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 # for calling outside of FAI, first
  22 # - source /a/bin/fai-wrapper
  23 # - set any appropriate classes with: fai-setclass OPT1... which sets CLASS_OPT1=true...
  24 #   or run eval-fai-classfile FILE
  25 #
  26 # OPTIONS:
  27 #
  28 # environment variables:
  29 #
  30 # HOSTNAME: if frodo, we exclude 2 devices from the /boot array, which
  31 # the bios does not see. if demohost, we set the luks password to just
  32 # 'x'.
  33 #
  34 # SPECIAL_DISK: For use outside of fai. A base disk name like
  35 # /dev/sdk. If set, we just cryptsetup and partition this one disk then
  36 # exit. This is useful for partitioning a disk in preparation to replace
  37 # a failed or failing disk from a raid10 array.
  38 #
  39 # classes:
  40 #
  41 # REPARTITION: forces repartitioning even if we detect the proper amount
  42 # of partitions already exist.
  43 #
  44 # ROTATIONAL: forces to install onto hdds instead of sdds. normally sdds
  45 # are chosen if they exist.
  46 #
  47 # PARTITION_PROMPT: command line prompt before partitioning
  48 #
  49 # RAID0: forces raid0 filesystem. Normally with 4+ devices, we use
  50 # raid10.
  51 # RAID1: forces raid1 filesystem.
  52
  53 if [[ $SPECIAL_DISK ]]; then
  54   export CLASS_REPARTITION=true
  55 fi
  56
  57 # # fai's setup-storage won't do btrfs on luks,
  58 # # so we do it ourself :)
  59 # inspiration taken from files in fai-setup-storage package
  60
  61 # if we are not running in fai, skiptask won't be defined, so carry on.
  62 skiptask partition || ! type skiptask
  63
  64 if ! type -p devbyid; then
  65   for d in $FAI/distro-install-common \
  66              /a/bin/fai/fai/config/distro-install-common $FAI $PWD; do
  67     [[ -d $d ]] || continue
  68     if [[ -e $d/devbyid ]]; then
  69       devbyid=$d/devbyid
  70       devbyid() { $devbyid "$@"; }
  71       break
  72     fi
  73   done
  74   if [[ ! $devbyid ]]; then
  75     echo "$0: error: failed to find devbyid script" >&2
  76     exit 1
  77   fi
  78 fi
  79
  80
  81
  82 #### begin configuration
  83
  84 rootn=1
  85 swapn=2
  86 bootn=3
  87 # ext partition so grub can write persistent variables,
  88 # so it can do a one time boot. grub can't write to
  89 # btrfs or any cow fs because it's more
  90 # more complicated to do and they don't want to.
  91 grub_extn=4
  92 # bios boot partition,
  93 # https://wiki.archlinux.org/index.php/GRUB
  94 bios_grubn=5
  95 lastn=$bios_grubn
  96 # this is larger than needed for several /boot subvols,
  97 # becuase I keep a minimal debian install on it, for
  98 # recovery needs, and for doing pxe-kexec.
  99 boot_mib=10000
 100
 101
 102 ##### end configuration
 103
 104
 105 add-part() { # add partition suffix to $dev
 106   local d ret
 107   if [[ $# == 1 ]]; then
 108     d=$dev
 109     part=$1
 110   else
 111     d=$1
 112     part=$2
 113   fi
 114   if [[ $d == /dev/disk/by-id/* ]]; then
 115     ret=$d-part$part
 116   else
 117     ret=$d$part
 118   fi
 119   echo $ret
 120 }
 121
 122 bootdev() { add-part $@ $bootn; }
 123 rootdev() { add-part $@ $rootn; }
 124 swapdev() { add-part $@ $swapn; }
 125 grub_extdev() { add-part $@ $grub_extn; }
 126 bios_grubdev() { add-part $@ $bios_grubn; }
 127
 128 crypt-dev() { echo /dev/mapper/crypt_dev_${1##*/}; }
 129 crypt-name() { echo crypt_dev_${1##*/}; }
 130 root-cryptdev() { crypt-dev $(rootdev $@); }
 131 swap-cryptdev() { crypt-dev $(swapdev $@); }
 132 root-cryptname() { crypt-name $(rootdev $@); }
 133 swap-cryptname() { crypt-name $(swapdev $@); }
 134
 135 ##### end function defs
 136
 137 if ifclass REPARTITION;then
 138   partition=true # force a full wipe
 139 else
 140   partition=false # change to true to force a full wipe
 141 fi
 142
 143
 144
 145 hdds=()
 146 ssds=()
 147 cd /sys/block
 148 for disk in [sv]d[a-z]; do
 149   case $(cat $disk/queue/rotational) in
 150     0) ssds+=(/dev/$disk) ;;
 151     1) hdds+=(/dev/$disk) ;;
 152     *) echo "$0: error: unknown /sys/block/$disk/queue/rotational: \
 153 $(cat $disk/queue/rotational)"; exit 1 ;;
 154   esac
 155 done
 156
 157 # install all ssds, or if there are none, all hdds.
 158 # Note, usb flash disks are seen as rotational, which is
 159 # very odd, but convenient for ignoring them here.
 160 # TODO: find a reliable way to ignore them.
 161 if ! ifclass ROTATIONAL && (( ${#ssds[@]} > 0 )); then
 162   short_devs=( ${ssds[@]} )
 163 else
 164   short_devs=( ${hdds[@]} )
 165 fi
 166
 167 # check if the partitions exist have the right filesystems
 168 #blkid="$(blkid -s TYPE)"
 169 for dev in ${short_devs[@]}; do
 170   if $partition; then break; fi
 171   y=$(readlink -f $dev)
 172   arr=($y[0-9])
 173   [[ ${#arr[@]}  == "${lastn}" ]] || partition=true
 174   for (( i=1; i <= lastn; i++ )); do
 175     [[ -e ${dev}$i ]] || partition=true
 176   done
 177   # On one system, blkid is missing some partitions.
 178   # maybe we need a flag, like FUZZY_BLKID or something, so we
 179   # can check that at least some exist.
 180   # for x in "`rootdev`: TYPE=\"crypto_LUKS\"" "`bootdev`: TYPE=\"btrfs\""; do
 181   #     echo "$blkid" | grep -Fx "$x" &>/dev/null || partition=true
 182   # done
 183 done
 184
 185 if $partition && ifclass PARTITION_PROMPT; then
 186   echo "Press any key except ctrl-c to continue and partition these drives:"
 187   echo "  ${short_devs[*]}"
 188   read -r
 189 fi
 190
 191 devs=()
 192 shopt -s extglob
 193 for short_dev in ${short_devs[@]}; do
 194   devs+=($(devbyid $short_dev))
 195 done
 196 if [[ ! ${devs[@]} ]]; then
 197   echo "$0: error: failed to detect devs" >&2
 198   exit 1
 199 fi
 200
 201
 202 first=false
 203 boot_devs=()
 204 for dev in ${devs[@]}; do
 205   if ifclass frodo; then
 206     # I ran into a machine where the bios doesn't know about some disks,
 207     # so 1st stage of grub also doesn't know about them.
 208     # Also, grub does not support mounting degraded btrfs as far as
 209     # I can tell with some googling.
 210     # From within an arch install env, I could detect them by noting
 211     # their partitions were mixed with the next disk in /dev/disk/by-path,
 212     # and I have mixed model disks, and I could see the 8 models which showed
 213     # up in the bios, and thus see which 2 models were missing.
 214     # hdparm -I /dev/sdh will give model info in linux.
 215     # However, in fai on jessie, /dev/disk/by-path dir doesn't exist,
 216     # and I don't see another way, so I'm hardcoding them.
 217     # We still put grub on them and partition them the same, for uniformity
 218     # and in case they get moved to a system that can recognize them,
 219     # we just exclude them from the boot filesystem.
 220     cd /dev/disk/by-id/
 221     bad_disk=false
 222     for id in ata-TOSHIBA_MD04ACA500_8539K4TQFS9A \
 223                 ata-TOSHIBA_MD04ACA500_Y5IFK6IJFS9A; do
 224       if [[ $(readlink -f $id) == "$(readlink -f $dev)" ]]; then
 225         bad_disk=true
 226         break
 227       fi
 228     done
 229     $bad_disk || boot_devs+=(`bootdev`)
 230   else
 231     boot_devs+=(`bootdev`)
 232   fi
 233   if [[ $boot_devs && $first ]]; then
 234     first_grub_extdev=`grub_extdev`
 235     first=false
 236   fi
 237 done
 238
 239 if ifclass RAID0 || (( ${#boot_devs[@]} == 1 )); then
 240   raid_level=0
 241 elif ifclass RAID1 || (( ${#boot_devs[@]} <= 3 )); then
 242   raid_level=1
 243 else
 244   raid_level=10
 245   # need double the space if we are raid 10, and then
 246   # might as well give some extra.
 247   boot_mib=$((boot_mib * 3))
 248 fi
 249
 250
 251
 252 if [[ ! $DISTRO ]]; then
 253   if ifclass VOL_STRETCH_BOOTSTRAP; then
 254     DISTRO=debianstretch_bootstrap
 255   elif ifclass VOL_STRETCH; then
 256     DISTRO=debianstretch
 257   elif ifclass VOL_TESTING; then
 258     DISTRO=debiantesting
 259   elif ifclass VOL_XENIAL; then
 260     DISTRO=ubuntuxenial
 261   elif ifclass VOL_BELENOS; then
 262     DISTRO=trisquelbelenos
 263   elif ifclass VOL_FLIDAS; then
 264     DISTRO=trisquelflidas
 265   else
 266     echo "PARTITIONER ERROR: no distro class/var set" >&2
 267     exit 1
 268   fi
 269 fi
 270 first_boot_dev=${boot_devs[0]}
 271
 272
 273 bpart() { # btrfs a partition
 274   case $raid_level in
 275     0) mkfs.btrfs -f $@ ;;
 276     1) mkfs.btrfs -f -m raid1 -d raid1 $@ ;;
 277     10) mkfs.btrfs -f -m raid10 -d raid10 $@ ;;
 278   esac
 279 }
 280
 281
 282 # keyfiles generated like:
 283 # head -c 2048 /dev/urandom | od | s dd of=/q/root/luks/host-demohost
 284 luks_dir=${LUKS_DIR:-/var/lib/fai/config/distro-install-common/luks}
 285
 286 if [[ ! -e $luks_dir/host-$HOSTNAME ]]; then
 287   echo "$0: error: no key for hostname at $luks_dir/host-$HOSTNAME" >&2
 288   exit 1
 289 fi
 290
 291 lukspw=$(cat $luks_dir/iank)
 292 # # ian: disabled by chaning to tpnew while I use the tp host.
 293 # # note, corresponding changes in /b/ds/keyscript-{on,off}
 294 if ifclass tpnew; then
 295   lukspw=$(cat $luks_dir/traci)
 296 fi
 297 if ifclass demohost; then
 298   lukspw=x
 299 fi
 300
 301
 302 first_root_crypt=$(root-cryptdev ${devs[0]})
 303
 304 # 1.5 x based on https://access.redhat.com/documentation/en-US/Red_Hat_Enterprise_Linux/7/html/Installation_Guide/sect-disk-partitioning-setup-x86.html#sect-custom-partitioning-x86
 305 swap_mib=$(( $(grep ^MemTotal: /proc/meminfo | \
 306                  awk '{print $2}') * 3/(${#devs[@]} * 2 ) / 1024 ))
 307
 308 mkdir -p /tmp/fai
 309 root_devs=()
 310 for dev in ${devs[@]}; do
 311   root_devs+=(`rootdev`)
 312 done
 313 shopt -s nullglob
 314 if $partition; then
 315
 316   ### begin wipefs
 317   if [[ ! $SPECIAL_DISK ]]; then
 318     for dev in ${devs[@]}; do
 319       # if we repartition to the same as an old partition,
 320       # we don't want any old fses hanging around.
 321       for (( i=1; i <= lastn; i++ )); do
 322         x=$(add-part $dev $i)
 323         [[ -e $x ]] || continue
 324         count_down=10
 325         # wipefs has failed, manual run works, google suggests timing issue
 326         while ! wipefs -a $x; do
 327           sleep 2
 328           count_down=$((count_down - 1))
 329           (( count_down > 0 )) || exit 1
 330         done
 331       done
 332     done
 333   fi
 334   ### end wipefs
 335
 336   for dev in ${devs[@]}; do
 337     if [[ $SPECIAL_DISK ]]; then
 338       dev=$(devbyid $SPECIAL_DISK)
 339     fi
 340
 341     # parted will round up the disk size. Do -1 so we can have
 342     # fully 1MiB unit partitions for easy resizing of the last partition.
 343     # Otherwise we would pass in -0 for the end argument for the last partition.
 344     #
 345     # Note: parted print error output is expected. example:
 346     # Error: /dev/vda: unrecognised disk label
 347     disk_mib=$(( $(parted -m $dev unit MiB print | \
 348                      sed -nr "s#^/dev/[^:]+:([0-9]+).*#\1#p") - 1))
 349     root_end=$(( disk_mib - swap_mib - boot_mib /  ${#boot_devs[@]} ))
 350     swap_end=$(( root_end + swap_mib))
 351
 352     parted -s $dev mklabel gpt
 353     # MiB because parted complains about alignment otherwise.
 354     pcmd="parted -a optimal -s -- $dev"
 355     $pcmd mkpart primary ext3 12MiB ${root_end}MiB
 356     # without naming, systemd gives us misc errors like:
 357     # dev-disk-by\x2dpartlabel-primary.device: Dev dev-disk-by\x2dpartlabel-primary.device appeared twice
 358     $pcmd name $rootn root
 359     # normally a swap is type "linux-swap", but this is encrypted swap. using that
 360     # label will confuse systemd.
 361     $pcmd mkpart primary "" ${root_end}MiB ${swap_end}MiB
 362     $pcmd name $swapn swap
 363     $pcmd mkpart primary "" ${swap_end}MiB ${disk_mib}MiB
 364     $pcmd name $bootn boot
 365     # i only need a few k, but googling min size,
 366     # I found someone saying that gparted required
 367     # required at least 8 because of their hard drive cylinder size.
 368     # And 8 is still very tiny.
 369     $pcmd mkpart primary "ext2" 4MiB 12MiB
 370     $pcmd name $grub_extn grubext
 371     # gpt ubuntu cloud image uses ~4 mb for this partition. fai uses 1 MiB.
 372     # so, I use 3, whatever.
 373     # note: parted manual saying cheap flash media
 374     # should to start at 4.
 375     $pcmd mkpart primary "" 1MiB 4MiB
 376     $pcmd name $bios_grubn biosgrub
 377     $pcmd set $bios_grubn bios_grub on
 378     $pcmd set $bootn boot on # generally not needed on modern systems
 379     # the mkfs failed before on a vm, which prompted me to add
 380     # sleep .1
 381     # then it failed again on a physical machine
 382     # with:
 383     # Device /dev/disk/by-id/foo doesn't exist or access denied,
 384     # so I added a wait until it existed.
 385     # Then I added the mkfs.ext2, which claimed to succeed,
 386     # but then couldn't be found upon reboot. In that case we didn't
 387     # wait at all. So I've added a 3 second minimum wait.
 388     sleep 3
 389     secs=0
 390     while [[ ! -e `rootdev` ]] && (( secs < 10 )); do
 391       sleep 1
 392       secs=$((secs +1))
 393     done
 394     # Holds just a single file, rarely written, so
 395     # use ext2, like was often used for the /boot partition.
 396     # This exists because grub can only persist data to a non-cow fs.
 397     # And we use persisting a var in grub to do a one time boot.
 398     # We could pass the data on the kernel command line and persist it
 399     # to grubenv after booting, but that relies on the boot always succeeding.
 400     # This is just a bit more robust, and it could work for booting
 401     # into ipxe which can't persist data, if we ever got that working.
 402     mkfs.ext2 `grub_extdev`
 403     # when we move to newer than trisquel 8, we can remove
 404     # --type luks1. We can also check on cryptsetup --help | less /compil
 405     # to see about the other settings. Default in debian 9 is luks2.
 406     # You can convert from luks2 to luks 1 by adding a temporary key:
 407     # cryptsetup luksAddKey --pbkdf pbkdf2
 408     # then remove the new format keys with cryptsetup luksRemoveKey
 409     # then cryptsetup convert DEV --type luks1, then readd old keys and remove temp.
 410     yes YES | cryptsetup luksFormat `rootdev` $luks_dir/host-$HOSTNAME \
 411                          --type luks1 -c aes-cbc-essiv:sha256 -s 256 || [[ $? == 141 ]]
 412     yes "$lukspw" | \
 413       cryptsetup luksAddKey --key-file $luks_dir/host-$HOSTNAME \
 414                  `rootdev` || [[ $? == 141 ]]
 415     # background: Keyfile and password are treated just
 416     # like 2 ways to input a passphrase, so we don't actually need to have
 417     # different contents of keyfile and passphrase, but it makes some
 418     # security sense to a really big randomly generated passphrase
 419     # as much as possible, so we have both.
 420     #
 421     # This would remove the keyfile.
 422     #    yes 'test' | cryptsetup luksRemoveKey /dev/... \
 423       #                            /key/file || [[ $? == 141 ]]
 424
 425     cryptsetup luksOpen `rootdev` `root-cryptname` \
 426                --key-file $luks_dir/host-$HOSTNAME
 427
 428     if [[ $SPECIAL_DISK ]]; then
 429       exit 0
 430     fi
 431   done
 432   ls -la /dev/btrfs-control # this was probably for debugging...
 433   sleep 1
 434   bpart $(for dev in ${devs[@]}; do root-cryptdev; done)
 435   bpart ${boot_devs[@]}
 436 else
 437   for dev in ${devs[@]}; do
 438     mkfs.ext2 `grub_extdev`
 439     cryptsetup luksOpen `rootdev` `root-cryptname` \
 440                --key-file $luks_dir/host-$HOSTNAME
 441   done
 442   sleep 1
 443 fi
 444
 445
 446 if [[ $DISTRO != debianstretch_bootstrap ]]; then
 447   # bootstrap distro doesn't use separate encrypted root.
 448   mount -o subvolid=0 $first_root_crypt /mnt
 449   # systemd creates subvolumes we want to delete.
 450   s=($(btrfs subvolume list --sort=-path /mnt |
 451          sed -rn "s#^.*path\s*(root_$DISTRO/\S+)\s*\$#\1#p"))
 452   for subvol in ${s[@]}; do btrfs subvolume delete /mnt/$subvol; done
 453   btrfs subvolume set-default 0 /mnt
 454   [[ ! -e /mnt/root_$DISTRO ]] || btrfs subvolume delete /mnt/root_$DISTRO
 455
 456   ## create subvols ##
 457   cd /mnt
 458
 459   btrfs subvolume create root_$DISTRO
 460
 461   mkdir -p /mnt/root_$DISTRO/boot
 462   # could set default subvol like this, but no reason to.
 463   # btrfs subvolume set-default \
 464     #       $(btrfs subvolume list . | grep "root_$DISTRO$" | awk '{print $2}') .
 465
 466   # no cow on the root filesystem. it's setup is fully scripted,
 467   # if it's messed up, we will just recreated it,
 468   # and we can get better perf with this.
 469   # I can't remember exactly why, but this is preferable to mounting with
 470   # -o nodatacow, I think because subvolumes inherit that.
 471   chattr -Rf +C root_$DISTRO
 472   cd /
 473   umount /mnt
 474 fi
 475
 476 mount -o subvolid=0 $first_boot_dev /mnt
 477 cd /mnt
 478 btrfs subvolume set-default 0 /mnt # already default, just ensuring it.
 479
 480 # for libreboot systems. grub2 only reads from subvolid=0
 481 mkdir -p /mnt/grub2
 482 cp $FAI/distro-install-common/libreboot_grub.cfg /mnt/grub2
 483
 484 if [[ $DISTRO == debianstretch_bootstrap ]]; then
 485   # this is just convenience for the libreboot_grub config
 486   # so we can glob the other ones easier.
 487   boot_vol=$DISTRO
 488 else
 489   boot_vol=boot_$DISTRO
 490 fi
 491 [[ ! -e /mnt/$boot_vol ]] || btrfs subvolume delete /mnt/$boot_vol
 492 btrfs subvolume create $boot_vol
 493 cd /
 494 umount /mnt
 495 ## end create subvols ##
 496
 497 dev=${boot_devs[0]}
 498 mount $first_grub_extdev /mnt
 499 grub-editenv /mnt/grubenv set did_fai_check=true
 500 grub-editenv /mnt/grubenv set last_boot=/$boot_vol
 501 umount /mnt
 502
 503 if [[ $DISTRO == debianstretch_bootstrap ]]; then
 504   cat > /tmp/fai/fstab <<EOF
 505 $first_boot_dev  /  btrfs  noatime,subvol=$boot_vol  0 0
 506 EOF
 507   cat >/tmp/fai/disk_var.sh <<EOF
 508 BOOT_DEVICE="${short_devs[@]}"
 509 ROOT_PARTITION=$first_boot_dev
 510 EOF
 511 else
 512   # note, fai creates the mountpoints listed here
 513   cat > /tmp/fai/fstab <<EOF
 514 $first_root_crypt  /  btrfs  noatime,subvol=root_$DISTRO  0 0
 515 $first_root_crypt  /mnt/root  btrfs  noatime,subvolid=0  0 0
 516 $first_boot_dev  /boot  btrfs  noatime,subvol=$boot_vol  0 0
 517 EOF
 518   swaps=()
 519   for dev in ${devs[@]}; do
 520     swaps+=(`swap-cryptname`)
 521     cat >>/tmp/fai/crypttab <<EOF
 522 `root-cryptname`  `rootdev`  none  keyscript=/root/keyscript,discard,luks
 523 `swap-cryptname` `swapdev`  /dev/urandom  swap,cipher=aes-xts-plain64,size=256,hash=ripemd160
 524 EOF
 525     cat >> /tmp/fai/fstab <<EOF
 526 `swap-cryptdev`  none  swap  sw  0 0
 527 EOF
 528   done
 529
 530   # fai would do this:
 531   #BOOT_DEVICE=\${BOOT_DEVICE:-"${devs[0]}"}
 532
 533   # note: swaplist seems to do nothing.
 534   cat >/tmp/fai/disk_var.sh <<EOF
 535 BOOT_DEVICE="${short_devs[@]}"
 536 BOOT_PARTITION=\${BOOT_PARTITION:-$first_boot_dev}
 537 # ROOT_PARTITIONS is added by me, used in arch setup.
 538 ROOT_PARTITIONS="${root_devs[@]}"
 539 ROOT_PARTITION=\${ROOT_PARTITION:-$first_root_crypt}
 540 SWAPLIST=\${SWAPLIST:-"${swaps[@]}"}
 541 EOF
 542 fi