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 PS4='+ $LINENO '
  19 set -eE -o pipefail
  20 trap 'echo "$0:$LINENO:error: \"$BASH_COMMAND\" returned $?" >&2' ERR
  21
  22 if [[ $EUID != 0 ]]; then
  23   echo "$0: error: need to run as root" >&2
  24   exit 1
  25 fi
  26
  27 # for calling outside of FAI:
  28 # fai-redep
  29 #
  30 # source /b/fai/fai-wrapper
  31 # - set any appropriate classes with: fai-setclass OPT1... which sets CLASS_OPT1=true...
  32 #   or run eval-fai-classfile FILE.
  33 # - Set a VOL_DISTROVER (if not doing mkroot2) eg:
  34 #   fai-setclass VOL_NABIA
  35 #
  36 # OPTIONS:
  37 #
  38 # mkroot2: for running outside of fai and setting up the root2/boot2 luks and btrfs
  39 #
  40 # environment variables:
  41 #
  42 # HOSTNAME: if frodo, we exclude 2 devices from the /boot array, which
  43 # the bios does not see. if demohost, we set the luks password to just
  44 # 'x'.
  45 #
  46 # SPECIAL_DISK: For use outside of fai. A base disk name like
  47 # /dev/sdk. If set, we just cryptsetup and partition this one disk then
  48 # exit. This is useful for partitioning a disk in preparation to replace
  49 # a failed or failing disk from a raid10 array.
  50 #
  51 # classes:
  52 #
  53 # REPARTITION: forces repartitioning even if we detect the proper amount
  54 # of partitions already exist.
  55 #
  56 # NOWIPE: use existing subvolumes if they exist
  57 #
  58 # ROTATIONAL: forces to install onto hdds instead of sdds. normally sdds
  59 # are chosen if they exist.
  60 #
  61 # PARTITION_PROMPT: command line prompt before partitioning
  62 #
  63 # RAID0: forces raid0 filesystem. Normally with 4+ devices, we use
  64 # raid10.
  65 # RAID1: forces raid1 filesystem.
  66
  67 mkroot2=false
  68 if [[ $1 ]]; then
  69   case $1 in
  70     mkroot2)
  71       mkroot2=true
  72       ;;
  73     *)
  74       echo "$0: error: unsupported arg: $1" >&2
  75       exit 1
  76       ;;
  77   esac
  78 fi
  79
  80
  81 if [[ $SPECIAL_DISK ]]; then
  82   export CLASS_REPARTITION=true
  83 fi
  84
  85 # # fai's setup-storage won't do btrfs on luks,
  86 # # so we do it ourself :)
  87 # inspiration taken from files in fai-setup-storage package
  88
  89 # if we are not running in fai, skiptask won't be defined, so carry on.
  90 skiptask partition || ! type skiptask
  91
  92 if ! type -p devbyid; then
  93   for d in $FAI/distro-install-common \
  94                /a/bin/fai/fai/config/distro-install-common $FAI $PWD; do
  95     [[ -d $d ]] || continue
  96     if [[ -e $d/devbyid ]]; then
  97       devbyid=$d/devbyid
  98       devbyid() { $devbyid "$@"; }
  99       break
 100     fi
 101   done
 102   if [[ ! $devbyid ]]; then
 103     echo "$0: error: failed to find devbyid script" >&2
 104     exit 1
 105   fi
 106 fi
 107
 108
 109
 110 #### begin configuration
 111
 112 # this is the ordering of the /dev/sdaX, but
 113 # the ordering of the partition layout goes like this:
 114 # bios_grub
 115 # grub_ext
 116 # efi
 117 # root
 118 # swap
 119 # boot
 120
 121 rootn=1
 122 root2n=2
 123 swapn=3
 124 bootn=4
 125 boot2n=5
 126 efin=6
 127 # ext partition so grub can write persistent variables,
 128 # so it can do a one time boot. grub can't write to
 129 # btrfs or any cow fs because it's more
 130 # more complicated to do and they don't want to.
 131 grub_extn=7
 132 # bios boot partition,
 133 # https://wiki.archlinux.org/index.php/GRUB
 134 bios_grubn=8
 135 even_bign=9
 136 lastn=$bios_grubn
 137
 138
 139
 140 ##### end configuration
 141
 142
 143 add-part() { # add partition suffix to $dev
 144   local d part
 145   if [[ $# == 1 ]]; then
 146     d=$dev
 147     part=$1
 148   else
 149     d=$1
 150     part=$2
 151   fi
 152   echo $d-part$part
 153 }
 154
 155 rootdev() { add-part $@ $rootn; }
 156 root2dev() { add-part $@ $root2n; }
 157 swapdev() { add-part $@ $swapn; }
 158 bootdev() { add-part $@ $bootn; }
 159 boot2dev() { add-part $@ $boot2n; }
 160 efidev() { add-part $@ $efin; }
 161 grub_extdev() { add-part $@ $grub_extn; }
 162 bios_grubdev() { add-part $@ $bios_grubn; }
 163 even_bigdev() { add-part $@ $even_bign; }
 164
 165 crypt-dev() { echo /dev/mapper/crypt_dev_${1##*/}; }
 166 crypt-name() { echo crypt_dev_${1##*/}; }
 167 root-cryptdev() { crypt-dev $(rootdev $@); }
 168 root2-cryptdev() { crypt-dev $(root2dev $@); }
 169 swap-cryptdev() { crypt-dev $(swapdev $@); }
 170 root-cryptname() { crypt-name $(rootdev $@); }
 171 root2-cryptname() { crypt-name $(root2dev $@); }
 172 swap-cryptname() { crypt-name $(swapdev $@); }
 173
 174 dev-mib() {
 175   local d=${1:-$dev}
 176   echo $(( $(parted -m $d unit MiB print | \
 177                sed -nr "s#^/dev/[^:]+:([0-9]+).*#\1#p") - 1))
 178 }
 179
 180 luks-setup() {
 181   local luksdev="$1"
 182   # when we move to newer than trisquel 9, we can remove
 183   # --type luks1. We can also check on cryptsetup --help | less /compil
 184   # to see about the other settings. Default in debian 9 is luks2.
 185   # You can convert from luks2 to luks 1 by adding a temporary key:
 186   # cryptsetup luksAddKey --pbkdf pbkdf2
 187   # then remove the new format keys with cryptsetup luksRemoveKey
 188   # then cryptsetup convert DEV --type luks1, then readd old keys and remove temp.
 189   yes YES | cryptsetup luksFormat $luksdev $luks_file || [[ $? == 141 ]]
 190   yes "$lukspw" | \
 191     cryptsetup luksAddKey --key-file $luks_file \
 192                $luksdev || [[ $? == 141 ]]
 193   # background: Keyfile and password are treated just
 194   # like 2 ways to input a passphrase, so we don't actually need to have
 195   # different contents of keyfile and passphrase, but it makes some
 196   # security sense to a really big randomly generated passphrase
 197   # as much as possible, so we have both.
 198   #
 199   # This would remove the keyfile.
 200   #    yes 'test' | cryptsetup luksRemoveKey /dev/... \
 201     #                            /key/file || [[ $? == 141 ]]
 202   cryptsetup luksOpen $luksdev $(crypt-name $luksdev) --key-file $luks_file
 203 }
 204
 205 ##### end function defs
 206
 207 if ifclass REPARTITION; then
 208   partition=true # force a full wipe
 209 else
 210   partition=false # change to true to force a full wipe
 211 fi
 212 if ifclass NOWIPE; then
 213   wipe=false
 214 else
 215   wipe=true
 216 fi
 217
 218 if (($(nproc) > 2)); then
 219   mopts=,compress=zstd
 220 fi
 221
 222 declare -A disk_excludes
 223 if ! $mkroot2 && ! ifclass USE_MOUNTED; then
 224   ## ignore disks that are mounted, eg when running from fai-cd
 225   while read -r l; do
 226     eval "$l"
 227     if [[ ! $PKNAME ]]; then
 228       PKNAME="$KNAME"
 229     fi
 230     if [[ $MOUNTPOINT ]]; then
 231       disk_excludes[$PKNAME]=true
 232     fi
 233   done < <(lsblk -nP -o KNAME,MOUNTPOINT,PKNAME)
 234 fi
 235
 236 hdds=()
 237 ssds=()
 238 # this excludes usb. note: i may encounter some other type in the future.
 239 for disk in $(lsblk -do name,tran -n | awk '$2 ~ "^(sata|nvme)$" { print $1 }'); do
 240   if [[ ${disk_excludes[$disk]} ]]; then
 241     continue
 242   fi
 243   case $(cat /sys/block/$disk/queue/rotational) in
 244     0) ssds+=(/dev/$disk) ;;
 245     1) hdds+=(/dev/$disk) ;;
 246     *) echo "$0: error: unknown /sys/block/$disk/queue/rotational: \
 247 $(cat $disk/queue/rotational)"; exit 1 ;;
 248   esac
 249 done
 250
 251 # install all ssds, or if there are none, all hdds.
 252 # Note, usb flash disks are seen as rotational, which is
 253 # very odd, but convenient for ignoring them here.
 254 # TODO: find a reliable way to ignore them.
 255 if ! ifclass ROTATIONAL && (( ${#ssds[@]} > 0 )); then
 256   short_devs=( ${ssds[@]} )
 257 else
 258   short_devs=( ${hdds[@]} )
 259 fi
 260
 261 # check if the partitions exist have the right filesystems
 262 #blkid="$(blkid -s TYPE)"
 263 for dev in ${short_devs[@]}; do
 264   if $partition; then break; fi
 265   y=$(readlink -f $dev)
 266   arr=($y?*)
 267   if (( ${#arr[@]} < lastn )); then
 268     partition=true
 269   fi
 270   # On one system, blkid is missing some partitions.
 271   # maybe we need a flag, like FUZZY_BLKID or something, so we
 272   # can check that at least some exist.
 273   # for x in "`rootdev`: TYPE=\"crypto_LUKS\"" "`bootdev`: TYPE=\"btrfs\""; do
 274   #     echo "$blkid" | grep -Fx "$x" &>/dev/null || partition=true
 275   # done
 276 done
 277
 278 if $partition && ifclass PARTITION_PROMPT; then
 279   echo "Press any key except ctrl-c to continue and partition these drives:"
 280   echo "  ${short_devs[*]}"
 281   read -r
 282 fi
 283
 284 devs=()
 285 shopt -s extglob
 286 for short_dev in ${short_devs[@]}; do
 287   devs+=($(devbyid $short_dev))
 288 done
 289 if [[ ! ${devs[0]} ]]; then
 290   echo "$0: error: failed to detect devs" >&2
 291   exit 1
 292 fi
 293
 294 boot_space=0
 295 first=true
 296 boot_devs=()
 297 boot2_devs=()
 298 for dev in ${devs[@]}; do
 299   if ifclass frodo; then
 300     # I ran into a machine where the bios doesn't know about some disks,
 301     # so 1st stage of grub also doesn't know about them.
 302     # Also, grub does not support mounting degraded btrfs as far as
 303     # I can tell with some googling.
 304     # From within an arch install env, I could detect them by noting
 305     # their partitions were mixed with the next disk in /dev/disk/by-path,
 306     # and I have mixed model disks, and I could see the 8 models which showed
 307     # up in the bios, and thus see which 2 models were missing.
 308     # hdparm -I /dev/sdh will give model info in linux.
 309     # However, in fai on jessie, /dev/disk/by-path dir doesn't exist,
 310     # and I don't see another way, so I'm hardcoding them.
 311     # We still put grub on them and partition them the same, for uniformity
 312     # and in case they get moved to a system that can recognize them,
 313     # we just exclude them from the boot filesystem.
 314     cd /dev/disk/by-id/
 315     bad_disk=false
 316     for id in ata-TOSHIBA_MD04ACA500_8539K4TQFS9A \
 317                 ata-TOSHIBA_MD04ACA500_Y5IFK6IJFS9A; do
 318       if [[ $(readlink -f $id) == "$(readlink -f $dev)" ]]; then
 319         bad_disk=true
 320         break
 321       fi
 322     done
 323     if ! $bad_disk; then
 324       boot_devs+=($(bootdev))
 325       boot2_devs+=($(boot2dev))
 326     fi
 327   else
 328     boot_space=$(( boot_space + $(parted -m $dev unit MiB print | \
 329                                     sed -nr "s#^/dev/[^:]+:([0-9]+).*#\1#p") - 1))
 330     boot_devs+=($(bootdev))
 331     boot2_devs+=($(boot2dev))
 332   fi
 333   if $first && [[ $boot_devs ]]; then
 334     first_efi=$(efidev)
 335     first_grub_extdev=$(grub_extdev)
 336     first=false
 337   fi
 338 done
 339 first_boot_dev=${boot_devs[0]}
 340
 341 even_raid=false
 342 if ifclass RAID0 || (( ${#boot_devs[@]} == 1 )); then
 343   raid_level=0
 344 elif ifclass RAID1 || (( ${#boot_devs[@]} <= 3 )); then
 345   if (( ${#boot_devs[@]} == 2 )); then
 346     even_raid=true
 347   fi
 348   raid_level=1
 349 else
 350   raid_level=10
 351 fi
 352
 353 ### Begin calculate boot partition space
 354 # due to raid duplication
 355 case $raid_level in
 356   1*) boot_space=$(( boot_space / 2 )) ;;
 357 esac
 358 if (( boot_space > 60000 )); then
 359   # this is larger than needed for several /boot subvols,
 360   # becuase I keep a minimal debian install on it for
 361   # recovery needs and for doing pxe-kexec.
 362   boot_mib=10000
 363   root2_mib=1000000
 364   boot2_mib=5000
 365 elif (( boot_spa_ce > 30000 )); then
 366   boot_mib=$(( 5000 + (boot_space - 30000) / 2 ))
 367   root2_mib=100
 368   boot2_mib=100
 369 else
 370   # Small vms don't have room for /boot recovery.  With 3 kernels
 371   # installed, i'm using 132M on t8, so this seems like plenty of
 372   # room. note: rhel 8 recomments 1g for /boot.  u20.04 with 3 kernels =
 373   # 308 mb, so things have grown significantly
 374   boot_mib=1000
 375   root2_mib=100
 376   boot2_mib=100
 377 fi
 378 case $raid_level in
 379   1*)
 380     boot_mib=$(( boot_mib * 2 ))
 381     boot2_mib=$(( boot2_mib * 2 ))
 382     root2_mib=$(( root2_mib * 2 ))
 383     ;;
 384 esac
 385 ### end calculate boot partition space
 386
 387
 388
 389 if [[ ! $DISTRO ]]; then
 390   if ifclass VOL_BULLSEYE_BOOTSTRAP; then
 391     DISTRO=debianbullseye_bootstrap
 392   elif ifclass VOL_STRETCH; then
 393     DISTRO=debianstretch
 394   elif ifclass VOL_BUSTER; then
 395     DISTRO=debianbuster
 396   elif ifclass VOL_BULLSEYE; then
 397     DISTRO=debianbullseye
 398   elif ifclass VOL_BOOKWORM; then
 399     DISTRO=debianbookworm
 400   elif ifclass VOL_TESTING; then
 401     DISTRO=debiantesting
 402   elif ifclass VOL_XENIAL; then
 403     DISTRO=ubuntuxenial
 404   elif ifclass VOL_BIONIC; then
 405     DISTRO=ubuntubionic
 406   elif ifclass VOL_FOCAL; then
 407     DISTRO=ubuntufocal
 408   elif ifclass VOL_FLIDAS; then
 409     DISTRO=trisquelflidas
 410   elif ifclass VOL_ETIONA; then
 411     DISTRO=trisqueletiona
 412   elif ifclass VOL_NABIA; then
 413     DISTRO=trisquelnabia
 414   elif $mkroot2; then
 415     :
 416   else
 417     echo "PARTITIONER ERROR: no distro class/var set" >&2
 418     exit 1
 419   fi
 420 fi
 421
 422
 423 bpart() { # btrfs a partition
 424   case $raid_level in
 425     0) mkfs.btrfs -f $@ ;;
 426     1) mkfs.btrfs -f -m raid1 -d raid1 $@ ;;
 427     10) mkfs.btrfs -f -m raid10 -d raid10 $@ ;;
 428   esac
 429 }
 430
 431
 432 if [[ ! $luks_dir ]]; then
 433   # see README for docs about how to create these
 434   luks_dir=$FAI/distro-install-common/luks
 435   if [[ ! -d $luks_dir ]]; then
 436     luks_dir=/q/root/luks
 437   fi
 438   if [[ ! -d $luks_dir ]]; then
 439     echo "$0: error: no luks_dir found" >&2
 440     exit 1
 441   fi
 442 fi
 443
 444 luks_file=$luks_dir/host-$HOSTNAME
 445 if [[ ! -e $luks_file ]]; then
 446   hostkeys=($luks_dir/host-*)
 447   # if there is only one key, we might be deploying somewhere
 448   # where dhcp doesnt give us a proper hostname, so use that.
 449   if [[ ${#hostkeys[@]} == 1 && -e ${hostkeys[0]} ]]; then
 450     luks_file=${hostkeys[0]}
 451   else
 452     echo "$0: error: no key for hostname at $luks_file" >&2
 453     exit 1
 454   fi
 455 fi
 456
 457 # # note, corresponding changes in /b/ds/keyscript-{on,off}
 458 if ifclass demohost; then
 459   lukspw=x
 460 elif [[ -e $luks_dir/$HOSTNAME ]]; then
 461   lukspw=$(cat $luks_dir/$HOSTNAME)
 462 else
 463   lukspw=$(cat $luks_dir/iank)
 464 fi
 465
 466
 467 first_root_crypt=$(root-cryptdev ${devs[0]})
 468
 469 # 1.5 x based on https://access.redhat.com/documentation/en-US/Red_Hat_Enterprise_Linux/7/html/Installation_Guide/sect-disk-partitioning-setup-x86.html#sect-custom-partitioning-x86
 470 swap_mib=$(( $(grep ^MemTotal: /proc/meminfo | \
 471                  awk '{print $2}') * 3/(${#devs[@]} * 2 ) / 1024 ))
 472
 473 mkdir -p /tmp/fai
 474 root_devs=()
 475 for dev in ${devs[@]}; do
 476   root_devs+=($(rootdev))
 477 done
 478 shopt -s nullglob
 479
 480 # We write to these files instead of just /etc/fstab, /etc/crypttab,
 481 # because these are filesystems created after our current root, and so
 482 # this allows us to update other root filesystems too.
 483 rm -f /mnt/root/root2-{fs,crypt}tab
 484 if $mkroot2; then
 485   if $partition; then
 486     echo $0: error: found partition=true but have mkroot2 arg
 487     exit 1
 488   fi
 489   for dev in ${devs[@]}; do
 490     luks_file=$luks_dir/host-amy
 491     lukspw=$(cat $luks_dir/amy)
 492     luks-setup $(root2dev)
 493     cat >>/mnt/root/root2-crypttab <<EOF
 494 $(root2-cryptname) $(root2dev)  $luks_file  discard,luks,initramfs
 495 EOF
 496   done
 497   bpart $(for dev in ${devs[@]}; do root2-cryptdev; done)
 498   bpart ${boot2_devs[@]}
 499   mkdir -p /mnt/root2 /mnt/boot2
 500   cat >>/mnt/root/root2-fstab <<EOF
 501 $(root2-cryptdev ${devs[0]}) /mnt/root2  btrfs  nofail,x-systemd.device-timeout=30s,x-systemd.mount-timeout=30s,noatime,subvolid=0$mopts  0 0
 502 ${boot2_devs[0]} /mnt/boot2  btrfs    nofail,x-systemd.device-timeout=30s,x-systemd.mount-timeout=30s,noatime,subvolid=0  0 0
 503 EOF
 504   exit 0
 505 fi
 506
 507
 508 if $partition; then
 509   ### begin wipefs
 510   if [[ ! $SPECIAL_DISK ]]; then
 511     for dev in ${devs[@]}; do
 512       # if we repartition to the same as an old partition,
 513       # we don't want any old fses hanging around.
 514       for (( i=1; i <= lastn; i++ )); do
 515         x=$(add-part $i)
 516         [[ -e $x ]] || continue
 517         count_down=10
 518         # wipefs has failed, manual run works, google suggests timing issue
 519         while ! wipefs -a $x; do
 520           sleep 2
 521           count_down=$((count_down - 1))
 522           (( count_down > 0 )) || exit 1
 523         done
 524       done
 525     done
 526   fi
 527   ### end wipefs
 528
 529
 530   # When we have 2 disks of at least 100g difference in size,
 531   # make an extra partition on the end of the bigger one.
 532   even_big_part=false
 533   even_diff_min=100000
 534   if $even_raid; then
 535     smalli=0
 536     bigi=1
 537     if (( $(dev-mib ${devs[0]}) >= $(dev-mib ${devs[1]}) )); then
 538       smalli=1
 539       bigi=0
 540     fi
 541     disk_mib=$(dev-mib ${devs[smalli]})
 542     even_big_dev=${devs[bigi]}
 543     even_big_mib=$(dev-mib $even_big_dev)
 544     if (( even_big_mib - disk_mib > even_diff_min )); then
 545       even_big_part=true
 546     fi
 547   fi
 548
 549   for dev in ${devs[@]}; do
 550     if [[ $SPECIAL_DISK ]]; then
 551       dev=$(devbyid $SPECIAL_DISK)
 552     fi
 553
 554     # parted will round up the disk size. Do -1 so we can have
 555     # fully 1MiB unit partitions for easy resizing of the last partition.
 556     # Otherwise we would pass in -0 for the end argument for the last partition.
 557     #
 558     # Note: parted print error output is expected. example:
 559     # Error: /dev/vda: unrecognised disk label
 560     if ! $even_raid; then
 561       disk_mib=$(dev-mib)
 562     fi
 563
 564     boot_part_mib=$(( boot_mib / ${#boot_devs[@]} ))
 565     boot2_part_mib=$(( boot2_mib / ${#boot_devs[@]} ))
 566     root2_part_mib=$(( root2_mib / ${#root_devs[@]} ))
 567     root_end=$(( disk_mib - root2_part_mib - swap_mib - boot_part_mib - boot2_part_mib ))
 568     root2_end=$(( root_end + root2_part_mib ))
 569     swap_end=$(( root2_end + swap_mib ))
 570     boot_end=$(( swap_end + boot_part_mib ))
 571
 572     parted -s $dev mklabel gpt
 573     # MiB because parted complains about alignment otherwise.
 574     pcmd="parted -a optimal -s -- $dev"
 575     # root partition, the main big one
 576     $pcmd mkpart primary ext3 524MiB ${root_end}MiB
 577     # without naming, systemd gives us misc errors like:
 578     # dev-disk-by\x2dpartlabel-primary.device: Dev dev-disk-by\x2dpartlabel-primary.device appeared twice
 579     $pcmd name $rootn root
 580     # root2 partition
 581     $pcmd mkpart primary ext3 ${root_end}MiB ${root2_end}MiB
 582     $pcmd name $root2n root2
 583     # normally a swap is type "linux-swap", but this is encrypted swap. using that
 584     # label will confuse systemd.
 585     # swap partition
 586     $pcmd mkpart primary "" ${root2_end}MiB ${swap_end}MiB
 587     $pcmd name $swapn swap
 588     # boot partition
 589     $pcmd mkpart primary "" ${swap_end}MiB ${boot_end}MiB
 590     $pcmd name $bootn boot
 591     # boot2 partition
 592     $pcmd mkpart primary "" ${boot_end}MiB ${disk_mib}MiB
 593     $pcmd name $boot2n boot2
 594     # uefi partition. efi sucks, half a gig, rediculous.
 595     $pcmd mkpart primary "fat32" 12MiB 524MiB
 596     $pcmd name $efin efi
 597     $pcmd set $efin esp on
 598     # note, this is shown here: https://support.system76.com/articles/bootloader/
 599     # but not mentioned https://wiki.archlinux.org/index.php/EFI_system_partition
 600     # probably not needed
 601     $pcmd set $bootn boot on
 602     $pcmd set $boot2n boot on
 603     # i only need a few k, but googling min size,
 604     # I found someone saying that gparted required
 605     # required at least 8 because of their hard drive cylinder size.
 606     # And 8 is still very tiny.
 607     # grub_ext partition
 608     $pcmd mkpart primary "ext2" 4MiB 12MiB
 609     $pcmd name $grub_extn grubext
 610     # gpt ubuntu cloud image uses ~4 mb for this partition. fai uses 1 MiB.
 611     # so, I use 3, whatever.
 612     # note: parted manual saying cheap flash media
 613     # should to start at 4.
 614     # biols grub partition
 615     $pcmd mkpart primary "" 1MiB 4MiB
 616     $pcmd name $bios_grubn biosgrub
 617     $pcmd set $bios_grubn bios_grub on
 618     if $even_big_part  && [[ $dev == "$even_big_dev" ]]; then
 619       $pcmd mkpart primary ext3 ${disk_mib}MiB ${even_big_mib}MiB
 620       $pcmd name $even_bign even_big
 621     fi
 622
 623     # the mkfs failed before on a vm, which prompted me to add
 624     # sleep .1
 625     # then it failed again on a physical machine
 626     # with:
 627     # Device /dev/disk/by-id/foo doesn't exist or access denied,
 628     # so I added a wait until it existed.
 629     # Then I added the mkfs.ext2, which claimed to succeed,
 630     # but then couldn't be found upon reboot. In that case we didn't
 631     # wait at all. So I've added a 3 second minimum wait.
 632     secs=0
 633     while [[ ! -e $(bios_grubdev) ]] && (( secs < 10 )); do
 634       sleep 1
 635       secs=$((secs +1))
 636     done
 637     sleep 3
 638
 639     mkfs.fat -F32 $(efidev)
 640
 641     if $even_big_part  && [[ $dev == "$even_big_dev" ]]; then
 642       luks-setup $(even_bigdev)
 643       mkfs.btrfs -f $(crypt-dev $(even_bigdev))
 644     fi
 645
 646     # Holds just a single file, rarely written, so
 647     # use ext2, like was often used for the /boot partition.
 648     # This exists because grub can only persist data to a non-cow fs.
 649     # And we use persisting a var in grub to do a one time boot.
 650     # We could pass the data on the kernel command line and persist it
 651     # to grubenv after booting, but that relies on the boot always succeeding.
 652     # This is just a bit more robust, and it could work for booting
 653     # into ipxe which can't persist data, if we ever got that working.
 654     mkfs.ext2 $(grub_extdev)
 655     luks-setup $(rootdev)
 656
 657     if [[ $SPECIAL_DISK ]]; then
 658       exit 0
 659     fi
 660   done
 661   ls -la /dev/btrfs-control # this was probably for debugging...
 662   sleep 1
 663   bpart $(for dev in ${devs[@]}; do root-cryptdev; done)
 664   bpart ${boot_devs[@]}
 665 else
 666   for dev in ${devs[@]}; do
 667     if [[ -e /dev/mapper/$(root-cryptname) ]]; then
 668       continue
 669     fi
 670     cryptsetup luksOpen $(rootdev) $(root-cryptname) \
 671                --key-file $luks_file
 672   done
 673   sleep 1
 674 fi
 675
 676
 677 if $wipe && [[ $DISTRO != debianbullseye_bootstrap ]]; then
 678   # bootstrap distro doesn't use separate encrypted root.
 679   mount -o subvolid=0 $first_root_crypt /mnt
 680   # systemd creates subvolumes we want to delete.
 681   s=($(btrfs subvolume list --sort=-path /mnt |
 682          sed -rn "s#^.*path\s*(root_$DISTRO/\S+)\s*\$#\1#p"))
 683   for subvol in ${s[@]}; do btrfs subvolume delete /mnt/$subvol; done
 684   btrfs subvolume set-default 0 /mnt
 685   [[ ! -e /mnt/root_$DISTRO ]] || btrfs subvolume delete /mnt/root_$DISTRO
 686
 687   ## create subvols ##
 688   cd /mnt
 689
 690   btrfs subvolume create root_$DISTRO
 691
 692   # could set default subvol like this, but no reason to.
 693   # btrfs subvolume set-default \
 694     #       $(btrfs subvolume list . | grep "root_$DISTRO$" | awk '{print $2}') .
 695
 696   # For raid systems, cow allows for error correction, for non-raid systems,
 697   # protects root fs from having the plug pulled. Reprovisioning a root
 698   # subvol is not my favorite thing to do.
 699   # # no cow on the root filesystem. it's setup is fully scripted,
 700   # # if it's messed up, we will just recreated it,
 701   # # and we can get better perf with this.
 702   # # I can't remember exactly why, but this is preferable to mounting with
 703   # # -o nodatacow, I think because subvolumes inherit that.
 704   # chattr -Rf +C root_$DISTRO
 705   cd /
 706   umount /mnt
 707 fi
 708
 709 mount -o subvolid=0 $first_boot_dev /mnt
 710 cd /mnt
 711 btrfs subvolume set-default 0 /mnt # already default, just ensuring it.
 712
 713 # for libreboot systems. grub2 only reads from subvolid=0
 714 mkdir -p /mnt/grub2
 715 cp $FAI/distro-install-common/libreboot_grub.cfg /mnt/grub2
 716
 717 if [[ $DISTRO == debianbullseye_bootstrap ]]; then
 718   # this is just convenience for the libreboot_grub config
 719   # so we can glob the other ones easier.
 720   boot_vol=$DISTRO
 721 else
 722   boot_vol=boot_$DISTRO
 723 fi
 724 if $wipe && [[ -e /mnt/$boot_vol ]]; then
 725   btrfs subvolume delete /mnt/$boot_vol
 726 fi
 727 if [[ ! -e /mnt/$boot_vol ]]; then
 728   btrfs subvolume create $boot_vol
 729 fi
 730 cd /
 731 umount /mnt
 732 ## end create subvols ##
 733
 734 dev=${boot_devs[0]}
 735 mount $first_grub_extdev /mnt
 736 grub-editenv /mnt/grubenv set did_fai_check=true
 737 grub-editenv /mnt/grubenv set last_boot=/$boot_vol
 738 umount /mnt
 739
 740 fstabstd=x-systemd.device-timeout=30s,x-systemd.mount-timeout=30s
 741 if [[ $DISTRO == debianbullseye_bootstrap ]]; then
 742   cat > /tmp/fai/fstab <<EOF
 743 $first_boot_dev  /  btrfs  noatime,subvol=$boot_vol  0 0
 744 $first_efi  /boot/efi  vfat          nofail,$fstabstd  0 0
 745 EOF
 746   cat >/tmp/fai/disk_var.sh <<EOF
 747 BOOT_DEVICE="${short_devs[@]}"
 748 ROOT_PARTITION=$first_boot_dev
 749 EOF
 750 else
 751   # note, fai creates the mountpoints listed here
 752   cat > /tmp/fai/fstab <<EOF
 753 $first_root_crypt  /  btrfs $fstabstdopts,noatime,subvol=root_$DISTRO$mopts  0 0
 754 $first_root_crypt  /mnt/root  btrfs  nofail,$fstabstd,noatime,subvolid=0$mopts  0 0
 755 $first_boot_dev  /boot  btrfs        nofail,$fstabstd,noatime,subvol=$boot_vol  0 0
 756 $first_efi  /boot/efi  vfat          nofail,$fstabstd  0 0
 757 $first_boot_dev  /mnt/boot  btrfs    nofail,$fstabstd,noatime,subvolid=0  0 0
 758 EOF
 759   swaps=()
 760   rm -f /tmp/fai/crypttab
 761   for dev in ${devs[@]}; do
 762     swaps+=($(swap-cryptname))
 763     cat >>/tmp/fai/crypttab <<EOF
 764 $(root-cryptname) $(rootdev)  none  keyscript=/root/keyscript,discard,luks,initramfs
 765 $(swap-cryptname) $(swapdev)  /dev/urandom  swap,cipher=aes-xts-plain64,size=256,hash=ripemd160
 766 EOF
 767     cat >> /tmp/fai/fstab <<EOF
 768 $(swap-cryptdev)  none  swap  nofail,$fstabstd,sw  0 0
 769 EOF
 770   done
 771
 772   # fai would do this:
 773   #BOOT_DEVICE=\${BOOT_DEVICE:-"${devs[0]}"}
 774
 775   # note: swaplist seems to do nothing.
 776   cat >/tmp/fai/disk_var.sh <<EOF
 777 BOOT_DEVICE="${short_devs[@]}"
 778 BOOT_PARTITION=\${BOOT_PARTITION:-$first_boot_dev}
 779 # ROOT_PARTITIONS is added by me, used in arch setup.
 780 ROOT_PARTITIONS="${root_devs[@]}"
 781 ROOT_PARTITION=\${ROOT_PARTITION:-$first_root_crypt}
 782 SWAPLIST=\${SWAPLIST:-"${swaps[@]}"}
 783 EOF
 784
 785   if [[ $HOSTNAME == kd ]]; then
 786     # note, having these with keyscript and initramfs causes a luks error in fai.log,
 787     # but it is safely ignorable and gets us the ability to just type our password
 788     # in once at boot. A downside is that they are probably needed to be plugged in to boot.
 789     cat >>/tmp/fai/crypttab <<EOF
 790 crypt_dev_ata-Samsung_SSD_870_QVO_8TB_S5VUNG0N900656V-part${even_bign} /dev/disk/by-id/ata-Samsung_SSD_870_QVO_8TB_S5VUNG0N900656V-part7  none  keyscript=decrypt_keyctl,discard,luks,initramfs
 791 crypt_dev_ata-TOSHIBA_MD04ACA500_84R2K773FS9A-part1 /dev/disk/by-id/ata-TOSHIBA_MD04ACA500_84R2K773FS9A-part1  none  keyscript=decrypt_keyctl,discard,luks,initramfs
 792 crypt_dev_ata-ST6000DM001-1XY17Z_Z4D29EBL-part1 /dev/disk/by-id/ata-ST6000DM001-1XY17Z_Z4D29EBL-part1  none  keyscript=decrypt_keyctl,discard,luks,initramfs
 793 EOF
 794     cat >> /tmp/fai/fstab <<EOF
 795 # r7 = root partition7. it isnt actually #7 anymore, not a great name, but whatever
 796 /dev/mapper/crypt_dev_ata-Samsung_SSD_870_QVO_8TB_S5VUNG0N900656V-part${even_bign}  /mnt/r7  btrfs  nofail,$fstabstd,noatime,compress=zstd,subvolid=0  0 0
 797 /dev/mapper/crypt_dev_ata-TOSHIBA_MD04ACA500_84R2K773FS9A-part1  /mnt/rust1  btrfs  nofail,$fstabstd,noatime,compress=zstd,subvolid=0  0 0
 798 /dev/mapper/crypt_dev_ata-ST6000DM001-1XY17Z_Z4D29EBL-part1  /mnt/rust2  btrfs  nofail,$fstabstd,noatime,compress=zstd,subvolid=0  0 0
 799 EOF
 800   fi
 801
 802 fi
 803
 804 # initial setup of extra data fs, mounted,
 805 # btrfs subvol create nocow
 806 # chattr +C nocow
 807 # chown iank.iank nocow
 808