#!/bin/bash

# ============================================================================
#  LibrePortal Task Processor — v2
#
#  Design goals:
#    1. Single source of truth: the task file (`<id>.json`). `status` is the
#       only field anyone reads to know what a task is doing.
#    2. Single instance: enforced via flock. A second process exits cleanly.
#    3. Push-based: the daemon blocks on a FIFO read and wakes within
#       milliseconds when Node writes a task id to the FIFO.
#    4. Subprocess isolation: every command runs inside `setsid bash -c …`
#       so that a non-zero exit (or signal, or `exit 1` from checkSuccess in
#       non-interactive mode) kills the subprocess, not the processor.
#    5. Atomic writes: every state transition is `tmp + mv`. Readers never
#       see a half-written file.
#    6. Heartbeat: the processor stamps `heartbeat_at` every 5s while a task
#       runs. Stale heartbeats (>60s on a `running` task) are recovered to
#       `failed` on the next idle cycle.
#    7. Cancellable: a `<id>.cancel` marker file triggers SIGTERM → SIGKILL
#       to the task's process group.
# ============================================================================

script_task_processor_flag="$1"
# Source guard — DO NOT remove. mainLoop is an infinite loop.
[[ "$script_task_processor_flag" != "start_script" ]] && return 0 2>/dev/null

# --- Load the privilege helpers + docker-type config ------------------------
# systemd launches this script standalone, so the de-sudo helpers
# (runFileOp/runFileWrite) and the config they key off (rooted vs rootless) are
# NOT otherwise in scope. Without them every privileged write into the
# docker-install-owned task dir fails ("command not found") and tasks loop
# forever. Load them here — these files are pure function/var defs, safe to
# source, no side effects.
# Self-locate the scripts dir (where systemd launched us) so we can load the
# relocatable path roots — paths.sh sets docker_dir/containers_dir/configs_dir/…
# (the data + backup roots can't be derived from our own location).
LP_SELF_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" 2>/dev/null && pwd)"
LP_SCRIPTS="${install_scripts_dir:-$(cd "$LP_SELF_DIR/../.." 2>/dev/null && pwd)/}"
[[ -f "${LP_SCRIPTS}source/paths.sh" ]] && source "${LP_SCRIPTS}source/paths.sh"
LP_SCRIPTS="${install_scripts_dir:-$LP_SCRIPTS}"
LP_DOCKER_CFG="${configs_dir:-/libreportal-system/configs/}general/general_docker_install"
[[ -f "$LP_DOCKER_CFG" ]] && \
  eval "$(grep -E '^CFG_DOCKER_INSTALL_(TYPE|USER)=' "$LP_DOCKER_CFG" | sed 's/[[:space:]]*#.*//')"
: "${sudo_user_name:=libreportal}"
: "${containers_dir:=/libreportal-containers/}"
: "${docker_dir:=/libreportal-system}"
for _lp_f in docker/command/run_privileged.sh \
             docker/command/docker_run_install.sh \
             checks/requirements/check_install_type.sh; do
  [[ -f "${LP_SCRIPTS}${_lp_f}" ]] && source "${LP_SCRIPTS}${_lp_f}"
done
command -v resolveDockerInstallUser >/dev/null 2>&1 && resolveDockerInstallUser

# ============================================================================
#  PATHS & CONSTANTS
# ============================================================================

TASK_DIR="${TASK_DIR:-${containers_dir:-/libreportal-containers/}libreportal/frontend/data/tasks}"
LOCK_FILE="$TASK_DIR/.processor.lock"
FIFO="$TASK_DIR/.queue.fifo"
LOG_FILE="$TASK_DIR/task_processor.log"

# Cron's default PATH is minimal (/usr/bin:/bin) and doesn't include the
# /usr/local/bin where the libreportal CLI lives. Augment it so eval'd commands
# can find their binaries.
export PATH="/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:$PATH"

HEARTBEAT_INTERVAL=5
HEARTBEAT_STALE_SECS=60
IDLE_POLL_SECS=3    # max time we'll block on FIFO before re-scanning anyway

# ============================================================================
#  LOGGING
# ============================================================================

logInfo()  { echo "$(date '+%Y-%m-%d %H:%M:%S') [INFO ] $*" | runFileWrite -a "$LOG_FILE"; }
logError() { echo "$(date '+%Y-%m-%d %H:%M:%S') [ERROR] $*" | runFileWrite -a "$LOG_FILE"; }
logDebug() { [[ "${LOG_LEVEL:-INFO}" == "DEBUG" ]] && echo "$(date '+%Y-%m-%d %H:%M:%S') [DEBUG] $*" | runFileWrite -a "$LOG_FILE"; }

# ============================================================================
#  FILE OPS  — atomic writes, JSON helpers
# ============================================================================

# Atomically replace a file's content. POSIX rename(2) is atomic on the same
# filesystem, so concurrent readers either see the old or the new file —
# never a half-written one.
writeAtomic() {
  local target="$1"; local content="$2"
  local tmp="${target}.tmp.$$"
  printf '%s' "$content" | runFileWrite "$tmp"
  runFileOp chmod 644 "$tmp" 2>/dev/null
  runFileOp mv "$tmp" "$target"
}

# Update one or more fields on a task file, atomically. Args after taskFile
# are alternating <key> <value> pairs.
#   updateTaskFields "$file" status running heartbeat_at "$(date -Iseconds)"
updateTaskFields() {
  local taskFile="$1"; shift
  [[ -f "$taskFile" ]] || return 1
  command -v jq >/dev/null 2>&1 || return 1

  local jqArgs=() jqExpr='.'
  while [[ $# -ge 2 ]]; do
    local key="$1"; local value="$2"; shift 2
    jqArgs+=(--arg "$key" "$value")
    jqExpr="$jqExpr | .$key = \$$key"
  done

  # The processor runs as the manager user, which CANNOT create files in the
  # docker-install-owned TASK_DIR. A plain `jq … > "$tmp"` redirect therefore
  # failed silently, so neither `status: running` nor `status: completed` was
  # ever written and the task was reprocessed forever. Capture jq's output and
  # write the temp via runFileWrite (the privileged helper, like writeAtomic),
  # then chmod + atomic mv as before.
  local updated
  updated=$(jq "${jqArgs[@]}" "$jqExpr" "$taskFile" 2>/dev/null) || return 1
  [[ -z "$updated" ]] && return 1
  local tmp="${taskFile}.tmp.$$"
  printf '%s' "$updated" | runFileWrite "$tmp" \
    && runFileOp chmod 644 "$tmp" 2>/dev/null \
    && runFileOp mv "$tmp" "$taskFile"
}

readTaskField() {
  local taskFile="$1"; local field="$2"
  [[ -f "$taskFile" ]] || return 1
  command -v jq >/dev/null 2>&1 || return 1
  jq -r --arg f "$field" '.[$f] // empty' "$taskFile" 2>/dev/null
}

# ============================================================================
#  SETUP / SINGLE INSTANCE
# ============================================================================

setupTaskDir() {
  runFileOp mkdir -p "$TASK_DIR"
  if [[ ! -p "$FIFO" ]]; then
    # Remove any stale non-FIFO file at that path before creating the FIFO.
    [[ -e "$FIFO" ]] && runFileOp rm -f "$FIFO"
    runFileOp mkfifo "$FIFO" 2>/dev/null
  fi
  runFileOp chmod 666 "$FIFO" 2>/dev/null
  runFileOp chmod 755 "$TASK_DIR" 2>/dev/null
  # The processor (manager user) can't create files in the docker-install-owned
  # task dir, so pre-create the lock AS the dir owner, world-writable, so the
  # `exec 200>"$LOCK_FILE"` in acquireSingletonLock (run as the manager) can open
  # it. Create-if-absent to keep a stable inode for flock across restarts.
  [[ -e "$LOCK_FILE" ]] || runFileOp install -m 666 /dev/null "$LOCK_FILE" 2>/dev/null
  runFileOp chmod 666 "$LOCK_FILE" 2>/dev/null
  # Establish ownership via the root-owned helper: the unprivileged dir owner
  # can't reclaim files an earlier run left root/manager-owned (e.g. a root-owned
  # task_processor.log), which would then block the daemon's log appends.
  runOwnership taskdir 2>/dev/null
}

# Open the FIFO in read-write mode on fd 3. With <>, Linux returns from open()
# immediately (process is both reader and writer) so subsequent reads with a
# timeout actually time out instead of hanging forever waiting for a writer.
openFifoReader() {
  if [[ -p "$FIFO" ]]; then
    exec 3<> "$FIFO" || return 1
    return 0
  fi
  return 1
}

acquireSingletonLock() {
  exec 200>"$LOCK_FILE"
  if ! flock -n 200; then
    logInfo "Another task processor is already running; this instance will exit."
    exit 0
  fi
  echo $$ | runFileWrite "${LOCK_FILE}.pid" 2>/dev/null
}

# ============================================================================
#  ORPHAN RECOVERY
# ============================================================================
# `running` tasks whose heartbeat is older than HEARTBEAT_STALE_SECS, or
# tasks that have no heartbeat and a started_at older than the same threshold,
# are treated as dead (the processor that owned them is gone).

recoverOrphans() {
  command -v jq >/dev/null 2>&1 || return 0
  local now; now=$(date +%s)

  # Batch the read: one `jq` invocation across every task file emits
  # "<file>\t<heartbeat-or-started-or-empty>" for each running task. The
  # previous loop spawned 1-2 jq processes per file, which dominated startup
  # and idle-tick latency once a few dozen task files were on disk.
  shopt -s nullglob
  local files=( "$TASK_DIR"/task_*.json )
  shopt -u nullglob
  [[ ${#files[@]} -eq 0 ]] && return 0

  while IFS=$'\t' read -r file stamp; do
    [[ -n "$file" ]] || continue
    [[ -n "$stamp" ]] || continue
    local stampEpoch; stampEpoch=$(date -d "$stamp" +%s 2>/dev/null) || continue
    if (( now - stampEpoch > HEARTBEAT_STALE_SECS )); then
      local id; id=$(basename "$file" .json)
      logInfo "Orphan recovery: $id (heartbeat age $((now - stampEpoch))s) -> failed"
      updateTaskFields "$file" \
        status failed \
        error_message "Task interrupted — processor died mid-run." \
        completed_at "$(date -Iseconds)" \
        updated_at "$(date -Iseconds)"
    fi
  done < <(jq -r 'select(.status == "running") | "\(input_filename)\t\(.heartbeat_at // .started_at // "")"' "${files[@]}" 2>/dev/null)
}

# ============================================================================
#  TASK EXECUTION
# ============================================================================
# A single task runs to completion in this function. The eval'd command lives
# in a `setsid bash -c` subshell so an `exit 1` from inside it cannot kill us.

runTask() {
  local taskFile="$1"
  local taskId; taskId=$(basename "$taskFile" .json)
  local logFile="$TASK_DIR/${taskId}.log"
  local cancelMarker="$TASK_DIR/${taskId}.cancel"

  local command; command=$(readTaskField "$taskFile" command)
  if [[ -z "$command" || "$command" == "null" ]]; then
    logError "Task $taskId has no command; marking failed."
    updateTaskFields "$taskFile" \
      status failed \
      error_message "Task file has no command field." \
      completed_at "$(date -Iseconds)" \
      updated_at "$(date -Iseconds)"
    return 1
  fi

  logInfo "Task $taskId starting: $command"

  # Any pre-existing cancel marker is stale; remove it.
  [[ -f "$cancelMarker" ]] && runFileOp rm -f "$cancelMarker"

  # Mark running with initial heartbeat.
  local now; now=$(date -Iseconds)
  updateTaskFields "$taskFile" \
    status running \
    started_at "$now" \
    heartbeat_at "$now" \
    updated_at "$now"

  # Initialise the log file owned by the daemon's user so the bash redirection
  # `>>"$logFile"` (which runs as that user, NOT under sudo) can append to it.
  # Previously this used `sudo truncate` + `sudo chmod 644` which left the file
  # root-owned and unwritable to the daemon, so the redirection failed and the
  # task immediately exited with rc=1.
  # TASK_DIR is owned by the docker install user, so the manager-user processor
  # can't create the log there directly. Create/truncate it AS the dir owner via
  # runFileOp and leave it world-writable for the run so the eval's
  # `>>"$logFile"` append (which runs as the manager, NOT under sudo) succeeds.
  # Re-owned to the dir owner when the task finishes.
  runFileOp install -m 666 /dev/null "$logFile" 2>/dev/null \
    || runFileOp truncate -s 0 "$logFile" 2>/dev/null
  runFileOp chmod 666 "$logFile" 2>/dev/null

  export LIBREPORTAL_NONINTERACTIVE=1

  # Run the command in a subshell so:
  #   * `eval` inherits the daemon's full env (PATH, functions, vars). Switching
  #     to `bash -c` broke this because the libreportal CLI lives at
  #     /usr/local/bin/libreportal which a fresh bash subshell can't always find.
  #   * an `exit 1` from inside (e.g. checkSuccess in non-interactive mode)
  #     only terminates the subshell — the daemon keeps running.
  # `set -m` gives the backgrounded subshell its own process group so we can
  # kill -TERM the whole tree on cancel via `kill -TERM -$cmdPid`.
  set -m
  ( cd "$HOME" 2>/dev/null || cd /; eval "$command" ) </dev/null >>"$logFile" 2>&1 &
  local cmdPid=$!
  set +m

  # Heartbeat + cancel watcher. Runs alongside the command and exits as soon
  # as the command's pid is gone.
  #
  # `set -m` so the watcher subshell becomes its own process group leader —
  # we need that below to SIGKILL the bash subshell *and* any in-flight
  # jq / sudo children with a single signal, so a heartbeat update that
  # started just before the command exited can't write `status=running`
  # back over our `status=completed` write.
  #
  # Inside the loop we poll in 1-second chunks instead of one
  # `sleep $HEARTBEAT_INTERVAL`. With the long sleep the watcher only
  # noticed the command had finished after the next 5s tick, which was the
  # queued -> completed lag the user was seeing. Same trick gives us ~1s
  # cancel-marker response instead of up to 5s.
  set -m
  (
    while kill -0 "$cmdPid" 2>/dev/null; do
      updateTaskFields "$taskFile" heartbeat_at "$(date -Iseconds)" >/dev/null 2>&1
      for ((i=0; i<HEARTBEAT_INTERVAL; i++)); do
        kill -0 "$cmdPid" 2>/dev/null || exit 0
        if [[ -f "$cancelMarker" ]]; then
          logInfo "Task $taskId cancel requested; sending SIGTERM."
          kill -TERM "-$cmdPid" 2>/dev/null
          sleep 5
          if kill -0 "$cmdPid" 2>/dev/null; then
            logInfo "Task $taskId did not exit on TERM; sending SIGKILL."
            kill -KILL "-$cmdPid" 2>/dev/null
          fi
          runFileOp rm -f "$cancelMarker"
          exit 0
        fi
        sleep 1
      done
    done
  ) &
  local watcherPid=$!
  set +m

  # Wait for the command to finish (or to be killed by the watcher).
  wait "$cmdPid"
  local rc=$?

  # Stop the watcher. SIGKILL on the *process group* so any jq/sudo child of
  # the watcher that's mid-write also dies — otherwise their `sudo mv` can
  # land after our final status write and stomp it back to running.
  kill -KILL -- "-$watcherPid" 2>/dev/null
  wait "$watcherPid" 2>/dev/null

  local finalStatus="completed"
  local errorMessage=""

  if [[ -f "$cancelMarker" ]] || [[ $rc -eq 143 ]] || [[ $rc -eq 137 ]]; then
    finalStatus="cancelled"
    errorMessage="Cancelled by user."
    runFileOp rm -f "$cancelMarker" 2>/dev/null
  elif [[ $rc -ne 0 ]]; then
    finalStatus="failed"
    errorMessage="Command exited with status $rc."
  fi

  updateTaskFields "$taskFile" \
    status "$finalStatus" \
    completed_at "$(date -Iseconds)" \
    updated_at "$(date -Iseconds)" \
    exit_code "$rc" \
    error_message "$errorMessage"

  if [[ -n "$docker_install_user" ]]; then
    runFileOp chown "$docker_install_user":"$docker_install_user" "$logFile" 2>/dev/null
  fi

  case "$finalStatus" in
    completed) logInfo "Task $taskId completed (exit $rc)";;
    cancelled) logInfo "Task $taskId cancelled";;
    failed)    logError "Task $taskId failed (exit $rc)";;
  esac
}

# ============================================================================
#  SCAN & DISPATCH
# ============================================================================
# Process every queued task in created-at order. We keep going until the
# directory has no more `queued` files; FIFO wake-ups will deliver new ones.

dispatchPending() {
  command -v jq >/dev/null 2>&1 || { logError "jq not available; cannot dispatch."; return 1; }

  while true; do
    # Sort by mtime (oldest first) so FIFO order is preserved, then run a
    # single `jq` over all of them — emits the path of every queued/pending
    # task, in argument order. We pick the first; runTask consumes it; loop.
    # This replaces the previous N-jq-per-pass scan that scaled poorly with
    # how many task files were on disk and was the dominant queued ->
    # running latency on healthy systems.
    local listing
    listing=$(ls -tr "$TASK_DIR"/task_*.json 2>/dev/null) || listing=""
    [[ -z "$listing" ]] && return 0

    local IFS=$'\n'
    local files=( $listing )
    unset IFS

    local nextFile
    nextFile=$(jq -r 'select(.status == "queued" or .status == "pending") | input_filename' "${files[@]}" 2>/dev/null | head -n 1)
    [[ -z "$nextFile" ]] && return 0

    runTask "$nextFile"
  done
}

# Fast path used by the FIFO wake-up. Skips the whole-dir scan + per-file jq
# spawn that dispatchPending does — with many tasks on disk that scan was the
# dominant queued -> running latency.
dispatchSpecific() {
  local taskId="$1"
  [[ -z "$taskId" ]] && return 0
  # Guard against junk in the FIFO.
  [[ "$taskId" =~ ^task_[0-9]+_[A-Za-z0-9]+$ ]] || return 0
  local taskFile="$TASK_DIR/${taskId}.json"
  [[ -f "$taskFile" ]] || return 0
  local s; s=$(readTaskField "$taskFile" status)
  if [[ "$s" == "queued" || "$s" == "pending" ]]; then
    runTask "$taskFile"
  fi
}

# ============================================================================
#  HOUSEKEEPING
# ============================================================================

cleanupZeroByteFiles() {
  # Use the bash builtin `-s` (file size > 0) instead of forking `stat` per
  # file — at 100+ task files the stat fork was a measurable share of the
  # idle-tick cost.
  for f in "$TASK_DIR"/task_*; do
    [[ -f "$f" ]] || continue
    [[ -s "$f" ]] && continue
    runFileOp rm -f "$f"
  done
}

# ============================================================================
#  MAIN LOOP
# ============================================================================
# Block on the FIFO with a timeout. Two paths:
#
#   * FIFO wake-up — Node poked us with the new task id. Skip the full
#     housekeeping pass (orphan recovery + zero-byte cleanup re-iterate every
#     task file and shell out to jq for each; with N tasks on disk that's the
#     dominant queued -> running latency). Dispatch the specific task by id,
#     then drain any other queued ones, then back to read.
#
#   * Idle timeout — no Node activity. Run the full housekeeping pass; this
#     is what catches orphaned-running tasks from a previously-killed
#     processor and tasks that were created while the FIFO was unhealthy.
#
# With a healthy Node side, queued -> running latency is now bounded by one
# `read`, one jq invocation, and the runTask setup — typically well under
# 100ms regardless of how many task files are on disk.

mainLoop() {
  logInfo "Task processor v2 entering main loop"

  # Open the FIFO once and reuse the descriptor. `< "$FIFO"` per-iteration would
  # block at open() waiting for a writer, defeating the read timeout entirely.
  local fifoReady=0
  if openFifoReader; then fifoReady=1; fi

  # One full housekeeping pass at startup so anything left behind from a
  # previous processor gets noticed before we enter the fast path.
  recoverOrphans
  dispatchPending
  cleanupZeroByteFiles

  local signal rc
  while true; do
    signal=""
    if [[ "$fifoReady" -eq 1 ]]; then
      # `read -u 3` reads from the fd we opened in <> mode. -t is the timeout.
      read -t "$IDLE_POLL_SECS" -u 3 signal
      rc=$?
    else
      sleep "$IDLE_POLL_SECS"
      rc=1
      # Try to (re)open the FIFO if it's appeared since startup.
      [[ -p "$FIFO" ]] && openFifoReader && fifoReady=1
    fi

    if [[ $rc -eq 0 && -n "$signal" ]]; then
      # Fast path: dispatch the specific task we were poked about, then drain
      # anything else that piled up while we were running it.
      dispatchSpecific "$signal"
      dispatchPending
    else
      # Idle timeout (or FIFO unhealthy) — full housekeeping pass.
      recoverOrphans
      dispatchPending
      cleanupZeroByteFiles
    fi
  done
}

# ============================================================================
#  ENTRY POINT
# ============================================================================

run_task_processor() {
  logInfo "=== LibrePortal task processor starting ==="
  setupTaskDir
  acquireSingletonLock
  cleanupZeroByteFiles
  recoverOrphans
  mainLoop
}

run_task_processor