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

PATH=/usr/local/bin:/usr/bin:/usr/sbin:/sbin:/bin
export PATH
umask 077

PREFIX=${TALLYNOTE_INSTALL_PREFIX:-/opt/tallynote}
DATA_DIR=${TALLYNOTE_DATA_DIR:-/var/lib/tallynote}
CONFIG_DIR=${TALLYNOTE_CONFIG_DIR:-/etc/tallynote}
CONFIG_FILE="$CONFIG_DIR/tallynote.env"
REQUEST_FILE="$DATA_DIR/update-request.json"
CURRENT_LINK="$PREFIX/current"
STATE_FILE="$PREFIX/.update-state"
LOCK_FILE="$PREFIX/.update-runner.lock"
RUNNER_LOG="$PREFIX/.update-runner.log"
SERVICE_NAME=${TALLYNOTE_SERVICE_NAME:-tallynote.service}
HOST=${TALLYNOTE_HOST:-127.0.0.1}
PORT=${TALLYNOTE_PORT:-3000}
HEALTH_HOST=$HOST
if [[ "$HEALTH_HOST" == 0.0.0.0 ]]; then HEALTH_HOST=127.0.0.1; fi
if [[ "$HEALTH_HOST" == :: ]]; then HEALTH_HOST=::1; fi
if [[ "$HEALTH_HOST" == *:* && "$HEALTH_HOST" != \[* ]]; then HEALTH_HOST="[$HEALTH_HOST]"; fi

die() { printf 'tallynote update runner: %s\n' "$*" >&2; exit 1; }

node_is_usable() {
  local candidate=$1 major
  [[ -n "$candidate" && -x "$candidate" ]] || return 1
  major=$("$candidate" -p 'process.versions.node.split(".")[0]' 2>/dev/null || true)
  [[ "$major" =~ ^[0-9]+$ && "$major" -ge 24 ]]
}

resolve_node() {
  local candidate=${TALLYNOTE_NODE:-}
  if [[ -z "$candidate" && -f "$CONFIG_FILE" && ! -L "$CONFIG_FILE" ]]; then
    candidate=$(sed -n 's/^TALLYNOTE_NODE=//p' "$CONFIG_FILE" | head -n 1)
  fi
  if node_is_usable "$candidate"; then
    printf '%s' "$candidate"
    return 0
  fi
  candidate=$(command -v node || true)
  node_is_usable "$candidate" || return 1
  printf '%s' "$candidate"
}

# The runner may exit during any of the checks below. Install its EXIT cleanup
# before doing privileged preflight so a partial invocation never leaves a
# heartbeat or lock behind.
STATE_CREATED=0
heartbeat_pid=''
heartbeat_owner=$$
RUNNER_LOCK_FD=9
RUNNER_LOCK_MODE=''
stop_heartbeat() {
  if [[ -n "$heartbeat_pid" ]]; then
    kill "$heartbeat_pid" 2>/dev/null || true
    wait "$heartbeat_pid" 2>/dev/null || true
    heartbeat_pid=''
  fi
}
# shellcheck disable=SC2329 # invoked indirectly by the EXIT trap
release_runner_lock() {
  if [[ "$RUNNER_LOCK_MODE" == flock ]]; then
    flock -u "$RUNNER_LOCK_FD" 2>/dev/null || true
    eval "exec ${RUNNER_LOCK_FD}>&-" 2>/dev/null || true
  elif [[ "$RUNNER_LOCK_MODE" == mkdir ]]; then
    rmdir -- "$LOCK_FILE.d" 2>/dev/null || true
  fi
}
# shellcheck disable=SC2329 # invoked indirectly by the EXIT trap
early_cleanup() {
  local result=$?
  stop_heartbeat
  if (( result != 0 )); then
    # A preflight failure happens before the normal phase-specific trap is
    # installed. Remove only the one-shot request marker; never remove an
    # existing recovery marker unless this invocation created it.
    rm -f -- "$REQUEST_FILE" 2>/dev/null || true
    if (( STATE_CREATED == 1 )); then rm -f -- "$STATE_FILE" 2>/dev/null || true; fi
  fi
  release_runner_lock
  return "$result"
}
trap early_cleanup EXIT

[[ ${EUID:-$(id -u)} -eq 0 ]] || die 'must run as root'
[[ -d "$PREFIX" ]] || die 'install prefix is missing'
if command -v flock >/dev/null 2>&1; then
  exec 9>"$LOCK_FILE" || die '无法打开更新运行锁'
  flock -n "$RUNNER_LOCK_FD" || exit 0
  RUNNER_LOCK_MODE=flock
else
  # macOS development fixtures do not ship util-linux; retain an atomic lock
  # fallback there while Linux production uses flock above.
  mkdir "$LOCK_FILE.d" 2>/dev/null || exit 0
  RUNNER_LOCK_MODE='mkdir'
fi
[[ -f "$REQUEST_FILE" || -f "$STATE_FILE" ]] || exit 0
[[ -L "$CURRENT_LINK" ]] || die 'current release link is missing'

old_target=$(readlink -f -- "$CURRENT_LINK")
[[ "$old_target" == "$PREFIX/releases/"* && -d "$old_target" ]] || die 'current release target is invalid'

# Capture the service state before any download/apply work. The value is
# persisted in the recovery marker so a later runner process can restore the
# operator's original state after a crash (the service is normally inactive by
# the time recovery starts).
was_active=0
if systemctl is-active --quiet "$SERVICE_NAME"; then was_active=1; fi

request_operation='apply'
if [[ -f "$REQUEST_FILE" && ! -L "$REQUEST_FILE" ]]; then
  request_operation=$(sed -n 's/.*"operation"[[:space:]]*:[[:space:]]*"\(download\|apply\)".*/\1/p' "$REQUEST_FILE" | head -n 1)
  [[ "$request_operation" == download || "$request_operation" == apply ]] || request_operation='apply'
fi

# Capture the request id before any privileged preflight can fail. The
# request file is an application-owned one-shot marker; removing it on an
# early runner failure lets the server-side lease reaper release the DB row.
job_id=''
if [[ -f "$REQUEST_FILE" && ! -L "$REQUEST_FILE" ]]; then
  job_id=$(sed -n 's/.*"jobId"[[:space:]]*:[[:space:]]*"\([0-9a-f-]*\)".*/\1/p' "$REQUEST_FILE" | head -n 1)
fi
write_recovery_state() {
  local phase=$1 temporary
  temporary="$PREFIX/.update-state-$$-${RANDOM}.tmp"
  [[ ! -e "$temporary" && ! -L "$temporary" ]] || return 1
  printf 'job_id=%s\nold_target=%s\nphase=%s\ninitial_active=%s\n' "$job_id" "$old_target" "$phase" "$was_active" > "$temporary"
  chmod 600 "$temporary"
  mv -Tf -- "$temporary" "$STATE_FILE"
  STATE_CREATED=1
}

clear_recovery_state() {
  [[ ! -L "$STATE_FILE" ]] || return 1
  rm -f -- "$STATE_FILE"
  STATE_CREATED=0
}

heartbeat() {
  # Keep the lease fresh during long downloads/backups, but stop on a hard
  # runner kill so an orphaned child cannot keep the recovery marker alive.
  while kill -0 "$heartbeat_owner" 2>/dev/null; do
    sleep 10 || exit 0
    [[ -f "$STATE_FILE" && ! -L "$STATE_FILE" ]] || exit 0
    touch "$STATE_FILE" 2>/dev/null || exit 0
  done
}

start_heartbeat() {
  stop_heartbeat
  heartbeat &
  heartbeat_pid=$!
}

# This trap covers failures before the normal apply cleanup trap is installed,
# including a missing runtime, an invalid current link, and a failed service
# stop. It deliberately does not remove a pre-existing recovery marker.
# shellcheck disable=SC2329 # invoked indirectly by the EXIT trap
preflight_cleanup() {
  local result=$?
  stop_heartbeat
  release_runner_lock
  if (( result != 0 )); then
    rm -f -- "$REQUEST_FILE" 2>/dev/null || true
    if (( STATE_CREATED == 1 )); then clear_recovery_state || true; fi
  fi
  return "$result"
}
trap preflight_cleanup EXIT

DOWNLOAD_TIMEOUT_SECONDS=${TALLYNOTE_UPDATE_DOWNLOAD_TIMEOUT_SECONDS:-${TALLYNOTE_UPDATE_RUNNER_DOWNLOAD_TIMEOUT_SECONDS:-1800}}
APPLY_TIMEOUT_SECONDS=${TALLYNOTE_UPDATE_APPLY_TIMEOUT_SECONDS:-${TALLYNOTE_UPDATE_RUNNER_APPLY_TIMEOUT_SECONDS:-1800}}
FINALIZE_TIMEOUT_SECONDS=${TALLYNOTE_UPDATE_FINALIZE_TIMEOUT_SECONDS:-${TALLYNOTE_UPDATE_RUNNER_FINALIZE_TIMEOUT_SECONDS:-30}}
TIMEOUT_BIN=$(command -v timeout || true)

run_update_cli() {
  local node=$1 timeout_seconds=$2 label=$3 result
  shift 3
  [[ "$timeout_seconds" =~ ^[1-9][0-9]*$ ]] || die "${label} timeout must be a positive integer"
  {
    printf '\n[%s] %s (timeout=%ss)\ncommand:' "$(date -u '+%Y-%m-%dT%H:%M:%SZ')" "$label" "$timeout_seconds"
    printf ' %q' "$node" "$CURRENT_LINK/dist/server/cli/update.js" "$@"
    printf '\n'
  } >>"$RUNNER_LOG"
  if [[ -n "$TIMEOUT_BIN" ]]; then
    "$TIMEOUT_BIN" --foreground --signal=TERM --kill-after=10s "${timeout_seconds}s" \
      "$node" "$CURRENT_LINK/dist/server/cli/update.js" "$@" >>"$RUNNER_LOG" 2>&1
    result=$?
  elif "$node" "$CURRENT_LINK/dist/server/cli/update.js" "$@" >>"$RUNNER_LOG" 2>&1; then
    result=0
  else
    result=$?
  fi
  printf '[%s] %s exited with status %s\n' "$(date -u '+%Y-%m-%dT%H:%M:%SZ')" "$label" "$result" >>"$RUNNER_LOG"
  return "$result"
}

# Downloading is intentionally handled while the main service remains up.
# The CLI persists the validated payload under the root-owned workspace and
# leaves the job staged for a later apply request.
if [[ "$request_operation" == download ]]; then
  # A previous download runner may have been interrupted after creating its
  # marker. Clear only that download marker and retry the idempotent request.
  if [[ -f "$STATE_FILE" && ! -L "$STATE_FILE" ]] && grep -q '^phase=download$' "$STATE_FILE"; then
    clear_recovery_state || die '无法清理上一次下载状态'
  fi
  write_recovery_state download || die '无法写入更新恢复状态'
# shellcheck disable=SC2329 # invoked indirectly by the EXIT trap
  cleanup_download() {
    local result=$?
    stop_heartbeat
    if (( result != 0 )); then
      # The CLI normally records failed itself. If it died before opening the
      # database, the expired marker/request will be reconciled by the app.
      rm -f -- "$REQUEST_FILE" 2>/dev/null || true
    fi
    clear_recovery_state || true
    release_runner_lock
    return "$result"
  }
  trap cleanup_download EXIT
  trap 'exit 143' TERM
  trap 'exit 130' INT
  start_heartbeat
  node_bin=$(resolve_node) || die 'Node.js 24+ not found'
  cli="$CURRENT_LINK/dist/server/cli/update.js"
  [[ -f "$cli" ]] || die 'update CLI not found in current release'
  set +e
  run_update_cli "$node_bin" "$DOWNLOAD_TIMEOUT_SECONDS" download --request-file "$REQUEST_FILE"
  download_result=$?
  set -e
  if (( download_result != 0 )); then
    # The CLI normally records failed itself. Retry the explicit finalization
    # for failures that happen before its catch handler can persist the row,
    # then remove the one-shot request so a failed download cannot keep the
    # path unit in a permanently triggered state.
    download_job_id=$(sed -n 's/.*"jobId"[[:space:]]*:[[:space:]]*"\([0-9a-f-]*\)".*/\1/p' "$REQUEST_FILE" | head -n 1)
    if [[ "$download_job_id" =~ ^[0-9a-f-]{36}$ ]]; then
      for _ in 1 2 3; do
        if run_update_cli "$node_bin" "$FINALIZE_TIMEOUT_SECONDS" finalize-download --finalize-job "$download_job_id" --finalize-status failed --message '更新下载失败'; then break; fi
        sleep 1
      done
    fi
    rm -f -- "$REQUEST_FILE"
    exit "$download_result"
  fi
  rm -f -- "$REQUEST_FILE"
  exit 0
fi

# shellcheck disable=SC2329 # invoked indirectly by the EXIT trap below
restore_initial_service() {
  local result=$?
  stop_heartbeat
  release_runner_lock
  if (( result != 0 )); then
    rm -f -- "$REQUEST_FILE" 2>/dev/null || true
    if (( STATE_CREATED == 1 )); then clear_recovery_state || true; fi
  fi
  if (( was_active )); then systemctl start "$SERVICE_NAME" || true; fi
  return "$result"
}
trap restore_initial_service EXIT
old_node=$(resolve_node) || die 'Node.js 24+ not found'
handled=0

write_update_state() { write_recovery_state "$1"; }
clear_update_state() { clear_recovery_state; }

finalize_state_job() {
  local node=$1 status=$2 state_job=$3
  [[ "$state_job" =~ ^[0-9a-f-]{36}$ && -n "$node" ]] || return 1
  [[ -f "$CURRENT_LINK/dist/server/cli/update.js" ]] || return 1
  run_update_cli "$node" "$FINALIZE_TIMEOUT_SECONDS" finalize-recovery --finalize-job "$state_job" --finalize-status "$status" --message '新版本健康检查失败，已恢复上一版本'
}

recover_stale_state() {
  local state_job state_old state_phase state_initial_active current_target recovery_node rollback_link state_mode state_uid
  [[ -f "$STATE_FILE" && ! -L "$STATE_FILE" ]] || die 'update state file is invalid'
  state_uid=$(stat -c '%u' "$STATE_FILE" 2>/dev/null || stat -f '%u' "$STATE_FILE")
  state_mode=$(stat -c '%a' "$STATE_FILE" 2>/dev/null || stat -f '%Lp' "$STATE_FILE")
  [[ "$state_uid" == 0 && "$state_mode" =~ ^[0-7]+$ && $((8#$state_mode & 077)) -eq 0 ]] || die 'update state file permissions are invalid'
  state_job=$(sed -n 's/^job_id=//p' "$STATE_FILE" | head -n 1)
  state_old=$(sed -n 's/^old_target=//p' "$STATE_FILE" | head -n 1)
  state_phase=$(sed -n 's/^phase=//p' "$STATE_FILE" | head -n 1)
  state_initial_active=$(sed -n 's/^initial_active=//p' "$STATE_FILE" | head -n 1)
  [[ "$state_job" =~ ^[0-9a-f-]{36}$ ]] || die 'update state job id is invalid'
  [[ "$state_old" == "$PREFIX/releases/"* && -d "$state_old" && ! -L "$state_old" ]] || die 'update state target is invalid'
  if [[ -z "$state_initial_active" ]]; then
    # Markers from older releases did not persist this field. Preserve their
    # historical conservative behavior instead of rejecting recovery.
    state_initial_active=0
  fi
  [[ "$state_initial_active" == 0 || "$state_initial_active" == 1 ]] || die 'update state initial service state is invalid'
  was_active=$state_initial_active
  current_target=$(readlink -f -- "$CURRENT_LINK" 2>/dev/null || true)
  if [[ "$state_phase" == download && "$current_target" == "$state_old" ]]; then
    # Downloading never changes the active release. If the runner was killed
    # after the CLI staged its payload but before it removed the recovery
    # marker, keep the request available for an idempotent retry. Treating
    # every stale download marker as a failed apply would discard a usable
    # staged payload and leave the browser showing a misleading failure.
    clear_update_state || true
    return 0
  fi
  if [[ "$state_phase" == finalizing && "$current_target" != "$state_old" ]]; then
    recovery_node=$(resolve_node) || die 'Node.js 24+ not found'
    for _ in 1 2 3; do
      if finalize_state_job "$recovery_node" completed "$state_job"; then
        rm -f -- "$REQUEST_FILE" 2>/dev/null || true
        clear_update_state || true
        return 10
      fi
      sleep 1
    done
    return 1
  fi
  if [[ "$current_target" == "$state_old" ]]; then
    # The process may have restored the old release before it was killed. In
    # that case the old link is already safe to serve, but the database row
    # can still be `applying`; finish it as failed before clearing recovery
    # markers so the UI does not poll forever.
    recovery_node=$(resolve_node) || die 'Node.js 24+ not found'
    if finalize_state_job "$recovery_node" failed "$state_job"; then
      rm -f -- "$REQUEST_FILE" 2>/dev/null || true
      clear_update_state || true
      return 11
    fi
    # A crash before the CLI created its job row is safe to retry. Preserve
    # the request while dropping only the stale state marker.
    if [[ -f "$REQUEST_FILE" && ! -L "$REQUEST_FILE" ]]; then
      clear_update_state || true
      return 0
    fi
    clear_update_state || true
    return 0
  fi
  if [[ "$current_target" != "$state_old" ]]; then
    rollback_link="$PREFIX/.current-recovery-$$-${RANDOM}.tmp"
    [[ ! -e "$rollback_link" && ! -L "$rollback_link" ]] || return 1
    ln -s -- "$state_old" "$rollback_link" || return 1
    if ! mv -Tf -- "$rollback_link" "$CURRENT_LINK"; then
      rm -f -- "$rollback_link" 2>/dev/null || true
      return 1
    fi
    recovery_node=$(resolve_node) || die 'Node.js 24+ not found'
    if ! finalize_state_job "$recovery_node" failed "$state_job"; then
      # If the original queue is still present, retry it from the restored old
      # release; a crash before the CLI wrote its job row is recoverable this
      # way. Without a queue there is no safe operation to replay.
      if [[ -f "$REQUEST_FILE" && ! -L "$REQUEST_FILE" ]]; then
        clear_update_state || true
        return 0
      fi
      return 1
    fi
    rm -f -- "$REQUEST_FILE" 2>/dev/null || true
    clear_update_state || true
    return 11
  fi
  clear_update_state || true
  return 0
}

if [[ -e "$STATE_FILE" ]]; then
  recovery_result=0
  set +e
  recover_stale_state
  recovery_result=$?
  set -e
  case "$recovery_result" in
    10) if (( was_active )); then systemctl start "$SERVICE_NAME" || true; fi; exit 0 ;;
    11) if (( was_active )); then systemctl start "$SERVICE_NAME" || true; fi; exit 1 ;;
    0) : ;;
    *) if (( was_active )); then systemctl start "$SERVICE_NAME" || true; fi; exit 1 ;;
  esac
fi

[[ -f "$REQUEST_FILE" && ! -L "$REQUEST_FILE" ]] || exit 0

# Only create the marker for this invocation after any marker from a previous
# interrupted run has been reconciled. Otherwise the freshly-created `running`
# marker is indistinguishable from stale recovery state and the runner can
# finalize its own queued job as failed before the update CLI starts.
if [[ ! -e "$STATE_FILE" ]]; then
  write_recovery_state running || die '无法写入更新恢复状态'
fi
start_heartbeat
if ! systemctl stop "$SERVICE_NAME"; then
  die '无法停止 TallyNote 服务'
fi

rollback_current() {
  local current_target rollback_link
  current_target=$(readlink -f -- "$CURRENT_LINK" 2>/dev/null || true)
  if [[ "$current_target" == "$old_target" ]]; then
    # An earlier failure branch may already have restored the link. Keep the
    # marker truthful so the EXIT trap can still finalize the job.
    return 0
  fi
  rollback_link="$PREFIX/.current-rollback-$$-${RANDOM}.tmp"
  [[ ! -e "$rollback_link" && ! -L "$rollback_link" ]] || return 1
  ln -s -- "$old_target" "$rollback_link" || return 1
  if ! mv -Tf -- "$rollback_link" "$CURRENT_LINK"; then
    rm -f -- "$rollback_link" 2>/dev/null || true
    return 1
  fi
}

finalize_failed_job() {
  [[ "$job_id" =~ ^[0-9a-f-]{36}$ ]] || return 0
  [[ -n "$old_node" && -f "$CURRENT_LINK/dist/server/cli/update.js" ]] || return 1
  # Give SQLite a moment to release a transient lock before declaring the
  # recovery itself failed.
  for _ in 1 2 3; do
    if run_update_cli "$old_node" "$FINALIZE_TIMEOUT_SECONDS" finalize-failed --finalize-job "$job_id" --finalize-status failed --message '新版本健康检查失败，已恢复上一版本'; then
      return 0
    fi
    sleep 1
  done
  return 1
}

finalize_completed_job() {
  [[ "$job_id" =~ ^[0-9a-f-]{36}$ ]] || return 0
  [[ -n "$final_node" ]] || return 1
  run_update_cli "$final_node" "$FINALIZE_TIMEOUT_SECONDS" finalize-completed --finalize-job "$job_id" --finalize-status completed
}

# shellcheck disable=SC2329 # invoked indirectly by the EXIT trap below
cleanup_after_update() {
  local result=$? rollback_ok=1
  stop_heartbeat
  if (( result != 0 && handled == 0 )); then
    if ! rollback_current; then rollback_ok=0; fi
    # Once the old release is active again, always try to close the job. The
    # previous marker could remain set when an earlier branch had already
    # rolled back before entering this EXIT trap, leaving `applying` forever.
    if (( rollback_ok == 1 )); then
      if finalize_failed_job; then
        rm -f -- "$REQUEST_FILE"
        clear_update_state || true
      fi
    fi
  fi
  if (( was_active )); then
    systemctl start "$SERVICE_NAME" || true
  else
    systemctl stop "$SERVICE_NAME" || true
  fi
  release_runner_lock
  return "$result"
}
trap cleanup_after_update EXIT

node_bin=$(resolve_node) || die 'Node.js 24+ not found'
cli="$CURRENT_LINK/dist/server/cli/update.js"
[[ -f "$cli" ]] || die 'update CLI not found in current release'

set +e
run_update_cli "$node_bin" "$APPLY_TIMEOUT_SECONDS" apply --request-file "$REQUEST_FILE" --defer-completion
update_result=$?
set -e
if (( update_result != 0 )); then
  exit "$update_result"
fi

write_update_state health-check || exit 1

systemctl start "$SERVICE_NAME"
healthy=0
for _ in $(seq 1 30); do
  if curl --proto '=http' --max-time 2 --silent --show-error "http://$HEALTH_HOST:$PORT/health" >/dev/null 2>&1; then healthy=1; break; fi
  sleep 1
done

if (( healthy == 0 )); then
  systemctl stop "$SERVICE_NAME" || true
  rollback_current || die '无法恢复上一版本链接'
  if (( was_active )); then systemctl start "$SERVICE_NAME" || true; fi
  if ! finalize_failed_job; then
    exit 1
  fi
  rm -f -- "$REQUEST_FILE"
  clear_update_state || true
  handled=1
  trap - EXIT
  exit 1
fi

# Preserve an operator's intentionally stopped service after validating the
# new release in a temporary start.
if (( was_active == 0 )); then
  systemctl stop "$SERVICE_NAME"
fi

write_update_state finalizing || exit 1
final_node=$(resolve_node) || die 'Node.js 24+ not found'
if [[ "$job_id" =~ ^[0-9a-f-]{36}$ ]]; then
  finalized=0
  for _ in 1 2 3; do
    if finalize_completed_job; then finalized=1; break; fi
    sleep 1
  done
  (( finalized == 1 )) || exit 1
fi
rm -f -- "$REQUEST_FILE"
clear_update_state || true
handled=1
trap - EXIT
exit 0