watcher.hpp

#ifndef W973564ED9F278A21F3E12037288412FBAF175F889
#define W973564ED9F278A21F3E12037288412FBAF175F889

#include <array>
#include <charconv>
#include <chrono>
#include <filesystem>
#include <functional>
#include <ios>
#include <limits>
#include <memory>
#include <string>
#include <string_view>
#include <type_traits>

namespace wtr {
inline namespace watcher {

/*  The `event` object is used to carry information about
    filesystem events to the user through the (user-supplied)
    callback given to `watch`.

    The `event` object will contain the:
      - `path_name`: The path to the event.
      - `path_type`: One of:
        - dir
        - file
        - hard_link
        - sym_link
        - watcher
        - other
      - `effect_type`: One of:
        - rename
        - modify
        - create
        - destroy
        - owner
        - other
      - `effect_time`:
        The time of the event in nanoseconds since epoch.

    The `watcher` type is special.
    Events with this type will include messages from
    the watcher. You may recieve error messages or
    important status updates.

    The first event always has a `create` value for the
    `effect_type`, a `watcher` value for the `path_type`,
    and a status message in the `path_name` field; Either
    "s/self/live@{some path}" or "e/self/live@{some path}".

    Similarly, the last event always carries `destroy` and
    `watcher` in the `effect_type` and `path_type` fields.
    The `path_name` field will have the same message as the
    first event, except for "die" instead of "live". */

struct event {

private:
  /*  I like these names. Very human. */
  using Nanos = std::chrono::nanoseconds;
  using Clock = std::chrono::system_clock;
  using TimePoint = std::chrono::time_point<Clock>;

public:
  /*  Ensure the user's callback can receive
      events and will return nothing. */
  using callback = std::function<void(event const&)>;

  /*  Represents "what happened" to a path. */
  enum class effect_type {
    rename,
    modify,
    create,
    destroy,
    owner,
    other,
  };

  /*  The essential types of paths. */
  enum class path_type {
    dir,
    file,
    hard_link,
    sym_link,
    watcher,
    other,
  };

  std::filesystem::path const path_name{};

  enum effect_type const effect_type {};

  enum path_type const path_type {};

  long long const effect_time{std::chrono::duration_cast<Nanos>(
                                TimePoint{Clock::now()}.time_since_epoch())
                                .count()};

  std::unique_ptr<event> const associated{nullptr};

  inline event(event const& from) noexcept
      : path_name{from.path_name}
      , effect_type{from.effect_type}
      , path_type{from.path_type}
      , effect_time{from.effect_time}
      , associated{
          from.associated ? std::make_unique<event>(*from.associated)
                          : nullptr} {};

  inline event(
    std::filesystem::path const& path_name,
    enum effect_type effect_type,
    enum path_type path_type) noexcept
      : path_name{path_name}
      , effect_type{effect_type}
      , path_type{path_type} {};

  inline event(event const& base, event&& associated) noexcept
      : path_name{base.path_name}
      , effect_type{base.effect_type}
      , path_type{base.path_type}
      , associated{std::make_unique<event>(std::forward<event>(associated))} {};

  inline ~event() noexcept = default;

  /*  An equality comparison for all the fields in this object.
      Includes the `effect_time`, which might not be wanted,
      because the `effect_time` is typically (not always) unique. */
  inline friend auto operator==(event const& l, event const& r) noexcept -> bool
  {
    return l.path_name == r.path_name && l.effect_time == r.effect_time
        && l.path_type == r.path_type && l.effect_type == r.effect_type
        && (l.associated && r.associated ? *l.associated == *r.associated
                                         : ! l.associated && ! r.associated);
  }

  inline friend auto operator!=(event const& l, event const& r) noexcept -> bool
  {
    return ! (l == r);
  }
};

} /*  namespace watcher */

// clang-format off

namespace {

#define wtr_effect_type_to_str_lit(Char, from, Lit) \
  switch (from) { \
    case ::wtr::event::effect_type::rename  : return Lit"rename"; \
    case ::wtr::event::effect_type::modify  : return Lit"modify"; \
    case ::wtr::event::effect_type::create  : return Lit"create"; \
    case ::wtr::event::effect_type::destroy : return Lit"destroy"; \
    case ::wtr::event::effect_type::owner   : return Lit"owner"; \
    case ::wtr::event::effect_type::other   : return Lit"other"; \
    default                                 : return Lit"other"; \
  }

#define wtr_path_type_to_str_lit(Char, from, Lit) \
  switch (from) { \
    case ::wtr::event::path_type::dir       : return Lit"dir"; \
    case ::wtr::event::path_type::file      : return Lit"file"; \
    case ::wtr::event::path_type::hard_link : return Lit"hard_link"; \
    case ::wtr::event::path_type::sym_link  : return Lit"sym_link"; \
    case ::wtr::event::path_type::watcher   : return Lit"watcher"; \
    case ::wtr::event::path_type::other     : return Lit"other"; \
    default                                 : return Lit"other"; \
  }

#define wtr_event_to_str_cls_as_json(Char, from, Lit)                         \
  using Cls = std::basic_string<Char>;                                        \
  auto&& etm = Lit"\"" + to<Cls>(from.effect_time) + Lit"\"";                 \
  auto&& ety = Lit"\"" + to<Cls>(from.effect_type) + Lit"\"";                 \
  auto&& pnm = Lit"\"" + to<Cls>(from.path_name)   + Lit"\"";                 \
  auto&& pty = Lit"\"" + to<Cls>(from.path_type)   + Lit"\"";                 \
  return {                         etm + Lit":{"                              \
         + Lit"\"effect_type\":" + ety + Lit","                               \
         + Lit"\"path_name\":"   + pnm + Lit","                               \
         + Lit"\"path_type\":"   + pty                                        \
         + [&]() -> Cls {                                                     \
              if (! from.associated) return Cls{};                            \
              auto asc = from.associated.get();                               \
              auto ttl = Cls{Lit",\"associated\""};                           \
              auto ety = Lit"\"" + to<Cls>(asc->effect_type) + Lit"\"";       \
              auto pnm = Lit"\"" + to<Cls>(asc->path_name)   + Lit"\"";       \
              auto pty = Lit"\"" + to<Cls>(asc->path_type)   + Lit"\"";       \
              return { ttl                                                    \
                     + Lit":{"                                                \
                     + Lit"\"effect_type\":" + ety + Lit","                   \
                     + Lit"\"path_name\":"   + pnm + Lit","                   \
                     + Lit"\"path_type\":"   + pty + Lit"}"                   \
                     };                                                       \
           }()                                                                \
         + Lit"}"                                                             \
         };

/*  For types larger than char and/or char8_t, we can just cast
    each element in our `char` buffer to an `unsigned char`, and
    then zero-extend the elements to any of `wchar_t`, `char16_t`
    and/or `char32_t`.
    We can use something like `format_to(chararray, "{}", from)`
    when library support is more common.
    If we support C++20 later on, we should parameterize `from`
    as `std::integral auto from` and add `char8_t` to the list
    of allowed narrow character types. */
template<class Char>
inline auto num_to_str(long long from) noexcept -> std::basic_string<Char> {
  static_assert(std::is_integral_v<decltype(from)>);
  static constexpr bool is_sys_sane_narrow_char_size =
      (sizeof(char) == sizeof(signed char))
      && (sizeof(char) == sizeof(unsigned char))
  ;
  static constexpr bool is_narrow_char = (
    is_sys_sane_narrow_char_size
    && (  std::is_same_v<Char, char>
       || std::is_same_v<Char, signed char>
       || std::is_same_v<Char, unsigned char>
    )
  );
  static constexpr bool is_wide_char = (
       std::is_same_v<Char, wchar_t>
    || std::is_same_v<Char, char16_t>
    || std::is_same_v<Char, char32_t>
  );
  static_assert(is_narrow_char || is_wide_char);

  static constexpr auto buflen = std::numeric_limits<decltype(from)>::digits10 + 1;
  auto buf = std::array<char, buflen>{0};
  auto [_, ec] = std::to_chars(buf.data(), buf.data() + buf.size(), from);
  if (ec != std::errc{}) {
    return {};
  }
  else if constexpr (is_narrow_char) {
    return {reinterpret_cast<Char const*>(buf.data()), buflen};
  }
  else if constexpr (is_wide_char) {
    auto xbuf = std::array<Char, buflen>{0};
    for (std::size_t i = 0; i < buflen; ++i)
      xbuf[i] = static_cast<Char>(static_cast<unsigned char>(buf[i]));
    return {xbuf.data(), buflen};
  }
};

} /*  namespace */

/*  We use function templates, as opposed to class templates with an
    overloaded () operator, so that we can avoid the extra () or {}
    when (needlessly) "constructing" the class and to make extending
    the templates once or more for the same type (a-la parameterize the
    return type *and* the arguments) possible.

    The `template<template<class...> class... Ts> auto to(auto) noexcept`
    is just for the user. It allows specialization outside of this file,
    and has no effect within this file (it does not guide deduction here).

    Some downsides:
    `std::hash` takes the class template approach. That's a (well?) known
    api. It's not ideal that we miss out on a user's existing familiarity
    with that pattern.
    We need to "declare" a function template before "defining" it:
    `template<class T> auto to(OurTypeOrConcept) -> T` must come before
    the `template<> auto to(...) -> T { ... }`. I'm not sure why class
    templates can get away without declaring a template like that, or
    why we need to do that at all.
    Another downside to the function templates is that we need to "declare"
    function templates before "defining" them. That's not exactly intuitive
    to me, I'm fuzzy on why that is, and users probably shouldn't need to
    remember that whenever they specialize this template.
    We need to declare a template before defining it:
    `template<class T> auto to(OurTypeOrConcept) -> T` must come before
    the `template<> auto to(...) -> T { ... }`. I'm not sure why class
    templates can get away without declaring a template like that, or
    why we need to do that at all.

    Beware of linker errors when using a concept as an argument type.
    AFAICT This is something that can't be done, or I haven't figured out
    how just yet.

    The upsides to function templates are:
    - We get to a concise api, `to<Type>(from_thing)`
    - The user can further specialize the function templates, i.e. We can
      have a specialization `to<string>(type_in_file_a)` and, somewhere else,
      `to<string>(type_in_file_b)` as well. Class templates cannot do that. */

template<class T, class... Ts> auto to(T) noexcept -> decltype(auto);

template<class T> auto to (long long                                from) noexcept -> T ;
template<class T> auto to (decltype(::wtr::event::effect_time)      from) noexcept -> T ;
template<class T> auto to (enum     ::wtr::event::effect_type       from) noexcept -> T ;
template<class T> auto to (decltype(::wtr::event::path_name) const& from) noexcept -> T ;
template<class T> auto to (enum     ::wtr::event::path_type         from) noexcept -> T ;
template<class T> auto to (         ::wtr::event             const& from) noexcept -> T ;

template<> inline constexpr auto to<std::basic_string_view<char>>   (enum     ::wtr::event::effect_type       from) noexcept -> std::basic_string_view<char>    { wtr_effect_type_to_str_lit(  char,    from,   ""          );  }
template<> inline constexpr auto to<std::basic_string_view<char>>   (enum     ::wtr::event::path_type         from) noexcept -> std::basic_string_view<char>    { wtr_path_type_to_str_lit(    char,    from,   ""          );  }
template<> inline           auto to<std::basic_string     <char>>   (decltype(::wtr::event::path_name) const& from) noexcept -> std::basic_string<char>         { return {                     from.string()                };  }
template<> inline           auto to<std::basic_string     <char>>   (enum     ::wtr::event::path_type         from) noexcept -> std::basic_string<char>         { wtr_path_type_to_str_lit(    char,    from,   ""          );  }
template<> inline           auto to<std::basic_string     <char>>   (decltype(::wtr::event::effect_time)      from) noexcept -> std::basic_string<char>         { return num_to_str           <char>(   from                );  }
template<> inline           auto to<std::basic_string     <char>>   (enum     ::wtr::event::effect_type       from) noexcept -> std::basic_string<char>         { wtr_effect_type_to_str_lit(  char,    from,   ""          );  }
template<> inline           auto to<std::basic_string     <char>>   (         ::wtr::event             const& from) noexcept -> std::basic_string<char>         { wtr_event_to_str_cls_as_json(char,    from,   ""          );  }
template<> inline constexpr auto to<std::basic_string_view<wchar_t>>(enum     ::wtr::event::effect_type       from) noexcept -> std::basic_string_view<wchar_t> { wtr_effect_type_to_str_lit(  wchar_t, from,  L""          );  }
template<> inline constexpr auto to<std::basic_string_view<wchar_t>>(enum     ::wtr::event::path_type         from) noexcept -> std::basic_string_view<wchar_t> { wtr_path_type_to_str_lit(    wchar_t, from,  L""          );  }
template<> inline           auto to<std::basic_string     <wchar_t>>(enum     ::wtr::event::path_type         from) noexcept -> std::basic_string<wchar_t>      { wtr_path_type_to_str_lit(    wchar_t, from,  L""          );  }
template<> inline           auto to<std::basic_string     <wchar_t>>(decltype(::wtr::event::path_name) const& from) noexcept -> std::basic_string<wchar_t>      { return {                     from.wstring()               };  }
template<> inline           auto to<std::basic_string     <wchar_t>>(decltype(::wtr::event::effect_time)      from) noexcept -> std::basic_string<wchar_t>      { return num_to_str           <wchar_t>(from                );  }
template<> inline           auto to<std::basic_string     <wchar_t>>(enum     ::wtr::event::effect_type       from) noexcept -> std::basic_string<wchar_t>      { wtr_effect_type_to_str_lit(  wchar_t, from,  L""          );  }
template<> inline           auto to<std::basic_string     <wchar_t>>(         ::wtr::event             const& from) noexcept -> std::basic_string<wchar_t>      { wtr_event_to_str_cls_as_json(wchar_t, from,  L""          );  }
/*
template<> inline constexpr auto to<std::basic_string_view<char8_t>>(enum     ::wtr::event::effect_type       from) noexcept -> std::basic_string_view<char8_t> { wtr_effect_type_to_str_lit(  char8_t, from, u8""          );  }
template<> inline constexpr auto to<std::basic_string_view<char8_t>>(enum     ::wtr::event::path_type         from) noexcept -> std::basic_string_view<char8_t> { wtr_path_type_to_str_lit(    char8_t, from, u8""          );  }
template<> inline           auto to<std::basic_string     <char8_t>>(enum     ::wtr::event::path_type         from) noexcept -> std::basic_string<char8_t>      { wtr_path_type_to_str_lit(    char8_t, from, u8""          );  }
template<> inline           auto to<std::basic_string     <char8_t>>(decltype(::wtr::event::path_name) const& from) noexcept -> std::basic_string<char8_t>      { return {                     from.u8string()              };  }
template<> inline           auto to<std::basic_string     <char8_t>>(decltype(::wtr::event::effect_time)      from) noexcept -> std::basic_string<char8_t>      { return num_to_str           <char8_t>(from                );  }
template<> inline           auto to<std::basic_string     <char8_t>>(enum     ::wtr::event::effect_type       from) noexcept -> std::basic_string<char8_t>      { wtr_effect_type_to_str_lit(  char8_t, from, u8""          );  }
template<> inline           auto to<std::basic_string     <char8_t>>(         ::wtr::event             const& from) noexcept -> std::basic_string<char8_t>      { wtr_event_to_str_cls_as_json(char8_t, from, u8""          );  }
*/

#undef wtr_effect_type_to_str_lit
#undef wtr_path_type_to_str_lit
#undef wtr_event_to_str_cls_as_json

template<class Char, class Traits>
inline auto operator<<(
  std::basic_ostream<Char, Traits>& into,
  enum ::wtr::event::effect_type from) noexcept
-> std::basic_ostream<Char, Traits>&
{ return into << to<std::basic_string<Char, Traits>>(from); };

template<class Char, class Traits>
inline auto operator<<(
  std::basic_ostream<Char, Traits>& into,
  enum ::wtr::event::path_type from) noexcept
-> std::basic_ostream<Char, Traits>&
{ return into << to<std::basic_string<Char, Traits>>(from); };

/*  Streams out `path_name`, `effect_type` and `path_type`.
    Formats the stream as a json object.
    Looks like this (without line breaks)
      "1678046920675963000":{
       "effect_type":"create",
       "path_name":"/some_file.txt",
       "path_type":"file"
      } */
template<class Char, class Traits>
inline auto operator<<(
  std::basic_ostream<Char, Traits>& into,
  ::wtr::event const& from) noexcept
-> std::basic_ostream<Char, Traits>&
{ return into << to<std::basic_string<Char, Traits>>(from); };

// clang-format on

} /*  namespace wtr   */

#include <atomic>

#ifdef __linux__
#include <errno.h>
#include <stdint.h>
#include <string.h>
#include <sys/eventfd.h>
#include <unistd.h>
#elif defined(__APPLE__)
#include <dispatch/dispatch.h>
#endif

namespace detail::wtr::watcher {

/*  A semaphore-like construct which can be
    used with the "native" async I/O APIs
    on (currently) Linux and Darwin.

    On Darwin, this is a semaphore which is
    schedulable with the dispatch library.

    On Linux, this is an eventfd in semaphore
    mode. The file descriptor is exposed for
    use with poll and friends.

    On other platforms, this is an atomic flag
    which can be checked in a sleep, wake loop,
    ideally with a generous sleep time.
*/

class semabin {
public:
  enum state { pending, released, error };

private:
  mutable std::atomic<state> is = pending;

public:
#if defined(__linux__)

  int const fd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK | EFD_SEMAPHORE);

  inline auto release() noexcept -> state
  {
    auto write_ev = [this]()
    {
      if (eventfd_write(this->fd, 1) == 0)
        return released;
      else
        return error;
    };

    if (this->is == released)
      return released;
    else
      return this->is = write_ev();
  }

  inline auto state() const noexcept -> state
  {
    auto read_ev = [this]()
    {
      uint64_t _ = 0;
      if (eventfd_read(this->fd, &_) == 0)
        return released;
      else if (errno == EAGAIN)
        return pending;
      else
        return error;
    };

    if (this->is == pending)
      return pending;
    else
      return this->is = read_ev();
  }

  inline ~semabin() noexcept { close(this->fd); }

#elif defined(__APPLE__)

  mutable dispatch_semaphore_t sem = dispatch_semaphore_create(0);

  inline auto release() noexcept -> state
  {
    auto exchange_when = pending;
    auto was_exchanged = this->is.compare_exchange_strong(
      exchange_when,
      released,
      std::memory_order_release,
      std::memory_order_acquire);
    if (was_exchanged) dispatch_semaphore_signal(this->sem);
    return released;
  }

  inline auto state() const noexcept -> state
  {
    return this->is.load(std::memory_order_acquire);
  }

  inline ~semabin() noexcept { this->release(); }

#else

  inline auto release() noexcept -> enum state { return this->is = released; }

  inline auto state() const noexcept -> enum state { return this->is; }

#endif
};

} /*  namespace detail::wtr::watcher */

#if defined(__APPLE__)

#include <CoreFoundation/CoreFoundation.h>
#include <CoreServices/CoreServices.h>
#include <dispatch/dispatch.h>
#include <filesystem>
#include <mutex>
#include <string>
#include <unordered_set>

namespace detail {
namespace wtr {
namespace watcher {
namespace adapter {
namespace {

// clang-format off

/*  If we want less "sleepy" time after a period of time
    without receiving filesystem events, we could OR like:
    `fsev_flag_listen | kFSEventStreamCreateFlagNoDefer`.
    We're talking about saving a maximum latency of `delay_s`
    after some period of inactivity, which is not likely to
    be noticeable. I'm not sure what Darwin sets the "period
    of inactivity" to, and I'm not sure it matters. */
inline constexpr unsigned fsev_listen_for
  = kFSEventStreamCreateFlagFileEvents;
inline constexpr auto fsev_listen_since
  = kFSEventStreamEventIdSinceNow;
inline constexpr unsigned fsev_flag_path_file
  = kFSEventStreamEventFlagItemIsFile;
inline constexpr unsigned fsev_flag_path_dir
  = kFSEventStreamEventFlagItemIsDir;
inline constexpr unsigned fsev_flag_path_sym_link
  = kFSEventStreamEventFlagItemIsSymlink;
inline constexpr unsigned fsev_flag_path_hard_link
  = kFSEventStreamEventFlagItemIsHardlink
  | kFSEventStreamEventFlagItemIsLastHardlink;
inline constexpr unsigned fsev_flag_effect_create
  = kFSEventStreamEventFlagItemCreated;
inline constexpr unsigned fsev_flag_effect_remove
  = kFSEventStreamEventFlagItemRemoved;
inline constexpr unsigned fsev_flag_effect_modify_any
  = kFSEventStreamEventFlagItemModified
  | kFSEventStreamEventFlagItemChangeOwner
  | kFSEventStreamEventFlagItemXattrMod
  | kFSEventStreamEventFlagItemFinderInfoMod
  | kFSEventStreamEventFlagItemInodeMetaMod;
inline constexpr unsigned fsev_flag_effect_modify
  = fsev_flag_effect_modify_any
  & ~kFSEventStreamEventFlagItemInodeMetaMod;
inline constexpr unsigned fsev_flag_effect_rename
  = kFSEventStreamEventFlagItemRenamed;
inline constexpr unsigned fsev_flag_effect_any
  = fsev_flag_effect_create
  | fsev_flag_effect_remove
  | fsev_flag_effect_modify_any
  | fsev_flag_effect_rename;

// clang-format on

struct ContextData {
  using fspath = std::filesystem::path;
  /*  `fs::path` has no hash function, so we use this. */
  using pathset = std::unordered_set<std::string>;

  ::wtr::watcher::event::callback const& callback{};
  pathset* seen_created_paths{nullptr};
  fspath* last_rename_path{nullptr};
  std::mutex* mtx{nullptr};
};

inline auto event_recv_one(ContextData& ctx, char const* path, unsigned flags)
{
  using pty = enum ::wtr::watcher::event::path_type;
  using ety = enum ::wtr::watcher::event::effect_type;

  /*  A single path won't have different "types". */
  auto pt = flags & fsev_flag_path_file      ? pty::file
          : flags & fsev_flag_path_dir       ? pty::dir
          : flags & fsev_flag_path_sym_link  ? pty::sym_link
          : flags & fsev_flag_path_hard_link ? pty::hard_link
                                             : pty::other;
  /*  More than one thing can happen to the same path.
      (So these `if`s are mostly not exclusive.)
      We want to report odd events (even with an empty path)
      but we can bail early if we don't recognize the effect
      because everything else we do depends on that. */
  if (! (flags & fsev_flag_effect_any)) {
    ctx.callback({path, ety::other, pt});
    return;
  }
  if (flags & fsev_flag_effect_create) {
    auto at = ctx.seen_created_paths->find(path);
    if (at == ctx.seen_created_paths->end()) {
      ctx.seen_created_paths->emplace(path);
      ctx.callback({path, ety::create, pt});
    }
  }
  if (flags & fsev_flag_effect_remove) {
    auto at = ctx.seen_created_paths->find(path);
    if (at != ctx.seen_created_paths->end()) {
      ctx.seen_created_paths->erase(at);
      ctx.callback({path, ety::destroy, pt});
    }
  }
  if (flags & fsev_flag_effect_modify) {
    ctx.callback({path, ety::modify, pt});
  }
  if (flags & fsev_flag_effect_rename) {
    /*  Assumes that the last "renamed-from" path
        is "honestly" correlated to the current
        "rename-to" path.
        For non-destructive rename events, we
        usually receive events in this order:
          1. A rename event on the "from-path"
          2. A rename event on the "to-path"
        As long as that pattern holds, we can
        store the first path in a set, look it
        up, test it against the current path
        for inequality, and check that it no
        longer exists -- In which case, we can
        say that we were renamed from that path
        to the current path.
        We want to store the last rename-from
        path in a set on the heap because the
        rename events might not be batched, and
        we don't want to trample on some other
        watcher with a static.
        This pattern breaks down if there are
        intervening rename events.
        For thoughts on recognizing destructive
        rename events, see this directory's
        notes (in the `notes.md` file).
    */
    auto lr_path = *ctx.last_rename_path;
    auto differs = ! lr_path.empty() && lr_path != path;
    auto missing = access(lr_path.c_str(), F_OK) == -1;
    if (differs && missing)
      ctx.callback({
        {lr_path, ety::rename, pt},
        {   path, ety::rename, pt}
      }),
        ctx.last_rename_path->clear();
    else
      *ctx.last_rename_path = path;
  }
}

/*  Sometimes events are batched together and re-sent
    (despite having already been sent).
    Example:
      [first batch of events from the os]
      file 'a' created
      -> create event for 'a' is sent
      [some tiny delay, 1 ms or so]
      [second batch of events from the os]
      file 'a' destroyed
      -> create event for 'a' is sent
      -> destroy event for 'a' is sent
    So, we filter out duplicate events when they're sent
    in a batch. We do this by storing and pruning the
    set of paths which we've seen created. */
inline auto event_recv(
  ConstFSEventStreamRef,      /*  `ConstFS..` is important */
  void* maybe_ctx,            /*  Arguments passed to us */
  unsigned long count,        /*  Event count */
  void* maybe_paths,          /*  Paths with events */
  unsigned const* flags,      /*  Event flags */
  FSEventStreamEventId const* /*  A unique stream id */
  ) -> void
{
  /*  These checks are unfortunate, but they are also necessary.
      Once in a blue moon, some of them are missing. */
  auto data_ok = maybe_ctx && maybe_paths && flags;
  if (! data_ok) return;
  auto paths = static_cast<char const**>(maybe_paths);
  auto ctx = *static_cast<ContextData*>(maybe_ctx);
  if (! ctx.mtx->try_lock()) return;
  for (unsigned long i = 0; i < count; i++)
    event_recv_one(ctx, paths[i], flags[i]);
  ctx.mtx->unlock();
}

static_assert(event_recv == FSEventStreamCallback{event_recv});

inline auto open_event_stream(std::filesystem::path const& path, void* ctx)
  -> FSEventStreamRef
{
  static auto queue = dispatch_get_global_queue(QOS_CLASS_USER_INITIATED, 0);
  auto context = FSEventStreamContext{
    .version = 0,               /*  FSEvents.h: "Only valid value is zero." */
    .info = ctx,                /*  The context; Our "argument pointer". */
    .retain = nullptr,          /*  Not needed; We manage the lifetimes. */
    .release = nullptr,         /*  Same reason as `.retain` */
    .copyDescription = nullptr, /*  Optional string for debugging. */
  };
  /*  path_cfstring and path_array appear to be managed by the
      FSEventStream, are always null when we go to close the
      stream, and shouldn't be freed before then. Would seem
      to me like we'd need to release them (in re. the Create
      rule), but we don't appear to. */
  void const* path_cfstring =
    CFStringCreateWithCString(nullptr, path.c_str(), kCFStringEncodingUTF8);
  /*  A predefined structure which is (from CFArray.h) --
      "appropriate when the values in a CFArray are CFTypes" */
  static auto const cf_arr_ty = kCFTypeArrayCallBacks;
  CFArrayRef path_array = CFArrayCreate(
    nullptr,        /*  A custom allocator is optional */
    &path_cfstring, /*  Data: A ptr-ptr of (in our case) strings */
    1,              /*  We're just storing one path here */
    &cf_arr_ty      /*  The type of the data we're storing */
  );
  /*  Request a filesystem event stream for `path` from the
      kernel. The event stream will call `event_recv` with
      `context` and some details about each filesystem event
      the kernel sees for the paths in `path_array`. */
  auto stream = FSEventStreamCreate(
    nullptr,           /*  A custom allocator is optional */
    &event_recv,       /*  A callable to invoke on changes */
    &context,          /*  The callable's arguments (context) */
    path_array,        /*  The path(s) we were asked to watch */
    fsev_listen_since, /*  The time "since when" we watch */
    0.016,             /*  Seconds between scans *after inactivity* */
    fsev_listen_for    /*  Which event types to send up to us */
  );
  auto data_ok = stream && queue && path_cfstring && path_array;
  if (! data_ok) {
    if (stream) FSEventStreamRelease(stream);
    if (path_cfstring) CFRelease(path_cfstring);
    if (path_array) CFRelease(path_array);
    return nullptr;
  }
  FSEventStreamSetDispatchQueue(stream, queue);
  FSEventStreamStart(stream);
  return stream;
}

inline auto wait(semabin const& sb)
{
  auto s = sb.state();
  if (s != semabin::pending) return s;
  dispatch_semaphore_wait(sb.sem, DISPATCH_TIME_FOREVER);
  return semabin::released;
}

/*  Bugs, footguns

    1.
    When we flood the filesystem with events, Darwin may choose,
    for reason I don't fully understand, to tell us about events
    after we are long gone. Maybe the FSEvent stream (which should
    be f'ing closed) sometimes ignores us having asking it to stop.
    Maybe some tasks in the dispatch queue are not being cleared
    properly. I'm not sure.

    Creating events as quickly as possible while a few short-lived
    FSEvents streams are open is usually enough to reproduce this.
    The recipe to reproduce seems to be a ton of filesystem events
    ongoing while some FSEvent streams are open, then closing the
    streams either during or shortly after the events stop happening
    to the filesystem, but before they have all been reported.

    A minimal-ish reproducer is in /etc/wip-fsevents-issue.

    Whatever the reason, sometimes, Darwin seems happy call into
    our event handler with resource after we have left the memory
    space those resources belong to. I consider that a bug somewhere
    in FSEvents, Dispatch or maybe something deeper.

    At one point, I thought that purging events for the device
    would help. Even that fails under sufficiently high load.
    The positive side effect may have effectively been a sleep.
    Sometimes I even consider adding a deliberate  sleep here.
    Because time is not a synchronization primitive, that will
    also eventually fail. Though, if it's the best we can do,
    despite the kernel, maybe it's worth it. I'm not sure.

    Before that, I added a bunch of synchronization primitives
    to the context and made the lifetime of the context a bit
    transactional. We'd check on atomic flags which in the event
    reception loop to be sure we were alive there, after which
    we'd set an idle flag which the `watch` function would wait
    on before cleaning up. All well and good, but then again,
    it's not like any of that memory *even exists* when Darwin
    calls into it after we've asked it to stop and left.

    There's a minimal-ish reproducer in /etc/wip-fsevents-issue.

    "Worse is better."

    2.
    We want to handle any outstanding events before closing,
    so we flush the event stream before stopping it.
    `FSEventStreamInvalidate()` only needs to be called
    if we scheduled via `FSEventStreamScheduleWithRunLoop()`.
    That scheduling function is deprecated (as of macOS 13).
    Calling `FSEventStreamInvalidate()` fails an assertion
    and produces a warning in the console. However, calling
    `FSEventStreamRelease()` without first invalidating via
    `FSEventStreamInvalidate()` also fails an assertion and
    produces a warning. Releasing seems safer than not, so
    we'll do that.
*/
inline auto
close_event_stream(FSEventStreamRef stream, ContextData& ctx) -> bool
{
  if (! stream) return false;
  auto _ = std::scoped_lock{*ctx.mtx};
  FSEventStreamFlushSync(stream);
  FSEventStreamStop(stream);
  FSEventStreamInvalidate(stream);
  FSEventStreamRelease(stream);
  stream = nullptr;
  return true;
}

} /*  namespace */

/*  Lifetimes
    We will ensure that the queue, context and callback
    are alive at least until we close the event stream.
    They are shared between us and the system. The system
    is self-inconsistent but well-meaning. Sometimes, Apple
    will use our resources after we've asked it not to. I
    consider that a bug somewhere in FSEvents, Dispatch or
    maybe something deeper. There's a minimal-ish reproducer
    in the `/etc/wip-fsevents-issue` directory.  */
inline auto watch(
  std::filesystem::path const& path,
  ::wtr::watcher::event::callback const& cb,
  semabin const& living) -> bool
{
  auto seen_created_paths = ContextData::pathset{};
  auto last_rename_path = ContextData::fspath{};
  auto mtx = std::mutex{};
  auto ctx = ContextData{cb, &seen_created_paths, &last_rename_path, &mtx};
  auto fsevs = open_event_stream(path, &ctx);
  auto state_ok = wait(living) == semabin::released;
  auto close_ok = close_event_stream(fsevs, ctx);
  return state_ok && close_ok;
}

} /*  namespace adapter */
} /*  namespace watcher */
} /*  namespace wtr   */
} /*  namespace detail */

#endif

#if (defined(__linux__) || __ANDROID_API__) \
  && ! defined(WATER_WATCHER_USE_WARTHOG)

#include <dirent.h>
#include <errno.h>
#include <limits.h>
#include <stdio.h>
#include <string.h>
#include <string>
#include <sys/epoll.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <unistd.h>
#include <utility>

namespace detail::wtr::watcher::adapter {

/*  Must maintain that an enumerated value
    less than complete is either pending or
    a warning (which means "keep running"),
    that complete is a successful, terminal
    result (means "stop, ok"), and that a
    value greater than complete is an error
    which we should send along to the user
    after we (attempt to) clean up. */
enum class result : unsigned short {
  pending = 0,
  w,
  w_sys_not_watched,
  w_sys_phantom,
  w_sys_bad_fd,
  w_sys_bad_meta,
  w_sys_q_overflow,
  complete,
  e,
  e_sys_api_inotify,
  e_sys_api_fanotify,
  e_sys_api_epoll,
  e_sys_api_read,
  e_sys_api_eventfd,
  e_sys_ret,
  e_sys_lim_kernel_version,
  e_self_noent,
  e_self_ev_recv,
};

struct ep {
  /*  Number of events allowed to be
      given to do_event_recv. The same
      value is usually returned by a
      call to `epoll_wait`). Any number
      between 1 and INT_MAX should be
      fine... But lower is fine for us.
      We don't lose events if we 'miss'
      them, the events are still waiting
      in the next call to `epoll_wait`.
  */
  static constexpr auto q_ulim = 64;
  /*  The delay, in milliseconds, while
      `epoll_wait` will 'pause' us for
      until we are woken up. We check if
      we are still alive through the fd
      from our semaphore-like eventfd.
  */
  static constexpr auto wake_ms = -1;

  int fd = -1;
  epoll_event interests[q_ulim]{};
};

inline constexpr auto to_str(result r)
{
  // clang-format off
  switch (r) {
    case result::pending:                            return "pending@";
    case result::w:                                  return "w@";
    case result::w_sys_not_watched:                  return "w/sys/not_watched@";
    case result::w_sys_phantom:                      return "w/sys/phantom@";
    case result::w_sys_bad_fd:                       return "w/sys/bad_fd@";
    case result::w_sys_bad_meta:                     return "w/sys/bad_meta@";
    case result::w_sys_q_overflow:                   return "w/sys/q_overflow@";
    case result::complete:                           return "complete@";
    case result::e:                                  return "e@";
    case result::e_sys_api_inotify:                  return "e/sys/api/inotify@";
    case result::e_sys_api_fanotify:                 return "e/sys/api/fanotify@";
    case result::e_sys_api_epoll:                    return "e/sys/api/epoll@";
    case result::e_sys_api_read:                     return "e/sys/api/read@";
    case result::e_sys_api_eventfd:                  return "e/sys/api/eventfd@";
    case result::e_sys_ret:                          return "e/sys/ret@";
    case result::e_sys_lim_kernel_version:           return "e/sys/lim/kernel_version@";
    case result::e_self_noent:                       return "e/self/noent@";
    case result::e_self_ev_recv:                     return "e/self/ev_recv@";
    default:                                         return "e/unknown@";
  }
  // clang-format on
};

inline auto send_msg = [](result r, auto path, auto const& cb)
{
  using et = enum ::wtr::watcher::event::effect_type;
  using pt = enum ::wtr::watcher::event::path_type;
  auto msg = std::string{to_str(r)};
  cb({msg + path, et::other, pt::watcher});
};

inline auto make_ep(int ev_fs_fd, int ev_il_fd) -> ep
{
#if __ANDROID_API__
  int fd = epoll_create(1);
#else
  int fd = epoll_create1(EPOLL_CLOEXEC);
#endif
  auto want_ev_fs = epoll_event{.events = EPOLLIN, .data{.fd = ev_fs_fd}};
  auto want_ev_il = epoll_event{.events = EPOLLIN, .data{.fd = ev_il_fd}};
  bool ctl_ok = fd >= 0
             && epoll_ctl(fd, EPOLL_CTL_ADD, ev_fs_fd, &want_ev_fs) >= 0
             && epoll_ctl(fd, EPOLL_CTL_ADD, ev_il_fd, &want_ev_il) >= 0;
  if (! ctl_ok && fd >= 0) close(fd), fd = -1;
  return ep{.fd = fd};
}

inline auto is_dir(char const* const path) -> bool
{
  struct stat s;
  return stat(path, &s) == 0 && S_ISDIR(s.st_mode);
}

/*  $ echo time wtr.watcher / -ms 1
      | sudo bash -E
      ...
      real 0m25.094s
      user 0m4.091s
      sys  0m20.856s

    $ sudo find / -type d
      | wc -l
      ...
      784418

    We could parallelize this, but
    it's never going to be instant.

    It might be worth it to start
    watching before we're done this
    hot find-and-mark path, despite
    not having a full picture.
*/
template<class Fn>
inline auto walkdir_do(char const* const path, Fn const& f) -> void
{
  if (DIR* d = opendir(path)) {
    f(path);
    while (dirent* de = readdir(d)) {
      char next[PATH_MAX];
      char real[PATH_MAX];
      if (de->d_type != DT_DIR) continue;
      if (strcmp(de->d_name, ".") == 0) continue;
      if (strcmp(de->d_name, "..") == 0) continue;
      if (snprintf(next, PATH_MAX, "%s/%s", path, de->d_name) <= 0) continue;
      if (! realpath(next, real)) continue;
      walkdir_do(real, f);
    }
    (void)closedir(d);
  }
}

} /*  namespace detail::wtr::watcher::adapter */

#endif

#if (defined(__linux__) || __ANDROID_API__) \
  && ! defined(WATER_WATCHER_USE_WARTHOG)

#include <linux/version.h>

#if (KERNEL_VERSION(5, 9, 0) <= LINUX_VERSION_CODE) && ! __ANDROID_API__

#include <errno.h>
#include <fcntl.h>
#include <limits.h>
#include <stdio.h>
#include <string.h>
#include <string>
#include <sys/epoll.h>
#include <sys/fanotify.h>
#include <unistd.h>

namespace detail::wtr::watcher::adapter::fanotify {

/*  We request post-event reporting, non-blocking
    IO and unlimited marks for fanotify. We need
    sudo mode for the unlimited marks. If we were
    making a filesystem auditor, we might use:
      FAN_CLASS_PRE_CONTENT
      | FAN_UNLIMITED_QUEUE
      | FAN_UNLIMITED_MARKS
    and some writable fields within the callback.
*/

// clang-format off
struct ke_fa_ev {
  static constexpr auto buf_len = 4096;
  static constexpr auto c_ulim = buf_len / sizeof(fanotify_event_metadata);
  static constexpr auto init_io_flags
    = O_RDONLY
    | O_CLOEXEC
    | O_NONBLOCK;
  static constexpr auto init_flags
    = FAN_CLASS_NOTIF
    | FAN_REPORT_FID
    | FAN_REPORT_DIR_FID
    | FAN_REPORT_NAME
    | FAN_UNLIMITED_QUEUE
    | FAN_UNLIMITED_MARKS
    | FAN_NONBLOCK;
  /*  todo: Support change of ownership w/ FAN_ATTRIB */
  static constexpr auto recv_flags
    = FAN_ONDIR
    | FAN_CREATE
    | FAN_MODIFY
    | FAN_MOVE
    | FAN_DELETE
    | FAN_EVENT_ON_CHILD;

  int fd = -1;
  alignas(fanotify_event_metadata) char buf[buf_len]{0};
};

// clang-format on

struct sysres {
  result ok = result::e;
  ke_fa_ev ke{};
  semabin const& il{};
  adapter::ep ep{};
};

inline auto do_mark =
  [](char const* const dirpath, int fa_fd, auto const& cb) -> result
{
  auto e = result::w_sys_not_watched;
  char real[PATH_MAX];
  int anonymous_wd =
    realpath(dirpath, real) && is_dir(real)
      ? fanotify_mark(fa_fd, FAN_MARK_ADD, ke_fa_ev::recv_flags, AT_FDCWD, real)
      : -1;
  if (anonymous_wd == 0)
    return result::complete;
  else
    return send_msg(e, dirpath, cb), e;
};

/*  Grabs the resources we need from `fanotify`
    and `epoll`. Marks itself invalid on errors,
    sends diagnostics on warnings and errors.
    Walks the given base path, recursively,
    marking each directory along the way. */
inline auto make_sysres = [](
                            char const* const base_path,
                            auto const& cb,
                            semabin const& living) -> sysres
{
  int fa_fd = fanotify_init(ke_fa_ev::init_flags, ke_fa_ev::init_io_flags);
  if (fa_fd < 1) return sysres{.ok = result::e_sys_api_fanotify, .il = living};
  walkdir_do(base_path, [&](auto dir) { do_mark(dir, fa_fd, cb); });
  auto ep = make_ep(fa_fd, living.fd);
  if (ep.fd < 1)
    return close(fa_fd), sysres{.ok = result::e_sys_api_epoll, .il = living};
  return sysres{
    .ok = result::pending,
    .ke{.fd = fa_fd},
    .il = living,
    .ep = ep,
  };
};

/*  Parses a full path from an event's metadata.
    The shenanigans we do here depend on this event being
    `FAN_EVENT_INFO_TYPE_DFID_NAME`. The kernel passes us
    some info about the directory and the directory entry
    (the filename) of this event that doesn't exist when
    other event types are reported. In particular, we need
    a file descriptor to the directory (which we use
    `readlink` on) and a character string representing the
    name of the directory entry.
    TLDR: We need information for the full path of the
    event, information which is only reported inside this
    `if`.
    From the kernel:
    Variable size struct for
    dir file handle + child file handle + name
    [ Omitting definition of `fanotify_info` here ]
    (struct fanotify_fh)
      dir_fh starts at
        buf[0] (optional)
      dir2_fh starts at
        buf[dir_fh_totlen] (optional)
      file_fh starts at
        buf[dir_fh_totlen+dir2_fh_totlen]
      name starts at
        buf[dir_fh_totlen+dir2_fh_totlen+file_fh_totlen]
      ...
    The kernel guarentees that there is a null-terminated
    character string to the event's directory entry
    after the file handle to the directory.
    Confusing, right? */
inline auto
pathof(fanotify_event_metadata const* const mtd, int* ec) -> std::string
{
  constexpr size_t path_ulim = PATH_MAX - sizeof('\0');
  constexpr int ofl = O_RDONLY | O_CLOEXEC | O_PATH;
  auto dir_info = (fanotify_event_info_fid*)(mtd + 1);
  auto dir_fh = (file_handle*)(dir_info->handle);
  char path_buf[PATH_MAX];
  ssize_t file_name_offset = 0;
  /*  Directory name */
  int fd = open_by_handle_at(AT_FDCWD, dir_fh, ofl);
  if (fd <= 0) {
    *ec = -errno;
    return {};
  }
  char fs_ev_pidpath[32] = {0};
  snprintf(fs_ev_pidpath, sizeof(fs_ev_pidpath), "/proc/self/fd/%d", fd);
  file_name_offset = readlink(fs_ev_pidpath, path_buf, path_ulim);
  close(fd);
  path_buf[file_name_offset] = 0;
  /*  File name ("Directory entry")
      If we wrote the directory name before here, we
      can start writing the file name after its offset. */
  if (file_name_offset > 0) {
    char* file_name = ((char*)dir_fh->f_handle + dir_fh->handle_bytes);
    auto not_selfdir = strcmp(file_name, ".") != 0;
    auto beg = path_buf + file_name_offset;
    auto end = PATH_MAX - file_name_offset;
    if (file_name && not_selfdir) snprintf(beg, end, "/%s", file_name);
  }
  return {path_buf};
}

inline auto peek(fanotify_event_metadata const* const m, size_t read_len)
  -> fanotify_event_metadata const*
{
  if (m) {
    auto ev_len = m->event_len;
    auto next = (fanotify_event_metadata*)((char*)m + ev_len);
    return FAN_EVENT_OK(next, read_len - ev_len) ? next : nullptr;
  }
  else
    return nullptr;
}

struct Parsed {
  ::wtr::watcher::event ev;
  fanotify_event_metadata const* next = nullptr;
  unsigned this_len = 0;
};

// The error code is -errno if there was a system error which prevents
// correctly parsing the event, or a positive error if the event was
// malformed or unexpected (without a system error).
inline auto
parse_ev(fanotify_event_metadata const* const m, size_t read_len, int* ec)
  -> Parsed
{
  using ev = ::wtr::watcher::event;
  using ev_pt = enum ev::path_type;
  using ev_et = enum ev::effect_type;
  auto n = peek(m, read_len);
  auto pt = m->mask & FAN_ONDIR ? ev_pt::dir : ev_pt::file;
  auto et = m->mask & FAN_CREATE ? ev_et::create
          : m->mask & FAN_DELETE ? ev_et::destroy
          : m->mask & FAN_MODIFY ? ev_et::modify
          : m->mask & FAN_MOVE   ? ev_et::rename
                                 : ev_et::other;
  auto isfromto = [et](unsigned a, unsigned b) -> bool
  { return et == ev_et::rename && a & FAN_MOVED_FROM && b & FAN_MOVED_TO; };
  auto one = [&](auto* m) -> Parsed
  { return {ev(pathof(m, ec), et, pt), n, m->event_len}; };
  auto assoc = [&](auto* m, auto* n) -> Parsed
  {
    auto nn = peek(n, read_len);
    auto here_to_nnn = m->event_len + n->event_len;
    auto e = ev(ev(pathof(m, ec), et, pt), ev(pathof(n, ec), et, pt));
    return {e, nn, here_to_nnn};
  };
  return et != ev_et::rename        ? one(m)
       : ! n                        ? (*ec = 2, one(m))
       : isfromto(m->mask, n->mask) ? assoc(m, n)
       : isfromto(n->mask, m->mask) ? assoc(n, m)
       : (*ec = 2, one(m));
}

inline auto is_newdir = [](::wtr::watcher::event const& ev) -> bool
{
  return ev.effect_type == ::wtr::watcher::event::effect_type::create
      && ev.path_type == ::wtr::watcher::event::path_type::dir;
};

/*  Read some events from what fanotify gives
    us. Sends (the valid) events to the user's
    callback. Send a diagnostic to the user on
    warnings and errors. Returns false on errors.
    True otherwise.
    Notes:
    The `metadata->fd` field contains either a
    file descriptor or the value `FAN_NOFD`. File
    descriptors are always greater than 0 (but we
    will get the FAN_NOFD value, and we can grab
    other information from the /proc/self/<fd>
    filesystem, which we do in `pathof()`).
    `FAN_NOFD` represents an event queue overflow
    for `fanotify` listeners which are *not*
    monitoring file handles, such as mount/volume
    watchers. The file handle is in the metadata
    when an `fanotify` listener is monitoring
    events by their file handles.
    The `metadata->vers` field may differ between
    kernel versions, so we check it against the
    version we were compiled with. */
inline auto do_ev_recv = [](auto const& cb, sysres& sr) -> result
{
  auto ev_info = [](fanotify_event_metadata const* const m)
  { return (fanotify_event_info_fid*)(m + 1); };
  auto ev_has_dirname = [&](fanotify_event_metadata const* const m) -> bool
  { return ev_info(m)->hdr.info_type == FAN_EVENT_INFO_TYPE_DFID_NAME; };

  unsigned read_ev_count = 0;
  int read_len = read(sr.ke.fd, sr.ke.buf, sr.ke.buf_len);
  auto const* mtd = (fanotify_event_metadata*)(sr.ke.buf);
  if (read_len <= 0 && errno != EAGAIN)
    return result::pending;
  else if (read_len < 0)
    return result::e_sys_api_read;
  else
    while (mtd && FAN_EVENT_OK(mtd, read_len))
      if (read_ev_count++ > sr.ke.c_ulim)
        return result::e_sys_ret;
      else if (mtd->vers != FANOTIFY_METADATA_VERSION)
        return result::e_sys_lim_kernel_version;
      else if (mtd->fd != FAN_NOFD)
        return result::w_sys_bad_fd;
      else if (mtd->mask & FAN_Q_OVERFLOW)
        return result::w_sys_q_overflow;
      else if (! ev_has_dirname(mtd))
        return result::w_sys_bad_meta;
      else {
        int ec = 0;
        auto r = parse_ev(mtd, read_len, &ec);
        if (ec < 0) return result::w_sys_bad_fd;
        if (is_newdir(r.ev))
          walkdir_do(r.ev.path_name.c_str(), [&](auto dir) {
            do_mark(dir, sr.ke.fd, cb);
            cb({dir, r.ev.effect_type, r.ev.path_type});
          });
        else if (ec == 0)
          cb(r.ev);
        mtd = r.next;
        read_len -= r.this_len;
      }
  return result::pending;
};

} /*  namespace detail::wtr::watcher::adapter::fanotify */

#endif
#endif

#if (defined(__linux__) || __ANDROID_API__) \
  && ! defined(WATER_WATCHER_USE_WARTHOG)

#include <linux/version.h>

#if (KERNEL_VERSION(2, 7, 0) <= LINUX_VERSION_CODE) || __ANDROID_API__

#include <filesystem>
#include <limits.h>
#include <string.h>
#include <sys/epoll.h>
#include <sys/inotify.h>
#include <unistd.h>
#include <unordered_map>
#include <utility>

namespace detail::wtr::watcher::adapter::inotify {

// clang-format off
struct ke_in_ev {
  /*  The maximum length of an inotify
      event. Inotify will add at least
      one null byte to the end of the
      event to terminate the path name,
      but *also* to align the next event
      on an 8-byte boundary. We add one
      here to account for that byte, in
      this case for alignment.
  */
  static constexpr unsigned one_ulim = sizeof(inotify_event) + NAME_MAX + 1;
  static_assert(one_ulim % 8 == 0, "alignment");
  /*  Practically, this buffer is large
      enough for most read calls we would
      ever need to make. It's many times
      the size of the largest possible
      event, and those events are rare.
      Note that this isn't scaled with
      our `ep_q_sz` because this buffer
      length has nothing to do with an
      `epoll` event loop and everything
      to do with `read()` calls which
      fill a buffer with batched events.
  */
  static constexpr unsigned buf_len = 4096;
  static_assert(buf_len > one_ulim * 8, "capacity");
  /*  The upper limit of how many events
      we could possibly read into a buffer
      which is `buf_len` bytes large.
      Practically, if we come half-way
      close to this value, we should
      be skeptical of the `read`.
  */
  static constexpr unsigned c_ulim = buf_len / sizeof(inotify_event);
  static_assert(c_ulim == 256);
  /*  These are the kinds of events which
      we're intersted in. Inotify *should*
      only send us these events, but that's
      not always the case. What's more, we
      sometimes receive events not in the
      IN_ALL_EVENTS mask. We'll ask inotify
      for these events and filter more if
      needed later on.
      These are all the supported events:
        IN_ACCESS
        IN_ATTRIB
        IN_CLOSE_WRITE
        IN_CLOSE_NOWRITE
        IN_CREATE
        IN_DELETE
        IN_DELETE_SELF
        IN_MODIFY
        IN_MOVE_SELF
        IN_MOVED_FROM
        IN_MOVED_TO
        IN_OPEN
  */
  static constexpr unsigned recv_mask
    = IN_CREATE
    | IN_DELETE
    | IN_DELETE_SELF
    | IN_MODIFY
    | IN_MOVE_SELF
    | IN_MOVED_FROM
    | IN_MOVED_TO;

  int fd = -1;
  using paths = std::unordered_map<int, std::filesystem::path>;
  paths dm{};
  alignas(inotify_event) char ev_buf[buf_len]{0};
  int rm_wd_buf[buf_len]{0};

  static_assert(sizeof(ev_buf) % sizeof(inotify_event) == 0, "alignment");
};

// clang-format on

struct sysres {
  result ok = result::e;
  ke_in_ev ke{};
  semabin const& il{};
  adapter::ep ep{};
};

inline auto known_pathof_wd_or_default =
  [](ke_in_ev::paths const& dm, int wd) -> std::filesystem::path
{
  auto dmhit = dm.find(wd);
  return dmhit != dm.end() ? dmhit->second : "";
};

inline auto do_mark =
  [](char const* const dirpath, int dirfd, auto& dm, auto const& cb) -> result
{
  auto e = result::w_sys_not_watched;
  char real[PATH_MAX];
  int wd = realpath(dirpath, real) && is_dir(real)
           ? inotify_add_watch(dirfd, real, ke_in_ev::recv_mask)
           : -1;
  if (wd > 0)
    return dm.emplace(wd, real), result::complete;
  else
    return send_msg(e, dirpath, cb), e;
};

inline auto make_sysres = [](
                            char const* const base_path,
                            auto const& cb,
                            semabin const& living) -> sysres
{
  auto make_inotify = [](result* ok) -> int
  {
    if (*ok >= result::e) return -1;
    int in_fd = inotify_init1(IN_CLOEXEC | IN_NONBLOCK);
    if (in_fd < 0) *ok = result::e_sys_api_inotify;
    return in_fd;
  };

  auto make_dm = [&](result* ok, int in_fd) -> ke_in_ev::paths
  {
    auto dm = ke_in_ev::paths{};
    if (*ok >= result::e) return dm;
    walkdir_do(base_path, [&](auto dir) { do_mark(dir, in_fd, dm, cb); });
    if (dm.empty()) *ok = result::e_self_noent;
    return dm;
  };

  auto make_ep = [&](result* ok, int in_fd, int il_fd) -> ep
  {
    if (*ok >= result::e) return ep{};
    auto ep = adapter::make_ep(in_fd, il_fd);
    if (ep.fd < 0) *ok = result::e_sys_api_epoll;
    return ep;
  };

  auto ok = result::pending;
  auto in_fd = make_inotify(&ok);
  auto dm = make_dm(&ok, in_fd);
  auto ep = make_ep(&ok, in_fd, living.fd);
  return sysres{
    .ok = ok,
    .ke{
        .fd = in_fd,
        .dm = std::move(dm),
        },
    .il = living,
    .ep = ep,
  };
};

inline auto peek = [](
                     inotify_event const* const in_ev,
                     inotify_event const* const ev_tail) -> inotify_event*
{
  auto len_to_next = sizeof(inotify_event) + (in_ev ? in_ev->len : 0);
  auto next = (inotify_event*)((char*)in_ev + len_to_next);
  return next < ev_tail ? next : nullptr;
};

struct parsed {
  ::wtr::watcher::event ev;
  inotify_event* next = nullptr;
};

inline auto parse_ev = [](
                         ke_in_ev::paths const& dm,
                         inotify_event const* const in,
                         inotify_event const* const tail,
                         int* ec) -> parsed
{
  using ev = ::wtr::watcher::event;
  using ev_pt = enum ev::path_type;
  using ev_et = enum ev::effect_type;
  auto pathof = [&](inotify_event const* const m)
  { return known_pathof_wd_or_default(dm, m->wd) / m->name; };
  auto calc_pt =
    [&](uint32_t ev_mask, std::filesystem::path const& ev_path) -> ev_pt
  {
    return ev_mask & IN_ISDIR  ? ev_pt::dir
         : is_symlink(ev_path) ? ev_pt::sym_link
                               : ev_pt::file;
  };
  auto et = in->mask & IN_CREATE ? ev_et::create
          : in->mask & IN_DELETE ? ev_et::destroy
          : in->mask & IN_MOVE   ? ev_et::rename
          : in->mask & IN_MODIFY ? ev_et::modify
                                 : ev_et::other;
  auto isassoc = [&](auto* a, auto* b) -> bool
  { return b && b->cookie && b->cookie == a->cookie && et == ev_et::rename; };
  auto isfromto = [](auto* a, auto* b) -> bool
  { return (a->mask & IN_MOVED_FROM) && (b->mask & IN_MOVED_TO); };
  auto one = [&](auto* next) -> parsed
  {
    auto in_path = pathof(in);
    return {ev(in_path, et, calc_pt(in->mask, in_path)), next};
  };
  auto assoc = [&](auto* a, auto* b) -> parsed
  {
    auto b_path = pathof(b);
    auto pt_b = calc_pt(b->mask, b_path);
    return {ev(ev(pathof(a), et, pt_b), ev(b_path, et, pt_b)), peek(b, tail)};
  };
  auto next = peek(in, tail);
  return ! isassoc(in, next) ? one(next)
       : isfromto(in, next)  ? assoc(in, next)
       : isfromto(next, in)  ? assoc(next, in)
                             : (*ec = 1, one(next));
};

struct defer_dm_rm_wd {
  ke_in_ev& ke;
  size_t back_idx = 0;

  /*  It is impossible to exceed our buffer
      because we store as many events as are
      possible to receive from `read()`.
      If we do exceed the indices, then some
      system invariant has been violated. */
  inline auto push(int wd) -> bool
  {
    if (back_idx < ke.buf_len)
      return ke.rm_wd_buf[back_idx++] = wd, true;
    else
      return false;
  };

  inline defer_dm_rm_wd(ke_in_ev& ke)
      : ke{ke} {};

  inline ~defer_dm_rm_wd()
  {
    for (size_t i = 0; i < back_idx; ++i) {
      auto wd = ke.rm_wd_buf[i];
      auto at = ke.dm.find(wd);
      if (at != ke.dm.end()) ke.dm.erase(at);
    }
  };
};

/*  Parses each event's path name,
    path type and effect.
    Looks for the directory path
    in a map of watch descriptors
    to directory paths.
    Updates the path map, adding
    new directories as they are
    created, and removing them
    as they are destroyed.

    Forward events and errors to
    the user. Returns on errors
    and when eventless.

    Event notes:
    Phantom Events --
    An event from an unmarked path,
    or a path which we didn't mark,
    was somehow reported to us.
    These events are rare, but they
    do happen. We won't be able to
    parse this event's real path.
    The ->name field seems to be
    null on these events, and we
    won't have a directory path to
    prepend it with. I'm not sure
    if we should try to parse the
    other fields, or if they would
    be interesting to the user.
    This may change.

    Impossible Events --
    These events are relatively
    rare, but they happen more
    than I think they should. We
    usually see these during
    high-throughput flurries of
    events. Maybe there is an
    error in our implementation?

    Deferred Events --
    Inotify closes the removed watch
    descriptors itself. We want to
    keep parity with inotify in our
    path map. That way, we can be
    in agreement about which watch
    descriptors map to which paths.
    We need to postpone removing
    this watch and, possibly, its
    watch descriptor from our path
    map until we're done with this
    event batch. Self-destroy events
    might come before we read other
    events that would map the watch
    descriptor to a path. Because
    we need to look the directory
    path up in the path map, we
    will defer its removal.

    Other notes:
    Sometimes we can fail to mark
    a new directory on its create
    event. This can happen if the
    directory is removed quickly
    after being created.
    In that case, we are unlikely
    to lose any path names on
    future events because events
    won't happen in that directory.
    If this happens for some other
    reason, we're in trouble.
*/
inline auto do_ev_recv = [](auto const& cb, sysres& sr) -> result
{
  auto is_parity_lost = [](unsigned msk) -> bool
  { return msk & IN_DELETE_SELF && ! (msk & IN_MOVE_SELF); };
  auto is_real_event = [](unsigned msk) -> bool
  {
    bool has_any = msk & ke_in_ev::recv_mask;
    bool is_self_info = msk & (IN_IGNORED | IN_DELETE_SELF | IN_MOVE_SELF);
    return has_any && ! is_self_info;
  };

  auto read_len = read(sr.ke.fd, sr.ke.ev_buf, sizeof(sr.ke.ev_buf));
  if (read_len < 0 && errno != EAGAIN)
    return result::e_sys_api_read;
  else {
    auto const* in_ev = (inotify_event*)(sr.ke.ev_buf);
    auto const* const in_ev_tail = (inotify_event*)(sr.ke.ev_buf + read_len);
    unsigned in_ev_c = 0;
    auto dmrm = defer_dm_rm_wd{sr.ke};
    while (in_ev && in_ev < in_ev_tail) {
      auto in_ev_next = peek(in_ev, in_ev_tail);
      unsigned msk = in_ev->mask;
      if (in_ev_c++ > ke_in_ev::c_ulim)
        return result::e_sys_ret;
      else if (is_parity_lost(msk) && ! dmrm.push(in_ev->wd))
        return result::e_sys_ret;
      else if (msk & IN_Q_OVERFLOW)
        send_msg(result::w_sys_q_overflow, "", cb);
      else if (is_real_event(msk)) {
        int ec = 0;
        auto parsed = parse_ev(sr.ke.dm, in_ev, in_ev_tail, &ec);
        if (msk & IN_ISDIR && msk & IN_CREATE)
          walkdir_do(parsed.ev.path_name.c_str(), [&](auto dir) {
            do_mark(dir, sr.ke.fd, sr.ke.dm, cb);
            cb({dir, parsed.ev.effect_type, parsed.ev.path_type});
          });
        else if (! ec)
          cb(parsed.ev);
        in_ev_next = parsed.next;
      }
      in_ev = in_ev_next;
    }
    return result::pending;
  }
};

} /*  namespace detail::wtr::watcher::adapter::inotify */

#endif
#endif

#if (defined(__linux__) || __ANDROID_API__) \
  && ! defined(WATER_WATCHER_USE_WARTHOG)

#include <linux/version.h>

#if KERNEL_VERSION(2, 7, 0) > LINUX_VERSION_CODE
#error "Define 'WATER_WATCHER_USE_WARTHOG' on kernel versions < 2.7.0"
#endif

#include <unistd.h>

namespace detail::wtr::watcher::adapter {

inline auto watch =
  [](auto const& path, auto const& cb, auto const& living) -> bool
{
  auto platform_watch = [&](auto make_sysres, auto do_ev_recv) -> result
  {
    auto sr = make_sysres(path.c_str(), cb, living);
    auto is_ev_of = [&](int nth, int fd) -> bool
    { return sr.ep.interests[nth].data.fd == fd; };

    while (sr.ok < result::complete) {
      int ep_c =
        epoll_wait(sr.ep.fd, sr.ep.interests, sr.ep.q_ulim, sr.ep.wake_ms);
      if (ep_c < 0) {
        if (errno != EINTR) sr.ok = result::e_sys_api_epoll;
      }
      else
        for (int n = 0; n < ep_c; ++n)
          if (is_ev_of(n, sr.il.fd))
            sr.ok = result::complete;
          else if (is_ev_of(n, sr.ke.fd))
            sr.ok = do_ev_recv(cb, sr);
          else
            sr.ok = result::e_sys_api_epoll;
    }

    /*  We aren't worried about losing data after
        a failed call to `close()` (whereas closing
        a file descriptor in use for, say, writing
        would be a problem). Linux will eventually
        close the file descriptor regardless of the
        return value of `close()`.
        We are only interested in failures about
        bad file descriptors. We would probably
        hit that if we failed to create a valid
        file descriptor, on, say, an out-of-fds
        device or a machine running some odd OS.
        Everything else is fine to pass on, and
        the out-of-file-descriptors case will be
        handled in `make_sysres()`.
    */
    return close(sr.ke.fd), close(sr.ep.fd), sr.ok;
  };

  /*  e_sys_api_fanotify

      It is possible to be unable to use fanotify, even
      with the CAP_SYS_ADMIN capability. For example,
      the system calls we need may be limited because
      we're running within a cgroup.
      IOW -- We're probably inside a container if we're
      root but lack permission to use fanotify.

      The same error applies to being built on a kernel
      which doesn't have the fanotify api.
  */
  auto try_fanotify = [&]()
  {
#if (KERNEL_VERSION(5, 9, 0) <= LINUX_VERSION_CODE) && ! __ANDROID_API__
    if (geteuid() == 0)
      return platform_watch(fanotify::make_sysres, fanotify::do_ev_recv);
#endif
    return result::e_sys_api_fanotify;
  };

  auto r = try_fanotify();
  if (r == result::e_sys_api_fanotify)
    r = platform_watch(inotify::make_sysres, inotify::do_ev_recv);
  if (r >= result::e)
    return send_msg(r, path.c_str(), cb), false;
  else
    return true;
};

} /*  namespace detail::wtr::watcher::adapter */

#endif

#if defined(_WIN32)

#include <chrono>
#include <filesystem>
#include <string>
#include <thread>
#include <windows.h>

namespace detail::wtr::watcher::adapter {
namespace {

/*  Hold resources necessary to recieve and send filesystem events. */
class watch_event_proxy {
public:
  /*  Timeout for the completion port to wait for events,
      giving us some room to check if we're still alive. */
  static constexpr auto delay_ms =
    static_cast<DWORD>(std::chrono::milliseconds(16).count());
  /*  I think the default page size in Windows is 64kb,
      so 65536 might also work well. */
  static constexpr auto event_buf_len_max = 8192;
  bool is_valid = true;
  std::filesystem::path path = {};
  wchar_t path_name[256] = {L""};
  HANDLE path_handle = nullptr;
  HANDLE event_completion_token = nullptr;
  HANDLE event_token = CreateEventW(nullptr, true, false, nullptr);
  OVERLAPPED event_overlap = {};
  FILE_NOTIFY_INFORMATION event_buf[event_buf_len_max] = {0};
  DWORD event_buf_len_ready = 0;

  watch_event_proxy(std::filesystem::path const& path) noexcept
      : path{path}
  {
    memcpy(path_name, path.c_str(), path.string().size());
    this->path_handle = CreateFileW(
      path.c_str(),
      FILE_LIST_DIRECTORY,
      FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE,
      nullptr,
      OPEN_EXISTING,
      FILE_FLAG_BACKUP_SEMANTICS | FILE_FLAG_OVERLAPPED,
      nullptr);
    if (path_handle)
      this->event_completion_token =
        CreateIoCompletionPort(INVALID_HANDLE_VALUE, nullptr, 0, 0);
    if (event_completion_token)
      this->is_valid = CreateIoCompletionPort(
                         path_handle,
                         event_completion_token,
                         (ULONG_PTR)path_handle,
                         1)
                    && ResetEvent(event_token);
  }

  ~watch_event_proxy() noexcept
  {
    if (event_token) CloseHandle(event_token);
    if (event_completion_token) CloseHandle(event_completion_token);
  }
};

inline auto has_event(watch_event_proxy& w) noexcept -> bool
{
  return w.event_buf_len_ready != 0;
}

inline auto do_event_recv(
  watch_event_proxy& w,
  ::wtr::watcher::event::callback const& callback) noexcept -> bool
{
  w.event_buf_len_ready = 0;
  DWORD bytes_returned = 0;
  memset(&w.event_overlap, 0, sizeof(OVERLAPPED));
  if (! w.is_valid) return false;
  auto read_ok = ReadDirectoryChangesW(
    w.path_handle,
    w.event_buf,
    w.event_buf_len_max,
    true,
    FILE_NOTIFY_CHANGE_SECURITY | FILE_NOTIFY_CHANGE_CREATION
      | FILE_NOTIFY_CHANGE_LAST_WRITE | FILE_NOTIFY_CHANGE_SIZE
      | FILE_NOTIFY_CHANGE_ATTRIBUTES | FILE_NOTIFY_CHANGE_DIR_NAME
      | FILE_NOTIFY_CHANGE_FILE_NAME,
    &bytes_returned,
    &w.event_overlap,
    nullptr);
  if (read_ok) {
    w.event_buf_len_ready = bytes_returned > 0 ? bytes_returned : 0;
    return true;
  }
  else {
    switch (GetLastError()) {
      case ERROR_IO_PENDING :
        w.event_buf_len_ready = 0;
        w.is_valid = false;
        callback(
          {"e/sys/read/pending",
           ::wtr::event::effect_type::other,
           ::wtr::event::path_type::watcher});
        break;
      default :
        callback(
          {"e/sys/read",
           ::wtr::event::effect_type::other,
           ::wtr::event::path_type::watcher});
        break;
    }
    return false;
  }
}

inline auto do_event_send(
  watch_event_proxy& w,
  ::wtr::watcher::event::callback const& callback) noexcept -> bool
{
  using namespace ::wtr::watcher;

  struct RenameEventTracker {
    std::filesystem::path path_name;
    enum event::effect_type effect_type;
    enum event::path_type path_type;
    bool set = false;
  };

  FILE_NOTIFY_INFORMATION* buf = w.event_buf;
  /*  Rename events on Windows send two individual messages
      that correspond with the old data and the new data.
      While it is believed that these are sent sequentially
      with the old data first, there is no guarantee in the documentation.
      These trackers are used to ensure all data is available for the callback
      regardless of the order. */
  RenameEventTracker old_tracker;
  RenameEventTracker new_tracker;
  auto const trigger_rename_callback = [&]()
  {
    auto renamed_from = event{
      old_tracker.path_name,
      old_tracker.effect_type,
      old_tracker.path_type};
    auto renamed_to = event{
      new_tracker.path_name,
      new_tracker.effect_type,
      new_tracker.path_type};
    callback({renamed_from, std::move(renamed_to)});
    /*  Reset for the possibility of more events. */
    old_tracker = {};
    new_tracker = {};
  };

  if (! w.is_valid) return false;
  while (buf + sizeof(FILE_NOTIFY_INFORMATION) <= buf + w.event_buf_len_ready) {
    if (buf->FileNameLength % 2 == 0) {
      auto path_name =
        w.path / std::wstring{buf->FileName, buf->FileNameLength / 2};

      auto effect_type = [&buf]() noexcept
      {
        switch (buf->Action) {
          case FILE_ACTION_MODIFIED : return event::effect_type::modify;
          case FILE_ACTION_ADDED : return event::effect_type::create;
          case FILE_ACTION_REMOVED : return event::effect_type::destroy;
          case FILE_ACTION_RENAMED_OLD_NAME : return event::effect_type::rename;
          case FILE_ACTION_RENAMED_NEW_NAME : return event::effect_type::rename;
          default : return event::effect_type::other;
        }
      }();

      auto path_type = [&path_name]()
      {
        try {
          return std::filesystem::is_directory(path_name)
                 ? event::path_type::dir
                 : event::path_type::file;
        } catch (...) {
          return event::path_type::other;
        }
      }();

      if (buf->Action == FILE_ACTION_RENAMED_OLD_NAME) {
        old_tracker.path_name = path_name;
        old_tracker.effect_type = effect_type;
        old_tracker.path_type = path_type;
        old_tracker.set = true;
        if (new_tracker.set) trigger_rename_callback();
      }
      else if (buf->Action == FILE_ACTION_RENAMED_NEW_NAME) {
        new_tracker.path_name = path_name;
        new_tracker.effect_type = effect_type;
        new_tracker.path_type = path_type;
        new_tracker.set = true;
        if (old_tracker.set) trigger_rename_callback();
      }
      else {
        callback({path_name, effect_type, path_type});
      }
      if (buf->NextEntryOffset == 0)
        break;
      else
        buf = (FILE_NOTIFY_INFORMATION*)((uint8_t*)buf + buf->NextEntryOffset);
    }
  }
  return true;
}

}  // namespace

/*  while living
    watch for events
    return when dead
    true if no errors */
inline auto watch(
  std::filesystem::path const& path,
  ::wtr::watcher::event::callback const& callback,
  semabin const& living) noexcept -> bool
{
  using namespace ::wtr::watcher;
  auto w = watch_event_proxy{path};
  if (! w.is_valid) return false;
  do_event_recv(w, callback);
  while (w.is_valid && has_event(w)) do_event_send(w, callback);
  while (living.state() == semabin::state::pending) {
    ULONG_PTR completion_key = 0;
    LPOVERLAPPED overlap = nullptr;
    bool complete = GetQueuedCompletionStatus(
      w.event_completion_token,
      &w.event_buf_len_ready,
      &completion_key,
      &overlap,
      w.delay_ms);
    if (complete && overlap) {
      while (w.is_valid && has_event(w)) {
        do_event_send(w, callback);
        do_event_recv(w, callback);
      }
    }
  }
  return true;
}

} /*  namespace detail::wtr::watcher::adapter  */

#endif

/*  A sturdy, universal adapter.

    This is the fallback adapter on platforms that either
      - Only support `kqueue` (`warthog` beats `kqueue`)
      - Only support the C++ standard library */

#if ! defined(__linux__) && ! defined(__ANDROID_API__) && ! defined(__APPLE__) \
  && ! defined(_WIN32)

#include <chrono>
#include <filesystem>
#include <functional>
#include <string>
#include <system_error>
#include <thread>
#include <unordered_map>

namespace detail {
namespace wtr {
namespace watcher {
namespace adapter {
namespace {

inline constexpr std::filesystem::directory_options scan_dir_opts =
  std::filesystem::directory_options::skip_permission_denied
  & std::filesystem::directory_options::follow_directory_symlink;

using bucket_type =
  std::unordered_map<std::string, std::filesystem::file_time_type>;

/*  - Scans `path` for changes.
    - Updates our bucket to match the changes.
    - Calls `send_event` when changes happen.
    - Returns false if the file tree cannot be scanned. */
inline bool scan(
  std::filesystem::path const& path,
  auto const& send_event,
  bucket_type& bucket) noexcept
{
  /*  - Scans a (single) file for changes.
      - Updates our bucket to match the changes.
      - Calls `send_event` when changes happen.
      - Returns false if the file cannot be scanned. */
  auto scan_file =
    [&](std::filesystem::path const& file, auto const& send_event) -> bool
  {
    using namespace ::wtr::watcher;
    using namespace std::filesystem;

    if (exists(file) && is_regular_file(file)) {
      auto ec = std::error_code{};
      /*  grabbing the file's last write time */
      auto timestamp = last_write_time(file, ec);
      if (ec) {
        /*  the file changed while we were looking at it.
            so, we call the closure, indicating destruction,
            and remove it from the bucket. */
        send_event(
          event{file, event::effect_type::destroy, event::path_type::file});
        if (bucket.contains(file)) bucket.erase(file);
      }
      /*  if it's not in our bucket, */
      else if (! bucket.contains(file)) {
        /*  we put it in there and call the closure,
            indicating creation. */
        bucket[file] = timestamp;
        send_event(
          event{file, event::effect_type::create, event::path_type::file});
      }
      /*  otherwise, it is already in our bucket. */
      else {
        /*  we update the file's last write time, */
        if (bucket[file] != timestamp) {
          bucket[file] = timestamp;
          /*  and call the closure on them,
              indicating modification */
          send_event(
            event{file, event::effect_type::modify, event::path_type::file});
        }
      }
      return true;
    }
    else
      return false; /*  if the path doesn't exist, nudge the caller. */
  };

  /*  - Scans a (single) directory for changes.
      - Updates our bucket to match the changes.
      - Calls `send_event` when changes happen.
      - Returns false if the directory cannot be scanned. */
  auto const& scan_directory =
    [&](std::filesystem::path const& dir, auto const& send_event) -> bool
  {
    using namespace std::filesystem;
    if (is_directory(dir)) {
      auto ec = std::error_code{};
      for (auto file : recursive_directory_iterator(dir, scan_dir_opts, ec))
        if (ec)
          return false;
        else
          scan_file(file.path(), send_event);
      return true;
    }
    else
      return false;
  };

  return scan_directory(path, send_event) ? true
       : scan_file(path, send_event)      ? true
                                          : false;
};

/*  If the bucket is empty, try to populate it.
    otherwise, prune it. */
inline bool tend_bucket(
  std::filesystem::path const& path,
  auto const& send_event,
  bucket_type& bucket) noexcept
{
  /*  Creates a file map, the "bucket", from `path`. */
  auto populate = [&](std::filesystem::path const& path) -> bool
  {
    using namespace std::filesystem;
    /*  this happens when a path was changed while we were reading it.
        there is nothing to do here; we prune later. */
    auto dir_ec = std::error_code{};
    auto lwt_ec = std::error_code{};
    if (! exists(path))
      return false;
    else if (! is_directory(path))
      bucket[path] = last_write_time(path);
    else {
      for (auto file :
           recursive_directory_iterator(path, scan_dir_opts, dir_ec)) {
        if (! dir_ec) {
          auto lwt = last_write_time(file, lwt_ec);
          if (! lwt_ec)
            bucket[file.path()] = lwt;
          else
            bucket[file.path()] = last_write_time(path);
        }
      }
      return true;
    }
  };

  /*  Removes files which no longer exist from our bucket. */
  auto prune =
    [&](std::filesystem::path const& path, auto const& send_event) -> bool
  {
    using namespace ::wtr::watcher;
    using namespace std::filesystem;

    auto bucket_it = bucket.begin();
    /*  while looking through the bucket's contents, */
    while (bucket_it != bucket.end()) {
      /*  check if the stuff in our bucket exists anymore. */
      exists(bucket_it->first)
        /*  if so, move on. */
        ? std::advance(bucket_it, 1)
        /*  if not, call the closure,
            indicating destruction,
            and remove it from our bucket. */
        : [&]()
      {
        send_event(event{
          bucket_it->first,
          event::effect_type::destroy,
          is_regular_file(path) ? event::path_type::file
          : is_directory(path)  ? event::path_type::dir
          : is_symlink(path)    ? event::path_type::sym_link
                                : event::path_type::other});
        /*  bucket, erase it! */
        bucket_it = bucket.erase(bucket_it);
      }();
    }
    return true;
  };

  return bucket.empty() ? populate(path)          ? true
                        : prune(path, send_event) ? true
                                                  : false
                        : true;
};

} /* namespace */

inline auto watch(
  std::filesystem::path const& path,
  ::wtr::watcher::event::callback const& callback,
  semabin const& living) noexcept -> bool
{
  using std::this_thread::sleep_for;
  using namespace std::chrono_literals;
  auto bucket = bucket_type{};
  /*  Sleep, checking if we're alive every little while.
      Keep running if
        - We are alive
        - The bucket is doing well
        - No errors occured while scanning
      Otherwise, stop and return false. */
  while (living.state() == semabin::state::pending) {
    if (! tend_bucket(path, callback, bucket)) return false;
    if (! scan(path, callback, bucket)) return false;
    sleep_for(16ms);
  }
  return true;
}

} /*  namespace adapter */
} /*  namespace watcher */
} /*  namespace wtr */
} /*  namespace detail */

#endif

#include <filesystem>
#include <future>

namespace wtr {
inline namespace watcher {

/*  An asynchronous filesystem watcher.

    Begins watching when constructed.

    Stops watching when the object's lifetime ends
    or when `.close()` is called.

    Closing the watcher is the only blocking operation.

    @param path:
      The root path to watch for filesystem events.

    @param callback:
      Something (such as a closure) to be called when events
      occur in the path being watched.

    This is an adaptor "switch" that chooses the ideal
    adaptor for the host platform.

    Every adapter monitors `path` for changes and invokes
    the `callback` with an `event` object when they occur.

    There are two things the user needs: `watch` and
    `event`.

    Typical use looks something like this:

      auto show(event e) {
        auto s = [](auto a) { return to<string>(a); };
        cout << s(e.effect_type) + ' '
              + s(e.path_type)   + ' '
              + s(e.path_name)
              + (e.associated
                 ? " -> " + s(e.associated->path_name)
                 : "")
             << endl;
      }

      auto w = watch(".", show);

    That's it.

    Happy hacking. */
class watch {
private:
  using sb = ::detail::wtr::watcher::semabin;
  sb living{};
  std::future<bool> watching{};

public:
  inline watch(
    std::filesystem::path const& path,
    event::callback const& callback) noexcept
      : watching{std::async(
          std::launch::async,
          [this, path, callback]
          {
            using ::detail::wtr::watcher::adapter::watch;
            auto ec = std::error_code{};
            auto abs_path = std::filesystem::absolute(path, ec);
            auto pre_ok = ! ec && std::filesystem::is_directory(abs_path, ec)
                       && ! ec && this->living.state() == sb::state::pending;
            auto live_msg =
              (pre_ok ? "s/self/live@" : "e/self/live@") + abs_path.string();
            callback(
              {live_msg,
               event::effect_type::create,
               event::path_type::watcher});
            auto post_ok = pre_ok && watch(abs_path, callback, this->living);
            auto die_msg =
              (post_ok ? "s/self/die@" : "e/self/die@") + abs_path.string();
            callback(
              {die_msg,
               event::effect_type::destroy,
               event::path_type::watcher});
            return pre_ok && post_ok;
          })}
  {}

  inline auto close() noexcept -> bool
  {
    return this->living.release() != sb::state::error
        && this->watching.valid() && this->watching.get();
  };

  inline ~watch() noexcept { this->close(); }
};

} /*  namespace watcher */
} /*  namespace wtr   */
#endif /* W973564ED9F278A21F3E12037288412FBAF175F889 */