1 /* 2 * Copyright (C) 2015 The Android Open Source Project 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * * Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * * Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in 12 * the documentation and/or other materials provided with the 13 * distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 16 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 17 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 18 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 19 * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 20 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, 21 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS 22 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 23 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 24 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 25 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 */ 28 29 #include <pthread.h> 30 #include <stdatomic.h> 31 #include <stdint.h> 32 33 #include "private/bionic_futex.h" 34 35 int pthread_barrierattr_init(pthread_barrierattr_t* attr) { 36 *attr = 0; 37 return 0; 38 } 39 40 int pthread_barrierattr_destroy(pthread_barrierattr_t* attr) { 41 *attr = 0; 42 return 0; 43 } 44 45 int pthread_barrierattr_getpshared(const pthread_barrierattr_t* attr, int* pshared) { 46 *pshared = (*attr & 1) ? PTHREAD_PROCESS_SHARED : PTHREAD_PROCESS_PRIVATE; 47 return 0; 48 } 49 50 int pthread_barrierattr_setpshared(pthread_barrierattr_t* attr, int pshared) { 51 if (pshared == PTHREAD_PROCESS_SHARED) { 52 *attr |= 1; 53 } else { 54 *attr &= ~1; 55 } 56 return 0; 57 } 58 59 enum BarrierState { 60 WAIT, 61 RELEASE, 62 }; 63 64 struct pthread_barrier_internal_t { 65 // One barrier can be used for unlimited number of cycles. In each cycle, [init_count] 66 // threads must call pthread_barrier_wait() before any of them successfully return from 67 // the call. It is undefined behavior if there are more than [init_count] threads call 68 // pthread_barrier_wait() in one cycle. 69 uint32_t init_count; 70 // Barrier state. It is WAIT if waiting for more threads to enter the barrier in this cycle, 71 // otherwise threads are leaving the barrier. 72 _Atomic(BarrierState) state; 73 // Number of threads having entered but not left the barrier in this cycle. 74 atomic_uint wait_count; 75 // Whether the barrier is shared across processes. 76 bool pshared; 77 uint32_t __reserved[4]; 78 }; 79 80 static_assert(sizeof(pthread_barrier_t) == sizeof(pthread_barrier_internal_t), 81 "pthread_barrier_t should actually be pthread_barrier_internal_t in implementation." 82 ); 83 84 static_assert(alignof(pthread_barrier_t) >= 4, 85 "pthread_barrier_t should fulfill the alignment of pthread_barrier_internal_t."); 86 87 static inline pthread_barrier_internal_t* __get_internal_barrier(pthread_barrier_t* barrier) { 88 return reinterpret_cast<pthread_barrier_internal_t*>(barrier); 89 } 90 91 int pthread_barrier_init(pthread_barrier_t* barrier_interface, const pthread_barrierattr_t* attr, 92 unsigned count) { 93 pthread_barrier_internal_t* barrier = __get_internal_barrier(barrier_interface); 94 if (count == 0) { 95 return EINVAL; 96 } 97 barrier->init_count = count; 98 atomic_init(&barrier->state, WAIT); 99 atomic_init(&barrier->wait_count, 0); 100 barrier->pshared = false; 101 if (attr != nullptr && (*attr & 1)) { 102 barrier->pshared = true; 103 } 104 return 0; 105 } 106 107 // According to POSIX standard, pthread_barrier_wait() synchronizes memory between participating 108 // threads. It means all memory operations made by participating threads before calling 109 // pthread_barrier_wait() can be seen by all participating threads after the function call. 110 // We establish this by making a happens-before relation between all threads entering the barrier 111 // with the last thread entering the barrier, and a happens-before relation between the last 112 // thread entering the barrier with all threads leaving the barrier. 113 int pthread_barrier_wait(pthread_barrier_t* barrier_interface) { 114 pthread_barrier_internal_t* barrier = __get_internal_barrier(barrier_interface); 115 116 // Wait until all threads for the previous cycle have left the barrier. This is needed 117 // as a participating thread can call pthread_barrier_wait() again before other 118 // threads have left the barrier. Use acquire operation here to synchronize with 119 // the last thread leaving the previous cycle, so we can read correct wait_count below. 120 while(atomic_load_explicit(&barrier->state, memory_order_acquire) == RELEASE) { 121 __futex_wait_ex(&barrier->state, barrier->pshared, RELEASE, false, nullptr); 122 } 123 124 uint32_t prev_wait_count = atomic_load_explicit(&barrier->wait_count, memory_order_relaxed); 125 while (true) { 126 // It happens when there are more than [init_count] threads trying to enter the barrier 127 // at one cycle. We read the POSIX standard as disallowing this, since additional arriving 128 // threads are not synchronized with respect to the barrier reset. We also don't know of 129 // any reasonable cases in which this would be intentional. 130 if (prev_wait_count >= barrier->init_count) { 131 return EINVAL; 132 } 133 // Use memory_order_acq_rel operation here to synchronize between all threads entering 134 // the barrier with the last thread entering the barrier. 135 if (atomic_compare_exchange_weak_explicit(&barrier->wait_count, &prev_wait_count, 136 prev_wait_count + 1u, memory_order_acq_rel, 137 memory_order_relaxed)) { 138 break; 139 } 140 } 141 142 int result = 0; 143 if (prev_wait_count + 1 == barrier->init_count) { 144 result = PTHREAD_BARRIER_SERIAL_THREAD; 145 if (prev_wait_count != 0) { 146 // Use release operation here to synchronize between the last thread entering the 147 // barrier with all threads leaving the barrier. 148 atomic_store_explicit(&barrier->state, RELEASE, memory_order_release); 149 __futex_wake_ex(&barrier->state, barrier->pshared, prev_wait_count); 150 } 151 } else { 152 // Use acquire operation here to synchronize between the last thread entering the 153 // barrier with all threads leaving the barrier. 154 while (atomic_load_explicit(&barrier->state, memory_order_acquire) == WAIT) { 155 __futex_wait_ex(&barrier->state, barrier->pshared, WAIT, false, nullptr); 156 } 157 } 158 // Use release operation here to make it not reordered with previous operations. 159 if (atomic_fetch_sub_explicit(&barrier->wait_count, 1, memory_order_release) == 1) { 160 // Use release operation here to synchronize with threads entering the barrier for 161 // the next cycle, or the thread calling pthread_barrier_destroy(). 162 atomic_store_explicit(&barrier->state, WAIT, memory_order_release); 163 __futex_wake_ex(&barrier->state, barrier->pshared, barrier->init_count); 164 } 165 return result; 166 } 167 168 int pthread_barrier_destroy(pthread_barrier_t* barrier_interface) { 169 pthread_barrier_internal_t* barrier = __get_internal_barrier(barrier_interface); 170 if (barrier->init_count == 0) { 171 return EINVAL; 172 } 173 // Use acquire operation here to synchronize with the last thread leaving the barrier. 174 // So we can read correct wait_count below. 175 while (atomic_load_explicit(&barrier->state, memory_order_acquire) == RELEASE) { 176 __futex_wait_ex(&barrier->state, barrier->pshared, RELEASE, false, nullptr); 177 } 178 if (atomic_load_explicit(&barrier->wait_count, memory_order_relaxed) != 0) { 179 return EBUSY; 180 } 181 barrier->init_count = 0; 182 return 0; 183 } 184