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