1 | // -*- C++ -*- header. |
2 | |
3 | // Copyright (C) 2020-2024 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/version.h> |
36 | |
37 | #if __glibcxx_atomic_wait |
38 | #include <cstdint> |
39 | #include <bits/functional_hash.h> |
40 | #include <bits/gthr.h> |
41 | #include <ext/numeric_traits.h> |
42 | |
43 | #ifdef _GLIBCXX_HAVE_LINUX_FUTEX |
44 | # include <cerrno> |
45 | # include <climits> |
46 | # include <unistd.h> |
47 | # include <syscall.h> |
48 | # include <bits/functexcept.h> |
49 | #endif |
50 | |
51 | # include <bits/std_mutex.h> // std::mutex, std::__condvar |
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 | # if ATOMIC_LONG_LOCK_FREE == 2 |
68 | using __platform_wait_t = unsigned long; |
69 | # else |
70 | using __platform_wait_t = unsigned int; |
71 | # endif |
72 | inline constexpr size_t __platform_wait_alignment |
73 | = __alignof__(__platform_wait_t); |
74 | #endif |
75 | } // namespace __detail |
76 | |
77 | template<typename _Tp> |
78 | inline constexpr bool __platform_wait_uses_type |
79 | #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT |
80 | = is_scalar_v<_Tp> |
81 | && ((sizeof(_Tp) == sizeof(__detail::__platform_wait_t)) |
82 | && (alignof(_Tp*) >= __detail::__platform_wait_alignment)); |
83 | #else |
84 | = false; |
85 | #endif |
86 | |
87 | namespace __detail |
88 | { |
89 | #ifdef _GLIBCXX_HAVE_LINUX_FUTEX |
90 | enum class __futex_wait_flags : int |
91 | { |
92 | #ifdef _GLIBCXX_HAVE_LINUX_FUTEX_PRIVATE |
93 | __private_flag = 128, |
94 | #else |
95 | __private_flag = 0, |
96 | #endif |
97 | __wait = 0, |
98 | __wake = 1, |
99 | __wait_bitset = 9, |
100 | __wake_bitset = 10, |
101 | __wait_private = __wait | __private_flag, |
102 | __wake_private = __wake | __private_flag, |
103 | __wait_bitset_private = __wait_bitset | __private_flag, |
104 | __wake_bitset_private = __wake_bitset | __private_flag, |
105 | __bitset_match_any = -1 |
106 | }; |
107 | |
108 | template<typename _Tp> |
109 | void |
110 | __platform_wait(const _Tp* __addr, __platform_wait_t __val) noexcept |
111 | { |
112 | auto __e = syscall (SYS_futex, static_cast<const void*>(__addr), |
113 | static_cast<int>(__futex_wait_flags::__wait_private), |
114 | __val, nullptr); |
115 | if (!__e || errno == EAGAIN) |
116 | return; |
117 | if (errno != EINTR) |
118 | __throw_system_error(errno); |
119 | } |
120 | |
121 | template<typename _Tp> |
122 | void |
123 | __platform_notify(const _Tp* __addr, bool __all) noexcept |
124 | { |
125 | syscall (SYS_futex, static_cast<const void*>(__addr), |
126 | static_cast<int>(__futex_wait_flags::__wake_private), |
127 | __all ? INT_MAX : 1); |
128 | } |
129 | #endif |
130 | |
131 | inline void |
132 | __thread_yield() noexcept |
133 | { |
134 | #if defined _GLIBCXX_HAS_GTHREADS && defined _GLIBCXX_USE_SCHED_YIELD |
135 | __gthread_yield(); |
136 | #endif |
137 | } |
138 | |
139 | inline void |
140 | __thread_relax() noexcept |
141 | { |
142 | #if defined __i386__ || defined __x86_64__ |
143 | __builtin_ia32_pause(); |
144 | #else |
145 | __thread_yield(); |
146 | #endif |
147 | } |
148 | |
149 | inline constexpr auto __atomic_spin_count_relax = 12; |
150 | inline constexpr auto __atomic_spin_count = 16; |
151 | |
152 | struct __default_spin_policy |
153 | { |
154 | bool |
155 | operator()() const noexcept |
156 | { return false; } |
157 | }; |
158 | |
159 | template<typename _Pred, |
160 | typename _Spin = __default_spin_policy> |
161 | bool |
162 | __atomic_spin(_Pred& __pred, _Spin __spin = _Spin{ }) noexcept |
163 | { |
164 | for (auto __i = 0; __i < __atomic_spin_count; ++__i) |
165 | { |
166 | if (__pred()) |
167 | return true; |
168 | |
169 | if (__i < __atomic_spin_count_relax) |
170 | __detail::__thread_relax(); |
171 | else |
172 | __detail::__thread_yield(); |
173 | } |
174 | |
175 | while (__spin()) |
176 | { |
177 | if (__pred()) |
178 | return true; |
179 | } |
180 | |
181 | return false; |
182 | } |
183 | |
184 | // return true if equal |
185 | template<typename _Tp> |
186 | bool __atomic_compare(const _Tp& __a, const _Tp& __b) |
187 | { |
188 | // TODO make this do the correct padding bit ignoring comparison |
189 | return __builtin_memcmp(&__a, &__b, sizeof(_Tp)) == 0; |
190 | } |
191 | |
192 | struct __waiter_pool_base |
193 | { |
194 | // Don't use std::hardware_destructive_interference_size here because we |
195 | // don't want the layout of library types to depend on compiler options. |
196 | static constexpr auto _S_align = 64; |
197 | |
198 | alignas(_S_align) __platform_wait_t _M_wait = 0; |
199 | |
200 | #ifndef _GLIBCXX_HAVE_PLATFORM_WAIT |
201 | mutex _M_mtx; |
202 | #endif |
203 | |
204 | alignas(_S_align) __platform_wait_t _M_ver = 0; |
205 | |
206 | #ifndef _GLIBCXX_HAVE_PLATFORM_WAIT |
207 | __condvar _M_cv; |
208 | #endif |
209 | __waiter_pool_base() = default; |
210 | |
211 | void |
212 | _M_enter_wait() noexcept |
213 | { __atomic_fetch_add(&_M_wait, 1, __ATOMIC_SEQ_CST); } |
214 | |
215 | void |
216 | _M_leave_wait() noexcept |
217 | { __atomic_fetch_sub(&_M_wait, 1, __ATOMIC_RELEASE); } |
218 | |
219 | bool |
220 | _M_waiting() const noexcept |
221 | { |
222 | __platform_wait_t __res; |
223 | __atomic_load(&_M_wait, &__res, __ATOMIC_SEQ_CST); |
224 | return __res != 0; |
225 | } |
226 | |
227 | void |
228 | _M_notify(__platform_wait_t* __addr, [[maybe_unused]] bool __all, |
229 | bool __bare) noexcept |
230 | { |
231 | #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT |
232 | if (__addr == &_M_ver) |
233 | { |
234 | __atomic_fetch_add(__addr, 1, __ATOMIC_SEQ_CST); |
235 | __all = true; |
236 | } |
237 | |
238 | if (__bare || _M_waiting()) |
239 | __platform_notify(__addr, __all); |
240 | #else |
241 | { |
242 | lock_guard<mutex> __l(_M_mtx); |
243 | __atomic_fetch_add(__addr, 1, __ATOMIC_RELAXED); |
244 | } |
245 | if (__bare || _M_waiting()) |
246 | _M_cv.notify_all(); |
247 | #endif |
248 | } |
249 | |
250 | static __waiter_pool_base& |
251 | _S_for(const void* __addr) noexcept |
252 | { |
253 | constexpr uintptr_t __ct = 16; |
254 | static __waiter_pool_base __w[__ct]; |
255 | auto __key = (uintptr_t(__addr) >> 2) % __ct; |
256 | return __w[__key]; |
257 | } |
258 | }; |
259 | |
260 | struct __waiter_pool : __waiter_pool_base |
261 | { |
262 | void |
263 | _M_do_wait(const __platform_wait_t* __addr, __platform_wait_t __old) noexcept |
264 | { |
265 | #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT |
266 | __platform_wait(__addr, val: __old); |
267 | #else |
268 | __platform_wait_t __val; |
269 | __atomic_load(__addr, &__val, __ATOMIC_SEQ_CST); |
270 | if (__val == __old) |
271 | { |
272 | lock_guard<mutex> __l(_M_mtx); |
273 | __atomic_load(__addr, &__val, __ATOMIC_RELAXED); |
274 | if (__val == __old) |
275 | _M_cv.wait(_M_mtx); |
276 | } |
277 | #endif // __GLIBCXX_HAVE_PLATFORM_WAIT |
278 | } |
279 | }; |
280 | |
281 | template<typename _Tp> |
282 | struct __waiter_base |
283 | { |
284 | using __waiter_type = _Tp; |
285 | |
286 | __waiter_type& _M_w; |
287 | __platform_wait_t* _M_addr; |
288 | |
289 | template<typename _Up> |
290 | static __platform_wait_t* |
291 | _S_wait_addr(const _Up* __a, __platform_wait_t* __b) |
292 | { |
293 | if constexpr (__platform_wait_uses_type<_Up>) |
294 | return reinterpret_cast<__platform_wait_t*>(const_cast<_Up*>(__a)); |
295 | else |
296 | return __b; |
297 | } |
298 | |
299 | static __waiter_type& |
300 | _S_for(const void* __addr) noexcept |
301 | { |
302 | static_assert(sizeof(__waiter_type) == sizeof(__waiter_pool_base)); |
303 | auto& res = __waiter_pool_base::_S_for(__addr); |
304 | return reinterpret_cast<__waiter_type&>(res); |
305 | } |
306 | |
307 | template<typename _Up> |
308 | explicit __waiter_base(const _Up* __addr) noexcept |
309 | : _M_w(_S_for(__addr)) |
310 | , _M_addr(_S_wait_addr(__addr, &_M_w._M_ver)) |
311 | { } |
312 | |
313 | void |
314 | _M_notify(bool __all, bool __bare = false) noexcept |
315 | { _M_w._M_notify(_M_addr, __all, __bare); } |
316 | |
317 | template<typename _Up, typename _ValFn, |
318 | typename _Spin = __default_spin_policy> |
319 | static bool |
320 | _S_do_spin_v(__platform_wait_t* __addr, |
321 | const _Up& __old, _ValFn __vfn, |
322 | __platform_wait_t& __val, |
323 | _Spin __spin = _Spin{ }) |
324 | { |
325 | auto const __pred = [=] |
326 | { return !__detail::__atomic_compare(__old, __vfn()); }; |
327 | |
328 | if constexpr (__platform_wait_uses_type<_Up>) |
329 | { |
330 | __builtin_memcpy(&__val, &__old, sizeof(__val)); |
331 | } |
332 | else |
333 | { |
334 | __atomic_load(__addr, &__val, __ATOMIC_ACQUIRE); |
335 | } |
336 | return __atomic_spin(__pred, __spin); |
337 | } |
338 | |
339 | template<typename _Up, typename _ValFn, |
340 | typename _Spin = __default_spin_policy> |
341 | bool |
342 | _M_do_spin_v(const _Up& __old, _ValFn __vfn, |
343 | __platform_wait_t& __val, |
344 | _Spin __spin = _Spin{ }) |
345 | { return _S_do_spin_v(_M_addr, __old, __vfn, __val, __spin); } |
346 | |
347 | template<typename _Pred, |
348 | typename _Spin = __default_spin_policy> |
349 | static bool |
350 | _S_do_spin(const __platform_wait_t* __addr, |
351 | _Pred __pred, |
352 | __platform_wait_t& __val, |
353 | _Spin __spin = _Spin{ }) |
354 | { |
355 | __atomic_load(__addr, &__val, __ATOMIC_ACQUIRE); |
356 | return __atomic_spin(__pred, __spin); |
357 | } |
358 | |
359 | template<typename _Pred, |
360 | typename _Spin = __default_spin_policy> |
361 | bool |
362 | _M_do_spin(_Pred __pred, __platform_wait_t& __val, |
363 | _Spin __spin = _Spin{ }) |
364 | { return _S_do_spin(_M_addr, __pred, __val, __spin); } |
365 | }; |
366 | |
367 | template<typename _EntersWait> |
368 | struct __waiter : __waiter_base<__waiter_pool> |
369 | { |
370 | using __base_type = __waiter_base<__waiter_pool>; |
371 | |
372 | template<typename _Tp> |
373 | explicit __waiter(const _Tp* __addr) noexcept |
374 | : __base_type(__addr) |
375 | { |
376 | if constexpr (_EntersWait::value) |
377 | _M_w._M_enter_wait(); |
378 | } |
379 | |
380 | ~__waiter() |
381 | { |
382 | if constexpr (_EntersWait::value) |
383 | _M_w._M_leave_wait(); |
384 | } |
385 | |
386 | template<typename _Tp, typename _ValFn> |
387 | void |
388 | _M_do_wait_v(_Tp __old, _ValFn __vfn) |
389 | { |
390 | do |
391 | { |
392 | __platform_wait_t __val; |
393 | if (__base_type::_M_do_spin_v(__old, __vfn, __val)) |
394 | return; |
395 | __base_type::_M_w._M_do_wait(addr: __base_type::_M_addr, old: __val); |
396 | } |
397 | while (__detail::__atomic_compare(__old, __vfn())); |
398 | } |
399 | |
400 | template<typename _Pred> |
401 | void |
402 | _M_do_wait(_Pred __pred) noexcept |
403 | { |
404 | do |
405 | { |
406 | __platform_wait_t __val; |
407 | if (__base_type::_M_do_spin(__pred, __val)) |
408 | return; |
409 | __base_type::_M_w._M_do_wait(addr: __base_type::_M_addr, old: __val); |
410 | } |
411 | while (!__pred()); |
412 | } |
413 | }; |
414 | |
415 | using __enters_wait = __waiter<std::true_type>; |
416 | using __bare_wait = __waiter<std::false_type>; |
417 | } // namespace __detail |
418 | |
419 | template<typename _Tp, typename _ValFn> |
420 | void |
421 | __atomic_wait_address_v(const _Tp* __addr, _Tp __old, |
422 | _ValFn __vfn) noexcept |
423 | { |
424 | __detail::__enters_wait __w(__addr); |
425 | __w._M_do_wait_v(__old, __vfn); |
426 | } |
427 | |
428 | template<typename _Tp, typename _Pred> |
429 | void |
430 | __atomic_wait_address(const _Tp* __addr, _Pred __pred) noexcept |
431 | { |
432 | __detail::__enters_wait __w(__addr); |
433 | __w._M_do_wait(__pred); |
434 | } |
435 | |
436 | // This call is to be used by atomic types which track contention externally |
437 | template<typename _Pred> |
438 | void |
439 | __atomic_wait_address_bare(const __detail::__platform_wait_t* __addr, |
440 | _Pred __pred) noexcept |
441 | { |
442 | #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT |
443 | do |
444 | { |
445 | __detail::__platform_wait_t __val; |
446 | if (__detail::__bare_wait::_S_do_spin(__addr, __pred, __val)) |
447 | return; |
448 | __detail::__platform_wait(__addr, __val); |
449 | } |
450 | while (!__pred()); |
451 | #else // !_GLIBCXX_HAVE_PLATFORM_WAIT |
452 | __detail::__bare_wait __w(__addr); |
453 | __w._M_do_wait(__pred); |
454 | #endif |
455 | } |
456 | |
457 | template<typename _Tp> |
458 | void |
459 | __atomic_notify_address(const _Tp* __addr, bool __all) noexcept |
460 | { |
461 | __detail::__bare_wait __w(__addr); |
462 | __w._M_notify(__all); |
463 | } |
464 | |
465 | // This call is to be used by atomic types which track contention externally |
466 | inline void |
467 | __atomic_notify_address_bare(const __detail::__platform_wait_t* __addr, |
468 | bool __all) noexcept |
469 | { |
470 | #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT |
471 | __detail::__platform_notify(__addr, __all); |
472 | #else |
473 | __detail::__bare_wait __w(__addr); |
474 | __w._M_notify(__all, true); |
475 | #endif |
476 | } |
477 | _GLIBCXX_END_NAMESPACE_VERSION |
478 | } // namespace std |
479 | #endif // __glibcxx_atomic_wait |
480 | #endif // _GLIBCXX_ATOMIC_WAIT_H |
481 | |