#!/usr/bin/env bash
set -euo pipefail

log() {
  printf '[spawn] %s\n' "$*"
}

usage() {
  cat <<'EOF'
Usage: ./run.sh [options]

Options:
  --name <name>        VM name (lowercase letters, digits, -)
  --vcpu <count>       vCPU count (default: 2)
  --ram <mib>          RAM in MiB (default: 1024)
  --overlay-size <size> Writable overlay size (e.g. 8G, 16384M)
  --rootfs <path>      Root filesystem image (default: ./rootfs-docker.ext4)
  --kernel <path>      Kernel image (default: ./vmlinux)
  --initrd <path>      Initrd image (optional)
  -h, --help           Show this help
EOF
}

log "starting"

DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
source "$DIR/dns.sh"
source "$DIR/packages.sh"
STATE="$DIR/state"
VM_ROOT="$STATE/vms"
mkdir -p "$VM_ROOT"

FC_BIN="$DIR/firecracker"
DEFAULT_KERNEL="$DIR/wtf/root/boot/vmlinux-6.8.0-94-generic"
DEFAULT_ROOTFS="$DIR/rootfs-docker.ext4"
DEFAULT_INITRD="$DIR/wtf/root/boot/initrd.img-6.8.0-94-generic"
SSH_KEY="$DIR/id_ed25519"
NAMEGEN="$DIR/namegen"

BR_DEV="br-fc"
BR_IP="172.16.0.1"
CIDR="24"

DEFAULT_VCPU=2
DEFAULT_RAM=1024
DEFAULT_OVERLAY_SIZE="8G"
MIN_VCPU=1
MAX_VCPU=16
MIN_RAM=256
MAX_RAM=32768
MAX_DISK_BYTES=$((128 * 1024 * 1024 * 1024))
DNS_SERVER="1.1.1.1"

VCPU_COUNT="$DEFAULT_VCPU"
RAM_MIB="$DEFAULT_RAM"
OVERLAY_SIZE="$DEFAULT_OVERLAY_SIZE"
KERNEL="$DEFAULT_KERNEL"
ROOTFS="$DEFAULT_ROOTFS"
INITRD="$DEFAULT_INITRD"
VM_NAME=""

shopt -s nullglob

name_taken() {
  local candidate="$1"
  local vm_json existing_name
  for vm_json in "$VM_ROOT"/*/vm.json; do
    existing_name="$(jq -r '.meta.name // empty' "$vm_json" 2>/dev/null)"
    if [[ "$existing_name" == "$candidate" ]]; then
      return 0
    fi
  done
  return 1
}

parse_disk_size() {
  local raw="$1"
  if [[ "$raw" =~ ^([0-9]+)([KMG])?$ ]]; then
    local num="${BASH_REMATCH[1]}"
    local unit="${BASH_REMATCH[2]}"
    case "$unit" in
      K) echo $((num * 1024)) ;;
      M|"") echo $((num * 1024 * 1024)) ;;
      G) echo $((num * 1024 * 1024 * 1024)) ;;
    esac
    return 0
  fi
  return 1
}

while [[ $# -gt 0 ]]; do
  case "$1" in
    --name)
      VM_NAME="${2:-}"
      shift 2
      ;;
    --vcpu)
      VCPU_COUNT="${2:-}"
      shift 2
      ;;
    --ram)
      RAM_MIB="${2:-}"
      shift 2
      ;;
    --overlay-size)
      OVERLAY_SIZE="${2:-}"
      shift 2
      ;;
    --rootfs)
      ROOTFS="${2:-}"
      shift 2
      ;;
    --kernel)
      KERNEL="${2:-}"
      shift 2
      ;;
    --initrd)
      INITRD="${2:-}"
      shift 2
      ;;
    -h|--help)
      usage
      exit 0
      ;;
    *)
      log "unknown option: $1"
      usage
      exit 1
      ;;
  esac
done

if ! [[ "$VCPU_COUNT" =~ ^[0-9]+$ ]]; then
  log "invalid --vcpu value: $VCPU_COUNT"
  exit 1
fi
if (( VCPU_COUNT < MIN_VCPU || VCPU_COUNT > MAX_VCPU )); then
  log "vcpu must be between $MIN_VCPU and $MAX_VCPU"
  exit 1
fi

if ! [[ "$RAM_MIB" =~ ^[0-9]+$ ]]; then
  log "invalid --ram value: $RAM_MIB"
  exit 1
fi
if (( RAM_MIB < MIN_RAM || RAM_MIB > MAX_RAM )); then
  log "ram must be between $MIN_RAM and $MAX_RAM MiB"
  exit 1
fi

if ! OVERLAY_BYTES="$(parse_disk_size "$OVERLAY_SIZE")"; then
  log "invalid --overlay-size value: $OVERLAY_SIZE"
  exit 1
fi
if (( OVERLAY_BYTES > MAX_DISK_BYTES )); then
  log "overlay-size exceeds max of $((MAX_DISK_BYTES / 1024 / 1024 / 1024))G"
  exit 1
fi

if [[ -z "$VM_NAME" ]]; then
  if [[ ! -x "$NAMEGEN" ]]; then
    log "name generator not found: $NAMEGEN"
    exit 1
  fi
  for _ in $(seq 1 20); do
    VM_NAME="$("$NAMEGEN")"
    if ! name_taken "$VM_NAME"; then
      break
    fi
  done
fi

if [[ -z "$VM_NAME" ]]; then
  log "failed to generate a unique name"
  exit 1
fi

if ! [[ "$VM_NAME" =~ ^[a-z0-9][a-z0-9-]{0,63}$ ]]; then
  log "invalid --name value: $VM_NAME"
  exit 1
fi
if [[ ! -f "$ROOTFS" ]]; then
  if [[ "$ROOTFS" == "$DEFAULT_ROOTFS" && -x "$DIR/make-rootfs.sh" ]]; then
    log "rootfs missing; building via make-rootfs.sh"
    "$DIR/make-rootfs.sh"
  fi
fi
if [[ ! -f "$ROOTFS" ]]; then
  log "rootfs not found: $ROOTFS"
  exit 1
fi
if [[ "$ROOTFS" == "$DEFAULT_ROOTFS" ]]; then
  ROOTFS_WARNING="$(banger_rootfs_manifest_warning "$ROOTFS" || true)"
  if [[ -n "$ROOTFS_WARNING" ]]; then
    log "warning: $ROOTFS_WARNING"
  fi
fi
if [[ ! -f "$KERNEL" ]]; then
  log "kernel not found: $KERNEL"
  exit 1
fi
if [[ -n "$INITRD" && ! -f "$INITRD" ]]; then
  log "initrd not found: $INITRD"
  exit 1
fi
if name_taken "$VM_NAME"; then
  log "name already in use: $VM_NAME"
  exit 1
fi

VM_ID="$(head -c 32 /dev/urandom | xxd -p -c 256)"
VM_TAG="${VM_ID:0:8}"
VM_DIR="$VM_ROOT/$VM_ID"
mkdir -p "$VM_DIR"

RUNTIME_DIR="${XDG_RUNTIME_DIR:-/run/user/$(id -u)}"
SOCK_DIR="$RUNTIME_DIR/banger"
sudo mkdir -p "$SOCK_DIR"
sudo chown "$(id -u):$(id -g)" "$SOCK_DIR"
API_SOCK="$SOCK_DIR/fc-$VM_TAG.sock"
LOG_FILE="$VM_DIR/firecracker.log"
TAP_DEV="tap-fc-$VM_TAG"

# Allocate guest IP
NEXT_IP_FILE="$STATE/next_ip"
NEXT_IP="$(cat "$NEXT_IP_FILE" 2>/dev/null || echo 2)"
GUEST_IP="172.16.0.$NEXT_IP"
echo "$((NEXT_IP + 1))" > "$NEXT_IP_FILE"

log "vm id: $VM_ID"
log "vm name: $VM_NAME"
log "allocated guest ip: $GUEST_IP"

sudo -v

VM_STARTED=0
CLEANUP_ON_EXIT=0
KEEP_VM_DIR_ON_FAIL=1
COW_FILE="$VM_DIR/cow.ext4"
BASE_LOOP=""
COW_LOOP=""
DM_NAME="fc-rootfs-$VM_TAG"
DM_DEV=""
DNS_NAME=""

cleanup() {
  local exit_code=$?
  if [[ "$VM_STARTED" -eq 1 && "$CLEANUP_ON_EXIT" -eq 0 ]]; then
    return
  fi
  log "cleaning up"
  if [[ -n "${FC_PID:-}" ]]; then
    sudo kill "$FC_PID" 2>/dev/null || true
  fi
  if [[ -n "${TAP_DEV:-}" ]]; then
    sudo ip link del "$TAP_DEV" 2>/dev/null || true
  fi
  rm -f "${API_SOCK:-}"
  banger_dns_remove_record_name "${DNS_NAME:-}"
  if [[ -n "${DM_NAME:-}" ]]; then
    sudo dmsetup remove "$DM_NAME" 2>/dev/null || true
  fi
  if [[ -n "${COW_LOOP:-}" ]]; then
    sudo losetup -d "$COW_LOOP" 2>/dev/null || true
  fi
  if [[ -n "${BASE_LOOP:-}" ]]; then
    sudo losetup -d "$BASE_LOOP" 2>/dev/null || true
  fi
  if [[ -n "${VM_DIR:-}" && -d "$VM_DIR" ]]; then
    if [[ "$KEEP_VM_DIR_ON_FAIL" -eq 0 || "$VM_STARTED" -eq 1 || "$CLEANUP_ON_EXIT" -eq 1 ]]; then
      rm -rf "$VM_DIR"
    fi
  fi
  return "$exit_code"
}

on_signal() {
  CLEANUP_ON_EXIT=1
  exit 1
}

trap cleanup EXIT
trap on_signal INT TERM

FC_USE_SUDO="${FC_USE_SUDO:-1}"
FC_RUN=("$FC_BIN")
CURL_CMD=(curl)
if [[ "$FC_USE_SUDO" == "1" ]]; then
  log "running firecracker with sudo (FC_USE_SUDO=1)"
  FC_RUN=(sudo -E "$FC_BIN")
  CURL_CMD=(sudo -E curl)
fi

if command -v setcap >/dev/null 2>&1; then
  if ! getcap "$FC_BIN" 2>/dev/null | rg -q "cap_net_admin"; then
    log "granting cap_net_admin to firecracker binary"
    sudo setcap cap_net_admin+ep "$FC_BIN"
  fi
else
  log "setcap not available; firecracker may need root to open TAP"
fi

if ! command -v dmsetup >/dev/null 2>&1 || ! command -v losetup >/dev/null 2>&1 || ! command -v blockdev >/dev/null 2>&1; then
  log "dmsetup, losetup, and blockdev are required for rootfs snapshots"
  exit 1
fi
if ! command -v e2cp >/dev/null 2>&1 || ! command -v e2rm >/dev/null 2>&1; then
  log "e2cp and e2rm are required to set hostname and resolv.conf"
  exit 1
fi
if ! command -v jq >/dev/null 2>&1; then
  log "jq is required to persist VM metadata"
  exit 1
fi

BASE_LOOP="$(sudo losetup -f --show --read-only "$ROOTFS")"
truncate -s "$OVERLAY_BYTES" "$COW_FILE"
COW_LOOP="$(sudo losetup -f --show "$COW_FILE")"
SECTORS="$(sudo blockdev --getsz "$BASE_LOOP")"
sudo dmsetup create "$DM_NAME" --table "0 $SECTORS snapshot $BASE_LOOP $COW_LOOP P 8"
DM_DEV="/dev/mapper/$DM_NAME"

RESOLV_TMP="$VM_DIR/resolv.conf"
HOSTNAME_TMP="$VM_DIR/hostname"
HOSTS_TMP="$VM_DIR/hosts"
printf 'nameserver %s\n' "$DNS_SERVER" >"$RESOLV_TMP"
printf '%s\n' "$VM_NAME" >"$HOSTNAME_TMP"
printf '127.0.0.1 localhost\n127.0.1.1 %s\n' "$VM_NAME" >"$HOSTS_TMP"
sudo e2rm "$DM_DEV:/etc/resolv.conf" >/dev/null 2>&1 || true
sudo e2rm "$DM_DEV:/etc/hostname" >/dev/null 2>&1 || true
sudo e2rm "$DM_DEV:/etc/hosts" >/dev/null 2>&1 || true
sudo e2cp "$RESOLV_TMP" "$DM_DEV:/etc/resolv.conf" >/dev/null 2>&1 || {
  log "failed to write /etc/resolv.conf into rootfs snapshot"
  exit 1
}
sudo e2cp "$HOSTNAME_TMP" "$DM_DEV:/etc/hostname" >/dev/null 2>&1 || {
  log "failed to write /etc/hostname into rootfs snapshot"
  exit 1
}
sudo e2cp "$HOSTS_TMP" "$DM_DEV:/etc/hosts" >/dev/null 2>&1 || {
  log "failed to write /etc/hosts into rootfs snapshot"
  exit 1
}

# Host bridge
if ! ip link show "$BR_DEV" >/dev/null 2>&1; then
  log "creating host bridge $BR_DEV ($BR_IP/$CIDR)"
  sudo ip link add name "$BR_DEV" type bridge
  sudo ip addr add "${BR_IP}/${CIDR}" dev "$BR_DEV"
  sudo ip link set "$BR_DEV" up
else
  log "host bridge $BR_DEV already exists"
  # Ensure existing bridge is up in case it was left down.
  sudo ip link set "$BR_DEV" up
fi

# Per-VM TAP
log "creating tap device $TAP_DEV"
TAP_USER="${SUDO_UID:-$(id -u)}"
TAP_GROUP="${SUDO_GID:-$(id -g)}"
sudo ip tuntap add dev "$TAP_DEV" mode tap user "$TAP_USER" group "$TAP_GROUP"
sudo ip link set "$TAP_DEV" master "$BR_DEV"
sudo ip link set "$TAP_DEV" up
sudo ip link set "$BR_DEV" up

# Start Firecracker
log "starting firecracker process"
rm -f "$API_SOCK"
nohup "${FC_RUN[@]}" --api-sock "$API_SOCK" >"$LOG_FILE" 2>&1 &
FC_PID="$!"
log "firecracker pid: $FC_PID"

# Wait for API socket
log "waiting for firecracker api socket"
for _ in $(seq 1 200); do
  [[ -S "$API_SOCK" ]] && break
  sleep 0.02
done
if [[ ! -S "$API_SOCK" ]]; then
  log "firecracker api socket not ready"
  if [[ -f "$LOG_FILE" ]]; then
    log "firecracker log (tail)"
    tail -n 20 "$LOG_FILE" || true
  fi
  exit 1
fi
log "api socket ready"

if [[ "$FC_USE_SUDO" == "1" ]]; then
  SUDO_CHILD_PID="$(pgrep -n -f "$API_SOCK" || true)"
  if [[ -n "$SUDO_CHILD_PID" ]]; then
    FC_PID="$SUDO_CHILD_PID"
    log "firecracker child pid: $FC_PID"
  fi
fi
echo "$FC_PID" > "$VM_DIR/pid"

# Machine config
log "configuring machine"
"${CURL_CMD[@]}" --unix-socket "$API_SOCK" -X PUT http://localhost/machine-config \
  -H "Content-Type: application/json" \
  -d "{
    \"vcpu_count\": $VCPU_COUNT,
    \"mem_size_mib\": $RAM_MIB,
    \"smt\": false
  }" >/dev/null

# Boot source
log "configuring boot source"
KCMD="console=ttyS0 reboot=k panic=1 pci=off root=/dev/vda rw ip=${GUEST_IP}::${BR_IP}:255.255.255.0::eth0:off:${DNS_SERVER} hostname=${VM_NAME} systemd.mask=home.mount systemd.mask=var.mount"

INITRD_JSON=""
if [[ -n "$INITRD" ]]; then
  INITRD_JSON=", \"initrd_path\": \"$INITRD\""
fi

"${CURL_CMD[@]}" --unix-socket "$API_SOCK" -X PUT http://localhost/boot-source \
  -H "Content-Type: application/json" \
  -d "{
    \"kernel_image_path\": \"$KERNEL\",
    \"boot_args\": \"$KCMD\"${INITRD_JSON}
  }" >/dev/null

# Root filesystem
log "attaching root filesystem"
"${CURL_CMD[@]}" --unix-socket "$API_SOCK" -X PUT http://localhost/drives/rootfs \
  -H "Content-Type: application/json" \
  -d "{
    \"drive_id\": \"rootfs\",
    \"path_on_host\": \"$DM_DEV\",
    \"is_root_device\": true,
    \"is_read_only\": false
  }" >/dev/null

# Network interface
log "configuring network interface"
"${CURL_CMD[@]}" --unix-socket "$API_SOCK" -X PUT http://localhost/network-interfaces/eth0 \
  -H "Content-Type: application/json" \
  -d "{
    \"iface_id\": \"eth0\",
    \"host_dev_name\": \"$TAP_DEV\"
  }" >/dev/null

# Start VM
log "starting virtual machine"
"${CURL_CMD[@]}" --unix-socket "$API_SOCK" -X PUT http://localhost/actions \
  -H "Content-Type: application/json" \
  -d '{ "action_type": "InstanceStart" }' >/dev/null
VM_CONFIG_JSON="$("${CURL_CMD[@]}" --unix-socket "$API_SOCK" -sS http://localhost/vm/config)"
CREATED_AT="$(date -Iseconds)"
DNS_NAME="$(banger_dns_name "$VM_NAME")"
banger_dns_write_record "$VM_NAME" "$GUEST_IP"
jq -n \
  --arg id "$VM_ID" \
  --arg name "$VM_NAME" \
  --arg pid "$FC_PID" \
  --arg created_at "$CREATED_AT" \
  --arg guest_ip "$GUEST_IP" \
  --arg tap "$TAP_DEV" \
  --arg api_sock "$API_SOCK" \
  --arg log "$LOG_FILE" \
  --arg rootfs "$ROOTFS" \
  --arg kernel "$KERNEL" \
  --arg base_loop "$BASE_LOOP" \
  --arg cow_file "$COW_FILE" \
  --arg cow_loop "$COW_LOOP" \
  --arg dm_name "$DM_NAME" \
  --arg dm_dev "$DM_DEV" \
  --arg dns_name "$DNS_NAME" \
  --argjson config "$VM_CONFIG_JSON" \
  '{meta:{id:$id,name:$name,pid:$pid,created_at:$created_at,guest_ip:$guest_ip,tap:$tap,api_sock:$api_sock,log:$log,rootfs:$rootfs,kernel:$kernel,base_loop:$base_loop,cow_file:$cow_file,cow_loop:$cow_loop,dm_name:$dm_name,dm_dev:$dm_dev,dns_name:$dns_name},config:$config}' \
  > "$VM_DIR/vm.json"
VM_STARTED=1

log "vm started successfully"
log "guest ip: $GUEST_IP"
log "ssh: ssh -i \"$SSH_KEY\" root@$GUEST_IP"
log "logs: $LOG_FILE"