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 # Commented because it's not used, but left because it
  73 # finishes the pattern and if we ever do need to use it, it's here.
  74 #bios_grubdev() { add-part $@ $bios_grubn; }
  75
  76 crypt-dev() { echo /dev/mapper/crypt_dev_${1##*/}; }
  77 crypt-name() { echo crypt_dev_${1##*/}; }
  78 root-cryptdev() { crypt-dev $(rootdev $@); }
  79 swap-cryptdev() { crypt-dev $(swapdev $@); }
  80 root-cryptname() { crypt-name $(rootdev $@); }
  81 swap-cryptname() { crypt-name $(swapdev $@); }
  82
  83
  84 ##### end function defs
  85
  86 if ifclass REPARTITION;then
  87     partition=true # force a full wipe
  88 else
  89     partition=false # change to true to force a full wipe
  90 fi
  91
  92
  93
  94 hdds=()
  95 ssds=()
  96 cd /sys/block
  97 for disk in [sv]d[a-z]; do
  98     case $(cat $disk/queue/rotational) in
  99         0) ssds+=(/dev/$disk) ;;
 100         1) hdds+=(/dev/$disk) ;;
 101         *) echo "$0: error: unknown /sys/block/$disk/queue/rotational: \
 102 $(cat $disk/queue/rotational)"; exit 1 ;;
 103     esac
 104 done
 105
 106 # install all ssds, or if there are none, all hdds
 107 if ! ifclass ROTATIONAL && (( ${#ssds[@]} > 0 )); then
 108     short_devs=( ${ssds[@]} )
 109 else
 110     short_devs=( ${hdds[@]} )
 111 fi
 112
 113 # check if the partitions exist have the right filesystems
 114 #blkid="$(blkid -s TYPE)"
 115 for dev in ${short_devs[@]}; do
 116     if $partition; then break; fi
 117     y=$(readlink -f $dev)
 118     x=($y[0-9])
 119     [[ ${#x[@]}  == "${lastn}" ]] || partition=true
 120     for (( i=1; i <= lastn; i++ )); do
 121         [[ -e ${dev}$i ]] || partition=true
 122     done
 123     # On one system, blkid is missing some partitions.
 124     # maybe we need a flag, like FUZZY_BLKID or something, so we
 125     # can check that at least some exist.
 126     # for x in "`rootdev`: TYPE=\"crypto_LUKS\"" "`bootdev`: TYPE=\"btrfs\""; do
 127     #     echo "$blkid" | grep -Fx "$x" &>/dev/null || partition=true
 128     # done
 129 done
 130
 131 if $partition && ifclass PARTITION_PROMPT; then
 132     echo "Press any key except ctrl-c to continue and partition these drives:"
 133     echo "  ${short_devs[*]}"
 134     read -r
 135 fi
 136
 137 devs=()
 138 shopt -s extglob
 139 for short_dev in ${short_devs[@]}; do
 140     devs+=($(devbyid $short_dev))
 141 done
 142
 143
 144 first=false
 145 boot_devs=()
 146 for dev in ${devs[@]}; do
 147     if ifclass frodo; then
 148         # I ran into a machine where the bios doesn't know about some disks,
 149         # so 1st stage of grub also doesn't know about them.
 150         # Also, grub does not support mounting degraded btrfs as far as
 151         # I can tell with some googling.
 152         # From within an arch install env, I could detect them by noting
 153         # their partitions were mixed with the next disk in /dev/disk/by-path,
 154         # and I have mixed model disks, and I could see the 8 models which showed
 155         # up in the bios, and thus see which 2 models were missing.
 156         # hdparm -I /dev/sdh will give model info in linux.
 157         # However, in fai on jessie, /dev/disk/by-path dir doesn't exist,
 158         # and I don't see another way, so I'm hardcoding them.
 159         # We still put grub on them and partition them the same, for uniformity
 160         # and in case they get moved to a system that can recognize them,
 161         # we just exclude them from the boot filesystem.
 162         cd /dev/disk/by-id/
 163         bad_disk=false
 164         for id in ata-TOSHIBA_MD04ACA500_8539K4TQFS9A \
 165                       ata-TOSHIBA_MD04ACA500_Y5IFK6IJFS9A; do
 166             if [[ $(readlink -f $id) == "$(readlink -f $dev)" ]]; then
 167                 bad_disk=true
 168                 break
 169             fi
 170         done
 171         $bad_disk || boot_devs+=(`bootdev`)
 172     else
 173         boot_devs+=(`bootdev`)
 174     fi
 175     if [[ $boot_devs && $first ]]; then
 176         first_grub_extdev=`grub_extdev`
 177         first=false
 178     fi
 179 done
 180
 181 if ifclass RAID0 || (( ${#boot_devs[@]} < 4 )); then
 182     raid_level=0
 183 else
 184     raid_level=10
 185     # need double the space if we are raid 10, and then
 186     # might as well give some extra.
 187     boot_mib=$((boot_mib * 3))
 188 fi
 189
 190
 191
 192 if [[ ! $DISTRO ]]; then
 193     if ifclass STABLE_BOOTSTRAP; then
 194         DISTRO=debianstable_bootstrap
 195     elif ifclass STRETCH64; then
 196         DISTRO=debiantesting
 197     elif ifclass STABLE; then
 198         DISTRO=debianstable
 199     elif ifclass XENIAL64; then
 200         DISTRO=ubuntuxenial
 201     elif ifclass BELENOS64; then
 202         DISTRO=trisquelbelenos
 203     else
 204         echo "PARTITIONER ERROR: no distro class/var set" >&2
 205         exit 1
 206     fi
 207 fi
 208 first_boot_dev=${boot_devs[0]}
 209
 210
 211 bpart() { # btrfs a partition
 212     case $raid_level in
 213         0) mkfs.btrfs -f $@ ;;
 214         10) mkfs.btrfs -f -m raid10 -d raid10 $@ ;;
 215     esac
 216 }
 217
 218
 219 # keyfiles generated like:
 220 # head -c 2048 /dev/urandom | od | s dd of=/q/root/luks/host-demohost
 221 luks_dir=${LUKS_DIR:-/var/lib/fai/config/distro-install-common/luks}
 222
 223 if [[ ! -e $luks_dir/host-$HOSTNAME ]]; then
 224     echo "$0: error: no key for hostname at $luks_dir/host-$HOSTNAME" >&2
 225     exit 1
 226 fi
 227
 228 if ifclass tp; then
 229     lukspw=$(cat $luks_dir/traci)
 230 else
 231     lukspw=$(cat $luks_dir/ian)
 232 fi
 233 if ifclass demohost; then
 234     lukspw=x
 235 fi
 236
 237
 238 first_root_crypt=$(root-cryptdev ${devs[0]})
 239
 240 # 1.5 x based on https://access.redhat.com/documentation/en-US/Red_Hat_Enterprise_Linux/7/html/Installation_Guide/sect-disk-partitioning-setup-x86.html#sect-custom-partitioning-x86
 241 swap_mib=$(( $(grep ^MemTotal: /proc/meminfo | \
 242                    awk '{print $2}') * 3/(${#devs[@]} * 2 ) / 1024 ))
 243
 244 mkdir -p /tmp/fai
 245 root_devs=()
 246 for dev in ${devs[@]}; do
 247     root_devs+=(`rootdev`)
 248 done
 249 shopt -s nullglob
 250 if $partition; then
 251     for dev in ${devs[@]}; do
 252         # if we repartition to the same as an old partition,
 253         # we don't want any old fses hanging around.
 254         for (( i=1; i <= lastn; i++ )); do
 255             x=$(add-part $dev $i)
 256             [[ -e $x ]] || continue
 257             count_down=10
 258             # wipefs has failed, manual run works, google suggests timing issue
 259             while ! wipefs -a $x; do
 260                 sleep 2
 261                 count_down=$((count_down - 1))
 262                 (( count_down > 0 )) || exit 1
 263             done
 264         done
 265     done
 266     for dev in ${devs[@]}; do
 267         # parted will round up the disk size. Do -1 so we can have
 268         # fully 1MiB unit partitions for easy resizing of the last partition.
 269         # Otherwise we would pass in -0 for the end argument for the last partition.
 270         #
 271         # parted print error output is expected. example:
 272         # Error: /dev/vda: unrecognised disk label
 273         disk_mib=$(( $(parted -m $dev unit MiB print | \
 274                            sed -nr "s#^/dev/[^:]+:([0-9]+).*#\1#p") - 1))
 275         root_end=$(( disk_mib - swap_mib - boot_mib /  ${#boot_devs[@]} ))
 276         swap_end=$(( root_end + swap_mib))
 277
 278         parted -s $dev mklabel gpt
 279         # MiB because parted complains about alignment otherwise.
 280         pcmd="parted -a optimal -s -- $dev"
 281         $pcmd mkpart primary "ext3" 12MiB ${root_end}MiB
 282         $pcmd mkpart primary "linux-swap" ${root_end}MiB ${swap_end}MiB
 283         $pcmd mkpart primary "" ${swap_end}MiB ${disk_mib}MiB
 284         # i only need a few k, but googling min size,
 285         # I found someone saying that gparted required
 286         # required at least 8 because of their hard drive cylinder size.
 287         # And 8 is still very tiny.
 288         $pcmd mkpart primary "ext2" 4MiB 12MiB
 289         # gpt ubuntu cloud image uses ~4 mb for this partition. fai uses 1 MiB.
 290         # so, I use 3, whatever.
 291         # note: parted manual saying cheap flash media
 292         # should to start at 4.
 293         $pcmd mkpart primary "" 1MiB 4MiB
 294         $pcmd set $bios_grubn bios_grub on
 295         $pcmd set $bootn boot on # generally not needed on modern systems
 296         # the mkfs failed before on a vm, which prompted me to add
 297         # sleep .1
 298         # then it failed again on a physical machine
 299         # with:
 300         # Device /dev/disk/by-id/foo doesn't exist or access denied,
 301         # so I added a wait until it existed.
 302         # Then I added the mkfs.ext2, which claimed to succeed,
 303         # but then couldn't be found upon reboot. In that case we didn't
 304         # wait at all. So I've added a 3 second minimum wait.
 305         sleep 3
 306         secs=0
 307         while [[ ! -e `rootdev` ]] && (( secs < 10 )); do
 308             sleep 1
 309             secs=$((secs +1))
 310         done
 311         # Holds just a single file, rarely written, so
 312         # use ext2, like was often used for the /boot partition.
 313         # This exists because grub can only persist data to a non-cow fs.
 314         # And we use persisting a var in grub to do a one time boot.
 315         # We could pass the data on the kernel command line and persist it
 316         # to grubenv after booting, but that relies on the boot always succeeding.
 317         # This is just a bit more robust, and it could work for booting
 318         # into ipxe which can't persist data, if we ever got that working.
 319         mkfs.ext2 `grub_extdev`
 320         yes YES | cryptsetup luksFormat `rootdev` $luks_dir/host-$HOSTNAME \
 321                              -c aes-cbc-essiv:sha256 -s 256 || [[ $? == 141 ]]
 322         yes "$lukspw" | \
 323             cryptsetup luksAddKey --key-file $luks_dir/host-$HOSTNAME \
 324                        `rootdev` || [[ $? == 141 ]]
 325         # background: Keyfile and password are treated just
 326         # like 2 ways to input a passphrase, so we don't actually need to have
 327         # different contents of keyfile and passphrase, but it makes some
 328         # security sense to a really big randomly generated passphrase
 329         # as much as possible, so we have both.
 330         #
 331         # This would remove the keyfile.
 332         #    yes 'test' | cryptsetup luksRemoveKey /dev/... \
 333             #                            /key/file || [[ $? == 141 ]]
 334
 335         cryptsetup luksOpen `rootdev` `root-cryptname` \
 336                    --key-file $luks_dir/host-$HOSTNAME
 337     done
 338     ls -la /dev/btrfs-control # this was probably for debugging...
 339     sleep 1
 340     bpart $(for dev in ${devs[@]}; do root-cryptdev; done)
 341     bpart ${boot_devs[@]}
 342 else
 343     for dev in ${devs[@]}; do
 344         mkfs.ext2 `grub_extdev`
 345         cryptsetup luksOpen `rootdev` `root-cryptname` \
 346                    --key-file $luks_dir/host-$HOSTNAME
 347     done
 348     sleep 1
 349 fi
 350
 351
 352 if [[ $DISTRO != debianstable_bootstrap ]]; then
 353     # bootstrap distro doesn't use separate encrypted root.
 354     mount -o subvolid=0 $first_root_crypt /mnt
 355     # systemd creates subvolumes we want to delete.
 356     s=($(btrfs subvolume list --sort=-path /mnt |
 357              sed -rn "s#^.*path\s*(root_$DISTRO/\S+)\s*\$#\1#p"))
 358     for subvol in ${s[@]}; do btrfs subvolume delete /mnt/$subvol; done
 359     btrfs subvolume set-default 0 /mnt
 360     [[ ! -e /mnt/root_$DISTRO ]] || btrfs subvolume delete /mnt/root_$DISTRO
 361
 362     ## create subvols ##
 363     cd /mnt
 364
 365     btrfs subvolume create root_$DISTRO
 366     [[ -e nocow ]] || btrfs subvolume create nocow
 367     chown root:1000 nocow
 368     chattr +C nocow
 369
 370     mkdir -p /mnt/root_$DISTRO/boot
 371     # could set default subvol like this, but no reason to.
 372     # btrfs subvolume set-default \
 373         #       $(btrfs subvolume list . | grep "root_$DISTRO$" | awk '{print $2}') .
 374
 375     # no cow on the root filesystem. it's setup is fully scripted,
 376     # if it's messed up, we will just recreated it,
 377     # and we can get better perf with this.
 378     # I can't remember exactly why, but this is preferable to mounting with
 379     # -o nodatacow, I think because subvolumes inherit that.
 380     chattr -Rf +C root_$DISTRO
 381     cd /
 382     umount /mnt
 383 fi
 384
 385 mount -o subvolid=0 $first_boot_dev /mnt
 386 cd /mnt
 387 btrfs subvolume set-default 0 /mnt # already default, just ensuring it.
 388
 389 # for libreboot systems. grub2 only reads from subvolid=0
 390 mkdir -p /mnt/grub2
 391 cp $FAI/distro-install-common/libreboot_grub.cfg /mnt/grub2
 392
 393 if [[ $DISTRO == debianstable_bootstrap ]]; then
 394     # this is just convenience for the libreboot_grub config
 395     # so we can glob the other ones easier.
 396     boot_vol=$DISTRO
 397 else
 398     boot_vol=boot_$DISTRO
 399 fi
 400 [[ ! -e /mnt/$boot_vol ]] || btrfs subvolume delete /mnt/$boot_vol
 401 btrfs subvolume create $boot_vol
 402 cd /
 403 umount /mnt
 404 ## end create subvols ##
 405
 406 dev=${boot_devs[0]}
 407 mount $first_grub_extdev /mnt
 408 grub-editenv /mnt/grubenv set did_fai_check=true
 409 grub-editenv /mnt/grubenv set last_boot=/$boot_vol
 410 umount /mnt
 411
 412 if [[ $DISTRO == debianstable_bootstrap ]]; then
 413     cat > /tmp/fai/fstab <<EOF
 414 $first_boot_dev  /  btrfs  noatime,subvol=$boot_vol  0 0
 415 EOF
 416     cat >/tmp/fai/disk_var.sh <<EOF
 417 BOOT_DEVICE="${short_devs[@]}"
 418 ROOT_PARTITION=$first_boot_dev
 419 EOF
 420 else
 421     # note, fai creates the mountpoints like /nocow
 422     cat > /tmp/fai/fstab <<EOF
 423 $first_root_crypt  /  btrfs  noatime,subvol=root_$DISTRO  0 0
 424 $first_root_crypt  /nocow  btrfs  noatime,subvol=nocow  0 0
 425 $first_root_crypt  /mnt/root  btrfs  noatime,subvolid=0  0 0
 426 $first_boot_dev  /boot  btrfs  noatime,subvol=$boot_vol  0 0
 427 EOF
 428     swaps=()
 429     for dev in ${devs[@]}; do
 430         swaps+=(`swap-cryptname`)
 431         cat >>/tmp/fai/crypttab <<EOF
 432 `root-cryptname`  `rootdev`  none  keyscript=/root/keyscript,discard,luks
 433 `swap-cryptname` `swapdev`  /dev/urandom  swap,cipher=aes-xts-plain64,size=256,hash=ripemd160
 434 EOF
 435         cat >> /tmp/fai/fstab <<EOF
 436 `swap-cryptdev`  none  swap  sw  0 0
 437 EOF
 438     done
 439
 440     # fai would do this:
 441     #BOOT_DEVICE=\${BOOT_DEVICE:-"${devs[0]}"}
 442
 443     # note: swaplist seems to do nothing.
 444     cat >/tmp/fai/disk_var.sh <<EOF
 445 BOOT_DEVICE="${short_devs[@]}"
 446 BOOT_PARTITION=\${BOOT_PARTITION:-$first_boot_dev}
 447 # ROOT_PARTITIONS is added by me, used in arch setup.
 448 ROOT_PARTITIONS="${root_devs[@]}"
 449 ROOT_PARTITION=\${ROOT_PARTITION:-$first_root_crypt}
 450 SWAPLIST=\${SWAPLIST:-"${swaps[@]}"}
 451 EOF
 452 fi