iankelling.org Git - automated-distro-installer/blob - fai/config/hooks/partition.DEFAULT

   1 #!/bin/bash -x
   2 # Copyright (C) 2016 Ian Kelling
   3
   4 # This program is free software; you can redistribute it and/or
   5 # modify it under the terms of the GNU General Public License
   6 # as published by the Free Software Foundation; either version 2
   7 # of the License, or (at your option) any later version.
   8
   9 # This program is distributed in the hope that it will be useful,
  10 # but WITHOUT ANY WARRANTY; without even the implied warranty of
  11 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12 # GNU General Public License for more details.
  13
  14 # You should have received a copy of the GNU General Public License
  15 # along with this program; if not, write to the Free Software
  16 # Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
  17
  18 set -eE -o pipefail
  19 trap 'echo "$0:$LINENO:error: \"$BASH_COMMAND\" returned $?" >&2' ERR
  20
  21 # # fai's setup-storage won't do btrfs on luks,
  22 # # so we do it ourself :)
  23 # inspiration taken from files in fai-setup-storage package
  24
  25
  26 skiptask partition || ! type skiptask # for running not in fai
  27
  28 #### begin configuration
  29
  30 rootn=1
  31 swapn=2
  32 bootn=3
  33 # ext partition so grub can write persistent variables,
  34 # so it can do a one time boot. grub can't write to
  35 # btrfs or any cow fs because it's more
  36 # more complicated to do and they don't want to.
  37 grub_extn=4
  38 # bios boot partition,
  39 # https://wiki.archlinux.org/index.php/GRUB
  40 bios_grubn=5
  41 lastn=$bios_grubn
  42 # this is larger than needed for several /boot subvols,
  43 # becuase I keep a minimal debian install on it, for
  44 # recovery needs, and for doing pxe-kexec.
  45 boot_mib=10000
  46
  47
  48 ##### end configuration
  49
  50
  51 add-part() { # add partition suffix to $dev
  52     local d ret
  53     if [[ $# == 1 ]]; then
  54         d=$dev
  55         part=$1
  56     else
  57         d=$1
  58         part=$2
  59     fi
  60     if [[ $d == /dev/disk/by-id/* ]]; then
  61         ret=$d-part$part
  62     else
  63         ret=$d$part
  64     fi
  65     echo $ret
  66 }
  67
  68 bootdev() { add-part $@ $bootn; }
  69 rootdev() { add-part $@ $rootn; }
  70 swapdev() { add-part $@ $swapn; }
  71 grub_extdev() { add-part $@ $grub_extn; }
  72 bios_grubdev() { add-part $@ $bios_grubn; }
  73
  74 crypt-dev() { echo /dev/mapper/crypt_dev_${1##*/}; }
  75 crypt-name() { echo crypt_dev_${1##*/}; }
  76 root-cryptdev() { crypt-dev $(rootdev $@); }
  77 swap-cryptdev() { crypt-dev $(swapdev $@); }
  78 root-cryptname() { crypt-name $(rootdev $@); }
  79 swap-cryptname() { crypt-name $(swapdev $@); }
  80 devbyid() {
  81     local f
  82     for f in $FAI/distro-install-common/devbyid \
  83                  /a/bin/fai/fai/config/distro-install-common/devbyid; do
  84         if [[ -e $f ]]; then $f "$@"; fi
  85     done
  86
  87 }
  88
  89 ##### end function defs
  90
  91 if ifclass REPARTITION;then
  92     partition=true # force a full wipe
  93 else
  94     partition=false # change to true to force a full wipe
  95 fi
  96
  97
  98
  99 hdds=()
 100 ssds=()
 101 cd /sys/block
 102 for disk in [sv]d[a-z]; do
 103     case $(cat $disk/queue/rotational) in
 104         0) ssds+=(/dev/$disk) ;;
 105         1) hdds+=(/dev/$disk) ;;
 106         *) echo "$0: error: unknown /sys/block/$disk/queue/rotational: \
 107 $(cat $disk/queue/rotational)"; exit 1 ;;
 108     esac
 109 done
 110
 111 # install all ssds, or if there are none, all hdds
 112 if ! ifclass ROTATIONAL && (( ${#ssds[@]} > 0 )); then
 113     short_devs=( ${ssds[@]} )
 114 else
 115     short_devs=( ${hdds[@]} )
 116 fi
 117
 118 # check if the partitions exist have the right filesystems
 119 #blkid="$(blkid -s TYPE)"
 120 for dev in ${short_devs[@]}; do
 121     if $partition; then break; fi
 122     y=$(readlink -f $dev)
 123     arr=($y[0-9])
 124     [[ ${#arr[@]}  == "${lastn}" ]] || partition=true
 125     for (( i=1; i <= lastn; i++ )); do
 126         [[ -e ${dev}$i ]] || partition=true
 127     done
 128     # On one system, blkid is missing some partitions.
 129     # maybe we need a flag, like FUZZY_BLKID or something, so we
 130     # can check that at least some exist.
 131     # for x in "`rootdev`: TYPE=\"crypto_LUKS\"" "`bootdev`: TYPE=\"btrfs\""; do
 132     #     echo "$blkid" | grep -Fx "$x" &>/dev/null || partition=true
 133     # done
 134 done
 135
 136 if $partition && ifclass PARTITION_PROMPT; then
 137     echo "Press any key except ctrl-c to continue and partition these drives:"
 138     echo "  ${short_devs[*]}"
 139     read -r
 140 fi
 141
 142 devs=()
 143 shopt -s extglob
 144 for short_dev in ${short_devs[@]}; do
 145     devs+=($(devbyid $short_dev))
 146 done
 147
 148
 149 first=false
 150 boot_devs=()
 151 for dev in ${devs[@]}; do
 152     if ifclass frodo; then
 153         # I ran into a machine where the bios doesn't know about some disks,
 154         # so 1st stage of grub also doesn't know about them.
 155         # Also, grub does not support mounting degraded btrfs as far as
 156         # I can tell with some googling.
 157         # From within an arch install env, I could detect them by noting
 158         # their partitions were mixed with the next disk in /dev/disk/by-path,
 159         # and I have mixed model disks, and I could see the 8 models which showed
 160         # up in the bios, and thus see which 2 models were missing.
 161         # hdparm -I /dev/sdh will give model info in linux.
 162         # However, in fai on jessie, /dev/disk/by-path dir doesn't exist,
 163         # and I don't see another way, so I'm hardcoding them.
 164         # We still put grub on them and partition them the same, for uniformity
 165         # and in case they get moved to a system that can recognize them,
 166         # we just exclude them from the boot filesystem.
 167         cd /dev/disk/by-id/
 168         bad_disk=false
 169         for id in ata-TOSHIBA_MD04ACA500_8539K4TQFS9A \
 170                       ata-TOSHIBA_MD04ACA500_Y5IFK6IJFS9A; do
 171             if [[ $(readlink -f $id) == "$(readlink -f $dev)" ]]; then
 172                 bad_disk=true
 173                 break
 174             fi
 175         done
 176         $bad_disk || boot_devs+=(`bootdev`)
 177     else
 178         boot_devs+=(`bootdev`)
 179     fi
 180     if [[ $boot_devs && $first ]]; then
 181         first_grub_extdev=`grub_extdev`
 182         first=false
 183     fi
 184 done
 185
 186 if ifclass RAID0 || (( ${#boot_devs[@]} < 4 )); then
 187     raid_level=0
 188 else
 189     raid_level=10
 190     # need double the space if we are raid 10, and then
 191     # might as well give some extra.
 192     boot_mib=$((boot_mib * 3))
 193 fi
 194
 195
 196
 197 if [[ ! $DISTRO ]]; then
 198     if ifclass VOL_STABLE_BOOTSTRAP; then
 199         DISTRO=debianstable_bootstrap
 200     elif ifclass VOL_STRETCH; then
 201         DISTRO=debiantesting
 202     elif ifclass VOL_STABLE; then
 203         DISTRO=debianstable
 204     elif ifclass VOL_XENIAL; then
 205         DISTRO=ubuntuxenial
 206     elif ifclass VOL_BELENOS; then
 207         DISTRO=trisquelbelenos
 208     elif ifclass VOL_FLIDAS; then
 209         DISTRO=trisquelflidas
 210     else
 211         echo "PARTITIONER ERROR: no distro class/var set" >&2
 212         exit 1
 213     fi
 214 fi
 215 first_boot_dev=${boot_devs[0]}
 216
 217
 218 bpart() { # btrfs a partition
 219     case $raid_level in
 220         0) mkfs.btrfs -f $@ ;;
 221         10) mkfs.btrfs -f -m raid10 -d raid10 $@ ;;
 222     esac
 223 }
 224
 225
 226 # keyfiles generated like:
 227 # head -c 2048 /dev/urandom | od | s dd of=/q/root/luks/host-demohost
 228 luks_dir=${LUKS_DIR:-/var/lib/fai/config/distro-install-common/luks}
 229
 230 if [[ ! -e $luks_dir/host-$HOSTNAME ]]; then
 231     echo "$0: error: no key for hostname at $luks_dir/host-$HOSTNAME" >&2
 232     exit 1
 233 fi
 234
 235 if ifclass tp; then
 236     lukspw=$(cat $luks_dir/traci)
 237 else
 238     lukspw=$(cat $luks_dir/iank)
 239 fi
 240 if ifclass demohost; then
 241     lukspw=x
 242 fi
 243
 244
 245 first_root_crypt=$(root-cryptdev ${devs[0]})
 246
 247 # 1.5 x based on https://access.redhat.com/documentation/en-US/Red_Hat_Enterprise_Linux/7/html/Installation_Guide/sect-disk-partitioning-setup-x86.html#sect-custom-partitioning-x86
 248 swap_mib=$(( $(grep ^MemTotal: /proc/meminfo | \
 249                    awk '{print $2}') * 3/(${#devs[@]} * 2 ) / 1024 ))
 250
 251 mkdir -p /tmp/fai
 252 root_devs=()
 253 for dev in ${devs[@]}; do
 254     root_devs+=(`rootdev`)
 255 done
 256 shopt -s nullglob
 257 if $partition; then
 258     for dev in ${devs[@]}; do
 259         # if we repartition to the same as an old partition,
 260         # we don't want any old fses hanging around.
 261         for (( i=1; i <= lastn; i++ )); do
 262             x=$(add-part $dev $i)
 263             [[ -e $x ]] || continue
 264             count_down=10
 265             # wipefs has failed, manual run works, google suggests timing issue
 266             while ! wipefs -a $x; do
 267                 sleep 2
 268                 count_down=$((count_down - 1))
 269                 (( count_down > 0 )) || exit 1
 270             done
 271         done
 272     done
 273     for dev in ${devs[@]}; do
 274         # parted will round up the disk size. Do -1 so we can have
 275         # fully 1MiB unit partitions for easy resizing of the last partition.
 276         # Otherwise we would pass in -0 for the end argument for the last partition.
 277         #
 278         # parted print error output is expected. example:
 279         # Error: /dev/vda: unrecognised disk label
 280         disk_mib=$(( $(parted -m $dev unit MiB print | \
 281                            sed -nr "s#^/dev/[^:]+:([0-9]+).*#\1#p") - 1))
 282         root_end=$(( disk_mib - swap_mib - boot_mib /  ${#boot_devs[@]} ))
 283         swap_end=$(( root_end + swap_mib))
 284
 285         parted -s $dev mklabel gpt
 286         # MiB because parted complains about alignment otherwise.
 287         pcmd="parted -a optimal -s -- $dev"
 288         $pcmd mkpart primary ext3 12MiB ${root_end}MiB
 289         # without naming, systemd gives us misc errors like:
 290         # dev-disk-by\x2dpartlabel-primary.device: Dev dev-disk-by\x2dpartlabel-primary.device appeared twice
 291         $pcmd name $rootn root
 292         # normally a swap is type "linux-swap", but this is encrypted swap. using that
 293         # label will confuse systemd.
 294         $pcmd mkpart primary "" ${root_end}MiB ${swap_end}MiB
 295         $pcmd name $swapn swap
 296         $pcmd mkpart primary "" ${swap_end}MiB ${disk_mib}MiB
 297         $pcmd name $bootn boot
 298         # i only need a few k, but googling min size,
 299         # I found someone saying that gparted required
 300         # required at least 8 because of their hard drive cylinder size.
 301         # And 8 is still very tiny.
 302         $pcmd mkpart primary "ext2" 4MiB 12MiB
 303         $pcmd name $grub_extn grubext
 304         # gpt ubuntu cloud image uses ~4 mb for this partition. fai uses 1 MiB.
 305         # so, I use 3, whatever.
 306         # note: parted manual saying cheap flash media
 307         # should to start at 4.
 308         $pcmd mkpart primary "" 1MiB 4MiB
 309         $pcmd name $bios_grubn biosgrub
 310         $pcmd set $bios_grubn bios_grub on
 311         $pcmd set $bootn boot on # generally not needed on modern systems
 312         # the mkfs failed before on a vm, which prompted me to add
 313         # sleep .1
 314         # then it failed again on a physical machine
 315         # with:
 316         # Device /dev/disk/by-id/foo doesn't exist or access denied,
 317         # so I added a wait until it existed.
 318         # Then I added the mkfs.ext2, which claimed to succeed,
 319         # but then couldn't be found upon reboot. In that case we didn't
 320         # wait at all. So I've added a 3 second minimum wait.
 321         sleep 3
 322         secs=0
 323         while [[ ! -e `rootdev` ]] && (( secs < 10 )); do
 324             sleep 1
 325             secs=$((secs +1))
 326         done
 327         # Holds just a single file, rarely written, so
 328         # use ext2, like was often used for the /boot partition.
 329         # This exists because grub can only persist data to a non-cow fs.
 330         # And we use persisting a var in grub to do a one time boot.
 331         # We could pass the data on the kernel command line and persist it
 332         # to grubenv after booting, but that relies on the boot always succeeding.
 333         # This is just a bit more robust, and it could work for booting
 334         # into ipxe which can't persist data, if we ever got that working.
 335         mkfs.ext2 `grub_extdev`
 336         yes YES | cryptsetup luksFormat `rootdev` $luks_dir/host-$HOSTNAME \
 337                              -c aes-cbc-essiv:sha256 -s 256 || [[ $? == 141 ]]
 338         yes "$lukspw" | \
 339             cryptsetup luksAddKey --key-file $luks_dir/host-$HOSTNAME \
 340                        `rootdev` || [[ $? == 141 ]]
 341         # background: Keyfile and password are treated just
 342         # like 2 ways to input a passphrase, so we don't actually need to have
 343         # different contents of keyfile and passphrase, but it makes some
 344         # security sense to a really big randomly generated passphrase
 345         # as much as possible, so we have both.
 346         #
 347         # This would remove the keyfile.
 348         #    yes 'test' | cryptsetup luksRemoveKey /dev/... \
 349             #                            /key/file || [[ $? == 141 ]]
 350
 351         cryptsetup luksOpen `rootdev` `root-cryptname` \
 352                    --key-file $luks_dir/host-$HOSTNAME
 353     done
 354     ls -la /dev/btrfs-control # this was probably for debugging...
 355     sleep 1
 356     bpart $(for dev in ${devs[@]}; do root-cryptdev; done)
 357     bpart ${boot_devs[@]}
 358 else
 359     for dev in ${devs[@]}; do
 360         mkfs.ext2 `grub_extdev`
 361         cryptsetup luksOpen `rootdev` `root-cryptname` \
 362                    --key-file $luks_dir/host-$HOSTNAME
 363     done
 364     sleep 1
 365 fi
 366
 367
 368 if [[ $DISTRO != debianstable_bootstrap ]]; then
 369     # bootstrap distro doesn't use separate encrypted root.
 370     mount -o subvolid=0 $first_root_crypt /mnt
 371     # systemd creates subvolumes we want to delete.
 372     s=($(btrfs subvolume list --sort=-path /mnt |
 373              sed -rn "s#^.*path\s*(root_$DISTRO/\S+)\s*\$#\1#p"))
 374     for subvol in ${s[@]}; do btrfs subvolume delete /mnt/$subvol; done
 375     btrfs subvolume set-default 0 /mnt
 376     [[ ! -e /mnt/root_$DISTRO ]] || btrfs subvolume delete /mnt/root_$DISTRO
 377
 378     ## create subvols ##
 379     cd /mnt
 380
 381     btrfs subvolume create root_$DISTRO
 382
 383     mkdir -p /mnt/root_$DISTRO/boot
 384     # could set default subvol like this, but no reason to.
 385     # btrfs subvolume set-default \
 386         #       $(btrfs subvolume list . | grep "root_$DISTRO$" | awk '{print $2}') .
 387
 388     # no cow on the root filesystem. it's setup is fully scripted,
 389     # if it's messed up, we will just recreated it,
 390     # and we can get better perf with this.
 391     # I can't remember exactly why, but this is preferable to mounting with
 392     # -o nodatacow, I think because subvolumes inherit that.
 393     chattr -Rf +C root_$DISTRO
 394     cd /
 395     umount /mnt
 396 fi
 397
 398 mount -o subvolid=0 $first_boot_dev /mnt
 399 cd /mnt
 400 btrfs subvolume set-default 0 /mnt # already default, just ensuring it.
 401
 402 # for libreboot systems. grub2 only reads from subvolid=0
 403 mkdir -p /mnt/grub2
 404 cp $FAI/distro-install-common/libreboot_grub.cfg /mnt/grub2
 405
 406 if [[ $DISTRO == debianstable_bootstrap ]]; then
 407     # this is just convenience for the libreboot_grub config
 408     # so we can glob the other ones easier.
 409     boot_vol=$DISTRO
 410 else
 411     boot_vol=boot_$DISTRO
 412 fi
 413 [[ ! -e /mnt/$boot_vol ]] || btrfs subvolume delete /mnt/$boot_vol
 414 btrfs subvolume create $boot_vol
 415 cd /
 416 umount /mnt
 417 ## end create subvols ##
 418
 419 dev=${boot_devs[0]}
 420 mount $first_grub_extdev /mnt
 421 grub-editenv /mnt/grubenv set did_fai_check=true
 422 grub-editenv /mnt/grubenv set last_boot=/$boot_vol
 423 umount /mnt
 424
 425 if [[ $DISTRO == debianstable_bootstrap ]]; then
 426     cat > /tmp/fai/fstab <<EOF
 427 $first_boot_dev  /  btrfs  noatime,subvol=$boot_vol  0 0
 428 EOF
 429     cat >/tmp/fai/disk_var.sh <<EOF
 430 BOOT_DEVICE="${short_devs[@]}"
 431 ROOT_PARTITION=$first_boot_dev
 432 EOF
 433 else
 434     # note, fai creates the mountpoints listed here
 435     cat > /tmp/fai/fstab <<EOF
 436 $first_root_crypt  /  btrfs  noatime,subvol=root_$DISTRO  0 0
 437 $first_root_crypt  /mnt/root  btrfs  noatime,subvolid=0  0 0
 438 $first_boot_dev  /boot  btrfs  noatime,subvol=$boot_vol  0 0
 439 EOF
 440     swaps=()
 441     for dev in ${devs[@]}; do
 442         swaps+=(`swap-cryptname`)
 443         cat >>/tmp/fai/crypttab <<EOF
 444 `root-cryptname`  `rootdev`  none  keyscript=/root/keyscript,discard,luks
 445 `swap-cryptname` `swapdev`  /dev/urandom  swap,cipher=aes-xts-plain64,size=256,hash=ripemd160
 446 EOF
 447         cat >> /tmp/fai/fstab <<EOF
 448 `swap-cryptdev`  none  swap  sw  0 0
 449 EOF
 450     done
 451
 452     # fai would do this:
 453     #BOOT_DEVICE=\${BOOT_DEVICE:-"${devs[0]}"}
 454
 455     # note: swaplist seems to do nothing.
 456     cat >/tmp/fai/disk_var.sh <<EOF
 457 BOOT_DEVICE="${short_devs[@]}"
 458 BOOT_PARTITION=\${BOOT_PARTITION:-$first_boot_dev}
 459 # ROOT_PARTITIONS is added by me, used in arch setup.
 460 ROOT_PARTITIONS="${root_devs[@]}"
 461 ROOT_PARTITION=\${ROOT_PARTITION:-$first_root_crypt}
 462 SWAPLIST=\${SWAPLIST:-"${swaps[@]}"}
 463 EOF
 464 fi