#!/usr/bin/env bash
# docker-nuke — capture diagnostics from a wedged Docker Desktop, then
# force-kill and restart it.
#
# Written for bash 3.2 compatibility (macOS /bin/bash).

set -euo pipefail

# --- configuration ----------------------------------------------------------
#
# The SC2034 directives below are per-line on purpose, not file-level: each
# constant is consumed by a later task, and its directive is deleted at that
# point. A file-level disable would silence SC2034 for the whole file forever,
# hiding a genuinely dead variable added later. By the final task none of these
# directives should remain.

HEALTH_TIMEOUT=6        # seconds to wait on `docker info` before calling it wedged
# shellcheck disable=SC2034
GRACEFUL_WAIT=8         # seconds to let Docker quit politely
# shellcheck disable=SC2034
TERM_WAIT=5             # seconds to wait after SIGTERM before SIGKILL
# shellcheck disable=SC2034
RESTART_TIMEOUT=90      # seconds to wait for the daemon to come back
# shellcheck disable=SC2034
POLL_INTERVAL=2         # seconds between daemon health polls
SAMPLE_DURATION=2       # seconds of stack sampling per process
SAMPLE_TIMEOUT=10       # hard cap on a single `sample` invocation
DUMP_RETENTION=10       # how many dump directories to keep
LOG_TAIL_LINES=2000     # lines of Docker log to copy (keeps dumps small)
LOG_CAPTURE_TIMEOUT=20  # hard cap on the whole capture_logs step
CRASH_CAPTURE_TIMEOUT=10 # hard cap on the whole capture_crashes step

DOCKER_APP="/Applications/Docker.app"
DOCKER_LOG_DIR="$HOME/Library/Containers/com.docker.docker/Data/log"
DOCKER_RAW="$HOME/Library/Containers/com.docker.docker/Data/vms/0/data/Docker.raw"
CRASH_DIR="$HOME/Library/Logs/DiagnosticReports"

# Overridable so tests never touch real state.
DUMP_ROOT="${DOCKER_NUKE_DUMP_ROOT:-$HOME/.local/state/docker-nuke}"

OPT_FORCE=0
OPT_NO_RESTART=0
OPT_DRY_RUN=0
OPT_DEEP=0

# --- helpers ----------------------------------------------------------------

bin_exists() { command -v "$1" >/dev/null 2>&1; }

# Run a command with a hard time limit. Returns 124 if it timed out.
# Call inside a conditional so `set -e` does not abort on a non-zero return.
run_timeout() {
  local secs="$1"
  shift
  # A shell function (e.g. disk_report) has no PATH-resolvable executable, so
  # an external timeout/gtimeout binary can never exec it directly (execvp
  # fails with rc=127). Route those straight to the pure-bash poller below,
  # which runs the command in this same shell instead of a child process.
  if bin_exists timeout && ! declare -F "$1" >/dev/null 2>&1; then
    timeout "$secs" "$@"
    return $?
  fi
  if bin_exists gtimeout && ! declare -F "$1" >/dev/null 2>&1; then
    gtimeout "$secs" "$@"
    return $?
  fi

  # Pure-bash fallback: run in background and poll.
  "$@" &
  local pid=$!
  local waited=0
  while kill -0 "$pid" 2>/dev/null; do
    if [ "$waited" -ge "$secs" ]; then
      kill -9 "$pid" 2>/dev/null || true
      wait "$pid" 2>/dev/null || true
      return 124
    fi
    sleep 1
    waited=$((waited + 1))
  done
  local rc=0
  wait "$pid" || rc=$?
  return "$rc"
}

# Exact process names to target, in addition to anything under Docker.app.
# NOTE: never use `pgrep -f docker` here — it matches this script's own
# command line and the tool would kill itself.
DOCKER_PROC_NAMES="com.docker.backend com.docker.build com.docker.krun \
com.docker.vpnkit com.docker.virtualization vpnkit docker"

docker_pids() {
  local self="$$"
  local parent="${PPID:-0}"
  local name pid comm
  {
    # `pgrep -f "$DOCKER_APP"` matches any process whose *command line*
    # mentions the bundle path — e.g. `tail -f /Applications/Docker.app/...`
    # matches even though that process is plainly not Docker. Post-filter
    # each candidate by its actual executable so only processes truly
    # launched from the bundle survive. Wrapped in `|| true`: under
    # `set -o pipefail`, an empty `pgrep -f` match (exit 1) would otherwise
    # make this whole statement fail and abort the function under `set -e`.
    pgrep -f "$DOCKER_APP" 2>/dev/null | while read -r pid; do
      if [ -z "$pid" ]; then continue; fi
      # A dead PID (process exited between pgrep and ps) or a `ps` failure
      # must yield an empty comm, which matches no case branch below —
      # never a crash, never a false positive.
      # NOTE: `ps -o comm=` reports the executable path as invoked, not
      # canonicalized — a Docker helper ever exec'd by relative path or bare
      # name would report a non-bundle comm here and be silently dropped.
      comm="$(ps -o comm= -p "$pid" 2>/dev/null)" || true
      case "$comm" in
        "$DOCKER_APP"/*) printf '%s\n' "$pid" ;;
      esac
    done || true

    # Exact-name matches are already precise — do NOT apply the comm
    # filter here. One job of this branch is catching a hung
    # /usr/local/bin/docker CLI, which legitimately lives outside the
    # bundle; filtering by $DOCKER_APP prefix would break that.
    for name in $DOCKER_PROC_NAMES; do
      pgrep -x "$name" 2>/dev/null || true
    done
  } | sort -un | while read -r pid; do
    if [ -z "$pid" ]; then continue; fi
    if [ "$pid" = "$self" ]; then continue; fi
    if [ "$pid" = "$parent" ]; then continue; fi
    printf '%s\n' "$pid"
  done
}

# Returns 0 if the daemon answers within HEALTH_TIMEOUT, else non-zero
# (124 specifically means it hung rather than errored).
docker_healthy() {
  local rc=0
  run_timeout "$HEALTH_TIMEOUT" docker info >/dev/null 2>&1 || rc=$?
  return "$rc"
}

DUMP_DIR=""
CAPTURE_FAILURES=""

# new_dump_dir — create and print a fresh timestamped directory under
# DUMP_ROOT. This is the one place in the dump path that fails loudly: if the
# directory cannot be created there is nowhere left to record a failure, so
# the caller must find out immediately rather than have every later capture
# silently write into a path that never existed.
new_dump_dir() {
  local stamp dir
  stamp="$(date -u '+%Y%m%dT%H%M%SZ')"
  dir="$DUMP_ROOT/$stamp"
  if ! mkdir -p "$dir" 2>/dev/null; then
    warn "docker-nuke: could not create dump directory $dir (dump root unwritable or full)"
    return 1
  fi
  printf '%s\n' "$dir"
}

# capture NAME SECONDS CMD... — write CMD output to $DUMP_DIR/NAME.
# Never fails: a timeout or error is recorded and execution continues, so a
# broken capture can never block recovery.
capture() {
  local name="$1" secs="$2"
  shift 2
  local out="$DUMP_DIR/$name"
  local rc=0
  run_timeout "$secs" "$@" >"$out" 2>&1 || rc=$?
  if [ "$rc" -ne 0 ]; then
    # The banner append itself must never be able to trip set -e: on an
    # unwritable or full dump root this redirection fails too, and without
    # the guard that failure -- not the original capture failure -- would be
    # what aborts the entire dump at the very first capture.
    printf '\n[docker-nuke] capture "%s" failed or timed out (rc=%s)\n' "$name" "$rc" >>"$out" 2>/dev/null || true
    CAPTURE_FAILURES="$CAPTURE_FAILURES $name"
  fi
  return 0
}

# Sample user-space stacks of the daemon-side processes. This is the one
# signal that cannot be reconstructed after the processes are killed.
capture_samples() {
  if ! bin_exists sample; then
    printf 'sample(1) not available\n' >"$DUMP_DIR/samples-skipped.txt"
    return 0
  fi
  local name pid
  for name in com.docker.backend com.docker.krun; do
    for pid in $(pgrep -x "$name" 2>/dev/null || true); do
      capture "sample-$name-$pid.txt" "$SAMPLE_TIMEOUT" \
        sample "$pid" "$SAMPLE_DURATION" -mayDie
    done
  done
}

capture_logs() {
  mkdir -p "$DUMP_DIR/logs"
  local f base
  for f in "$DOCKER_LOG_DIR"/host/*.log "$DOCKER_LOG_DIR"/vm/*.log; do
    if [ -f "$f" ]; then
      base="$(basename "$f")"
      tail -n "$LOG_TAIL_LINES" "$f" >"$DUMP_DIR/logs/$base" 2>/dev/null || true
    fi
  done
}

capture_crashes() {
  mkdir -p "$DUMP_DIR/crashes"
  if [ ! -d "$CRASH_DIR" ]; then
    return 0
  fi
  local f
  # grep exits 1 when no crash report matches (the common case: this dump
  # feature is expected to run against a healthy history most of the time).
  # Under `set -o pipefail` that would fail the whole pipeline and abort
  # run_dump via `set -e`. Group it with `|| true` so a clean "no matches"
  # never propagates as a pipeline failure.
  #
  # Grouping only the grep is not enough on its own: `head -5` closes its
  # read end once it has 5 lines, so with more than 5 matches upstream
  # `ls -1t` is later killed by SIGPIPE (rc=141) instead of exiting 0; an
  # unreadable CRASH_DIR gives `ls` rc=2 the same way. Either propagates as
  # this pipeline's exit status under pipefail and would abort the dump via
  # `set -e`, so the trailing `|| true` guards the pipeline as a whole, not
  # just the grep.
  # shellcheck disable=SC2012  # crash report names are macOS-generated, never contain newlines; ls -t gives the mtime order find cannot without extra plumbing
  ls -1t "$CRASH_DIR" 2>/dev/null | { grep -i docker || true; } | head -5 | while IFS= read -r f; do
    cp "$CRASH_DIR/$f" "$DUMP_DIR/crashes/" 2>/dev/null || true
  done || true
}

disk_report() {
  df -h
  printf '\n--- Docker.raw ---\n'
  ls -lh "$DOCKER_RAW" 2>/dev/null || printf 'Docker.raw not found\n'
  du -h -d0 "$DOCKER_RAW" 2>/dev/null || true
}

socket_report() {
  local sock="$HOME/.docker/run/docker.sock"
  printf 'socket: %s\n' "$sock"
  ls -l "$sock" 2>/dev/null || printf 'socket does not exist\n'
  printf '\n--- listeners ---\n'
  lsof "$sock" 2>/dev/null || printf 'no processes hold the socket\n'
}

run_dump() {
  local rc
  # DUMP_DIR is reset to empty first so a failed new_dump_dir never leaves it
  # pointing at a directory that does not exist (Fix 1). new_dump_dir is the
  # one place in this path that is allowed to fail loudly: on an unwritable
  # or full dump root it has already logged why via `warn` and returned
  # non-zero, and there is nowhere left to record a capture failure, so
  # run_dump stops here -- before any capture runs -- rather than limping
  # into the first capture and dying mid-dump.
  DUMP_DIR=""
  DUMP_DIR="$(new_dump_dir)" || return 1
  CAPTURE_FAILURES=""
  log "docker-nuke: capturing diagnostics to $DUMP_DIR"

  capture "ps.txt" 10 ps -eo pid,ppid,stat,%cpu,%mem,etime,command
  capture "docker-info.txt" "$HEALTH_TIMEOUT" docker info
  capture "docker-version.txt" "$HEALTH_TIMEOUT" docker version
  capture "disk.txt" 15 disk_report
  capture "memory.txt" 10 vm_stat
  capture "socket.txt" 10 socket_report

  # capture_logs and capture_crashes write files directly rather than
  # emitting stdout, so they are not routed through capture() (which would
  # redirect stdout to a single named artifact). They still must be
  # time-bounded like every other step, and run_timeout's `declare -F` guard
  # (see run_timeout above) routes them to the pure-bash poller since neither
  # is a PATH-resolvable executable. A timeout or error is recorded into
  # CAPTURE_FAILURES the same way capture() records one, so a wedged
  # filesystem here cannot hang the dump either.
  rc=0
  run_timeout "$LOG_CAPTURE_TIMEOUT" capture_logs || rc=$?
  if [ "$rc" -ne 0 ]; then
    CAPTURE_FAILURES="$CAPTURE_FAILURES logs"
  fi

  rc=0
  run_timeout "$CRASH_CAPTURE_TIMEOUT" capture_crashes || rc=$?
  if [ "$rc" -ne 0 ]; then
    CAPTURE_FAILURES="$CAPTURE_FAILURES crashes"
  fi

  capture_samples
}

prune_dumps() {
  local keep="${1:-$DUMP_RETENTION}"
  if [ ! -d "$DUMP_ROOT" ]; then
    return 0
  fi
  local d
  # An unreadable DUMP_ROOT gives `ls` a non-zero exit, which -- since this
  # pipeline is prune_dumps' last statement and therefore its return value --
  # would otherwise abort a `set -e` caller. The trailing `|| true` matches
  # the guard used in capture_crashes for the same class of problem; the
  # normal-path behavior (prune the oldest, keep $keep newest) is unchanged.
  # shellcheck disable=SC2012  # dump dir names are our own generated timestamps, never contain newlines; sort -r needs plain lines, not find -print0
  ls -1 "$DUMP_ROOT" 2>/dev/null | sort -r | tail -n "+$((keep + 1))" | while IFS= read -r d; do
    if [ -n "$d" ]; then
      rm -rf "${DUMP_ROOT:?}/$d"
    fi
  done || true
  return 0
}

log()  { printf '%s\n' "$*"; }
warn() { printf '%s\n' "$*" >&2; }

usage() {
  cat <<'EOF'
docker-nuke: force-restart a wedged Docker Desktop, capturing diagnostics first

Usage:
  docker-nuke [options]

Options:
  -f, --force        Skip the confirmation shown when Docker looks healthy
  -n, --no-restart   Kill Docker but do not relaunch it
      --dry-run      Report what would be killed and dumped; change nothing
      --deep         Also stop root launchd helpers (com.docker.vmnetd); needs sudo
  -h, --help         Show this help

Diagnostics are written to ~/.local/state/docker-nuke/<timestamp>/ before
anything is killed. The most recent 10 dumps are kept.

Exit codes:
  0  Docker killed and confirmed responsive again
  1  Daemon did not come back within the restart timeout
  2  Usage error, or aborted at the healthy-daemon prompt
EOF
}

parse_args() {
  OPT_FORCE=0
  OPT_NO_RESTART=0
  OPT_DRY_RUN=0
  OPT_DEEP=0
  while [ "$#" -gt 0 ]; do
    # shellcheck disable=SC2034  # OPT_* are read by main once it is wired up
    case "$1" in
      -f|--force)      OPT_FORCE=1 ;;
      -n|--no-restart) OPT_NO_RESTART=1 ;;
      --dry-run)       OPT_DRY_RUN=1 ;;
      --deep)          OPT_DEEP=1 ;;
      -h|--help)       usage; return 3 ;;
      *)
        warn "docker-nuke: unknown option '$1'"
        usage >&2
        return 2
        ;;
    esac
    shift
  done
  return 0
}

# --- entrypoint -------------------------------------------------------------

main() {
  local rc=0
  parse_args "$@" || rc=$?
  if [ "$rc" -eq 3 ]; then return 0; fi
  if [ "$rc" -ne 0 ]; then return "$rc"; fi
  log "docker-nuke: not yet implemented"
  return 0
}

# Only run when executed, not when sourced by the test harness.
if [ "${BASH_SOURCE[0]}" = "${0}" ]; then
  main "$@"
fi
