master
1//===----------------------------------------------------------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8
9#include <__thread/timed_backoff_policy.h>
10#include <atomic>
11#include <climits>
12#include <functional>
13#include <thread>
14
15#include "include/apple_availability.h"
16
17#ifdef __linux__
18
19# include <linux/futex.h>
20# include <sys/syscall.h>
21# include <unistd.h>
22
23// libc++ uses SYS_futex as a universal syscall name. However, on 32 bit architectures
24// with a 64 bit time_t, we need to specify SYS_futex_time64.
25# if !defined(SYS_futex) && defined(SYS_futex_time64)
26# define SYS_futex SYS_futex_time64
27# endif
28# define _LIBCPP_FUTEX(...) syscall(SYS_futex, __VA_ARGS__)
29
30#elif defined(__FreeBSD__)
31
32# include <sys/types.h>
33# include <sys/umtx.h>
34
35# define _LIBCPP_FUTEX(...) syscall(SYS_futex, __VA_ARGS__)
36
37#elif defined(__OpenBSD__)
38
39# include <sys/futex.h>
40
41// OpenBSD has no indirect syscalls
42# define _LIBCPP_FUTEX(...) futex(__VA_ARGS__)
43
44#else // <- Add other operating systems here
45
46// Baseline needs no new headers
47
48# define _LIBCPP_FUTEX(...) syscall(SYS_futex, __VA_ARGS__)
49
50#endif
51
52_LIBCPP_BEGIN_NAMESPACE_STD
53
54#ifdef __linux__
55
56static void
57__libcpp_platform_wait_on_address(__cxx_atomic_contention_t const volatile* __ptr, __cxx_contention_t __val) {
58 static constexpr timespec __timeout = {2, 0};
59 _LIBCPP_FUTEX(__ptr, FUTEX_WAIT_PRIVATE, __val, &__timeout, 0, 0);
60}
61
62static void __libcpp_platform_wake_by_address(__cxx_atomic_contention_t const volatile* __ptr, bool __notify_one) {
63 _LIBCPP_FUTEX(__ptr, FUTEX_WAKE_PRIVATE, __notify_one ? 1 : INT_MAX, 0, 0, 0);
64}
65
66#elif defined(__APPLE__) && defined(_LIBCPP_USE_ULOCK)
67
68extern "C" int __ulock_wait(
69 uint32_t operation, void* addr, uint64_t value, uint32_t timeout); /* timeout is specified in microseconds */
70extern "C" int __ulock_wake(uint32_t operation, void* addr, uint64_t wake_value);
71
72// https://github.com/apple/darwin-xnu/blob/2ff845c2e033bd0ff64b5b6aa6063a1f8f65aa32/bsd/sys/ulock.h#L82
73# define UL_COMPARE_AND_WAIT64 5
74# define ULF_WAKE_ALL 0x00000100
75
76static void
77__libcpp_platform_wait_on_address(__cxx_atomic_contention_t const volatile* __ptr, __cxx_contention_t __val) {
78 static_assert(sizeof(__cxx_atomic_contention_t) == 8, "Waiting on 8 bytes value");
79 __ulock_wait(UL_COMPARE_AND_WAIT64, const_cast<__cxx_atomic_contention_t*>(__ptr), __val, 0);
80}
81
82static void __libcpp_platform_wake_by_address(__cxx_atomic_contention_t const volatile* __ptr, bool __notify_one) {
83 static_assert(sizeof(__cxx_atomic_contention_t) == 8, "Waking up on 8 bytes value");
84 __ulock_wake(
85 UL_COMPARE_AND_WAIT64 | (__notify_one ? 0 : ULF_WAKE_ALL), const_cast<__cxx_atomic_contention_t*>(__ptr), 0);
86}
87
88#elif defined(__FreeBSD__) && __SIZEOF_LONG__ == 8
89/*
90 * Since __cxx_contention_t is int64_t even on 32bit FreeBSD
91 * platforms, we have to use umtx ops that work on the long type, and
92 * limit its use to architectures where long and int64_t are synonyms.
93 */
94
95static void
96__libcpp_platform_wait_on_address(__cxx_atomic_contention_t const volatile* __ptr, __cxx_contention_t __val) {
97 _umtx_op(const_cast<__cxx_atomic_contention_t*>(__ptr), UMTX_OP_WAIT, __val, nullptr, nullptr);
98}
99
100static void __libcpp_platform_wake_by_address(__cxx_atomic_contention_t const volatile* __ptr, bool __notify_one) {
101 _umtx_op(const_cast<__cxx_atomic_contention_t*>(__ptr), UMTX_OP_WAKE, __notify_one ? 1 : INT_MAX, nullptr, nullptr);
102}
103
104#else // <- Add other operating systems here
105
106// Baseline is just a timed backoff
107
108static void
109__libcpp_platform_wait_on_address(__cxx_atomic_contention_t const volatile* __ptr, __cxx_contention_t __val) {
110 __libcpp_thread_poll_with_backoff(
111 [=]() -> bool { return !__cxx_nonatomic_compare_equal(__cxx_atomic_load(__ptr, memory_order_relaxed), __val); },
112 __libcpp_timed_backoff_policy());
113}
114
115static void __libcpp_platform_wake_by_address(__cxx_atomic_contention_t const volatile*, bool) {}
116
117#endif // __linux__
118
119static constexpr size_t __libcpp_contention_table_size = (1 << 8); /* < there's no magic in this number */
120
121struct alignas(64) /* aim to avoid false sharing */ __libcpp_contention_table_entry {
122 __cxx_atomic_contention_t __contention_state;
123 __cxx_atomic_contention_t __platform_state;
124 inline constexpr __libcpp_contention_table_entry() : __contention_state(0), __platform_state(0) {}
125};
126
127static __libcpp_contention_table_entry __libcpp_contention_table[__libcpp_contention_table_size];
128
129static hash<void const volatile*> __libcpp_contention_hasher;
130
131static __libcpp_contention_table_entry* __libcpp_contention_state(void const volatile* p) {
132 return &__libcpp_contention_table[__libcpp_contention_hasher(p) & (__libcpp_contention_table_size - 1)];
133}
134
135/* Given an atomic to track contention and an atomic to actually wait on, which may be
136 the same atomic, we try to detect contention to avoid spuriously calling the platform. */
137
138static void __libcpp_contention_notify(__cxx_atomic_contention_t volatile* __contention_state,
139 __cxx_atomic_contention_t const volatile* __platform_state,
140 bool __notify_one) {
141 if (0 != __cxx_atomic_load(__contention_state, memory_order_seq_cst))
142 // We only call 'wake' if we consumed a contention bit here.
143 __libcpp_platform_wake_by_address(__platform_state, __notify_one);
144}
145static __cxx_contention_t
146__libcpp_contention_monitor_for_wait(__cxx_atomic_contention_t volatile* /*__contention_state*/,
147 __cxx_atomic_contention_t const volatile* __platform_state) {
148 // We will monitor this value.
149 return __cxx_atomic_load(__platform_state, memory_order_acquire);
150}
151static void __libcpp_contention_wait(__cxx_atomic_contention_t volatile* __contention_state,
152 __cxx_atomic_contention_t const volatile* __platform_state,
153 __cxx_contention_t __old_value) {
154 __cxx_atomic_fetch_add(__contention_state, __cxx_contention_t(1), memory_order_relaxed);
155 // https://github.com/llvm/llvm-project/issues/109290
156 // There are no platform guarantees of a memory barrier in the platform wait implementation
157 __cxx_atomic_thread_fence(memory_order_seq_cst);
158 // We sleep as long as the monitored value hasn't changed.
159 __libcpp_platform_wait_on_address(__platform_state, __old_value);
160 __cxx_atomic_fetch_sub(__contention_state, __cxx_contention_t(1), memory_order_release);
161}
162
163/* When the incoming atomic is the wrong size for the platform wait size, need to
164 launder the value sequence through an atomic from our table. */
165
166static void __libcpp_atomic_notify(void const volatile* __location) {
167 auto const __entry = __libcpp_contention_state(__location);
168 // The value sequence laundering happens on the next line below.
169 __cxx_atomic_fetch_add(&__entry->__platform_state, __cxx_contention_t(1), memory_order_seq_cst);
170 __libcpp_contention_notify(
171 &__entry->__contention_state,
172 &__entry->__platform_state,
173 false /* when laundering, we can't handle notify_one */);
174}
175_LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_one(void const volatile* __location) noexcept {
176 __libcpp_atomic_notify(__location);
177}
178_LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_all(void const volatile* __location) noexcept {
179 __libcpp_atomic_notify(__location);
180}
181_LIBCPP_EXPORTED_FROM_ABI __cxx_contention_t __libcpp_atomic_monitor(void const volatile* __location) noexcept {
182 auto const __entry = __libcpp_contention_state(__location);
183 return __libcpp_contention_monitor_for_wait(&__entry->__contention_state, &__entry->__platform_state);
184}
185_LIBCPP_EXPORTED_FROM_ABI void
186__libcpp_atomic_wait(void const volatile* __location, __cxx_contention_t __old_value) noexcept {
187 auto const __entry = __libcpp_contention_state(__location);
188 __libcpp_contention_wait(&__entry->__contention_state, &__entry->__platform_state, __old_value);
189}
190
191/* When the incoming atomic happens to be the platform wait size, we still need to use the
192 table for the contention detection, but we can use the atomic directly for the wait. */
193
194_LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_one(__cxx_atomic_contention_t const volatile* __location) noexcept {
195 __libcpp_contention_notify(&__libcpp_contention_state(__location)->__contention_state, __location, true);
196}
197_LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_all(__cxx_atomic_contention_t const volatile* __location) noexcept {
198 __libcpp_contention_notify(&__libcpp_contention_state(__location)->__contention_state, __location, false);
199}
200// This function is never used, but still exported for ABI compatibility.
201_LIBCPP_EXPORTED_FROM_ABI __cxx_contention_t
202__libcpp_atomic_monitor(__cxx_atomic_contention_t const volatile* __location) noexcept {
203 return __libcpp_contention_monitor_for_wait(&__libcpp_contention_state(__location)->__contention_state, __location);
204}
205_LIBCPP_EXPORTED_FROM_ABI void
206__libcpp_atomic_wait(__cxx_atomic_contention_t const volatile* __location, __cxx_contention_t __old_value) noexcept {
207 __libcpp_contention_wait(&__libcpp_contention_state(__location)->__contention_state, __location, __old_value);
208}
209
210_LIBCPP_END_NAMESPACE_STD