devmapper-pool-init

#!/bin/busybox sh
# devmapper-pool-init — Talos extension service (ADR-038)
#
# Creates dm-thin-pool from r-dm-data raw partition,
# writes devmapper CRI config, restarts CRI containerd.
# Runs once at boot (restart: never).
set -e

POOL_NAME="containerd-pool"
DM_PART="/dev/disk/by-partlabel/r-dm-data"
# Use /run (tmpfs) for metadata — bind mounts from /var are snapshots taken when
# the extension container starts, and at that early boot point /var is still
# the read-only squashfs base (EPHEMERAL mount hasn't propagated to the extension
# container's mount namespace yet). /run is tmpfs, always writable, always available.
# Metadata is lost on reboot, but that's fine — a fresh boot has no running
# containers, so we rebuild the pool from scratch. The data partition (r-dm-data)
# is persistent but the thin-pool metadata is regenerated each boot.
META_DIR="/run/devmapper"
META_FILE="${META_DIR}/meta"
# Talos CRI reads config from /etc/cri/conf.d/ which is an overlay:
#   lower = read-only rootfs, upper = /var/system/overlays/etc-cri-diff/
# Writing to the overlay upper dir makes files appear in /etc/cri/conf.d/
# when CRI reads it. This is under /var (mounted into this container).
CRI_CONF_DIR="/var/system/overlays/etc-cri-diff/conf.d"
CRI_CONF="${CRI_CONF_DIR}/20-devmapper.part"
HOST_PROC="/host-proc"
# Persistent reboot marker — must survive reboot to prevent loop.
# /var/log/ is writable very early in boot (Talos writes service logs there
# from Phase 1) and on EPHEMERAL partition (persists across reboots).
REBOOT_MARKER="/var/log/devmapper-rebooted"

log() { echo "devmapper-pool: $*"; }

# ── 1. Check if pool already exists ──
if /sbin/dmsetup info "$POOL_NAME" >/dev/null 2>&1; then
    log "pool exists"
    # Ensure CRI config exists
    if [ -f "$CRI_CONF" ]; then
        log "config exists — nothing to do"
        exit 0
    fi
    # Pool exists but config missing — write it and restart CRI
    log "pool exists but config missing — writing config"
fi

# ── 2. Create thin pool (if not exists) ──
if ! /sbin/dmsetup info "$POOL_NAME" >/dev/null 2>&1; then
    # Wait for udev to create the by-partlabel symlink.
    # The partition may exist but udev hasn't scanned it yet (~1-2s after creation).
    WAIT=0
    while [ ! -b "$DM_PART" ] && [ "$WAIT" -lt 30 ]; do
        /bin/busybox sleep 1
        WAIT=$((WAIT + 1))
    done
    if [ ! -b "$DM_PART" ]; then
        log "ERROR: $DM_PART not found after ${WAIT}s"
        exit 1
    fi
    log "partition ready after ${WAIT}s"

    DATA_DEV=$(/bin/busybox readlink -f "$DM_PART")
    log "data: $DATA_DEV"

    # Metadata: 4GB sparse file on EPHEMERAL (dm-thin requires separate meta device)
    # Use META_DIR variable (set to /var/devmapper, NOT /var/lib/devmapper which is a late overlay)
    /bin/busybox mkdir -p "$META_DIR" || { log "ERROR: mkdir $META_DIR failed"; exit 1; }
    log "meta dir ready: $META_DIR"
    if [ ! -f "$META_FILE" ]; then
        /bin/busybox dd if=/dev/zero of="$META_FILE" bs=1 count=0 seek=4294967296 2>/dev/null
        log "meta file created"
    fi

    META_LOOP=$(/bin/busybox losetup -f "$META_FILE" && /bin/busybox losetup -a | /bin/busybox grep "$META_FILE" | /bin/busybox cut -d: -f1)
    SECTORS=$(/bin/busybox blockdev --getsz "$DATA_DEV")
    log "creating pool: data=$DATA_DEV sectors=$SECTORS meta=$META_LOOP"

    /sbin/dmsetup create "$POOL_NAME" \
        --table "0 ${SECTORS} thin-pool ${META_LOOP} ${DATA_DEV} 512 32768 1 skip_block_zeroing"
    log "pool created"
fi

# ── 3. Write CRI devmapper config ──
/bin/busybox mkdir -p "$CRI_CONF_DIR" 2>/dev/null || true
/bin/busybox cat > "$CRI_CONF" << 'TOML'
[plugins."io.containerd.snapshotter.v1.devmapper"]
  pool_name = "containerd-pool"
  root_path = "/var/lib/containerd/devmapper"
  base_image_size = "10GB"
  discard_blocks = true
  fs_type = "ext4"
  async_remove = true

[plugins."io.containerd.cri.v1.runtime".containerd.runtimes.kata]
  runtime_type = "io.containerd.kata.v2"
  privileged_without_host_devices = true
  pod_annotations = ["io.katacontainers.*"]
  snapshotter = "devmapper"

[plugins."io.containerd.cri.v1.runtime".containerd.runtimes.kata.options]
  ConfigPath = "/usr/local/share/kata-containers/configuration.toml"
TOML
log "CRI config written"

# ── 4. Restart CRI to load devmapper snapshotter ──
# Find CRI containerd PID via host /proc
CRI_PID=""
for pid in $(/bin/busybox ls "$HOST_PROC" 2>/dev/null | /bin/busybox grep '^[0-9]'); do
    cmdline=$(/bin/busybox cat "$HOST_PROC/$pid/cmdline" 2>/dev/null | /bin/busybox tr '\0' ' ')
    case "$cmdline" in
        *containerd*--config*/etc/cri*)
            CRI_PID="$pid"
            break
            ;;
    esac
done

if [ -z "$CRI_PID" ]; then
    log "CRI not found — config ready for next CRI start"
    exit 0
fi

log "CRI already running (PID $CRI_PID) — needs restart to load devmapper plugin"

# Why reboot instead of kill?
# Linux kernel forbids entering a PARENT PID namespace via setns(2) — this is
# hardcoded in kernel/pid_namespace.c. Child namespaces can only descend, never
# ascend. Extension containers run in a child PID namespace of the host, so
# kill(2) from any form (busybox kill, nsenter --pid=FILE, nsenter --target N)
# will always fail with EINVAL. This is a kernel design rule, not a tool bug.
#
# reboot(2) syscall is different: it's global (kernel-wide), requires CAP_SYS_BOOT
# (which extensions have via AllGrantableCapabilities), and works from any namespace.
#
# Strategy: on first run, after creating pool + writing config, reboot the node.
# On second boot, CRI starts with pool already existing and config in place →
# devmapper plugin loads successfully. Marker file prevents reboot loop.

if [ -f "$REBOOT_MARKER" ]; then
    log "reboot marker exists — this is post-reboot run, devmapper should be loaded"
    log "extension finished (no reboot needed)"
    exit 0
fi

log "writing reboot marker + triggering node reboot to reload CRI with devmapper"
/bin/busybox touch "$REBOOT_MARKER"
/bin/busybox sync

# Try reboot command first (uses reboot(2) syscall via util-linux)
if command -v reboot >/dev/null 2>&1; then
    log "calling reboot command"
    reboot -f 2>&1 || true
fi

# Fallback: sysrq reboot via host /proc (requires /host-proc mounted rw)
log "fallback: sysrq-trigger reboot"
echo 1 > "$HOST_PROC/sys/kernel/sysrq" 2>/dev/null || true
echo b > "$HOST_PROC/sysrq-trigger" 2>/dev/null || true

# Fallback: busybox reboot
/bin/busybox reboot -f 2>&1 || true

# If we get here, reboot failed
log "ERROR: all reboot methods failed — manual intervention required"
log "Run: talosctl -n <node-ip> reboot"
exit 1