1" megs=$(ceil4 "$2") dev
  truncate -s ${megs}M "$imgfile" || return
  dev=$(losetup -f) && losetup "$dev" "$imgfile" || return
  echo "$dev"
}

open_samizdat_blockdev_from_loop()
{
  local imgfile="$1" keyfile="$2" dev
  dev=$(losetup -f) && losetup "$dev" "$imgfile" || return
  open_samizdat_blockdev "$dev" "$keyfile"
}

open_samizdat_blockdev()
{
  local dev="$1" keyfile="$2"
  local cryptname=samizdatcrypt decrypted_keyfile=/luks.secret

  gpg2 --verify "$keyfile" || return

  # TODO: we should be ensuring we can decrypt this secret key before even
  # offering the option to boot the encrypted filesystem

  # The first --decrypt merely strips the signature.  The option is
  # poorly named for that case.
  gpg2 --decrypt "$keyfile" | gpg2 --decrypt > "$decrypted_keyfile" || return

  cryptsetup --key-file "$decrypted_keyfile" luksOpen "$dev" "$cryptname" || return

  [ -b /dev/mapper/"$cryptname" ] || return

}

init_samizdat_blockdev()
{
  local dev="$1" keyfile="$2"
  local cryptname=samizdatcrypt

  [ ! -b /dev/mapper/"$cryptname" ] || return

  luks_secret >/dev/null
  luks_secret | gpg2 --default-recipient-self --encrypt --armor | gpg2 --clearsign --output "$keyfile" || return

  luks_secret | cryptsetup -v luksFormat "$dev" - || return
  cryptsetup luksDump "$dev" >&2
  luks_secret | cryptsetup --key-file - luksOpen "$dev" "$cryptname" || return

  [ -b /dev/mapper/"$cryptname" ] || return
}

majmin()
{
  local dev="$1" major minor
  eval $(stat -c 'major=%t minor=%T' "$dev") || return
  [ "$major" -a "$minor" ] || return
  printf '%d:%d\n' 0x$major 0x$minor
}

cryptdev_to_dev()
{
  local dev="$1" majmin
  majmin=$(majmin "$dev") || return
  set -- /sys/dev/block/$majmin/slaves/*
  [ $# = 1 ] || return

  cryptsetup status "$dev" |while read k v; do if [ "$k" = device: ]; then echo $v; break; fi; done
}

cryptdev_to_backing_file()
{
  local dev="$1" majmin result
  majmin="$(majmin "$dev")" || return
  set -- /sys/dev/block/$majmin/slaves/*
  [ $# = 1 ] || return
  read result < "$1"/loop/backing_file || return
  printf '%s\n' "$result"
}

lodev_to_file()
{
  local result majmin dev="$1"
  majmin="$(majmin "$dev")" || return
  read result < /sys/dev/block/$majmin/loop/backing_file || return
  printf '%s' "$result"
}

mountpoint_to_dev()
{
  local wantmp="$1" dev mp rest
  mountpoint -q "$wantmp" || return
  while read dev mp rest; do if [ "$mp" = "$wantmp" ]; then echo "$dev"; return; fi; done < /proc/mounts
  return 1
}

get_cdrom_sizelimit()
{
  # returns bytes
  local dev="$1" sectors
  sectors=$(blockdev --getsz "$dev") || return
  if dd count=2 if="$dev" bs=2048 skip=$((sectors/4 - 2)) of=/dev/null 2>/dev/null; then
    return
  else
    echo $(((sectors-8)*512))
  fi
}

init_gpg()
{
  export GNUPGHOME=/gpg/gnupghome
  mkdir -p "$GNUPGHOME"
  if [ -e /gnupghome.tar ]; then
      tar -C "$GNUPGHOME" -zxf /gnupghome.tar && bootdone samizdat-gpg
      return
  else
      bootwait samizdat-cdrom
      (umask 077; rsync --exclude '/luks-key*' --ignore-existing -rpP /cdrom/gnupghome/ "$GNUPGHOME")
      bootdone samizdat-gpg
  fi

  local LOG_DIR=/run/initramfs/samizdat/log
  if samizdat-password-agent > "$LOG_DIR"/samizdat-password-agent.log 2>&1; then
    clear
    true
  else
    echo 'samizdat-password-agent failed; continuing in hope of hope...'
    true # false
  fi
}

start_meter()
{
  local startmsg="$*"
  (exec >&4
  clear
  echo -n $startmsg
  set +x
  while sleep 2; do
    echo -n .
  done) &
  meterpid=$!
}

stop_meter()
{
  local endmsg="$*"
  kill $meterpid
  echo "  $endmsg" >&4
}