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 without args:
  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 # ARGS (only 1 is valid):
  37 #
  38 # mkroot2: for running outside of fai and setting up the root2/boot2 luks and btrfs and tab files
  39 #
  40 # mkroot2tab: for running outside of fai and setting up the root2/boot2 tab files, in case luks and btrfs
  41 # happen to already be setup.
  42 #
  43 # mktab: for running outside of fai and generating a crypttab for
  44 # the main root fs in /tmp/fai. Must run with env var, eg export DISTRO=trisquelnabia.
  45 #
  46 # Example use in a bootstrap distro:
  47 # scp /a/bin/fai/fai/config/{distro-install-common/devbyid,hooks/partition.DEFAULT} root@HOST:
  48 # sl HOST
  49 # export DISTRO=trisquelnabia; ./partition.DEFAULT mktab
  50 # ## cryptsetup wont take within a pipeline
  51 # mapfile -t lines < <(awk '! /swap/ {print $2,$1}' /tmp/fai/crypttab )
  52 # for l in "${lines[@]}"; do cryptsetup luksOpen $l; done
  53 #
  54 # # or alternatively, to avoid typing it many times:
  55 # read -r lukspw; for l in "${lines[@]}"; do yes "$lukspw" | cryptsetup luksOpen $l; done
  56
  57 ## potentially useful later:
  58 # sed 's#/root/keyscript,#decrypt_keyctl,#;s/$/,noauto/' /tmp/fai/crypttab >/etc/crypttab
  59 #
  60 # environment variables:
  61 #
  62 # HOSTNAME: if frodo, we exclude 2 devices from the /boot array, which
  63 # the bios does not see. if demohost, we set the luks password to just
  64 # 'x'.
  65 #
  66 # SPECIAL_DISK: For use outside of fai. A base disk name like
  67 # /dev/sdk. If set, we just cryptsetup and partition this one disk then
  68 # exit. This is useful for partitioning a disk in preparation to replace
  69 # a failed or failing disk from a raid10 array.
  70 #
  71 # classes:
  72 #
  73 # REPARTITION: forces repartitioning even if we detect the proper amount
  74 # of partitions already exist.
  75 #
  76 # NOWIPE: use existing subvolumes if they exist
  77 #
  78 # REROOTFS: Don't reuse the root filesystem, even if we normally would
  79 #
  80 # ROTATIONAL: forces to install onto hdds instead of sdds. normally sdds
  81 # are chosen if they exist.
  82 #
  83 # PARTITION_PROMPT: command line prompt before partitioning
  84 #
  85 # RAID0: forces raid0 filesystem. Normally with 4+ devices, we use
  86 # raid10.
  87 # RAID1: forces raid1 filesystem.
  88
  89 mkroot2tab=false
  90 mkroot2=false
  91 mktab=false
  92 if [[ $1 ]]; then
  93   ## duplicates fai-wrapper, for convenience of not needing it
  94   if ! type -t ifclass &>/dev/null; then
  95     ifclass() {
  96       local var=${1/#/CLASS_}
  97       [[ $HOSTNAME == $1 || ${!var} ]]
  98     }
  99   fi
 100
 101   case $1 in
 102     mkroot2)
 103       mkroot2=true
 104       ;;
 105     mkroot2tab)
 106       mkroot2tab=true
 107       ;;
 108     mktab)
 109       mktab=true
 110       ;;
 111     *)
 112       echo "$0: error: unsupported arg: $1" >&2
 113       exit 1
 114       ;;
 115   esac
 116 fi
 117
 118
 119 if [[ $SPECIAL_DISK ]]; then
 120   export CLASS_REPARTITION=true
 121 fi
 122
 123 # # fai's setup-storage won't do btrfs on luks,
 124 # # so we do it ourself :)
 125 # inspiration taken from files in fai-setup-storage package
 126
 127 # if we are not running in fai, skiptask won't be defined, so carry on.
 128 skiptask partition || ! type skiptask
 129
 130 if ! type -p devbyid; then
 131   for d in $FAI/distro-install-common \
 132              /a/bin/fai/fai/config/distro-install-common $FAI $PWD; do
 133     [[ -d $d ]] || continue
 134     if [[ -e $d/devbyid ]]; then
 135       devbyid=$d/devbyid
 136       devbyid() { $devbyid "$@"; }
 137       break
 138     fi
 139   done
 140   if [[ ! $devbyid ]]; then
 141     echo "$0: error: failed to find devbyid script" >&2
 142     exit 1
 143   fi
 144 fi
 145
 146
 147
 148 #### begin configuration
 149
 150 # this is the ordering of the /dev/sdaX, but
 151 # the ordering of the partition layout goes like this:
 152 # bios_grub
 153 # grub_ext
 154 # efi
 155 # root
 156 # swap
 157 # boot
 158
 159 rootn=1
 160 root2n=2
 161 swapn=3
 162 bootn=4
 163 boot2n=5
 164 efin=6
 165 # ext partition so grub can write persistent variables,
 166 # so it can do a one time boot. grub can't write to
 167 # btrfs or any cow fs because it's more
 168 # more complicated to do and they don't want to.
 169 grub_extn=7
 170 # bios boot partition,
 171 # https://wiki.archlinux.org/index.php/GRUB
 172 bios_grubn=8
 173 # for an even raid (raid 1), when one disk is bigger, this partition goes on the big disk
 174 even_bign=9
 175 lastn=$bios_grubn
 176
 177
 178
 179 ##### end configuration
 180
 181 ##### begin function defs
 182
 183 add-part() { # add partition suffix to $dev
 184   local d part
 185   if [[ $# == 1 ]]; then
 186     d=$dev
 187     part=$1
 188   else
 189     d=$1
 190     part=$2
 191   fi
 192   echo $d-part$part
 193 }
 194
 195 rootdev() { add-part $@ $rootn; }
 196 root2dev() { add-part $@ $root2n; }
 197
 198 # note, the following block could all have $@ like below
 199 # But it is not added since it is not used and shellcheck
 200 # rightly says args never used should not exist.
 201 ##swapdev() { add-part $swapn; }
 202
 203 swapdev() { add-part $swapn; }
 204 bootdev() { add-part $bootn; }
 205 boot2dev() { add-part $boot2n; }
 206 efidev() { add-part $efin; }
 207 grub_extdev() { add-part $grub_extn; }
 208 bios_grubdev() { add-part $bios_grubn; }
 209 even_bigdev() { add-part $even_bign; }
 210
 211 crypt-dev() { echo /dev/mapper/crypt_dev_${1##*/}; }
 212 crypt-name() { echo crypt_dev_${1##*/}; }
 213 root-cryptdev() { crypt-dev $(rootdev $@); }
 214 root2-cryptdev() { crypt-dev $(root2dev $@); }
 215
 216 # I omit a possible parameter since it is unused:
 217 ##swap-cryptdev() { crypt-dev $(swapdev $@); }
 218 swap-cryptdev() { crypt-dev $(swapdev); }
 219 root-cryptname() { crypt-name $(rootdev); }
 220 root2-cryptname() { crypt-name $(root2dev); }
 221 swap-cryptname() { crypt-name $(swapdev); }
 222
 223 dev-mib() {
 224   local d=${1:-$dev}
 225   echo $(( $(parted -m $d unit MiB print | \
 226                sed -nr "s#^/dev/[^:]+:([0-9]+).*#\1#p") - 1))
 227 }
 228
 229 luks-setup() {
 230   local luksdev="$1"
 231   # when we move to newer than trisquel 9, we can remove
 232   # --type luks1. We can also check on cryptsetup --help | less /compil
 233   # to see about the other settings. Default in debian 9 is luks2.
 234   # You can convert from luks2 to luks 1 by adding a temporary key:
 235   # cryptsetup luksAddKey --pbkdf pbkdf2
 236   # then remove the new format keys with cryptsetup luksRemoveKey
 237   # then cryptsetup convert DEV --type luks1, then readd old keys and remove temp.
 238   yes YES | cryptsetup luksFormat $luksdev $luks_file || [[ $? == 141 ]]
 239   yes "$lukspw" | \
 240     cryptsetup luksAddKey --key-file $luks_file \
 241                $luksdev || [[ $? == 141 ]]
 242   # background: Keyfile and password are treated just
 243   # like 2 ways to input a passphrase, so we don't actually need to have
 244   # different contents of keyfile and passphrase, but it makes some
 245   # security sense to a really big randomly generated passphrase
 246   # as much as possible, so we have both.
 247   #
 248   # This would remove the keyfile.
 249   #    yes 'test' | cryptsetup luksRemoveKey /dev/... \
 250     #                            /key/file || [[ $? == 141 ]]
 251   cryptsetup luksOpen $luksdev $(crypt-name $luksdev) --key-file $luks_file
 252 }
 253
 254 ##### end function defs
 255
 256
 257 ##### begin variable setup
 258 partition=false
 259 if ifclass REPARTITION; then
 260   partition=true # force a full wipe
 261 fi
 262 wipe=true
 263 if ifclass NOWIPE; then
 264   wipe=false
 265 fi
 266
 267 rerootfs=false
 268 if ifclass REROOTFS; then
 269   rerootfs=true
 270 fi
 271
 272 if (($(nproc) > 2)); then
 273   mopts=,compress=zstd
 274 fi
 275
 276 declare -A disk_excludes
 277 if ! $mkroot2 && ! $mkroot2tab && ! $mktab ! ifclass USE_MOUNTED; then
 278   ## ignore disks that are mounted, eg when running from fai-cd
 279   while read -r l; do
 280     eval "$l"
 281     if [[ ! $PKNAME ]]; then
 282       # shellcheck disable=SC2153 # not a misspelling
 283       PKNAME="$KNAME"
 284     fi
 285     if [[ $MOUNTPOINT ]]; then
 286       disk_excludes[$PKNAME]=true
 287     fi
 288   done < <(lsblk -nP -o KNAME,MOUNTPOINT,PKNAME)
 289 fi
 290
 291 hdds=()
 292 ssds=()
 293 # this excludes usb. note: i may encounter some other type in the future.
 294 for disk in $(lsblk -do name,tran -n | awk '$2 ~ "^(sata|nvme)$" { print $1 }'); do
 295   if [[ ${disk_excludes[$disk]} ]]; then
 296     continue
 297   fi
 298   case $(cat /sys/block/$disk/queue/rotational) in
 299     0) ssds+=("/dev/$disk") ;;
 300     1) hdds+=("/dev/$disk") ;;
 301     *) echo "$0: error: unknown /sys/block/$disk/queue/rotational: \
 302 $(cat $disk/queue/rotational)"; exit 1 ;;
 303   esac
 304 done
 305
 306 # install all ssds, or if there are none, all hdds.
 307 # Note, usb flash disks are seen as rotational, which is
 308 # very odd, but convenient for ignoring them here.
 309 # TODO: find a reliable way to ignore them.
 310 if ! ifclass ROTATIONAL && (( ${#ssds[@]} > 0 )); then
 311   read -ra short_devs<<<"${ssds[@]}"
 312 else
 313   read -ra short_devs<<<"${hdds[@]}"
 314 fi
 315
 316 # check if the partitions exist have the right filesystems
 317 #blkid="$(blkid -s TYPE)"
 318 for dev in ${short_devs[@]}; do
 319   if $partition; then break; fi
 320   y=$(readlink -f $dev)
 321   # shellcheck disable=SC2206 # globbing is intended
 322   arr=($y?*)
 323   if (( ${#arr[@]} < lastn )); then
 324     partition=true
 325   fi
 326   # On one system, blkid is missing some partitions.
 327   # maybe we need a flag, like FUZZY_BLKID or something, so we
 328   # can check that at least some exist.
 329   # for x in "`rootdev`: TYPE=\"crypto_LUKS\"" "`bootdev`: TYPE=\"btrfs\""; do
 330   #     echo "$blkid" | grep -Fx "$x" &>/dev/null || partition=true
 331   # done
 332 done
 333
 334 if $partition && ifclass PARTITION_PROMPT; then
 335   echo "Press any key except ctrl-c to continue and partition these drives:"
 336   echo "  ${short_devs[*]}"
 337   read -r
 338 fi
 339
 340 devs=()
 341 shopt -s extglob
 342 for short_dev in ${short_devs[@]}; do
 343   devs+=("$(devbyid $short_dev)")
 344 done
 345 if [[ ! ${devs[0]} ]]; then
 346   echo "$0: error: failed to detect devs" >&2
 347   exit 1
 348 fi
 349
 350 boot_space=0
 351 first=true
 352 boot_devs=()
 353 boot2_devs=()
 354 for dev in ${devs[@]}; do
 355   if ifclass frodo; then
 356     # I ran into a machine where the bios doesn't know about some disks,
 357     # so 1st stage of grub also doesn't know about them.
 358     # Also, grub does not support mounting degraded btrfs as far as
 359     # I can tell with some googling.
 360     # From within an arch install env, I could detect them by noting
 361     # their partitions were mixed with the next disk in /dev/disk/by-path,
 362     # and I have mixed model disks, and I could see the 8 models which showed
 363     # up in the bios, and thus see which 2 models were missing.
 364     # hdparm -I /dev/sdh will give model info in linux.
 365     # However, in fai on jessie, /dev/disk/by-path dir doesn't exist,
 366     # and I don't see another way, so I'm hardcoding them.
 367     # We still put grub on them and partition them the same, for uniformity
 368     # and in case they get moved to a system that can recognize them,
 369     # we just exclude them from the boot filesystem.
 370     cd /dev/disk/by-id/
 371     bad_disk=false
 372     for id in ata-TOSHIBA_MD04ACA500_8539K4TQFS9A \
 373                 ata-TOSHIBA_MD04ACA500_Y5IFK6IJFS9A; do
 374       if [[ $(readlink -f $id) == "$(readlink -f $dev)" ]]; then
 375         bad_disk=true
 376         break
 377       fi
 378     done
 379     if ! $bad_disk; then
 380       boot_devs+=("$(bootdev)")
 381       boot2_devs+=("$(boot2dev)")
 382     fi
 383   else
 384     boot_space=$(( boot_space + $(parted -m $dev unit MiB print | \
 385                                     sed -nr "s#^/dev/[^:]+:([0-9]+).*#\1#p") - 1))
 386     boot_devs+=("$(bootdev)")
 387     boot2_devs+=("$(boot2dev)")
 388   fi
 389   if $first && (( ${#boot_devs[@]} >= 1 )) ; then
 390     first_efi=$(efidev)
 391     first_grub_extdev=$(grub_extdev)
 392     first=false
 393   fi
 394 done
 395 first_boot_dev=${boot_devs[0]}
 396
 397 even_raid=false
 398 if ifclass RAID0 || (( ${#boot_devs[@]} == 1 )); then
 399   raid_level=0
 400 elif ifclass RAID1 || (( ${#boot_devs[@]} <= 3 )); then
 401   if (( ${#boot_devs[@]} == 2 )); then
 402     even_raid=true
 403   fi
 404   raid_level=1
 405 else
 406   raid_level=10
 407 fi
 408
 409
 410
 411 ### Begin calculate boot partition space
 412 # due to raid duplication
 413 case $raid_level in
 414   1*) boot_space=$(( boot_space / 2 )) ;;
 415 esac
 416 if (( boot_space > 60000 )); then
 417   # this is larger than needed for several /boot subvols,
 418   # becuase I keep a minimal debian install on it for
 419   # recovery needs and for doing pxe-kexec.
 420   boot_mib=10000
 421   root2_mib=1000000
 422   boot2_mib=5000
 423 elif (( boot_space > 30000 )); then
 424   boot_mib=$(( 5000 + (boot_space - 30000) / 2 ))
 425   root2_mib=100
 426   boot2_mib=100
 427 else
 428   # Small vms don't have room for /boot recovery.  With 3 kernels
 429   # installed, i'm using 132M on t8, so this seems like plenty of
 430   # room. note: rhel 8 recomments 1g for /boot.  u20.04 with 3 kernels =
 431   # 308 mb, so things have grown significantly
 432   boot_mib=1000
 433   root2_mib=100
 434   boot2_mib=100
 435 fi
 436 case $raid_level in
 437   1*)
 438     boot_mib=$(( boot_mib * 2 ))
 439     boot2_mib=$(( boot2_mib * 2 ))
 440     root2_mib=$(( root2_mib * 2 ))
 441     ;;
 442 esac
 443 ### end calculate boot partition space
 444
 445 bpart() { # btrfs a partition
 446   case $raid_level in
 447     0) mkfs.btrfs -f $@ ;;
 448     1) mkfs.btrfs -f -m raid1 -d raid1 $@ ;;
 449     10) mkfs.btrfs -f -m raid10 -d raid10 $@ ;;
 450   esac
 451 }
 452
 453
 454 if [[ ! $DISTRO ]]; then
 455   if ifclass VOL_BULLSEYE_BOOTSTRAP; then
 456     DISTRO=debianbullseye_bootstrap
 457   elif ifclass VOL_STRETCH; then
 458     DISTRO=debianstretch
 459   elif ifclass VOL_BUSTER; then
 460     DISTRO=debianbuster
 461   elif ifclass VOL_BULLSEYE; then
 462     DISTRO=debianbullseye
 463   elif ifclass VOL_BOOKWORM; then
 464     DISTRO=debianbookworm
 465   elif ifclass VOL_TESTING; then
 466     DISTRO=debiantesting
 467   elif ifclass VOL_XENIAL; then
 468     DISTRO=ubuntuxenial
 469   elif ifclass VOL_BIONIC; then
 470     DISTRO=ubuntubionic
 471   elif ifclass VOL_FOCAL; then
 472     DISTRO=ubuntufocal
 473   elif ifclass VOL_JAMMY; then
 474     DISTRO=ubuntujammy
 475   elif ifclass VOL_FLIDAS; then
 476     DISTRO=trisquelflidas
 477   elif ifclass VOL_ETIONA; then
 478     DISTRO=trisqueletiona
 479   elif ifclass VOL_NABIA; then
 480     DISTRO=trisquelnabia
 481   elif ifclass VOL_ARAMO; then
 482     DISTRO=trisquelaramo
 483   elif $mkroot2 || $mkroot2tab; then
 484     :
 485   else
 486     echo "PARTITIONER ERROR: no distro class/var set" >&2
 487     exit 1
 488   fi
 489 fi
 490
 491 if [[ $DISTRO == debianbullseye_bootstrap ]]; then
 492   # this is just convenience for the libreboot_grub config
 493   # so we can glob the other ones easier.
 494   boot_vol=$DISTRO
 495 else
 496   boot_vol=boot_$DISTRO
 497 fi
 498
 499
 500 first_root_crypt=$(root-cryptdev ${devs[0]})
 501
 502 # 1.5 x based on https://access.redhat.com/documentation/en-US/Red_Hat_Enterprise_Linux/7/html/Installation_Guide/sect-disk-partitioning-setup-x86.html#sect-custom-partitioning-x86
 503 swap_mib=$(( $(grep ^MemTotal: /proc/meminfo | \
 504                  awk '{print $2}') * 3/(${#devs[@]} * 2 ) / 1024 ))
 505
 506 root_devs=()
 507 for dev in ${devs[@]}; do
 508   root_devs+=("$(rootdev)")
 509 done
 510 shopt -s nullglob
 511 ##### end variable setup
 512
 513
 514
 515 mktab() {
 516   mkdir -p /tmp/fai
 517   dev=${boot_devs[0]}
 518   fstabstd="x-systemd.device-timeout=30s,x-systemd.mount-timeout=30s"
 519
 520   if [[ $DISTRO == debianbullseye_bootstrap ]]; then
 521     cat > /tmp/fai/fstab <<EOF
 522 $first_boot_dev  /  btrfs  noatime,subvol=$boot_vol  0 0
 523 $first_efi  /boot/efi  vfat          nofail,$fstabstd  0 0
 524 EOF
 525     cat >/tmp/fai/disk_var.sh <<EOF
 526 BOOT_DEVICE="${short_devs[@]}"
 527 ROOT_PARTITION=$first_boot_dev
 528 EOF
 529   else
 530     # note, fai creates the mountpoints listed here
 531     cat > /tmp/fai/fstab <<EOF
 532 $first_root_crypt  /  btrfs                 $fstabstd,noatime,subvol=root_$DISTRO$mopts  0 0
 533 $first_root_crypt  /mnt/root  btrfs  nofail,$fstabstd,noatime,subvolid=0$mopts  0 0
 534 $first_boot_dev  /boot  btrfs        nofail,$fstabstd,noatime,subvol=$boot_vol  0 0
 535 $first_efi  /boot/efi  vfat          nofail,$fstabstd  0 0
 536 $first_boot_dev  /mnt/boot  btrfs    nofail,$fstabstd,noatime,subvolid=0  0 0
 537 EOF
 538     swaps=()
 539     rm -f /tmp/fai/crypttab
 540     for dev in ${devs[@]}; do
 541       swaps+=("$(swap-cryptname)")
 542       cat >>/tmp/fai/crypttab <<EOF
 543 $(root-cryptname) $(rootdev)  none  keyscript=/root/keyscript,discard,luks,initramfs
 544 $(swap-cryptname) $(swapdev)  /dev/urandom  swap,cipher=aes-xts-plain64,size=256,hash=ripemd160
 545 EOF
 546       cat >> /tmp/fai/fstab <<EOF
 547 $(swap-cryptdev)  none  swap  nofail,$fstabstd,sw  0 0
 548 EOF
 549     done
 550
 551     # fai would do this:
 552     #BOOT_DEVICE=\${BOOT_DEVICE:-"${devs[0]}"}
 553
 554     # note: swaplist seems to do nothing.
 555     cat >/tmp/fai/disk_var.sh <<EOF
 556 BOOT_DEVICE="${short_devs[@]}"
 557 BOOT_PARTITION=\${BOOT_PARTITION:-$first_boot_dev}
 558 # ROOT_PARTITIONS is added by me, used in arch setup.
 559 ROOT_PARTITIONS="${root_devs[@]}"
 560 ROOT_PARTITION=\${ROOT_PARTITION:-$first_root_crypt}
 561 SWAPLIST=\${SWAPLIST:-"${swaps[@]}"}
 562 EOF
 563
 564     if [[ $HOSTNAME == kd ]]; then
 565       # note, having these with keyscript and initramfs causes a luks error in fai.log,
 566       # but it is safely ignorable and gets us the ability to just type our password
 567       # in once at boot. A downside is that they are probably needed to be plugged in to boot.
 568       cat >>/tmp/fai/crypttab <<EOF
 569 crypt_dev_ata-Samsung_SSD_870_QVO_8TB_S5VUNG0N900656V-part${even_bign} /dev/disk/by-id/ata-Samsung_SSD_870_QVO_8TB_S5VUNG0N900656V-part${even_bign}  none  keyscript=decrypt_keyctl,discard,luks,initramfs
 570 crypt_dev_ata-TOSHIBA_MD04ACA500_84R2K773FS9A-part1 /dev/disk/by-id/ata-TOSHIBA_MD04ACA500_84R2K773FS9A-part1  none  keyscript=decrypt_keyctl,discard,luks,initramfs
 571 crypt_dev_ata-ST6000DM001-1XY17Z_Z4D29EBL-part1 /dev/disk/by-id/ata-ST6000DM001-1XY17Z_Z4D29EBL-part1  none  keyscript=decrypt_keyctl,discard,luks,initramfs
 572 EOF
 573       cat >> /tmp/fai/fstab <<EOF
 574 # r7 = root partition7. it isnt actually #7 anymore, not a great name, but whatever
 575 /dev/mapper/crypt_dev_ata-Samsung_SSD_870_QVO_8TB_S5VUNG0N900656V-part${even_bign}  /mnt/r7  btrfs  nofail,$fstabstd,noatime,compress=zstd,subvolid=0  0 0
 576 /dev/mapper/crypt_dev_ata-TOSHIBA_MD04ACA500_84R2K773FS9A-part1  /mnt/rust1  btrfs  nofail,$fstabstd,noatime,compress=zstd,subvolid=0  0 0
 577 /dev/mapper/crypt_dev_ata-ST6000DM001-1XY17Z_Z4D29EBL-part1  /mnt/rust2  btrfs  nofail,$fstabstd,noatime,compress=zstd,subvolid=0  0 0
 578 EOF
 579     fi
 580   fi
 581 }
 582
 583
 584
 585 getluks() {
 586   if [[ ! $luks_dir ]]; then
 587     # see README for docs about how to create these
 588     luks_dir=$FAI/distro-install-common/luks
 589     if [[ ! -d $luks_dir ]]; then
 590       luks_dir=/q/root/luks
 591     fi
 592     if [[ ! -d $luks_dir ]]; then
 593       echo "$0: error: no luks_dir found" >&2
 594       exit 1
 595     fi
 596   fi
 597
 598   luks_file=$luks_dir/host-$HOSTNAME
 599   if [[ ! -e $luks_file ]]; then
 600     # shellcheck disable=SC2206 # globbing is intended
 601     hostkeys=($luks_dir/host-*)
 602     # if there is only one key, we might be deploying somewhere
 603     # where dhcp doesnt give us a proper hostname, so use that.
 604     if [[ ${#hostkeys[@]} == 1 && -e ${hostkeys[0]} ]]; then
 605       luks_file=${hostkeys[0]}
 606     else
 607       echo "$0: error: no key for hostname at $luks_file" >&2
 608       exit 1
 609     fi
 610   fi
 611
 612   # # note, corresponding changes in /b/ds/keyscript-{on,off}
 613   if ifclass demohost; then
 614     lukspw=x
 615   elif [[ -e $luks_dir/$HOSTNAME ]]; then
 616     lukspw=$(cat $luks_dir/$HOSTNAME)
 617   else
 618     lukspw=$(cat $luks_dir/iank)
 619   fi
 620
 621   if $mkroot2; then
 622     luks_file=$luks_dir/host-amy
 623     lukspw=$(cat $luks_dir/amy)
 624   fi
 625 }
 626
 627
 628 #### root2 non-fai run
 629 doroot2() {
 630
 631   # We write to these files instead of just /etc/fstab, /etc/crypttab,
 632   # because these are filesystems created after our current root, and so
 633   # this allows us to update other root filesystems too.
 634   rm -f /mnt/root/root2-{fs,crypt}tab
 635   if $partition; then
 636     echo $0: error: found partition=true but have mkroot2 arg
 637     exit 1
 638   fi
 639   for dev in ${devs[@]}; do
 640     if $mkroot2; then
 641       luks-setup $(root2dev)
 642     fi
 643     cat >>/mnt/root/root2-crypttab <<EOF
 644 $(root2-cryptname) $(root2dev)  $luks_file  discard,luks,initramfs
 645 EOF
 646   done
 647   if $mkroot2; then
 648     bpart $(for dev in ${devs[@]}; do root2-cryptdev; done)
 649     bpart ${boot2_devs[@]}
 650   fi
 651   mkdir -p /mnt/root2 /mnt/boot2
 652   cat >>/mnt/root/root2-fstab <<EOF
 653 $(root2-cryptdev ${devs[0]}) /mnt/root2  btrfs  nofail,x-systemd.device-timeout=30s,x-systemd.mount-timeout=30s,noatime,subvolid=0$mopts  0 0
 654 ${boot2_devs[0]} /mnt/boot2  btrfs    nofail,x-systemd.device-timeout=30s,x-systemd.mount-timeout=30s,noatime,subvolid=0  0 0
 655 EOF
 656   exit 0
 657 }
 658
 659 if $mkroot2 || $mkroot2tab; then
 660   getluks
 661   doroot2
 662 elif $mktab; then
 663   mktab
 664   exit 0
 665 else
 666   mktab
 667   getluks
 668 fi
 669
 670
 671 if $partition; then
 672   ### begin wipefs
 673   if [[ ! $SPECIAL_DISK ]]; then
 674     for dev in ${devs[@]}; do
 675       # if we repartition to the same as an old partition,
 676       # we don't want any old fses hanging around.
 677       for (( i=1; i <= lastn; i++ )); do
 678         x=$(add-part $i)
 679         [[ -e $x ]] || continue
 680         count_down=10
 681         # wipefs has failed, manual run works, google suggests timing issue
 682         while ! wipefs -a $x; do
 683           sleep 2
 684           count_down=$((count_down - 1))
 685           (( count_down > 0 )) || exit 1
 686         done
 687       done
 688     done
 689   fi
 690   ### end wipefs
 691
 692
 693   # When we have 2 disks of at least 100g difference in size,
 694   # make an extra partition on the end of the bigger one.
 695   even_big_part=false
 696   even_diff_min=100000
 697   if $even_raid; then
 698     smalli=0
 699     bigi=1
 700     if (( $(dev-mib ${devs[0]}) >= $(dev-mib ${devs[1]}) )); then
 701       smalli=1
 702       bigi=0
 703     fi
 704     disk_mib=$(dev-mib ${devs[smalli]})
 705     even_big_dev=${devs[bigi]}
 706     even_big_mib=$(dev-mib $even_big_dev)
 707     if (( even_big_mib - disk_mib > even_diff_min )); then
 708       even_big_part=true
 709     fi
 710   fi
 711
 712   for dev in ${devs[@]}; do
 713     if [[ $SPECIAL_DISK ]]; then
 714       dev=$(devbyid $SPECIAL_DISK)
 715     fi
 716
 717     # parted will round up the disk size. Do -1 so we can have
 718     # fully 1MiB unit partitions for easy resizing of the last partition.
 719     # Otherwise we would pass in -0 for the end argument for the last partition.
 720     #
 721     # Note: parted print error output is expected. example:
 722     # Error: /dev/vda: unrecognised disk label
 723     if ! $even_raid; then
 724       disk_mib=$(dev-mib)
 725     fi
 726
 727     boot_part_mib=$(( boot_mib / ${#boot_devs[@]} ))
 728     boot2_part_mib=$(( boot2_mib / ${#boot_devs[@]} ))
 729     root2_part_mib=$(( root2_mib / ${#root_devs[@]} ))
 730     root_end=$(( disk_mib - root2_part_mib - swap_mib - boot_part_mib - boot2_part_mib ))
 731     root2_end=$(( root_end + root2_part_mib ))
 732     swap_end=$(( root2_end + swap_mib ))
 733     boot_end=$(( swap_end + boot_part_mib ))
 734
 735     parted -s $dev mklabel gpt
 736     # MiB because parted complains about alignment otherwise.
 737     pcmd="parted -a optimal -s -- $dev"
 738     # root partition, the main big one
 739     $pcmd mkpart primary ext3 524MiB ${root_end}MiB
 740     # without naming, systemd gives us misc errors like:
 741     # dev-disk-by\x2dpartlabel-primary.device: Dev dev-disk-by\x2dpartlabel-primary.device appeared twice
 742     $pcmd name $rootn root
 743     # root2 partition
 744     $pcmd mkpart primary ext3 ${root_end}MiB ${root2_end}MiB
 745     $pcmd name $root2n root2
 746     # normally a swap is type "linux-swap", but this is encrypted swap. using that
 747     # label will confuse systemd.
 748     # swap partition
 749     $pcmd mkpart primary "" ${root2_end}MiB ${swap_end}MiB
 750     $pcmd name $swapn swap
 751     # boot partition
 752     $pcmd mkpart primary "" ${swap_end}MiB ${boot_end}MiB
 753     $pcmd name $bootn boot
 754     # boot2 partition
 755     $pcmd mkpart primary "" ${boot_end}MiB ${disk_mib}MiB
 756     $pcmd name $boot2n boot2
 757     # uefi partition. efi sucks, half a gig, rediculous.
 758     $pcmd mkpart primary "fat32" 12MiB 524MiB
 759     $pcmd name $efin efi
 760     $pcmd set $efin esp on
 761     # note, this is shown here: https://support.system76.com/articles/bootloader/
 762     # but not mentioned https://wiki.archlinux.org/index.php/EFI_system_partition
 763     # probably not needed
 764     $pcmd set $bootn boot on
 765     $pcmd set $boot2n boot on
 766     # i only need a few k, but googling min size,
 767     # I found someone saying that gparted required
 768     # required at least 8 because of their hard drive cylinder size.
 769     # And 8 is still very tiny.
 770     # grub_ext partition
 771     $pcmd mkpart primary "ext2" 4MiB 12MiB
 772     $pcmd name $grub_extn grubext
 773     # gpt ubuntu cloud image uses ~4 mb for this partition. fai uses 1 MiB.
 774     # so, I use 3, whatever.
 775     # note: parted manual saying cheap flash media
 776     # should to start at 4.
 777     # biols grub partition
 778     $pcmd mkpart primary "" 1MiB 4MiB
 779     $pcmd name $bios_grubn biosgrub
 780     $pcmd set $bios_grubn bios_grub on
 781     if $even_big_part  && [[ $dev == "$even_big_dev" ]]; then
 782       $pcmd mkpart primary ext3 ${disk_mib}MiB ${even_big_mib}MiB
 783       $pcmd name $even_bign even_big
 784     fi
 785
 786     # the mkfs failed before on a vm, which prompted me to add
 787     # sleep .1
 788     # then it failed again on a physical machine
 789     # with:
 790     # Device /dev/disk/by-id/foo doesn't exist or access denied,
 791     # so I added a wait until it existed.
 792     # Then I added the mkfs.ext2, which claimed to succeed,
 793     # but then couldn't be found upon reboot. In that case we didn't
 794     # wait at all. So I've added a 3 second minimum wait.
 795     secs=0
 796     while [[ ! -e $(bios_grubdev) ]] && (( secs < 10 )); do
 797       sleep 1
 798       secs=$((secs +1))
 799     done
 800     sleep 3
 801
 802     mkfs.fat -F32 $(efidev)
 803
 804     if $even_big_part  && [[ $dev == "$even_big_dev" ]]; then
 805       luks-setup $(even_bigdev)
 806       mkfs.btrfs -f $(crypt-dev $(even_bigdev))
 807     fi
 808
 809     # Holds just a single file, rarely written, so
 810     # use ext2, like was often used for the /boot partition.
 811     # This exists because grub can only persist data to a non-cow fs.
 812     # And we use persisting a var in grub to do a one time boot.
 813     # We could pass the data on the kernel command line and persist it
 814     # to grubenv after booting, but that relies on the boot always succeeding.
 815     # This is just a bit more robust, and it could work for booting
 816     # into ipxe which can't persist data, if we ever got that working.
 817     mkfs.ext2 $(grub_extdev)
 818     luks-setup $(rootdev)
 819
 820     if [[ $SPECIAL_DISK ]]; then
 821       exit 0
 822     fi
 823   done
 824   ls -la /dev/btrfs-control # this was probably for debugging...
 825   sleep 1
 826   bpart $(for dev in ${devs[@]}; do root-cryptdev; done)
 827   bpart ${boot_devs[@]}
 828 else
 829   for dev in ${devs[@]}; do
 830     if [[ -e /dev/mapper/$(root-cryptname) ]]; then
 831       continue
 832     fi
 833     if $rerootfs; then
 834       luks-setup $(rootdev)
 835       sleep 1
 836       bpart $(for dev in ${devs[@]}; do root-cryptdev; done)
 837     else
 838       cryptsetup luksOpen $(rootdev) $(root-cryptname) \
 839                  --key-file $luks_file
 840     fi
 841   done
 842
 843   if $rerootfs; then
 844     sleep 1
 845     bpart $(for dev in ${devs[@]}; do root-cryptdev; done)
 846   fi
 847   sleep 1
 848 fi
 849
 850
 851 if $wipe && [[ $DISTRO != debianbullseye_bootstrap ]]; then
 852   # bootstrap distro doesn't use separate encrypted root.
 853   mount -o subvolid=0 $first_root_crypt /mnt
 854   # systemd creates subvolumes we want to delete.
 855   mapfile -t s < <(btrfs subvolume list --sort=-path /mnt |
 856                      sed -rn "s#^.*path\s*(root_$DISTRO/\S+)\s*\$#\1#p")
 857   for subvol in ${s[@]}; do btrfs subvolume delete /mnt/$subvol; done
 858   btrfs subvolume set-default 0 /mnt
 859   [[ ! -e /mnt/root_$DISTRO ]] || btrfs subvolume delete /mnt/root_$DISTRO
 860
 861   ## create subvols ##
 862   cd /mnt
 863
 864   btrfs subvolume create root_$DISTRO
 865
 866   # could set default subvol like this, but no reason to.
 867   # btrfs subvolume set-default \
 868     #       $(btrfs subvolume list . | grep "root_$DISTRO$" | awk '{print $2}') .
 869
 870   # For raid systems, cow allows for error correction, for non-raid systems,
 871   # protects root fs from having the plug pulled. Reprovisioning a root
 872   # subvol is not my favorite thing to do.
 873   # # no cow on the root filesystem. it's setup is fully scripted,
 874   # # if it's messed up, we will just recreated it,
 875   # # and we can get better perf with this.
 876   # # I can't remember exactly why, but this is preferable to mounting with
 877   # # -o nodatacow, I think because subvolumes inherit that.
 878   # chattr -Rf +C root_$DISTRO
 879   cd /
 880   umount /mnt
 881 fi
 882
 883 mount -o subvolid=0 $first_boot_dev /mnt
 884 cd /mnt
 885 btrfs subvolume set-default 0 /mnt # already default, just ensuring it.
 886
 887 # for libreboot systems. grub2 only reads from subvolid=0
 888 mkdir -p /mnt/grub2
 889 cp $FAI/distro-install-common/libreboot_grub.cfg /mnt/grub2
 890
 891 if $wipe && [[ -e /mnt/$boot_vol ]]; then
 892   btrfs subvolume delete /mnt/$boot_vol
 893 fi
 894 if [[ ! -e /mnt/$boot_vol ]]; then
 895   btrfs subvolume create $boot_vol
 896 fi
 897 cd /
 898 umount /mnt
 899 ## end create subvols ##
 900
 901 mount $first_grub_extdev /mnt
 902 grub-editenv /mnt/grubenv set did_fai_check=true
 903 grub-editenv /mnt/grubenv set last_boot=/$boot_vol
 904 umount /mnt
 905
 906
 907 # initial setup of extra data fs, mounted,
 908 # btrfs subvol create nocow
 909 # chattr +C nocow
 910 # chown iank.iank nocow
 911