| 1 | // -*- C++ -*- header.  |
| 2 |   |
| 3 | // Copyright (C) 2020-2021 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 | static 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 | static 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 | constexpr auto __atomic_spin_count_relax = 12;  |
| 146 | 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 | #ifdef __cpp_lib_hardware_interference_size  |
| 191 | static constexpr auto _S_align = hardware_destructive_interference_size;  |
| 192 | #else  |
| 193 | static constexpr auto _S_align = 64;  |
| 194 | #endif  |
| 195 |   |
| 196 | alignas(_S_align) __platform_wait_t _M_wait = 0;  |
| 197 |   |
| 198 | #ifndef _GLIBCXX_HAVE_PLATFORM_WAIT  |
| 199 | mutex _M_mtx;  |
| 200 | #endif  |
| 201 |   |
| 202 | alignas(_S_align) __platform_wait_t _M_ver = 0;  |
| 203 |   |
| 204 | #ifndef _GLIBCXX_HAVE_PLATFORM_WAIT  |
| 205 | __condvar _M_cv;  |
| 206 | #endif  |
| 207 | __waiter_pool_base() = default;  |
| 208 |   |
| 209 | void  |
| 210 | _M_enter_wait() noexcept  |
| 211 | { __atomic_fetch_add(&_M_wait, 1, __ATOMIC_SEQ_CST); }  |
| 212 |   |
| 213 | void  |
| 214 | _M_leave_wait() noexcept  |
| 215 | { __atomic_fetch_sub(&_M_wait, 1, __ATOMIC_RELEASE); }  |
| 216 |   |
| 217 | bool  |
| 218 | _M_waiting() const noexcept  |
| 219 | {  |
| 220 | __platform_wait_t __res;  |
| 221 | __atomic_load(&_M_wait, &__res, __ATOMIC_SEQ_CST);  |
| 222 | return __res != 0;  |
| 223 | }  |
| 224 |   |
| 225 | void  |
| 226 | _M_notify(__platform_wait_t* __addr, [[maybe_unused]] bool __all,  |
| 227 | bool __bare) noexcept  |
| 228 | {  |
| 229 | #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT  |
| 230 | if (__addr == &_M_ver)  |
| 231 | {  |
| 232 | __atomic_fetch_add(__addr, 1, __ATOMIC_SEQ_CST);  |
| 233 | __all = true;  |
| 234 | }  |
| 235 |   |
| 236 | if (__bare || _M_waiting())  |
| 237 | __platform_notify(__addr, __all);  |
| 238 | #else  |
| 239 | {  |
| 240 | lock_guard<mutex> __l(_M_mtx);  |
| 241 | __atomic_fetch_add(__addr, 1, __ATOMIC_RELAXED);  |
| 242 | }  |
| 243 | if (__bare || _M_waiting())  |
| 244 | _M_cv.notify_all();  |
| 245 | #endif  |
| 246 | }  |
| 247 |   |
| 248 | static __waiter_pool_base&  |
| 249 | _S_for(const void* __addr) noexcept  |
| 250 | {  |
| 251 | constexpr uintptr_t __ct = 16;  |
| 252 | static __waiter_pool_base __w[__ct];  |
| 253 | auto __key = (uintptr_t(__addr) >> 2) % __ct;  |
| 254 | return __w[__key];  |
| 255 | }  |
| 256 | };  |
| 257 |   |
| 258 | struct __waiter_pool : __waiter_pool_base  |
| 259 | {  |
| 260 | void  |
| 261 | _M_do_wait(const __platform_wait_t* __addr, __platform_wait_t __old) noexcept  |
| 262 | {  |
| 263 | #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT  |
| 264 | __platform_wait(__addr, val: __old);  |
| 265 | #else  |
| 266 | __platform_wait_t __val;  |
| 267 | __atomic_load(__addr, &__val, __ATOMIC_SEQ_CST);  |
| 268 | if (__val == __old)  |
| 269 | {  |
| 270 | lock_guard<mutex> __l(_M_mtx);  |
| 271 | __atomic_load(__addr, &__val, __ATOMIC_RELAXED);  |
| 272 | if (__val == __old)  |
| 273 | _M_cv.wait(_M_mtx);  |
| 274 | }  |
| 275 | #endif // __GLIBCXX_HAVE_PLATFORM_WAIT  |
| 276 | }  |
| 277 | };  |
| 278 |   |
| 279 | template<typename _Tp>  |
| 280 | struct __waiter_base  |
| 281 | {  |
| 282 | using __waiter_type = _Tp;  |
| 283 |   |
| 284 | __waiter_type& _M_w;  |
| 285 | __platform_wait_t* _M_addr;  |
| 286 |   |
| 287 | template<typename _Up>  |
| 288 | static __platform_wait_t*  |
| 289 | _S_wait_addr(const _Up* __a, __platform_wait_t* __b)  |
| 290 | {  |
| 291 | if constexpr (__platform_wait_uses_type<_Up>)  |
| 292 | return reinterpret_cast<__platform_wait_t*>(const_cast<_Up*>(__a));  |
| 293 | else  |
| 294 | return __b;  |
| 295 | }  |
| 296 |   |
| 297 | static __waiter_type&  |
| 298 | _S_for(const void* __addr) noexcept  |
| 299 | {  |
| 300 | static_assert(sizeof(__waiter_type) == sizeof(__waiter_pool_base));  |
| 301 | auto& res = __waiter_pool_base::_S_for(__addr);  |
| 302 | return reinterpret_cast<__waiter_type&>(res);  |
| 303 | }  |
| 304 |   |
| 305 | template<typename _Up>  |
| 306 | explicit __waiter_base(const _Up* __addr) noexcept  |
| 307 | : _M_w(_S_for(__addr))  |
| 308 | , _M_addr(_S_wait_addr(__addr, &_M_w._M_ver))  |
| 309 | { }  |
| 310 |   |
| 311 | void  |
| 312 | _M_notify(bool __all, bool __bare = false) noexcept  |
| 313 | { _M_w._M_notify(_M_addr, __all, __bare); }  |
| 314 |   |
| 315 | template<typename _Up, typename _ValFn,  |
| 316 | typename _Spin = __default_spin_policy>  |
| 317 | static bool  |
| 318 | _S_do_spin_v(__platform_wait_t* __addr,  |
| 319 | const _Up& __old, _ValFn __vfn,  |
| 320 | __platform_wait_t& __val,  |
| 321 | _Spin __spin = _Spin{ })  |
| 322 | {  |
| 323 | auto const __pred = [=]  |
| 324 | { return !__detail::__atomic_compare(__old, __vfn()); };  |
| 325 |   |
| 326 | if constexpr (__platform_wait_uses_type<_Up>)  |
| 327 | {  |
| 328 | __builtin_memcpy(&__val, &__old, sizeof(__val));  |
| 329 | }  |
| 330 | else  |
| 331 | {  |
| 332 | __atomic_load(__addr, &__val, __ATOMIC_ACQUIRE);  |
| 333 | }  |
| 334 | return __atomic_spin(__pred, __spin);  |
| 335 | }  |
| 336 |   |
| 337 | template<typename _Up, typename _ValFn,  |
| 338 | typename _Spin = __default_spin_policy>  |
| 339 | bool  |
| 340 | _M_do_spin_v(const _Up& __old, _ValFn __vfn,  |
| 341 | __platform_wait_t& __val,  |
| 342 | _Spin __spin = _Spin{ })  |
| 343 | { return _S_do_spin_v(_M_addr, __old, __vfn, __val, __spin); }  |
| 344 |   |
| 345 | template<typename _Pred,  |
| 346 | typename _Spin = __default_spin_policy>  |
| 347 | static bool  |
| 348 | _S_do_spin(const __platform_wait_t* __addr,  |
| 349 | _Pred __pred,  |
| 350 | __platform_wait_t& __val,  |
| 351 | _Spin __spin = _Spin{ })  |
| 352 | {  |
| 353 | __atomic_load(__addr, &__val, __ATOMIC_ACQUIRE);  |
| 354 | return __atomic_spin(__pred, __spin);  |
| 355 | }  |
| 356 |   |
| 357 | template<typename _Pred,  |
| 358 | typename _Spin = __default_spin_policy>  |
| 359 | bool  |
| 360 | _M_do_spin(_Pred __pred, __platform_wait_t& __val,  |
| 361 | _Spin __spin = _Spin{ })  |
| 362 | { return _S_do_spin(_M_addr, __pred, __val, __spin); }  |
| 363 | };  |
| 364 |   |
| 365 | template<typename _EntersWait>  |
| 366 | struct __waiter : __waiter_base<__waiter_pool>  |
| 367 | {  |
| 368 | using __base_type = __waiter_base<__waiter_pool>;  |
| 369 |   |
| 370 | template<typename _Tp>  |
| 371 | explicit __waiter(const _Tp* __addr) noexcept  |
| 372 | : __base_type(__addr)  |
| 373 | {  |
| 374 | if constexpr (_EntersWait::value)  |
| 375 | _M_w._M_enter_wait();  |
| 376 | }  |
| 377 |   |
| 378 | ~__waiter()  |
| 379 | {  |
| 380 | if constexpr (_EntersWait::value)  |
| 381 | _M_w._M_leave_wait();  |
| 382 | }  |
| 383 |   |
| 384 | template<typename _Tp, typename _ValFn>  |
| 385 | void  |
| 386 | _M_do_wait_v(_Tp __old, _ValFn __vfn)  |
| 387 | {  |
| 388 | do  |
| 389 | {  |
| 390 | __platform_wait_t __val;  |
| 391 | if (__base_type::_M_do_spin_v(__old, __vfn, __val))  |
| 392 | return;  |
| 393 | __base_type::_M_w._M_do_wait(addr: __base_type::_M_addr, old: __val);  |
| 394 | }  |
| 395 | while (__detail::__atomic_compare(__old, __vfn()));  |
| 396 | }  |
| 397 |   |
| 398 | template<typename _Pred>  |
| 399 | void  |
| 400 | _M_do_wait(_Pred __pred) noexcept  |
| 401 | {  |
| 402 | do  |
| 403 | {  |
| 404 | __platform_wait_t __val;  |
| 405 | if (__base_type::_M_do_spin(__pred, __val))  |
| 406 | return;  |
| 407 | __base_type::_M_w._M_do_wait(addr: __base_type::_M_addr, old: __val);  |
| 408 | }  |
| 409 | while (!__pred());  |
| 410 | }  |
| 411 | };  |
| 412 |   |
| 413 | using __enters_wait = __waiter<std::true_type>;  |
| 414 | using __bare_wait = __waiter<std::false_type>;  |
| 415 | } // namespace __detail  |
| 416 |   |
| 417 | template<typename _Tp, typename _ValFn>  |
| 418 | void  |
| 419 | __atomic_wait_address_v(const _Tp* __addr, _Tp __old,  |
| 420 | _ValFn __vfn) noexcept  |
| 421 | {  |
| 422 | __detail::__enters_wait __w(__addr);  |
| 423 | __w._M_do_wait_v(__old, __vfn);  |
| 424 | }  |
| 425 |   |
| 426 | template<typename _Tp, typename _Pred>  |
| 427 | void  |
| 428 | __atomic_wait_address(const _Tp* __addr, _Pred __pred) noexcept  |
| 429 | {  |
| 430 | __detail::__enters_wait __w(__addr);  |
| 431 | __w._M_do_wait(__pred);  |
| 432 | }  |
| 433 |   |
| 434 | // This call is to be used by atomic types which track contention externally  |
| 435 | template<typename _Pred>  |
| 436 | void  |
| 437 | __atomic_wait_address_bare(const __detail::__platform_wait_t* __addr,  |
| 438 | _Pred __pred) noexcept  |
| 439 | {  |
| 440 | #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT  |
| 441 | do  |
| 442 | {  |
| 443 | __detail::__platform_wait_t __val;  |
| 444 | if (__detail::__bare_wait::_S_do_spin(__addr, __pred, __val))  |
| 445 | return;  |
| 446 | __detail::__platform_wait(__addr, __val);  |
| 447 | }  |
| 448 | while (!__pred());  |
| 449 | #else // !_GLIBCXX_HAVE_PLATFORM_WAIT  |
| 450 | __detail::__bare_wait __w(__addr);  |
| 451 | __w._M_do_wait(__pred);  |
| 452 | #endif  |
| 453 | }  |
| 454 |   |
| 455 | template<typename _Tp>  |
| 456 | void  |
| 457 | __atomic_notify_address(const _Tp* __addr, bool __all) noexcept  |
| 458 | {  |
| 459 | __detail::__bare_wait __w(__addr);  |
| 460 | __w._M_notify(__all);  |
| 461 | }  |
| 462 |   |
| 463 | // This call is to be used by atomic types which track contention externally  |
| 464 | inline void  |
| 465 | __atomic_notify_address_bare(const __detail::__platform_wait_t* __addr,  |
| 466 | bool __all) noexcept  |
| 467 | {  |
| 468 | #ifdef _GLIBCXX_HAVE_PLATFORM_WAIT  |
| 469 | __detail::__platform_notify(__addr, __all);  |
| 470 | #else  |
| 471 | __detail::__bare_wait __w(__addr);  |
| 472 | __w._M_notify(__all, true);  |
| 473 | #endif  |
| 474 | }  |
| 475 | _GLIBCXX_END_NAMESPACE_VERSION  |
| 476 | } // namespace std  |
| 477 | #endif // GTHREADS || LINUX_FUTEX  |
| 478 | #endif // _GLIBCXX_ATOMIC_WAIT_H  |
| 479 | |