Merge pull request #3486 from balena-os/mtoman/cryptsetup-fail-hard

initrdscripts: Make cryptsetup fail hard in unexpected conditions

meta-balena-common/recipes-core/initrdscripts/files/cryptsetup-efi-tpm

@@ -1,20 +1,44 @@
 #!/bin/sh
 
+# This script unlocks encrypted partitions on UEFI systems
+# with secure boot enabled and encryption key protected by a TPM2.0
+#
+# The partition layout of such system is:
+# sda         // Boot device
+# |-sda1      // Unencrypted EFI partition, contains the bootloader and encrypted disk passphrase
+# |-sda2      // Encrypted boot partition
+# | `-dm-0    // Unlocked boot partition, contains balena-specifics, e.g. config.json
+# |-sda3      // Encrypted rootA partition
+# | `-dm-1    // Unlocked rootA partition
+# |-sda4      // Encrypted rootB partition
+# | `-dm-2    // Unlocked rootB partition
+# |-sda5      // Encrypted state partition
+# | `-dm-3    // Unlocked state partition
+# `-sda6      // Encrypted data partition
+#   `-dm-4    // Unlocked data partition
+#
+# When this script executes, it should only find the encrypted partitions.
+# The script unlocks the partitions using the metadata stored in the EFI
+# partition in combination with secrets stored in the TPM.
+# As a result the dm-0..dm-5 devices are created.
+#
+# After successfully unlocking a partition, the default udev rules will
+# in fact create two instances of the unlocked device:
+# * /dev/dm-X - the "dm-X" part is referred to as "KNAME".
+#   This is what the kernel uses internally to identify the device
+# * /dev/mapper/luks-$UUID - the "luks-$UUID" part is referred to as "NAME".
+#   This is a user-defined name, we chose "luks-$UUID", the same as Fedora.
+# The two devices are identical and can be used interchangeably, but this
+# script tries to prefer KNAME as it seems more practical for scripting.
+
 # shellcheck disable=SC1091
 . /usr/libexec/os-helpers-logging
 . /usr/libexec/os-helpers-fs
 . /usr/libexec/os-helpers-tpm2
 . /usr/libexec/os-helpers-efi
 . /usr/sbin/balena-config-defaults
 
-# Give a chance to the by-state directory to appear
-# We do not expect any particular device or partition to come up
-# but if balenaOS is correctly configured on the device the by-state
-# directory will eventually be created by the custom udev rule.
-# This is useful if the rootfs is on a device that takes a while
-# to initialize such as a USB disk.
-wait4file "/dev/disk/by-state" "50"
-EFI_DEV=$(get_state_path_from_label "${BALENA_NONENC_BOOT_LABEL}")
+ENCRYPTED_PARTITIONS="boot rootA rootB state data"
 
 ensure_luks2() {
     LUKS_DEVICE="$1"
@@ -35,16 +59,7 @@ cryptsetup_enabled() {
 
     # Ensure that secure boot is enabled and in user mode before unlocking
     if ! user_mode_enabled; then
-	    return 1
-    fi
-
-    # Only run if the EFI partition is split
-    if [ ! -e "$EFI_DEV" ]; then
-        return 1
-    fi
-
-    # Check whether there are any LUKS partitions
-    if ! lsblk -nlo fstype | grep -q crypto_LUKS; then
+        info "Won't attempt to decrypt drives because secure boot is not enabled"
         return 1
     fi
 
@@ -55,6 +70,28 @@ cryptsetup_run() {
     # Die if anything fails here
     set -e
 
+    # Make sure there is only a single unencrypted partition.
+    # We will blindly mount it here with no further authentication,
+    # so if there are ambiguities, we prefer to bail out.
+    NONENC_PARTS="$(lsblk -nlo label | grep "${BALENA_NONENC_BOOT_LABEL}")"
+    if [ -z "${NONENC_PARTS}" ]; then
+        fail "Partition '${BALENA_NONENC_BOOT_LABEL}' not found"
+    fi
+
+    if [ "$(echo "${NONENC_PARTS}" | wc -l)" -gt 1 ]; then
+        fail "Multiple '${BALENA_NONENC_BOOT_LABEL}' partitions found"
+    fi
+
+    EFI_DEV=$(get_state_path_from_label "${BALENA_NONENC_BOOT_LABEL}")
+    if [ ! -e "$EFI_DEV" ]; then
+        fail "Won't attempt to decrypt drives because the EFI partition is not split"
+    fi
+
+    # Check whether there are any LUKS partitions
+    if ! lsblk -nlo fstype | grep -q crypto_LUKS; then
+        fail "There are no encrypted partitions"
+    fi
+
     EFI_MOUNT_DIR="/efi"
     mkdir "$EFI_MOUNT_DIR"
     mount "$EFI_DEV" "$EFI_MOUNT_DIR"
@@ -86,14 +123,30 @@ cryptsetup_run() {
 
     BOOT_DEVICE=$(lsblk -nlo pkname "${EFI_DEV}")
 
+    # Check that we have the expected amount of encrypted partitions on the boot device
+    LUKS_PARTITIONS=$(lsblk -nlo "kname,uuid,fstype,partlabel" "/dev/${BOOT_DEVICE}" | grep "crypto_LUKS")
+    if [ "$(echo "${LUKS_PARTITIONS}" | wc -l)" != "$(echo "${ENCRYPTED_PARTITIONS}" | wc -w)" ]; then
+        fail "An unexpected amount of encrypted partitions was found"
+    fi
+
     # Unlock all the partitions - cryptsetup luksOpen does not wait for udev processing
     # of the DM device to complete, it just kicks off the process and returns.
     # Since this is async, we can perform all the luksOpens here, note the device names
     # and wait for them in a separate loop later
     LUKS_UNLOCKED=""
-    for LUKS_UUID in $(lsblk -nlo uuid,fstype "/dev/${BOOT_DEVICE}" | grep crypto_LUKS | cut -d " " -f 1); do
+    for PART_NAME in ${ENCRYPTED_PARTITIONS}; do
+        LUKS_UUID=$(echo "${LUKS_PARTITIONS}" | grep " \(balena\|resin\)-${PART_NAME}$" | cut -d " " -f 2)
+
+        if [ -z "${LUKS_UUID}" ]; then
+            fail "Partition '${PART_NAME}' not found"
+        fi
+
+        if [ "$(echo "${LUKS_UUID}" | wc -l)" -gt 1 ]; then
+            fail "More than one '${PART_NAME}' partition found"
+        fi
+
         ensure_luks2 "/dev/disk/by-uuid/${LUKS_UUID}"
-        cryptsetup luksOpen --key-file $PASSPHRASE_FILE UUID="${LUKS_UUID}" luks-"${LUKS_UUID}"
+        cryptsetup luksOpen --key-file "${PASSPHRASE_FILE}" "UUID=${LUKS_UUID}" "luks-${LUKS_UUID}"
         LUKS_UNLOCKED="${LUKS_UNLOCKED} luks-${LUKS_UUID}"
     done
 
@@ -102,6 +155,56 @@ cryptsetup_run() {
         wait4udev "/dev/mapper/${DM_NAME}"
     done
 
+    # Perform sanity checks after unlocking.
+    # We know what the system should look like after the partitions are unlocked.
+    # We want to make sure that the newly unlocked partitions are the ones
+    # we are going to actually use, there is nothing missing and nothing extra.
+    for PART_NAME in ${ENCRYPTED_PARTITIONS}; do
+        LUKS_LINE=$(echo "${LUKS_PARTITIONS}" | grep " \(balena\|resin\)-${PART_NAME}$")
+
+        # Effectively the same checks have been done in the previous loop
+        # We want to be defensive, so we don't mind double-checking
+        if [ -z "${LUKS_LINE}" ]; then
+            fail "Partition '${PART_NAME}' not found"
+        fi
+
+        if [ "$(echo "${LUKS_LINE}" | wc -l)" -gt 1 ]; then
+            fail "More than one '${PART_NAME}' partition found"
+        fi
+
+        LUKS_KNAME=$(echo "${PART_LINE}" | cut -d " " -f 1)
+        LUKS_LABEL=$(echo "${PART_LINE}" | cut -d " " -f 4)
+
+        # Even though the following lsblk could be restricted to just /dev/${LUKS_KNAME}
+        # we intentionally do not want to do that. If there are multiple filesystems with
+        # the correct label, even though we know which one to use, we prefer to bail out,
+        # as this should never happen.
+        FS_LINE=$(lsblk -nlo "kname,pkname,label" | grep "${LUKS_LABEL}")
+        if [ -z "${FS_LINE}" ]; then
+            fail "Filesystem '${LUKS_LABEL}' not found after decryption"
+        fi
+
+        if [ "$(echo "${FS_LINE}" | wc -l)" -gt 1 ]; then
+            fail "More than one '${LUKS_LABEL}' filesystems found after decryption"
+        fi
+
+        FS_KNAME=$(echo "${FS_LINE}" | cut -d " " -f 1)
+        FS_PKNAME=$(echo "${FS_LINE}" | cut -d " " -f 2)
+
+        # We have found an encrypted partition with partlabel=${LUKS_LABEL} and a filesystem
+        # with label=${LUKS_LABEL}. Check that these are parent/child.
+        if [ "${FS_PKNAME}" != "${LUKS_KNAME}" ]; then
+            fail "LUKS partition '${LUKS_LABEL}' does not provide the expected filesystem"
+        fi
+
+        # The original idea was to check the /dev/disk/by-state symlinks here,
+        # but they do not all exist at this point. Because resin_update_state_probe
+        # only creates the symlinks for partitions that are on the same physical drive
+        # as the root partition, no by-state symlinks are created until rootA or rootB
+        # is unlocked. This is not an issue here and all further scripts that actually
+        # want to use the symlinks will trigger udev and get them created.
+    done
+
     rm -f "$PASSPHRASE_FILE"
     umount "$EFI_MOUNT_DIR"
     rmdir "$EFI_MOUNT_DIR"