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 | |
29 | #elif defined(__FreeBSD__) |
30 | |
31 | # include <sys/types.h> |
32 | # include <sys/umtx.h> |
33 | |
34 | #else // <- Add other operating systems here |
35 | |
36 | // Baseline needs no new headers |
37 | |
38 | #endif |
39 | |
40 | _LIBCPP_BEGIN_NAMESPACE_STD |
41 | |
42 | #ifdef __linux__ |
43 | |
44 | static void |
45 | __libcpp_platform_wait_on_address(__cxx_atomic_contention_t const volatile* __ptr, __cxx_contention_t __val) { |
46 | static constexpr timespec __timeout = {.tv_sec: 2, .tv_nsec: 0}; |
47 | syscall(SYS_futex, __ptr, FUTEX_WAIT_PRIVATE, __val, &__timeout, 0, 0); |
48 | } |
49 | |
50 | static void __libcpp_platform_wake_by_address(__cxx_atomic_contention_t const volatile* __ptr, bool __notify_one) { |
51 | syscall(SYS_futex, __ptr, FUTEX_WAKE_PRIVATE, __notify_one ? 1 : INT_MAX, 0, 0, 0); |
52 | } |
53 | |
54 | #elif defined(__APPLE__) && defined(_LIBCPP_USE_ULOCK) |
55 | |
56 | extern "C" int __ulock_wait( |
57 | uint32_t operation, void* addr, uint64_t value, uint32_t timeout); /* timeout is specified in microseconds */ |
58 | extern "C" int __ulock_wake(uint32_t operation, void* addr, uint64_t wake_value); |
59 | |
60 | # define UL_COMPARE_AND_WAIT 1 |
61 | # define ULF_WAKE_ALL 0x00000100 |
62 | |
63 | static void |
64 | __libcpp_platform_wait_on_address(__cxx_atomic_contention_t const volatile* __ptr, __cxx_contention_t __val) { |
65 | __ulock_wait(UL_COMPARE_AND_WAIT, const_cast<__cxx_atomic_contention_t*>(__ptr), __val, 0); |
66 | } |
67 | |
68 | static void __libcpp_platform_wake_by_address(__cxx_atomic_contention_t const volatile* __ptr, bool __notify_one) { |
69 | __ulock_wake( |
70 | UL_COMPARE_AND_WAIT | (__notify_one ? 0 : ULF_WAKE_ALL), const_cast<__cxx_atomic_contention_t*>(__ptr), 0); |
71 | } |
72 | |
73 | #elif defined(__FreeBSD__) && __SIZEOF_LONG__ == 8 |
74 | /* |
75 | * Since __cxx_contention_t is int64_t even on 32bit FreeBSD |
76 | * platforms, we have to use umtx ops that work on the long type, and |
77 | * limit its use to architectures where long and int64_t are synonyms. |
78 | */ |
79 | |
80 | static void |
81 | __libcpp_platform_wait_on_address(__cxx_atomic_contention_t const volatile* __ptr, __cxx_contention_t __val) { |
82 | _umtx_op(const_cast<__cxx_atomic_contention_t*>(__ptr), UMTX_OP_WAIT, __val, NULL, NULL); |
83 | } |
84 | |
85 | static void __libcpp_platform_wake_by_address(__cxx_atomic_contention_t const volatile* __ptr, bool __notify_one) { |
86 | _umtx_op(const_cast<__cxx_atomic_contention_t*>(__ptr), UMTX_OP_WAKE, __notify_one ? 1 : INT_MAX, NULL, NULL); |
87 | } |
88 | |
89 | #else // <- Add other operating systems here |
90 | |
91 | // Baseline is just a timed backoff |
92 | |
93 | static void |
94 | __libcpp_platform_wait_on_address(__cxx_atomic_contention_t const volatile* __ptr, __cxx_contention_t __val) { |
95 | __libcpp_thread_poll_with_backoff( |
96 | [=]() -> bool { return !__cxx_nonatomic_compare_equal(__cxx_atomic_load(__ptr, memory_order_relaxed), __val); }, |
97 | __libcpp_timed_backoff_policy()); |
98 | } |
99 | |
100 | static void __libcpp_platform_wake_by_address(__cxx_atomic_contention_t const volatile*, bool) {} |
101 | |
102 | #endif // __linux__ |
103 | |
104 | static constexpr size_t __libcpp_contention_table_size = (1 << 8); /* < there's no magic in this number */ |
105 | |
106 | struct alignas(64) /* aim to avoid false sharing */ __libcpp_contention_table_entry { |
107 | __cxx_atomic_contention_t __contention_state; |
108 | __cxx_atomic_contention_t __platform_state; |
109 | inline constexpr __libcpp_contention_table_entry() : __contention_state(0), __platform_state(0) {} |
110 | }; |
111 | |
112 | static __libcpp_contention_table_entry __libcpp_contention_table[__libcpp_contention_table_size]; |
113 | |
114 | static hash<void const volatile*> __libcpp_contention_hasher; |
115 | |
116 | static __libcpp_contention_table_entry* __libcpp_contention_state(void const volatile* p) { |
117 | return &__libcpp_contention_table[__libcpp_contention_hasher(p) & (__libcpp_contention_table_size - 1)]; |
118 | } |
119 | |
120 | /* Given an atomic to track contention and an atomic to actually wait on, which may be |
121 | the same atomic, we try to detect contention to avoid spuriously calling the platform. */ |
122 | |
123 | static void __libcpp_contention_notify(__cxx_atomic_contention_t volatile* __contention_state, |
124 | __cxx_atomic_contention_t const volatile* __platform_state, |
125 | bool __notify_one) { |
126 | if (0 != __cxx_atomic_load(__contention_state, memory_order_seq_cst)) |
127 | // We only call 'wake' if we consumed a contention bit here. |
128 | __libcpp_platform_wake_by_address(__platform_state, __notify_one); |
129 | } |
130 | static __cxx_contention_t |
131 | __libcpp_contention_monitor_for_wait(__cxx_atomic_contention_t volatile* /*__contention_state*/, |
132 | __cxx_atomic_contention_t const volatile* __platform_state) { |
133 | // We will monitor this value. |
134 | return __cxx_atomic_load(__platform_state, memory_order_acquire); |
135 | } |
136 | static void __libcpp_contention_wait(__cxx_atomic_contention_t volatile* __contention_state, |
137 | __cxx_atomic_contention_t const volatile* __platform_state, |
138 | __cxx_contention_t __old_value) { |
139 | __cxx_atomic_fetch_add(__contention_state, __cxx_contention_t(1), memory_order_seq_cst); |
140 | // We sleep as long as the monitored value hasn't changed. |
141 | __libcpp_platform_wait_on_address(__platform_state, __old_value); |
142 | __cxx_atomic_fetch_sub(__contention_state, __cxx_contention_t(1), memory_order_release); |
143 | } |
144 | |
145 | /* When the incoming atomic is the wrong size for the platform wait size, need to |
146 | launder the value sequence through an atomic from our table. */ |
147 | |
148 | static void __libcpp_atomic_notify(void const volatile* __location) { |
149 | auto const __entry = __libcpp_contention_state(p: __location); |
150 | // The value sequence laundering happens on the next line below. |
151 | __cxx_atomic_fetch_add(&__entry->__platform_state, __cxx_contention_t(1), memory_order_release); |
152 | __libcpp_contention_notify( |
153 | &__entry->__contention_state, |
154 | &__entry->__platform_state, |
155 | false /* when laundering, we can't handle notify_one */); |
156 | } |
157 | _LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_one(void const volatile* __location) { |
158 | __libcpp_atomic_notify(__location); |
159 | } |
160 | _LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_all(void const volatile* __location) { |
161 | __libcpp_atomic_notify(__location); |
162 | } |
163 | _LIBCPP_EXPORTED_FROM_ABI __cxx_contention_t __libcpp_atomic_monitor(void const volatile* __location) { |
164 | auto const __entry = __libcpp_contention_state(__location); |
165 | return __libcpp_contention_monitor_for_wait(&__entry->__contention_state, &__entry->__platform_state); |
166 | } |
167 | _LIBCPP_EXPORTED_FROM_ABI void __libcpp_atomic_wait(void const volatile* __location, __cxx_contention_t __old_value) { |
168 | auto const __entry = __libcpp_contention_state(p: __location); |
169 | __libcpp_contention_wait(&__entry->__contention_state, &__entry->__platform_state, __old_value); |
170 | } |
171 | |
172 | /* When the incoming atomic happens to be the platform wait size, we still need to use the |
173 | table for the contention detection, but we can use the atomic directly for the wait. */ |
174 | |
175 | _LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_one(__cxx_atomic_contention_t const volatile* __location) { |
176 | __libcpp_contention_notify(&__libcpp_contention_state(__location)->__contention_state, __location, true); |
177 | } |
178 | _LIBCPP_EXPORTED_FROM_ABI void __cxx_atomic_notify_all(__cxx_atomic_contention_t const volatile* __location) { |
179 | __libcpp_contention_notify(&__libcpp_contention_state(__location)->__contention_state, __location, false); |
180 | } |
181 | // This function is never used, but still exported for ABI compatibility. |
182 | _LIBCPP_EXPORTED_FROM_ABI __cxx_contention_t |
183 | __libcpp_atomic_monitor(__cxx_atomic_contention_t const volatile* __location) { |
184 | return __libcpp_contention_monitor_for_wait(&__libcpp_contention_state(__location)->__contention_state, __location); |
185 | } |
186 | _LIBCPP_EXPORTED_FROM_ABI void |
187 | __libcpp_atomic_wait(__cxx_atomic_contention_t const volatile* __location, __cxx_contention_t __old_value) { |
188 | __libcpp_contention_wait(&__libcpp_contention_state(__location)->__contention_state, __location, __old_value); |
189 | } |
190 | |
191 | _LIBCPP_END_NAMESPACE_STD |
192 | |