1 /* 2 * Copyright (c) 2015, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #ifndef SHARE_VM_GC_SHARED_TASKQUEUE_INLINE_HPP 26 #define SHARE_VM_GC_SHARED_TASKQUEUE_INLINE_HPP 27 28 #include "gc/shared/taskqueue.hpp" 29 #include "memory/allocation.inline.hpp" 30 #include "oops/oop.inline.hpp" 31 #include "runtime/atomic.inline.hpp" 32 #include "runtime/orderAccess.inline.hpp" 33 #include "utilities/debug.hpp" 34 #include "utilities/stack.inline.hpp" 35 36 template <class T, MEMFLAGS F> 37 inline GenericTaskQueueSet<T, F>::GenericTaskQueueSet(int n) : _n(n) { 38 typedef T* GenericTaskQueuePtr; 39 _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n, F); 40 for (int i = 0; i < n; i++) { 41 _queues[i] = NULL; 42 } 43 } 44 45 template<class E, MEMFLAGS F, unsigned int N> 46 inline void GenericTaskQueue<E, F, N>::initialize() { 47 _elems = ArrayAllocator<E, F>::allocate(N); 48 } 49 50 template<class E, MEMFLAGS F, unsigned int N> 51 inline GenericTaskQueue<E, F, N>::~GenericTaskQueue() { 52 assert(false, "This code is currently never called"); 53 ArrayAllocator<E, F>::free(const_cast<E*>(_elems), N); 54 } 55 56 template<class E, MEMFLAGS F, unsigned int N> 57 bool GenericTaskQueue<E, F, N>::push_slow(E t, uint dirty_n_elems) { 58 if (dirty_n_elems == N - 1) { 59 // Actually means 0, so do the push. 60 uint localBot = _bottom; 61 // g++ complains if the volatile result of the assignment is 62 // unused, so we cast the volatile away. We cannot cast directly 63 // to void, because gcc treats that as not using the result of the 64 // assignment. However, casting to E& means that we trigger an 65 // unused-value warning. So, we cast the E& to void. 66 (void)const_cast<E&>(_elems[localBot] = t); 67 OrderAccess::release_store(&_bottom, increment_index(localBot)); 68 TASKQUEUE_STATS_ONLY(stats.record_push()); 69 return true; 70 } 71 return false; 72 } 73 74 template<class E, MEMFLAGS F, unsigned int N> inline bool 75 GenericTaskQueue<E, F, N>::push(E t) { 76 uint localBot = _bottom; 77 assert(localBot < N, "_bottom out of range."); 78 idx_t top = _age.top(); 79 uint dirty_n_elems = dirty_size(localBot, top); 80 assert(dirty_n_elems < N, "n_elems out of range."); 81 if (dirty_n_elems < max_elems()) { 82 // g++ complains if the volatile result of the assignment is 83 // unused, so we cast the volatile away. We cannot cast directly 84 // to void, because gcc treats that as not using the result of the 85 // assignment. However, casting to E& means that we trigger an 86 // unused-value warning. So, we cast the E& to void. 87 (void) const_cast<E&>(_elems[localBot] = t); 88 OrderAccess::release_store(&_bottom, increment_index(localBot)); 89 TASKQUEUE_STATS_ONLY(stats.record_push()); 90 return true; 91 } else { 92 return push_slow(t, dirty_n_elems); 93 } 94 } 95 96 template <class E, MEMFLAGS F, unsigned int N> 97 inline bool OverflowTaskQueue<E, F, N>::push(E t) 98 { 99 if (!taskqueue_t::push(t)) { 100 overflow_stack()->push(t); 101 TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size())); 102 } 103 return true; 104 } 105 106 // pop_local_slow() is done by the owning thread and is trying to 107 // get the last task in the queue. It will compete with pop_global() 108 // that will be used by other threads. The tag age is incremented 109 // whenever the queue goes empty which it will do here if this thread 110 // gets the last task or in pop_global() if the queue wraps (top == 0 111 // and pop_global() succeeds, see pop_global()). 112 template<class E, MEMFLAGS F, unsigned int N> 113 bool GenericTaskQueue<E, F, N>::pop_local_slow(uint localBot, Age oldAge) { 114 // This queue was observed to contain exactly one element; either this 115 // thread will claim it, or a competing "pop_global". In either case, 116 // the queue will be logically empty afterwards. Create a new Age value 117 // that represents the empty queue for the given value of "_bottom". (We 118 // must also increment "tag" because of the case where "bottom == 1", 119 // "top == 0". A pop_global could read the queue element in that case, 120 // then have the owner thread do a pop followed by another push. Without 121 // the incrementing of "tag", the pop_global's CAS could succeed, 122 // allowing it to believe it has claimed the stale element.) 123 Age newAge((idx_t)localBot, oldAge.tag() + 1); 124 // Perhaps a competing pop_global has already incremented "top", in which 125 // case it wins the element. 126 if (localBot == oldAge.top()) { 127 // No competing pop_global has yet incremented "top"; we'll try to 128 // install new_age, thus claiming the element. 129 Age tempAge = _age.cmpxchg(newAge, oldAge); 130 if (tempAge == oldAge) { 131 // We win. 132 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity"); 133 TASKQUEUE_STATS_ONLY(stats.record_pop_slow()); 134 return true; 135 } 136 } 137 // We lose; a completing pop_global gets the element. But the queue is empty 138 // and top is greater than bottom. Fix this representation of the empty queue 139 // to become the canonical one. 140 _age.set(newAge); 141 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity"); 142 return false; 143 } 144 145 template<class E, MEMFLAGS F, unsigned int N> inline bool 146 GenericTaskQueue<E, F, N>::pop_local(volatile E& t) { 147 uint localBot = _bottom; 148 // This value cannot be N-1. That can only occur as a result of 149 // the assignment to bottom in this method. If it does, this method 150 // resets the size to 0 before the next call (which is sequential, 151 // since this is pop_local.) 152 uint dirty_n_elems = dirty_size(localBot, _age.top()); 153 assert(dirty_n_elems != N - 1, "Shouldn't be possible..."); 154 if (dirty_n_elems == 0) return false; 155 localBot = decrement_index(localBot); 156 _bottom = localBot; 157 // This is necessary to prevent any read below from being reordered 158 // before the store just above. 159 OrderAccess::fence(); 160 // g++ complains if the volatile result of the assignment is 161 // unused, so we cast the volatile away. We cannot cast directly 162 // to void, because gcc treats that as not using the result of the 163 // assignment. However, casting to E& means that we trigger an 164 // unused-value warning. So, we cast the E& to void. 165 (void) const_cast<E&>(t = _elems[localBot]); 166 // This is a second read of "age"; the "size()" above is the first. 167 // If there's still at least one element in the queue, based on the 168 // "_bottom" and "age" we've read, then there can be no interference with 169 // a "pop_global" operation, and we're done. 170 idx_t tp = _age.top(); // XXX 171 if (size(localBot, tp) > 0) { 172 assert(dirty_size(localBot, tp) != N - 1, "sanity"); 173 TASKQUEUE_STATS_ONLY(stats.record_pop()); 174 return true; 175 } else { 176 // Otherwise, the queue contained exactly one element; we take the slow 177 // path. 178 return pop_local_slow(localBot, _age.get()); 179 } 180 } 181 182 template <class E, MEMFLAGS F, unsigned int N> 183 bool OverflowTaskQueue<E, F, N>::pop_overflow(E& t) 184 { 185 if (overflow_empty()) return false; 186 t = overflow_stack()->pop(); 187 return true; 188 } 189 190 template<class E, MEMFLAGS F, unsigned int N> 191 bool GenericTaskQueue<E, F, N>::pop_global(volatile E& t) { 192 Age oldAge = _age.get(); 193 // Architectures with weak memory model require a barrier here 194 // to guarantee that bottom is not older than age, 195 // which is crucial for the correctness of the algorithm. 196 #if !(defined SPARC || defined IA32 || defined AMD64) 197 OrderAccess::fence(); 198 #endif 199 uint localBot = OrderAccess::load_acquire((volatile juint*)&_bottom); 200 uint n_elems = size(localBot, oldAge.top()); 201 if (n_elems == 0) { 202 return false; 203 } 204 205 // g++ complains if the volatile result of the assignment is 206 // unused, so we cast the volatile away. We cannot cast directly 207 // to void, because gcc treats that as not using the result of the 208 // assignment. However, casting to E& means that we trigger an 209 // unused-value warning. So, we cast the E& to void. 210 (void) const_cast<E&>(t = _elems[oldAge.top()]); 211 Age newAge(oldAge); 212 newAge.increment(); 213 Age resAge = _age.cmpxchg(newAge, oldAge); 214 215 // Note that using "_bottom" here might fail, since a pop_local might 216 // have decremented it. 217 assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity"); 218 return resAge == oldAge; 219 } 220 221 template<class T, MEMFLAGS F> bool 222 GenericTaskQueueSet<T, F>::steal_best_of_2(uint queue_num, int* seed, E& t) { 223 if (_n > 2) { 224 uint k1 = queue_num; 225 while (k1 == queue_num) k1 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n; 226 uint k2 = queue_num; 227 while (k2 == queue_num || k2 == k1) k2 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n; 228 // Sample both and try the larger. 229 uint sz1 = _queues[k1]->size(); 230 uint sz2 = _queues[k2]->size(); 231 if (sz2 > sz1) return _queues[k2]->pop_global(t); 232 else return _queues[k1]->pop_global(t); 233 } else if (_n == 2) { 234 // Just try the other one. 235 uint k = (queue_num + 1) % 2; 236 return _queues[k]->pop_global(t); 237 } else { 238 assert(_n == 1, "can't be zero."); 239 return false; 240 } 241 } 242 243 template<class T, MEMFLAGS F> bool 244 GenericTaskQueueSet<T, F>::steal(uint queue_num, int* seed, E& t) { 245 for (uint i = 0; i < 2 * _n; i++) { 246 if (steal_best_of_2(queue_num, seed, t)) { 247 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true)); 248 return true; 249 } 250 } 251 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(false)); 252 return false; 253 } 254 255 template <unsigned int N, MEMFLAGS F> 256 inline typename TaskQueueSuper<N, F>::Age TaskQueueSuper<N, F>::Age::cmpxchg(const Age new_age, const Age old_age) volatile { 257 return (size_t) Atomic::cmpxchg_ptr((intptr_t)new_age._data, 258 (volatile intptr_t *)&_data, 259 (intptr_t)old_age._data); 260 } 261 262 template<class E, MEMFLAGS F, unsigned int N> 263 template<class Fn> 264 inline void GenericTaskQueue<E, F, N>::iterate(Fn fn) { 265 uint iters = size(); 266 uint index = _bottom; 267 for (uint i = 0; i < iters; ++i) { 268 index = decrement_index(index); 269 fn(const_cast<E&>(_elems[index])); // cast away volatility 270 } 271 } 272 273 274 #endif // SHARE_VM_GC_SHARED_TASKQUEUE_INLINE_HPP