libstdc++
atomic_wait.h
Go to the documentation of this file.
1 // -*- C++ -*- header.
2 
3 // Copyright (C) 2020-2022 Free Software Foundation, Inc.
4 //
5 // This file is part of the GNU ISO C++ Library. This library is free
6 // software; you can redistribute it and/or modify it under the
7 // terms of the GNU General Public License as published by the
8 // Free Software Foundation; either version 3, or (at your option)
9 // any later version.
10 
11 // This library is distributed in the hope that it will be useful,
12 // but WITHOUT ANY WARRANTY; without even the implied warranty of
13 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 // GNU General Public License for more details.
15 
16 // Under Section 7 of GPL version 3, you are granted additional
17 // permissions described in the GCC Runtime Library Exception, version
18 // 3.1, as published by the Free Software Foundation.
19 
20 // You should have received a copy of the GNU General Public License and
21 // a copy of the GCC Runtime Library Exception along with this program;
22 // see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
23 // <http://www.gnu.org/licenses/>.
24 
25 /** @file bits/atomic_wait.h
26  * This is an internal header file, included by other library headers.
27  * Do not attempt to use it directly. @headername{atomic}
28  */
29 
30 #ifndef _GLIBCXX_ATOMIC_WAIT_H
31 #define _GLIBCXX_ATOMIC_WAIT_H 1
32 
33 #pragma GCC system_header
34 
35 #include <bits/c++config.h>
36 #if defined _GLIBCXX_HAS_GTHREADS || defined _GLIBCXX_HAVE_LINUX_FUTEX
37 #include <bits/functional_hash.h>
38 #include <bits/gthr.h>
39 #include <ext/numeric_traits.h>
40 
41 #ifdef _GLIBCXX_HAVE_LINUX_FUTEX
42 # include <cerrno>
43 # include <climits>
44 # include <unistd.h>
45 # include <syscall.h>
46 # include <bits/functexcept.h>
47 #endif
48 
49 # include <bits/std_mutex.h> // std::mutex, std::__condvar
50 
51 #define __cpp_lib_atomic_wait 201907L
52 
53 namespace std _GLIBCXX_VISIBILITY(default)
54 {
55 _GLIBCXX_BEGIN_NAMESPACE_VERSION
56  namespace __detail
57  {
58 #ifdef _GLIBCXX_HAVE_LINUX_FUTEX
59 #define _GLIBCXX_HAVE_PLATFORM_WAIT 1
60  using __platform_wait_t = int;
61  inline constexpr size_t __platform_wait_alignment = 4;
62 #else
63 // define _GLIBCX_HAVE_PLATFORM_WAIT and implement __platform_wait()
64 // and __platform_notify() if there is a more efficient primitive supported
65 // by the platform (e.g. __ulock_wait()/__ulock_wake()) which is better than
66 // a mutex/condvar based wait.
67  using __platform_wait_t = uint64_t;
68  inline constexpr size_t __platform_wait_alignment
69  = __alignof__(__platform_wait_t);
70 #endif
71  } // namespace __detail
72 
73  template<typename _Tp>
74  inline constexpr bool __platform_wait_uses_type
75 #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
76  = is_scalar_v<_Tp>
77  && ((sizeof(_Tp) == sizeof(__detail::__platform_wait_t))
78  && (alignof(_Tp*) >= __detail::__platform_wait_alignment));
79 #else
80  = false;
81 #endif
82 
83  namespace __detail
84  {
85 #ifdef _GLIBCXX_HAVE_LINUX_FUTEX
86  enum class __futex_wait_flags : int
87  {
88 #ifdef _GLIBCXX_HAVE_LINUX_FUTEX_PRIVATE
89  __private_flag = 128,
90 #else
91  __private_flag = 0,
92 #endif
93  __wait = 0,
94  __wake = 1,
95  __wait_bitset = 9,
96  __wake_bitset = 10,
97  __wait_private = __wait | __private_flag,
98  __wake_private = __wake | __private_flag,
99  __wait_bitset_private = __wait_bitset | __private_flag,
100  __wake_bitset_private = __wake_bitset | __private_flag,
101  __bitset_match_any = -1
102  };
103 
104  template<typename _Tp>
105  void
106  __platform_wait(const _Tp* __addr, __platform_wait_t __val) noexcept
107  {
108  auto __e = syscall (SYS_futex, static_cast<const void*>(__addr),
109  static_cast<int>(__futex_wait_flags::__wait_private),
110  __val, nullptr);
111  if (!__e || errno == EAGAIN)
112  return;
113  if (errno != EINTR)
114  __throw_system_error(errno);
115  }
116 
117  template<typename _Tp>
118  void
119  __platform_notify(const _Tp* __addr, bool __all) noexcept
120  {
121  syscall (SYS_futex, static_cast<const void*>(__addr),
122  static_cast<int>(__futex_wait_flags::__wake_private),
123  __all ? INT_MAX : 1);
124  }
125 #endif
126 
127  inline void
128  __thread_yield() noexcept
129  {
130 #if defined _GLIBCXX_HAS_GTHREADS && defined _GLIBCXX_USE_SCHED_YIELD
131  __gthread_yield();
132 #endif
133  }
134 
135  inline void
136  __thread_relax() noexcept
137  {
138 #if defined __i386__ || defined __x86_64__
139  __builtin_ia32_pause();
140 #else
141  __thread_yield();
142 #endif
143  }
144 
145  inline constexpr auto __atomic_spin_count_relax = 12;
146  inline constexpr auto __atomic_spin_count = 16;
147 
148  struct __default_spin_policy
149  {
150  bool
151  operator()() const noexcept
152  { return false; }
153  };
154 
155  template<typename _Pred,
156  typename _Spin = __default_spin_policy>
157  bool
158  __atomic_spin(_Pred& __pred, _Spin __spin = _Spin{ }) noexcept
159  {
160  for (auto __i = 0; __i < __atomic_spin_count; ++__i)
161  {
162  if (__pred())
163  return true;
164 
165  if (__i < __atomic_spin_count_relax)
166  __detail::__thread_relax();
167  else
168  __detail::__thread_yield();
169  }
170 
171  while (__spin())
172  {
173  if (__pred())
174  return true;
175  }
176 
177  return false;
178  }
179 
180  // return true if equal
181  template<typename _Tp>
182  bool __atomic_compare(const _Tp& __a, const _Tp& __b)
183  {
184  // TODO make this do the correct padding bit ignoring comparison
185  return __builtin_memcmp(&__a, &__b, sizeof(_Tp)) == 0;
186  }
187 
188  struct __waiter_pool_base
189  {
190  // Don't use std::hardware_destructive_interference_size here because we
191  // don't want the layout of library types to depend on compiler options.
192  static constexpr auto _S_align = 64;
193 
194  alignas(_S_align) __platform_wait_t _M_wait = 0;
195 
196 #ifndef _GLIBCXX_HAVE_PLATFORM_WAIT
197  mutex _M_mtx;
198 #endif
199 
200  alignas(_S_align) __platform_wait_t _M_ver = 0;
201 
202 #ifndef _GLIBCXX_HAVE_PLATFORM_WAIT
203  __condvar _M_cv;
204 #endif
205  __waiter_pool_base() = default;
206 
207  void
208  _M_enter_wait() noexcept
209  { __atomic_fetch_add(&_M_wait, 1, __ATOMIC_SEQ_CST); }
210 
211  void
212  _M_leave_wait() noexcept
213  { __atomic_fetch_sub(&_M_wait, 1, __ATOMIC_RELEASE); }
214 
215  bool
216  _M_waiting() const noexcept
217  {
218  __platform_wait_t __res;
219  __atomic_load(&_M_wait, &__res, __ATOMIC_SEQ_CST);
220  return __res != 0;
221  }
222 
223  void
224  _M_notify(__platform_wait_t* __addr, [[maybe_unused]] bool __all,
225  bool __bare) noexcept
226  {
227 #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
228  if (__addr == &_M_ver)
229  {
230  __atomic_fetch_add(__addr, 1, __ATOMIC_SEQ_CST);
231  __all = true;
232  }
233 
234  if (__bare || _M_waiting())
235  __platform_notify(__addr, __all);
236 #else
237  {
238  lock_guard<mutex> __l(_M_mtx);
239  __atomic_fetch_add(__addr, 1, __ATOMIC_RELAXED);
240  }
241  if (__bare || _M_waiting())
242  _M_cv.notify_all();
243 #endif
244  }
245 
246  static __waiter_pool_base&
247  _S_for(const void* __addr) noexcept
248  {
249  constexpr uintptr_t __ct = 16;
250  static __waiter_pool_base __w[__ct];
251  auto __key = (uintptr_t(__addr) >> 2) % __ct;
252  return __w[__key];
253  }
254  };
255 
256  struct __waiter_pool : __waiter_pool_base
257  {
258  void
259  _M_do_wait(const __platform_wait_t* __addr, __platform_wait_t __old) noexcept
260  {
261 #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
262  __platform_wait(__addr, __old);
263 #else
264  __platform_wait_t __val;
265  __atomic_load(__addr, &__val, __ATOMIC_SEQ_CST);
266  if (__val == __old)
267  {
268  lock_guard<mutex> __l(_M_mtx);
269  __atomic_load(__addr, &__val, __ATOMIC_RELAXED);
270  if (__val == __old)
271  _M_cv.wait(_M_mtx);
272  }
273 #endif // __GLIBCXX_HAVE_PLATFORM_WAIT
274  }
275  };
276 
277  template<typename _Tp>
278  struct __waiter_base
279  {
280  using __waiter_type = _Tp;
281 
282  __waiter_type& _M_w;
283  __platform_wait_t* _M_addr;
284 
285  template<typename _Up>
286  static __platform_wait_t*
287  _S_wait_addr(const _Up* __a, __platform_wait_t* __b)
288  {
289  if constexpr (__platform_wait_uses_type<_Up>)
290  return reinterpret_cast<__platform_wait_t*>(const_cast<_Up*>(__a));
291  else
292  return __b;
293  }
294 
295  static __waiter_type&
296  _S_for(const void* __addr) noexcept
297  {
298  static_assert(sizeof(__waiter_type) == sizeof(__waiter_pool_base));
299  auto& res = __waiter_pool_base::_S_for(__addr);
300  return reinterpret_cast<__waiter_type&>(res);
301  }
302 
303  template<typename _Up>
304  explicit __waiter_base(const _Up* __addr) noexcept
305  : _M_w(_S_for(__addr))
306  , _M_addr(_S_wait_addr(__addr, &_M_w._M_ver))
307  { }
308 
309  void
310  _M_notify(bool __all, bool __bare = false) noexcept
311  { _M_w._M_notify(_M_addr, __all, __bare); }
312 
313  template<typename _Up, typename _ValFn,
314  typename _Spin = __default_spin_policy>
315  static bool
316  _S_do_spin_v(__platform_wait_t* __addr,
317  const _Up& __old, _ValFn __vfn,
318  __platform_wait_t& __val,
319  _Spin __spin = _Spin{ })
320  {
321  auto const __pred = [=]
322  { return !__detail::__atomic_compare(__old, __vfn()); };
323 
324  if constexpr (__platform_wait_uses_type<_Up>)
325  {
326  __builtin_memcpy(&__val, &__old, sizeof(__val));
327  }
328  else
329  {
330  __atomic_load(__addr, &__val, __ATOMIC_ACQUIRE);
331  }
332  return __atomic_spin(__pred, __spin);
333  }
334 
335  template<typename _Up, typename _ValFn,
336  typename _Spin = __default_spin_policy>
337  bool
338  _M_do_spin_v(const _Up& __old, _ValFn __vfn,
339  __platform_wait_t& __val,
340  _Spin __spin = _Spin{ })
341  { return _S_do_spin_v(_M_addr, __old, __vfn, __val, __spin); }
342 
343  template<typename _Pred,
344  typename _Spin = __default_spin_policy>
345  static bool
346  _S_do_spin(const __platform_wait_t* __addr,
347  _Pred __pred,
348  __platform_wait_t& __val,
349  _Spin __spin = _Spin{ })
350  {
351  __atomic_load(__addr, &__val, __ATOMIC_ACQUIRE);
352  return __atomic_spin(__pred, __spin);
353  }
354 
355  template<typename _Pred,
356  typename _Spin = __default_spin_policy>
357  bool
358  _M_do_spin(_Pred __pred, __platform_wait_t& __val,
359  _Spin __spin = _Spin{ })
360  { return _S_do_spin(_M_addr, __pred, __val, __spin); }
361  };
362 
363  template<typename _EntersWait>
364  struct __waiter : __waiter_base<__waiter_pool>
365  {
366  using __base_type = __waiter_base<__waiter_pool>;
367 
368  template<typename _Tp>
369  explicit __waiter(const _Tp* __addr) noexcept
370  : __base_type(__addr)
371  {
372  if constexpr (_EntersWait::value)
373  _M_w._M_enter_wait();
374  }
375 
376  ~__waiter()
377  {
378  if constexpr (_EntersWait::value)
379  _M_w._M_leave_wait();
380  }
381 
382  template<typename _Tp, typename _ValFn>
383  void
384  _M_do_wait_v(_Tp __old, _ValFn __vfn)
385  {
386  do
387  {
388  __platform_wait_t __val;
389  if (__base_type::_M_do_spin_v(__old, __vfn, __val))
390  return;
391  __base_type::_M_w._M_do_wait(__base_type::_M_addr, __val);
392  }
393  while (__detail::__atomic_compare(__old, __vfn()));
394  }
395 
396  template<typename _Pred>
397  void
398  _M_do_wait(_Pred __pred) noexcept
399  {
400  do
401  {
402  __platform_wait_t __val;
403  if (__base_type::_M_do_spin(__pred, __val))
404  return;
405  __base_type::_M_w._M_do_wait(__base_type::_M_addr, __val);
406  }
407  while (!__pred());
408  }
409  };
410 
411  using __enters_wait = __waiter<std::true_type>;
412  using __bare_wait = __waiter<std::false_type>;
413  } // namespace __detail
414 
415  template<typename _Tp, typename _ValFn>
416  void
417  __atomic_wait_address_v(const _Tp* __addr, _Tp __old,
418  _ValFn __vfn) noexcept
419  {
420  __detail::__enters_wait __w(__addr);
421  __w._M_do_wait_v(__old, __vfn);
422  }
423 
424  template<typename _Tp, typename _Pred>
425  void
426  __atomic_wait_address(const _Tp* __addr, _Pred __pred) noexcept
427  {
428  __detail::__enters_wait __w(__addr);
429  __w._M_do_wait(__pred);
430  }
431 
432  // This call is to be used by atomic types which track contention externally
433  template<typename _Pred>
434  void
435  __atomic_wait_address_bare(const __detail::__platform_wait_t* __addr,
436  _Pred __pred) noexcept
437  {
438 #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
439  do
440  {
441  __detail::__platform_wait_t __val;
442  if (__detail::__bare_wait::_S_do_spin(__addr, __pred, __val))
443  return;
444  __detail::__platform_wait(__addr, __val);
445  }
446  while (!__pred());
447 #else // !_GLIBCXX_HAVE_PLATFORM_WAIT
448  __detail::__bare_wait __w(__addr);
449  __w._M_do_wait(__pred);
450 #endif
451  }
452 
453  template<typename _Tp>
454  void
455  __atomic_notify_address(const _Tp* __addr, bool __all) noexcept
456  {
457  __detail::__bare_wait __w(__addr);
458  __w._M_notify(__all);
459  }
460 
461  // This call is to be used by atomic types which track contention externally
462  inline void
463  __atomic_notify_address_bare(const __detail::__platform_wait_t* __addr,
464  bool __all) noexcept
465  {
466 #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT
467  __detail::__platform_notify(__addr, __all);
468 #else
469  __detail::__bare_wait __w(__addr);
470  __w._M_notify(__all, true);
471 #endif
472  }
473 _GLIBCXX_END_NAMESPACE_VERSION
474 } // namespace std
475 #endif // GTHREADS || LINUX_FUTEX
476 #endif // _GLIBCXX_ATOMIC_WAIT_H
ISO C++ entities toplevel namespace is std.