Thu, 26 Sep 2013 10:25:02 -0400
7195622: CheckUnhandledOops has limited usefulness now
Summary: Enable CHECK_UNHANDLED_OOPS in fastdebug builds across all supported platforms.
Reviewed-by: coleenp, hseigel, dholmes, stefank, twisti, ihse, rdurbin
Contributed-by: lois.foltan@oracle.com
duke@435 | 1 | /* |
sla@5237 | 2 | * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved. |
duke@435 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
duke@435 | 4 | * |
duke@435 | 5 | * This code is free software; you can redistribute it and/or modify it |
duke@435 | 6 | * under the terms of the GNU General Public License version 2 only, as |
duke@435 | 7 | * published by the Free Software Foundation. |
duke@435 | 8 | * |
duke@435 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
duke@435 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
duke@435 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
duke@435 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
duke@435 | 13 | * accompanied this code). |
duke@435 | 14 | * |
duke@435 | 15 | * You should have received a copy of the GNU General Public License version |
duke@435 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
duke@435 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
duke@435 | 18 | * |
trims@1907 | 19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
trims@1907 | 20 | * or visit www.oracle.com if you need additional information or have any |
trims@1907 | 21 | * questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
stefank@2314 | 25 | #include "precompiled.hpp" |
stefank@2314 | 26 | #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp" |
stefank@2314 | 27 | #include "gc_implementation/parNew/parNewGeneration.hpp" |
stefank@2314 | 28 | #include "gc_implementation/parNew/parOopClosures.inline.hpp" |
stefank@2314 | 29 | #include "gc_implementation/shared/adaptiveSizePolicy.hpp" |
stefank@2314 | 30 | #include "gc_implementation/shared/ageTable.hpp" |
johnc@3982 | 31 | #include "gc_implementation/shared/parGCAllocBuffer.hpp" |
sla@5237 | 32 | #include "gc_implementation/shared/gcHeapSummary.hpp" |
sla@5237 | 33 | #include "gc_implementation/shared/gcTimer.hpp" |
sla@5237 | 34 | #include "gc_implementation/shared/gcTrace.hpp" |
sla@5237 | 35 | #include "gc_implementation/shared/gcTraceTime.hpp" |
sla@5237 | 36 | #include "gc_implementation/shared/copyFailedInfo.hpp" |
stefank@2314 | 37 | #include "gc_implementation/shared/spaceDecorator.hpp" |
stefank@2314 | 38 | #include "memory/defNewGeneration.inline.hpp" |
stefank@2314 | 39 | #include "memory/genCollectedHeap.hpp" |
stefank@2314 | 40 | #include "memory/genOopClosures.inline.hpp" |
stefank@2314 | 41 | #include "memory/generation.hpp" |
stefank@2314 | 42 | #include "memory/generation.inline.hpp" |
stefank@2314 | 43 | #include "memory/referencePolicy.hpp" |
stefank@2314 | 44 | #include "memory/resourceArea.hpp" |
stefank@2314 | 45 | #include "memory/sharedHeap.hpp" |
stefank@2314 | 46 | #include "memory/space.hpp" |
stefank@2314 | 47 | #include "oops/objArrayOop.hpp" |
stefank@2314 | 48 | #include "oops/oop.inline.hpp" |
stefank@2314 | 49 | #include "oops/oop.pcgc.inline.hpp" |
stefank@2314 | 50 | #include "runtime/handles.hpp" |
stefank@2314 | 51 | #include "runtime/handles.inline.hpp" |
stefank@2314 | 52 | #include "runtime/java.hpp" |
stefank@2314 | 53 | #include "runtime/thread.hpp" |
stefank@2314 | 54 | #include "utilities/copy.hpp" |
stefank@2314 | 55 | #include "utilities/globalDefinitions.hpp" |
stefank@2314 | 56 | #include "utilities/workgroup.hpp" |
duke@435 | 57 | |
duke@435 | 58 | #ifdef _MSC_VER |
duke@435 | 59 | #pragma warning( push ) |
duke@435 | 60 | #pragma warning( disable:4355 ) // 'this' : used in base member initializer list |
duke@435 | 61 | #endif |
duke@435 | 62 | ParScanThreadState::ParScanThreadState(Space* to_space_, |
duke@435 | 63 | ParNewGeneration* gen_, |
duke@435 | 64 | Generation* old_gen_, |
duke@435 | 65 | int thread_num_, |
duke@435 | 66 | ObjToScanQueueSet* work_queue_set_, |
zgu@3900 | 67 | Stack<oop, mtGC>* overflow_stacks_, |
duke@435 | 68 | size_t desired_plab_sz_, |
duke@435 | 69 | ParallelTaskTerminator& term_) : |
ysr@1114 | 70 | _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_), |
duke@435 | 71 | _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false), |
jcoomes@2191 | 72 | _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL), |
duke@435 | 73 | _ageTable(false), // false ==> not the global age table, no perf data. |
duke@435 | 74 | _to_space_alloc_buffer(desired_plab_sz_), |
duke@435 | 75 | _to_space_closure(gen_, this), _old_gen_closure(gen_, this), |
duke@435 | 76 | _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this), |
duke@435 | 77 | _older_gen_closure(gen_, this), |
duke@435 | 78 | _evacuate_followers(this, &_to_space_closure, &_old_gen_closure, |
duke@435 | 79 | &_to_space_root_closure, gen_, &_old_gen_root_closure, |
duke@435 | 80 | work_queue_set_, &term_), |
duke@435 | 81 | _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this), |
duke@435 | 82 | _keep_alive_closure(&_scan_weak_ref_closure), |
duke@435 | 83 | _strong_roots_time(0.0), _term_time(0.0) |
duke@435 | 84 | { |
jcoomes@2065 | 85 | #if TASKQUEUE_STATS |
jcoomes@2065 | 86 | _term_attempts = 0; |
jcoomes@2065 | 87 | _overflow_refills = 0; |
jcoomes@2065 | 88 | _overflow_refill_objs = 0; |
jcoomes@2065 | 89 | #endif // TASKQUEUE_STATS |
jcoomes@2065 | 90 | |
duke@435 | 91 | _survivor_chunk_array = |
duke@435 | 92 | (ChunkArray*) old_gen()->get_data_recorder(thread_num()); |
duke@435 | 93 | _hash_seed = 17; // Might want to take time-based random value. |
duke@435 | 94 | _start = os::elapsedTime(); |
duke@435 | 95 | _old_gen_closure.set_generation(old_gen_); |
duke@435 | 96 | _old_gen_root_closure.set_generation(old_gen_); |
duke@435 | 97 | } |
duke@435 | 98 | #ifdef _MSC_VER |
duke@435 | 99 | #pragma warning( pop ) |
duke@435 | 100 | #endif |
duke@435 | 101 | |
duke@435 | 102 | void ParScanThreadState::record_survivor_plab(HeapWord* plab_start, |
duke@435 | 103 | size_t plab_word_size) { |
duke@435 | 104 | ChunkArray* sca = survivor_chunk_array(); |
duke@435 | 105 | if (sca != NULL) { |
duke@435 | 106 | // A non-null SCA implies that we want the PLAB data recorded. |
duke@435 | 107 | sca->record_sample(plab_start, plab_word_size); |
duke@435 | 108 | } |
duke@435 | 109 | } |
duke@435 | 110 | |
duke@435 | 111 | bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const { |
duke@435 | 112 | return new_obj->is_objArray() && |
duke@435 | 113 | arrayOop(new_obj)->length() > ParGCArrayScanChunk && |
duke@435 | 114 | new_obj != old_obj; |
duke@435 | 115 | } |
duke@435 | 116 | |
duke@435 | 117 | void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) { |
duke@435 | 118 | assert(old->is_objArray(), "must be obj array"); |
duke@435 | 119 | assert(old->is_forwarded(), "must be forwarded"); |
duke@435 | 120 | assert(Universe::heap()->is_in_reserved(old), "must be in heap."); |
ysr@1114 | 121 | assert(!old_gen()->is_in(old), "must be in young generation."); |
duke@435 | 122 | |
duke@435 | 123 | objArrayOop obj = objArrayOop(old->forwardee()); |
duke@435 | 124 | // Process ParGCArrayScanChunk elements now |
duke@435 | 125 | // and push the remainder back onto queue |
duke@435 | 126 | int start = arrayOop(old)->length(); |
duke@435 | 127 | int end = obj->length(); |
duke@435 | 128 | int remainder = end - start; |
duke@435 | 129 | assert(start <= end, "just checking"); |
duke@435 | 130 | if (remainder > 2 * ParGCArrayScanChunk) { |
duke@435 | 131 | // Test above combines last partial chunk with a full chunk |
duke@435 | 132 | end = start + ParGCArrayScanChunk; |
duke@435 | 133 | arrayOop(old)->set_length(end); |
duke@435 | 134 | // Push remainder. |
duke@435 | 135 | bool ok = work_queue()->push(old); |
duke@435 | 136 | assert(ok, "just popped, push must be okay"); |
duke@435 | 137 | } else { |
duke@435 | 138 | // Restore length so that it can be used if there |
duke@435 | 139 | // is a promotion failure and forwarding pointers |
duke@435 | 140 | // must be removed. |
duke@435 | 141 | arrayOop(old)->set_length(end); |
duke@435 | 142 | } |
coleenp@548 | 143 | |
duke@435 | 144 | // process our set of indices (include header in first chunk) |
coleenp@548 | 145 | // should make sure end is even (aligned to HeapWord in case of compressed oops) |
duke@435 | 146 | if ((HeapWord *)obj < young_old_boundary()) { |
duke@435 | 147 | // object is in to_space |
coleenp@548 | 148 | obj->oop_iterate_range(&_to_space_closure, start, end); |
duke@435 | 149 | } else { |
duke@435 | 150 | // object is in old generation |
coleenp@548 | 151 | obj->oop_iterate_range(&_old_gen_closure, start, end); |
duke@435 | 152 | } |
duke@435 | 153 | } |
duke@435 | 154 | |
duke@435 | 155 | |
duke@435 | 156 | void ParScanThreadState::trim_queues(int max_size) { |
duke@435 | 157 | ObjToScanQueue* queue = work_queue(); |
ysr@1114 | 158 | do { |
ysr@1114 | 159 | while (queue->size() > (juint)max_size) { |
ysr@1114 | 160 | oop obj_to_scan; |
ysr@1114 | 161 | if (queue->pop_local(obj_to_scan)) { |
ysr@1114 | 162 | if ((HeapWord *)obj_to_scan < young_old_boundary()) { |
ysr@1114 | 163 | if (obj_to_scan->is_objArray() && |
ysr@1114 | 164 | obj_to_scan->is_forwarded() && |
ysr@1114 | 165 | obj_to_scan->forwardee() != obj_to_scan) { |
ysr@1114 | 166 | scan_partial_array_and_push_remainder(obj_to_scan); |
ysr@1114 | 167 | } else { |
ysr@1114 | 168 | // object is in to_space |
ysr@1114 | 169 | obj_to_scan->oop_iterate(&_to_space_closure); |
ysr@1114 | 170 | } |
duke@435 | 171 | } else { |
ysr@1114 | 172 | // object is in old generation |
ysr@1114 | 173 | obj_to_scan->oop_iterate(&_old_gen_closure); |
duke@435 | 174 | } |
duke@435 | 175 | } |
duke@435 | 176 | } |
ysr@1114 | 177 | // For the case of compressed oops, we have a private, non-shared |
ysr@1114 | 178 | // overflow stack, so we eagerly drain it so as to more evenly |
ysr@1114 | 179 | // distribute load early. Note: this may be good to do in |
ysr@1114 | 180 | // general rather than delay for the final stealing phase. |
ysr@1114 | 181 | // If applicable, we'll transfer a set of objects over to our |
ysr@1114 | 182 | // work queue, allowing them to be stolen and draining our |
ysr@1114 | 183 | // private overflow stack. |
ysr@1114 | 184 | } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this)); |
ysr@1114 | 185 | } |
ysr@1114 | 186 | |
ysr@1114 | 187 | bool ParScanThreadState::take_from_overflow_stack() { |
ysr@1130 | 188 | assert(ParGCUseLocalOverflow, "Else should not call"); |
ysr@1114 | 189 | assert(young_gen()->overflow_list() == NULL, "Error"); |
ysr@1114 | 190 | ObjToScanQueue* queue = work_queue(); |
zgu@3900 | 191 | Stack<oop, mtGC>* const of_stack = overflow_stack(); |
jcoomes@2191 | 192 | const size_t num_overflow_elems = of_stack->size(); |
jcoomes@2191 | 193 | const size_t space_available = queue->max_elems() - queue->size(); |
jcoomes@2191 | 194 | const size_t num_take_elems = MIN3(space_available / 4, |
jcoomes@2191 | 195 | ParGCDesiredObjsFromOverflowList, |
jcoomes@2191 | 196 | num_overflow_elems); |
ysr@1114 | 197 | // Transfer the most recent num_take_elems from the overflow |
ysr@1114 | 198 | // stack to our work queue. |
ysr@1114 | 199 | for (size_t i = 0; i != num_take_elems; i++) { |
ysr@1114 | 200 | oop cur = of_stack->pop(); |
ysr@1114 | 201 | oop obj_to_push = cur->forwardee(); |
ysr@1114 | 202 | assert(Universe::heap()->is_in_reserved(cur), "Should be in heap"); |
ysr@1114 | 203 | assert(!old_gen()->is_in_reserved(cur), "Should be in young gen"); |
ysr@1114 | 204 | assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap"); |
ysr@1114 | 205 | if (should_be_partially_scanned(obj_to_push, cur)) { |
ysr@1114 | 206 | assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); |
ysr@1114 | 207 | obj_to_push = cur; |
ysr@1114 | 208 | } |
ysr@1114 | 209 | bool ok = queue->push(obj_to_push); |
ysr@1114 | 210 | assert(ok, "Should have succeeded"); |
duke@435 | 211 | } |
ysr@1114 | 212 | assert(young_gen()->overflow_list() == NULL, "Error"); |
ysr@1114 | 213 | return num_take_elems > 0; // was something transferred? |
ysr@1114 | 214 | } |
ysr@1114 | 215 | |
ysr@1114 | 216 | void ParScanThreadState::push_on_overflow_stack(oop p) { |
ysr@1130 | 217 | assert(ParGCUseLocalOverflow, "Else should not call"); |
ysr@1114 | 218 | overflow_stack()->push(p); |
ysr@1114 | 219 | assert(young_gen()->overflow_list() == NULL, "Error"); |
duke@435 | 220 | } |
duke@435 | 221 | |
duke@435 | 222 | HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) { |
duke@435 | 223 | |
duke@435 | 224 | // Otherwise, if the object is small enough, try to reallocate the |
duke@435 | 225 | // buffer. |
duke@435 | 226 | HeapWord* obj = NULL; |
duke@435 | 227 | if (!_to_space_full) { |
duke@435 | 228 | ParGCAllocBuffer* const plab = to_space_alloc_buffer(); |
duke@435 | 229 | Space* const sp = to_space(); |
duke@435 | 230 | if (word_sz * 100 < |
duke@435 | 231 | ParallelGCBufferWastePct * plab->word_sz()) { |
duke@435 | 232 | // Is small enough; abandon this buffer and start a new one. |
duke@435 | 233 | plab->retire(false, false); |
duke@435 | 234 | size_t buf_size = plab->word_sz(); |
duke@435 | 235 | HeapWord* buf_space = sp->par_allocate(buf_size); |
duke@435 | 236 | if (buf_space == NULL) { |
duke@435 | 237 | const size_t min_bytes = |
duke@435 | 238 | ParGCAllocBuffer::min_size() << LogHeapWordSize; |
duke@435 | 239 | size_t free_bytes = sp->free(); |
duke@435 | 240 | while(buf_space == NULL && free_bytes >= min_bytes) { |
duke@435 | 241 | buf_size = free_bytes >> LogHeapWordSize; |
duke@435 | 242 | assert(buf_size == (size_t)align_object_size(buf_size), |
duke@435 | 243 | "Invariant"); |
duke@435 | 244 | buf_space = sp->par_allocate(buf_size); |
duke@435 | 245 | free_bytes = sp->free(); |
duke@435 | 246 | } |
duke@435 | 247 | } |
duke@435 | 248 | if (buf_space != NULL) { |
duke@435 | 249 | plab->set_word_size(buf_size); |
duke@435 | 250 | plab->set_buf(buf_space); |
duke@435 | 251 | record_survivor_plab(buf_space, buf_size); |
duke@435 | 252 | obj = plab->allocate(word_sz); |
duke@435 | 253 | // Note that we cannot compare buf_size < word_sz below |
duke@435 | 254 | // because of AlignmentReserve (see ParGCAllocBuffer::allocate()). |
duke@435 | 255 | assert(obj != NULL || plab->words_remaining() < word_sz, |
duke@435 | 256 | "Else should have been able to allocate"); |
duke@435 | 257 | // It's conceivable that we may be able to use the |
duke@435 | 258 | // buffer we just grabbed for subsequent small requests |
duke@435 | 259 | // even if not for this one. |
duke@435 | 260 | } else { |
duke@435 | 261 | // We're used up. |
duke@435 | 262 | _to_space_full = true; |
duke@435 | 263 | } |
duke@435 | 264 | |
duke@435 | 265 | } else { |
duke@435 | 266 | // Too large; allocate the object individually. |
duke@435 | 267 | obj = sp->par_allocate(word_sz); |
duke@435 | 268 | } |
duke@435 | 269 | } |
duke@435 | 270 | return obj; |
duke@435 | 271 | } |
duke@435 | 272 | |
duke@435 | 273 | |
duke@435 | 274 | void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, |
duke@435 | 275 | size_t word_sz) { |
duke@435 | 276 | // Is the alloc in the current alloc buffer? |
duke@435 | 277 | if (to_space_alloc_buffer()->contains(obj)) { |
duke@435 | 278 | assert(to_space_alloc_buffer()->contains(obj + word_sz - 1), |
duke@435 | 279 | "Should contain whole object."); |
duke@435 | 280 | to_space_alloc_buffer()->undo_allocation(obj, word_sz); |
duke@435 | 281 | } else { |
jcoomes@916 | 282 | CollectedHeap::fill_with_object(obj, word_sz); |
duke@435 | 283 | } |
duke@435 | 284 | } |
duke@435 | 285 | |
sla@5237 | 286 | void ParScanThreadState::print_promotion_failure_size() { |
sla@5237 | 287 | if (_promotion_failed_info.has_failed() && PrintPromotionFailure) { |
sla@5237 | 288 | gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ", |
sla@5237 | 289 | _thread_num, _promotion_failed_info.first_size()); |
ysr@1580 | 290 | } |
ysr@1580 | 291 | } |
ysr@1580 | 292 | |
duke@435 | 293 | class ParScanThreadStateSet: private ResourceArray { |
duke@435 | 294 | public: |
duke@435 | 295 | // Initializes states for the specified number of threads; |
duke@435 | 296 | ParScanThreadStateSet(int num_threads, |
duke@435 | 297 | Space& to_space, |
duke@435 | 298 | ParNewGeneration& gen, |
duke@435 | 299 | Generation& old_gen, |
duke@435 | 300 | ObjToScanQueueSet& queue_set, |
zgu@3900 | 301 | Stack<oop, mtGC>* overflow_stacks_, |
duke@435 | 302 | size_t desired_plab_sz, |
duke@435 | 303 | ParallelTaskTerminator& term); |
jcoomes@2065 | 304 | |
jcoomes@2065 | 305 | ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); } |
jcoomes@2065 | 306 | |
ysr@1580 | 307 | inline ParScanThreadState& thread_state(int i); |
jcoomes@2065 | 308 | |
sla@5237 | 309 | void trace_promotion_failed(YoungGCTracer& gc_tracer); |
jmasa@3294 | 310 | void reset(int active_workers, bool promotion_failed); |
duke@435 | 311 | void flush(); |
jcoomes@2065 | 312 | |
jcoomes@2065 | 313 | #if TASKQUEUE_STATS |
jcoomes@2065 | 314 | static void |
jcoomes@2065 | 315 | print_termination_stats_hdr(outputStream* const st = gclog_or_tty); |
jcoomes@2065 | 316 | void print_termination_stats(outputStream* const st = gclog_or_tty); |
jcoomes@2065 | 317 | static void |
jcoomes@2065 | 318 | print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty); |
jcoomes@2065 | 319 | void print_taskqueue_stats(outputStream* const st = gclog_or_tty); |
jcoomes@2065 | 320 | void reset_stats(); |
jcoomes@2065 | 321 | #endif // TASKQUEUE_STATS |
jcoomes@2065 | 322 | |
duke@435 | 323 | private: |
duke@435 | 324 | ParallelTaskTerminator& _term; |
duke@435 | 325 | ParNewGeneration& _gen; |
duke@435 | 326 | Generation& _next_gen; |
jmasa@3294 | 327 | public: |
jmasa@3294 | 328 | bool is_valid(int id) const { return id < length(); } |
jmasa@3294 | 329 | ParallelTaskTerminator* terminator() { return &_term; } |
duke@435 | 330 | }; |
duke@435 | 331 | |
duke@435 | 332 | |
duke@435 | 333 | ParScanThreadStateSet::ParScanThreadStateSet( |
duke@435 | 334 | int num_threads, Space& to_space, ParNewGeneration& gen, |
duke@435 | 335 | Generation& old_gen, ObjToScanQueueSet& queue_set, |
zgu@3900 | 336 | Stack<oop, mtGC>* overflow_stacks, |
duke@435 | 337 | size_t desired_plab_sz, ParallelTaskTerminator& term) |
duke@435 | 338 | : ResourceArray(sizeof(ParScanThreadState), num_threads), |
jcoomes@2065 | 339 | _gen(gen), _next_gen(old_gen), _term(term) |
duke@435 | 340 | { |
duke@435 | 341 | assert(num_threads > 0, "sanity check!"); |
jcoomes@2191 | 342 | assert(ParGCUseLocalOverflow == (overflow_stacks != NULL), |
jcoomes@2191 | 343 | "overflow_stack allocation mismatch"); |
duke@435 | 344 | // Initialize states. |
duke@435 | 345 | for (int i = 0; i < num_threads; ++i) { |
duke@435 | 346 | new ((ParScanThreadState*)_data + i) |
duke@435 | 347 | ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set, |
jcoomes@2191 | 348 | overflow_stacks, desired_plab_sz, term); |
duke@435 | 349 | } |
duke@435 | 350 | } |
duke@435 | 351 | |
ysr@1580 | 352 | inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) |
duke@435 | 353 | { |
duke@435 | 354 | assert(i >= 0 && i < length(), "sanity check!"); |
duke@435 | 355 | return ((ParScanThreadState*)_data)[i]; |
duke@435 | 356 | } |
duke@435 | 357 | |
sla@5237 | 358 | void ParScanThreadStateSet::trace_promotion_failed(YoungGCTracer& gc_tracer) { |
sla@5237 | 359 | for (int i = 0; i < length(); ++i) { |
sla@5237 | 360 | if (thread_state(i).promotion_failed()) { |
sla@5237 | 361 | gc_tracer.report_promotion_failed(thread_state(i).promotion_failed_info()); |
sla@5237 | 362 | thread_state(i).promotion_failed_info().reset(); |
sla@5237 | 363 | } |
sla@5237 | 364 | } |
sla@5237 | 365 | } |
duke@435 | 366 | |
jmasa@3294 | 367 | void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed) |
duke@435 | 368 | { |
jmasa@3294 | 369 | _term.reset_for_reuse(active_threads); |
ysr@1580 | 370 | if (promotion_failed) { |
ysr@1580 | 371 | for (int i = 0; i < length(); ++i) { |
sla@5237 | 372 | thread_state(i).print_promotion_failure_size(); |
ysr@1580 | 373 | } |
ysr@1580 | 374 | } |
duke@435 | 375 | } |
duke@435 | 376 | |
jcoomes@2065 | 377 | #if TASKQUEUE_STATS |
jcoomes@2065 | 378 | void |
jcoomes@2065 | 379 | ParScanThreadState::reset_stats() |
jcoomes@2065 | 380 | { |
jcoomes@2065 | 381 | taskqueue_stats().reset(); |
jcoomes@2065 | 382 | _term_attempts = 0; |
jcoomes@2065 | 383 | _overflow_refills = 0; |
jcoomes@2065 | 384 | _overflow_refill_objs = 0; |
jcoomes@2065 | 385 | } |
jcoomes@2065 | 386 | |
jcoomes@2065 | 387 | void ParScanThreadStateSet::reset_stats() |
jcoomes@2065 | 388 | { |
jcoomes@2065 | 389 | for (int i = 0; i < length(); ++i) { |
jcoomes@2065 | 390 | thread_state(i).reset_stats(); |
jcoomes@2065 | 391 | } |
jcoomes@2065 | 392 | } |
jcoomes@2065 | 393 | |
jcoomes@2065 | 394 | void |
jcoomes@2065 | 395 | ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) |
jcoomes@2065 | 396 | { |
jcoomes@2065 | 397 | st->print_raw_cr("GC Termination Stats"); |
jcoomes@2065 | 398 | st->print_raw_cr(" elapsed --strong roots-- " |
jcoomes@2065 | 399 | "-------termination-------"); |
jcoomes@2065 | 400 | st->print_raw_cr("thr ms ms % " |
jcoomes@2065 | 401 | " ms % attempts"); |
jcoomes@2065 | 402 | st->print_raw_cr("--- --------- --------- ------ " |
jcoomes@2065 | 403 | "--------- ------ --------"); |
jcoomes@2065 | 404 | } |
jcoomes@2065 | 405 | |
jcoomes@2065 | 406 | void ParScanThreadStateSet::print_termination_stats(outputStream* const st) |
jcoomes@2065 | 407 | { |
jcoomes@2065 | 408 | print_termination_stats_hdr(st); |
jcoomes@2065 | 409 | |
jcoomes@2065 | 410 | for (int i = 0; i < length(); ++i) { |
jcoomes@2065 | 411 | const ParScanThreadState & pss = thread_state(i); |
jcoomes@2065 | 412 | const double elapsed_ms = pss.elapsed_time() * 1000.0; |
jcoomes@2065 | 413 | const double s_roots_ms = pss.strong_roots_time() * 1000.0; |
jcoomes@2065 | 414 | const double term_ms = pss.term_time() * 1000.0; |
jcoomes@2065 | 415 | st->print_cr("%3d %9.2f %9.2f %6.2f " |
jcoomes@2065 | 416 | "%9.2f %6.2f " SIZE_FORMAT_W(8), |
jcoomes@2065 | 417 | i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, |
jcoomes@2065 | 418 | term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts()); |
jcoomes@2065 | 419 | } |
jcoomes@2065 | 420 | } |
jcoomes@2065 | 421 | |
jcoomes@2065 | 422 | // Print stats related to work queue activity. |
jcoomes@2065 | 423 | void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) |
jcoomes@2065 | 424 | { |
jcoomes@2065 | 425 | st->print_raw_cr("GC Task Stats"); |
jcoomes@2065 | 426 | st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); |
jcoomes@2065 | 427 | st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); |
jcoomes@2065 | 428 | } |
jcoomes@2065 | 429 | |
jcoomes@2065 | 430 | void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st) |
jcoomes@2065 | 431 | { |
jcoomes@2065 | 432 | print_taskqueue_stats_hdr(st); |
jcoomes@2065 | 433 | |
jcoomes@2065 | 434 | TaskQueueStats totals; |
jcoomes@2065 | 435 | for (int i = 0; i < length(); ++i) { |
jcoomes@2065 | 436 | const ParScanThreadState & pss = thread_state(i); |
jcoomes@2065 | 437 | const TaskQueueStats & stats = pss.taskqueue_stats(); |
jcoomes@2065 | 438 | st->print("%3d ", i); stats.print(st); st->cr(); |
jcoomes@2065 | 439 | totals += stats; |
jcoomes@2065 | 440 | |
jcoomes@2065 | 441 | if (pss.overflow_refills() > 0) { |
jcoomes@2065 | 442 | st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills " |
jcoomes@2065 | 443 | SIZE_FORMAT_W(10) " overflow objects", |
jcoomes@2065 | 444 | pss.overflow_refills(), pss.overflow_refill_objs()); |
jcoomes@2065 | 445 | } |
jcoomes@2065 | 446 | } |
jcoomes@2065 | 447 | st->print("tot "); totals.print(st); st->cr(); |
jcoomes@2065 | 448 | |
jcoomes@2065 | 449 | DEBUG_ONLY(totals.verify()); |
jcoomes@2065 | 450 | } |
jcoomes@2065 | 451 | #endif // TASKQUEUE_STATS |
jcoomes@2065 | 452 | |
duke@435 | 453 | void ParScanThreadStateSet::flush() |
duke@435 | 454 | { |
ysr@1580 | 455 | // Work in this loop should be kept as lightweight as |
ysr@1580 | 456 | // possible since this might otherwise become a bottleneck |
ysr@1580 | 457 | // to scaling. Should we add heavy-weight work into this |
ysr@1580 | 458 | // loop, consider parallelizing the loop into the worker threads. |
duke@435 | 459 | for (int i = 0; i < length(); ++i) { |
ysr@1580 | 460 | ParScanThreadState& par_scan_state = thread_state(i); |
duke@435 | 461 | |
duke@435 | 462 | // Flush stats related to To-space PLAB activity and |
duke@435 | 463 | // retire the last buffer. |
duke@435 | 464 | par_scan_state.to_space_alloc_buffer()-> |
duke@435 | 465 | flush_stats_and_retire(_gen.plab_stats(), |
johnc@3982 | 466 | true /* end_of_gc */, |
johnc@3982 | 467 | false /* retain */); |
duke@435 | 468 | |
duke@435 | 469 | // Every thread has its own age table. We need to merge |
duke@435 | 470 | // them all into one. |
duke@435 | 471 | ageTable *local_table = par_scan_state.age_table(); |
duke@435 | 472 | _gen.age_table()->merge(local_table); |
duke@435 | 473 | |
duke@435 | 474 | // Inform old gen that we're done. |
duke@435 | 475 | _next_gen.par_promote_alloc_done(i); |
duke@435 | 476 | _next_gen.par_oop_since_save_marks_iterate_done(i); |
jcoomes@2065 | 477 | } |
duke@435 | 478 | |
ysr@1580 | 479 | if (UseConcMarkSweepGC && ParallelGCThreads > 0) { |
ysr@1580 | 480 | // We need to call this even when ResizeOldPLAB is disabled |
ysr@1580 | 481 | // so as to avoid breaking some asserts. While we may be able |
ysr@1580 | 482 | // to avoid this by reorganizing the code a bit, I am loathe |
ysr@1580 | 483 | // to do that unless we find cases where ergo leads to bad |
ysr@1580 | 484 | // performance. |
ysr@1580 | 485 | CFLS_LAB::compute_desired_plab_size(); |
ysr@1580 | 486 | } |
duke@435 | 487 | } |
duke@435 | 488 | |
duke@435 | 489 | ParScanClosure::ParScanClosure(ParNewGeneration* g, |
duke@435 | 490 | ParScanThreadState* par_scan_state) : |
coleenp@4037 | 491 | OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) |
duke@435 | 492 | { |
duke@435 | 493 | assert(_g->level() == 0, "Optimized for youngest generation"); |
duke@435 | 494 | _boundary = _g->reserved().end(); |
duke@435 | 495 | } |
duke@435 | 496 | |
coleenp@548 | 497 | void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); } |
coleenp@548 | 498 | void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); } |
coleenp@548 | 499 | |
coleenp@548 | 500 | void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); } |
coleenp@548 | 501 | void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); } |
coleenp@548 | 502 | |
coleenp@548 | 503 | void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); } |
coleenp@548 | 504 | void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); } |
coleenp@548 | 505 | |
coleenp@548 | 506 | void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); } |
coleenp@548 | 507 | void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); } |
coleenp@548 | 508 | |
duke@435 | 509 | ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g, |
duke@435 | 510 | ParScanThreadState* par_scan_state) |
duke@435 | 511 | : ScanWeakRefClosure(g), _par_scan_state(par_scan_state) |
coleenp@548 | 512 | {} |
coleenp@548 | 513 | |
coleenp@548 | 514 | void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); } |
coleenp@548 | 515 | void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); } |
duke@435 | 516 | |
duke@435 | 517 | #ifdef WIN32 |
duke@435 | 518 | #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */ |
duke@435 | 519 | #endif |
duke@435 | 520 | |
duke@435 | 521 | ParEvacuateFollowersClosure::ParEvacuateFollowersClosure( |
duke@435 | 522 | ParScanThreadState* par_scan_state_, |
duke@435 | 523 | ParScanWithoutBarrierClosure* to_space_closure_, |
duke@435 | 524 | ParScanWithBarrierClosure* old_gen_closure_, |
duke@435 | 525 | ParRootScanWithoutBarrierClosure* to_space_root_closure_, |
duke@435 | 526 | ParNewGeneration* par_gen_, |
duke@435 | 527 | ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_, |
duke@435 | 528 | ObjToScanQueueSet* task_queues_, |
duke@435 | 529 | ParallelTaskTerminator* terminator_) : |
duke@435 | 530 | |
duke@435 | 531 | _par_scan_state(par_scan_state_), |
duke@435 | 532 | _to_space_closure(to_space_closure_), |
duke@435 | 533 | _old_gen_closure(old_gen_closure_), |
duke@435 | 534 | _to_space_root_closure(to_space_root_closure_), |
duke@435 | 535 | _old_gen_root_closure(old_gen_root_closure_), |
duke@435 | 536 | _par_gen(par_gen_), |
duke@435 | 537 | _task_queues(task_queues_), |
duke@435 | 538 | _terminator(terminator_) |
duke@435 | 539 | {} |
duke@435 | 540 | |
duke@435 | 541 | void ParEvacuateFollowersClosure::do_void() { |
duke@435 | 542 | ObjToScanQueue* work_q = par_scan_state()->work_queue(); |
duke@435 | 543 | |
duke@435 | 544 | while (true) { |
duke@435 | 545 | |
duke@435 | 546 | // Scan to-space and old-gen objs until we run out of both. |
duke@435 | 547 | oop obj_to_scan; |
duke@435 | 548 | par_scan_state()->trim_queues(0); |
duke@435 | 549 | |
duke@435 | 550 | // We have no local work, attempt to steal from other threads. |
duke@435 | 551 | |
duke@435 | 552 | // attempt to steal work from promoted. |
duke@435 | 553 | if (task_queues()->steal(par_scan_state()->thread_num(), |
duke@435 | 554 | par_scan_state()->hash_seed(), |
duke@435 | 555 | obj_to_scan)) { |
duke@435 | 556 | bool res = work_q->push(obj_to_scan); |
duke@435 | 557 | assert(res, "Empty queue should have room for a push."); |
duke@435 | 558 | |
duke@435 | 559 | // if successful, goto Start. |
duke@435 | 560 | continue; |
duke@435 | 561 | |
duke@435 | 562 | // try global overflow list. |
duke@435 | 563 | } else if (par_gen()->take_from_overflow_list(par_scan_state())) { |
duke@435 | 564 | continue; |
duke@435 | 565 | } |
duke@435 | 566 | |
duke@435 | 567 | // Otherwise, offer termination. |
duke@435 | 568 | par_scan_state()->start_term_time(); |
duke@435 | 569 | if (terminator()->offer_termination()) break; |
duke@435 | 570 | par_scan_state()->end_term_time(); |
duke@435 | 571 | } |
ysr@969 | 572 | assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0, |
ysr@969 | 573 | "Broken overflow list?"); |
duke@435 | 574 | // Finish the last termination pause. |
duke@435 | 575 | par_scan_state()->end_term_time(); |
duke@435 | 576 | } |
duke@435 | 577 | |
duke@435 | 578 | ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen, |
duke@435 | 579 | HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) : |
duke@435 | 580 | AbstractGangTask("ParNewGeneration collection"), |
duke@435 | 581 | _gen(gen), _next_gen(next_gen), |
duke@435 | 582 | _young_old_boundary(young_old_boundary), |
duke@435 | 583 | _state_set(state_set) |
duke@435 | 584 | {} |
duke@435 | 585 | |
jmasa@3294 | 586 | // Reset the terminator for the given number of |
jmasa@3294 | 587 | // active threads. |
jmasa@3294 | 588 | void ParNewGenTask::set_for_termination(int active_workers) { |
jmasa@3294 | 589 | _state_set->reset(active_workers, _gen->promotion_failed()); |
jmasa@3294 | 590 | // Should the heap be passed in? There's only 1 for now so |
jmasa@3294 | 591 | // grab it instead. |
jmasa@3294 | 592 | GenCollectedHeap* gch = GenCollectedHeap::heap(); |
jmasa@3294 | 593 | gch->set_n_termination(active_workers); |
jmasa@3294 | 594 | } |
jmasa@3294 | 595 | |
jmasa@3357 | 596 | void ParNewGenTask::work(uint worker_id) { |
duke@435 | 597 | GenCollectedHeap* gch = GenCollectedHeap::heap(); |
duke@435 | 598 | // Since this is being done in a separate thread, need new resource |
duke@435 | 599 | // and handle marks. |
duke@435 | 600 | ResourceMark rm; |
duke@435 | 601 | HandleMark hm; |
duke@435 | 602 | // We would need multiple old-gen queues otherwise. |
ysr@1114 | 603 | assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen."); |
duke@435 | 604 | |
duke@435 | 605 | Generation* old_gen = gch->next_gen(_gen); |
duke@435 | 606 | |
jmasa@3357 | 607 | ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id); |
jmasa@3357 | 608 | assert(_state_set->is_valid(worker_id), "Should not have been called"); |
jmasa@3294 | 609 | |
duke@435 | 610 | par_scan_state.set_young_old_boundary(_young_old_boundary); |
duke@435 | 611 | |
coleenp@4037 | 612 | KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(), |
coleenp@4037 | 613 | gch->rem_set()->klass_rem_set()); |
coleenp@4037 | 614 | |
coleenp@4037 | 615 | int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache; |
coleenp@4037 | 616 | |
duke@435 | 617 | par_scan_state.start_strong_roots(); |
duke@435 | 618 | gch->gen_process_strong_roots(_gen->level(), |
jrose@1424 | 619 | true, // Process younger gens, if any, |
jrose@1424 | 620 | // as strong roots. |
jrose@1424 | 621 | false, // no scope; this is parallel code |
coleenp@4037 | 622 | true, // is scavenging |
coleenp@4037 | 623 | SharedHeap::ScanningOption(so), |
jrose@1424 | 624 | &par_scan_state.to_space_root_closure(), |
jrose@1424 | 625 | true, // walk *all* scavengable nmethods |
coleenp@4037 | 626 | &par_scan_state.older_gen_closure(), |
coleenp@4037 | 627 | &klass_scan_closure); |
duke@435 | 628 | par_scan_state.end_strong_roots(); |
duke@435 | 629 | |
duke@435 | 630 | // "evacuate followers". |
duke@435 | 631 | par_scan_state.evacuate_followers_closure().do_void(); |
duke@435 | 632 | } |
duke@435 | 633 | |
duke@435 | 634 | #ifdef _MSC_VER |
duke@435 | 635 | #pragma warning( push ) |
duke@435 | 636 | #pragma warning( disable:4355 ) // 'this' : used in base member initializer list |
duke@435 | 637 | #endif |
duke@435 | 638 | ParNewGeneration:: |
duke@435 | 639 | ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level) |
duke@435 | 640 | : DefNewGeneration(rs, initial_byte_size, level, "PCopy"), |
duke@435 | 641 | _overflow_list(NULL), |
duke@435 | 642 | _is_alive_closure(this), |
duke@435 | 643 | _plab_stats(YoungPLABSize, PLABWeight) |
duke@435 | 644 | { |
ysr@969 | 645 | NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;) |
ysr@969 | 646 | NOT_PRODUCT(_num_par_pushes = 0;) |
duke@435 | 647 | _task_queues = new ObjToScanQueueSet(ParallelGCThreads); |
duke@435 | 648 | guarantee(_task_queues != NULL, "task_queues allocation failure."); |
duke@435 | 649 | |
duke@435 | 650 | for (uint i1 = 0; i1 < ParallelGCThreads; i1++) { |
jcoomes@2020 | 651 | ObjToScanQueue *q = new ObjToScanQueue(); |
jcoomes@2020 | 652 | guarantee(q != NULL, "work_queue Allocation failure."); |
jcoomes@2020 | 653 | _task_queues->register_queue(i1, q); |
duke@435 | 654 | } |
duke@435 | 655 | |
duke@435 | 656 | for (uint i2 = 0; i2 < ParallelGCThreads; i2++) |
duke@435 | 657 | _task_queues->queue(i2)->initialize(); |
duke@435 | 658 | |
jcoomes@2191 | 659 | _overflow_stacks = NULL; |
jcoomes@2191 | 660 | if (ParGCUseLocalOverflow) { |
zgu@3900 | 661 | |
zgu@3900 | 662 | // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal |
zgu@3900 | 663 | // with ',' |
zgu@3900 | 664 | typedef Stack<oop, mtGC> GCOopStack; |
zgu@3900 | 665 | |
zgu@3900 | 666 | _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC); |
jcoomes@2191 | 667 | for (size_t i = 0; i < ParallelGCThreads; ++i) { |
zgu@3900 | 668 | new (_overflow_stacks + i) Stack<oop, mtGC>(); |
ysr@1130 | 669 | } |
ysr@1130 | 670 | } |
ysr@1130 | 671 | |
duke@435 | 672 | if (UsePerfData) { |
duke@435 | 673 | EXCEPTION_MARK; |
duke@435 | 674 | ResourceMark rm; |
duke@435 | 675 | |
duke@435 | 676 | const char* cname = |
duke@435 | 677 | PerfDataManager::counter_name(_gen_counters->name_space(), "threads"); |
duke@435 | 678 | PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None, |
duke@435 | 679 | ParallelGCThreads, CHECK); |
duke@435 | 680 | } |
duke@435 | 681 | } |
duke@435 | 682 | #ifdef _MSC_VER |
duke@435 | 683 | #pragma warning( pop ) |
duke@435 | 684 | #endif |
duke@435 | 685 | |
duke@435 | 686 | // ParNewGeneration:: |
duke@435 | 687 | ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) : |
duke@435 | 688 | DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {} |
duke@435 | 689 | |
coleenp@548 | 690 | template <class T> |
coleenp@548 | 691 | void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) { |
coleenp@548 | 692 | #ifdef ASSERT |
coleenp@548 | 693 | { |
coleenp@548 | 694 | assert(!oopDesc::is_null(*p), "expected non-null ref"); |
coleenp@548 | 695 | oop obj = oopDesc::load_decode_heap_oop_not_null(p); |
coleenp@548 | 696 | // We never expect to see a null reference being processed |
coleenp@548 | 697 | // as a weak reference. |
coleenp@548 | 698 | assert(obj->is_oop(), "expected an oop while scanning weak refs"); |
coleenp@548 | 699 | } |
coleenp@548 | 700 | #endif // ASSERT |
duke@435 | 701 | |
duke@435 | 702 | _par_cl->do_oop_nv(p); |
duke@435 | 703 | |
duke@435 | 704 | if (Universe::heap()->is_in_reserved(p)) { |
coleenp@548 | 705 | oop obj = oopDesc::load_decode_heap_oop_not_null(p); |
coleenp@548 | 706 | _rs->write_ref_field_gc_par(p, obj); |
duke@435 | 707 | } |
duke@435 | 708 | } |
duke@435 | 709 | |
coleenp@548 | 710 | void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); } |
coleenp@548 | 711 | void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); } |
coleenp@548 | 712 | |
duke@435 | 713 | // ParNewGeneration:: |
duke@435 | 714 | KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) : |
duke@435 | 715 | DefNewGeneration::KeepAliveClosure(cl) {} |
duke@435 | 716 | |
coleenp@548 | 717 | template <class T> |
coleenp@548 | 718 | void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) { |
coleenp@548 | 719 | #ifdef ASSERT |
coleenp@548 | 720 | { |
coleenp@548 | 721 | assert(!oopDesc::is_null(*p), "expected non-null ref"); |
coleenp@548 | 722 | oop obj = oopDesc::load_decode_heap_oop_not_null(p); |
coleenp@548 | 723 | // We never expect to see a null reference being processed |
coleenp@548 | 724 | // as a weak reference. |
coleenp@548 | 725 | assert(obj->is_oop(), "expected an oop while scanning weak refs"); |
coleenp@548 | 726 | } |
coleenp@548 | 727 | #endif // ASSERT |
duke@435 | 728 | |
duke@435 | 729 | _cl->do_oop_nv(p); |
duke@435 | 730 | |
duke@435 | 731 | if (Universe::heap()->is_in_reserved(p)) { |
coleenp@548 | 732 | oop obj = oopDesc::load_decode_heap_oop_not_null(p); |
coleenp@548 | 733 | _rs->write_ref_field_gc_par(p, obj); |
duke@435 | 734 | } |
duke@435 | 735 | } |
duke@435 | 736 | |
coleenp@548 | 737 | void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); } |
coleenp@548 | 738 | void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); } |
coleenp@548 | 739 | |
coleenp@548 | 740 | template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) { |
coleenp@548 | 741 | T heap_oop = oopDesc::load_heap_oop(p); |
coleenp@548 | 742 | if (!oopDesc::is_null(heap_oop)) { |
coleenp@548 | 743 | oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); |
duke@435 | 744 | if ((HeapWord*)obj < _boundary) { |
duke@435 | 745 | assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?"); |
coleenp@548 | 746 | oop new_obj = obj->is_forwarded() |
coleenp@548 | 747 | ? obj->forwardee() |
coleenp@548 | 748 | : _g->DefNewGeneration::copy_to_survivor_space(obj); |
coleenp@548 | 749 | oopDesc::encode_store_heap_oop_not_null(p, new_obj); |
duke@435 | 750 | } |
duke@435 | 751 | if (_gc_barrier) { |
duke@435 | 752 | // If p points to a younger generation, mark the card. |
duke@435 | 753 | if ((HeapWord*)obj < _gen_boundary) { |
duke@435 | 754 | _rs->write_ref_field_gc_par(p, obj); |
duke@435 | 755 | } |
duke@435 | 756 | } |
duke@435 | 757 | } |
duke@435 | 758 | } |
duke@435 | 759 | |
coleenp@548 | 760 | void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); } |
coleenp@548 | 761 | void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); } |
coleenp@548 | 762 | |
duke@435 | 763 | class ParNewRefProcTaskProxy: public AbstractGangTask { |
duke@435 | 764 | typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; |
duke@435 | 765 | public: |
duke@435 | 766 | ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen, |
duke@435 | 767 | Generation& next_gen, |
duke@435 | 768 | HeapWord* young_old_boundary, |
duke@435 | 769 | ParScanThreadStateSet& state_set); |
duke@435 | 770 | |
duke@435 | 771 | private: |
jmasa@3357 | 772 | virtual void work(uint worker_id); |
jmasa@3294 | 773 | virtual void set_for_termination(int active_workers) { |
jmasa@3294 | 774 | _state_set.terminator()->reset_for_reuse(active_workers); |
jmasa@3294 | 775 | } |
duke@435 | 776 | private: |
duke@435 | 777 | ParNewGeneration& _gen; |
duke@435 | 778 | ProcessTask& _task; |
duke@435 | 779 | Generation& _next_gen; |
duke@435 | 780 | HeapWord* _young_old_boundary; |
duke@435 | 781 | ParScanThreadStateSet& _state_set; |
duke@435 | 782 | }; |
duke@435 | 783 | |
duke@435 | 784 | ParNewRefProcTaskProxy::ParNewRefProcTaskProxy( |
duke@435 | 785 | ProcessTask& task, ParNewGeneration& gen, |
duke@435 | 786 | Generation& next_gen, |
duke@435 | 787 | HeapWord* young_old_boundary, |
duke@435 | 788 | ParScanThreadStateSet& state_set) |
duke@435 | 789 | : AbstractGangTask("ParNewGeneration parallel reference processing"), |
duke@435 | 790 | _gen(gen), |
duke@435 | 791 | _task(task), |
duke@435 | 792 | _next_gen(next_gen), |
duke@435 | 793 | _young_old_boundary(young_old_boundary), |
duke@435 | 794 | _state_set(state_set) |
duke@435 | 795 | { |
duke@435 | 796 | } |
duke@435 | 797 | |
jmasa@3357 | 798 | void ParNewRefProcTaskProxy::work(uint worker_id) |
duke@435 | 799 | { |
duke@435 | 800 | ResourceMark rm; |
duke@435 | 801 | HandleMark hm; |
jmasa@3357 | 802 | ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id); |
duke@435 | 803 | par_scan_state.set_young_old_boundary(_young_old_boundary); |
jmasa@3357 | 804 | _task.work(worker_id, par_scan_state.is_alive_closure(), |
duke@435 | 805 | par_scan_state.keep_alive_closure(), |
duke@435 | 806 | par_scan_state.evacuate_followers_closure()); |
duke@435 | 807 | } |
duke@435 | 808 | |
duke@435 | 809 | class ParNewRefEnqueueTaskProxy: public AbstractGangTask { |
duke@435 | 810 | typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; |
duke@435 | 811 | EnqueueTask& _task; |
duke@435 | 812 | |
duke@435 | 813 | public: |
duke@435 | 814 | ParNewRefEnqueueTaskProxy(EnqueueTask& task) |
duke@435 | 815 | : AbstractGangTask("ParNewGeneration parallel reference enqueue"), |
duke@435 | 816 | _task(task) |
duke@435 | 817 | { } |
duke@435 | 818 | |
jmasa@3357 | 819 | virtual void work(uint worker_id) |
duke@435 | 820 | { |
jmasa@3357 | 821 | _task.work(worker_id); |
duke@435 | 822 | } |
duke@435 | 823 | }; |
duke@435 | 824 | |
duke@435 | 825 | |
duke@435 | 826 | void ParNewRefProcTaskExecutor::execute(ProcessTask& task) |
duke@435 | 827 | { |
duke@435 | 828 | GenCollectedHeap* gch = GenCollectedHeap::heap(); |
duke@435 | 829 | assert(gch->kind() == CollectedHeap::GenCollectedHeap, |
duke@435 | 830 | "not a generational heap"); |
jmasa@3294 | 831 | FlexibleWorkGang* workers = gch->workers(); |
duke@435 | 832 | assert(workers != NULL, "Need parallel worker threads."); |
jmasa@3294 | 833 | _state_set.reset(workers->active_workers(), _generation.promotion_failed()); |
duke@435 | 834 | ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(), |
duke@435 | 835 | _generation.reserved().end(), _state_set); |
duke@435 | 836 | workers->run_task(&rp_task); |
jmasa@3294 | 837 | _state_set.reset(0 /* bad value in debug if not reset */, |
jmasa@3294 | 838 | _generation.promotion_failed()); |
duke@435 | 839 | } |
duke@435 | 840 | |
duke@435 | 841 | void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) |
duke@435 | 842 | { |
duke@435 | 843 | GenCollectedHeap* gch = GenCollectedHeap::heap(); |
jmasa@3294 | 844 | FlexibleWorkGang* workers = gch->workers(); |
duke@435 | 845 | assert(workers != NULL, "Need parallel worker threads."); |
duke@435 | 846 | ParNewRefEnqueueTaskProxy enq_task(task); |
duke@435 | 847 | workers->run_task(&enq_task); |
duke@435 | 848 | } |
duke@435 | 849 | |
duke@435 | 850 | void ParNewRefProcTaskExecutor::set_single_threaded_mode() |
duke@435 | 851 | { |
duke@435 | 852 | _state_set.flush(); |
duke@435 | 853 | GenCollectedHeap* gch = GenCollectedHeap::heap(); |
duke@435 | 854 | gch->set_par_threads(0); // 0 ==> non-parallel. |
duke@435 | 855 | gch->save_marks(); |
duke@435 | 856 | } |
duke@435 | 857 | |
duke@435 | 858 | ScanClosureWithParBarrier:: |
duke@435 | 859 | ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) : |
duke@435 | 860 | ScanClosure(g, gc_barrier) {} |
duke@435 | 861 | |
duke@435 | 862 | EvacuateFollowersClosureGeneral:: |
duke@435 | 863 | EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level, |
duke@435 | 864 | OopsInGenClosure* cur, |
duke@435 | 865 | OopsInGenClosure* older) : |
duke@435 | 866 | _gch(gch), _level(level), |
duke@435 | 867 | _scan_cur_or_nonheap(cur), _scan_older(older) |
duke@435 | 868 | {} |
duke@435 | 869 | |
duke@435 | 870 | void EvacuateFollowersClosureGeneral::do_void() { |
duke@435 | 871 | do { |
duke@435 | 872 | // Beware: this call will lead to closure applications via virtual |
duke@435 | 873 | // calls. |
duke@435 | 874 | _gch->oop_since_save_marks_iterate(_level, |
duke@435 | 875 | _scan_cur_or_nonheap, |
duke@435 | 876 | _scan_older); |
duke@435 | 877 | } while (!_gch->no_allocs_since_save_marks(_level)); |
duke@435 | 878 | } |
duke@435 | 879 | |
duke@435 | 880 | |
sla@5237 | 881 | // A Generation that does parallel young-gen collection. |
sla@5237 | 882 | |
duke@435 | 883 | bool ParNewGeneration::_avoid_promotion_undo = false; |
duke@435 | 884 | |
sla@5237 | 885 | void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set, ParNewTracer& gc_tracer) { |
sla@5237 | 886 | assert(_promo_failure_scan_stack.is_empty(), "post condition"); |
sla@5237 | 887 | _promo_failure_scan_stack.clear(true); // Clear cached segments. |
sla@5237 | 888 | |
sla@5237 | 889 | remove_forwarding_pointers(); |
sla@5237 | 890 | if (PrintGCDetails) { |
sla@5237 | 891 | gclog_or_tty->print(" (promotion failed)"); |
sla@5237 | 892 | } |
sla@5237 | 893 | // All the spaces are in play for mark-sweep. |
sla@5237 | 894 | swap_spaces(); // Make life simpler for CMS || rescan; see 6483690. |
sla@5237 | 895 | from()->set_next_compaction_space(to()); |
sla@5237 | 896 | gch->set_incremental_collection_failed(); |
sla@5237 | 897 | // Inform the next generation that a promotion failure occurred. |
sla@5237 | 898 | _next_gen->promotion_failure_occurred(); |
sla@5237 | 899 | |
sla@5237 | 900 | // Trace promotion failure in the parallel GC threads |
sla@5237 | 901 | thread_state_set.trace_promotion_failed(gc_tracer); |
sla@5237 | 902 | // Single threaded code may have reported promotion failure to the global state |
sla@5237 | 903 | if (_promotion_failed_info.has_failed()) { |
sla@5237 | 904 | gc_tracer.report_promotion_failed(_promotion_failed_info); |
sla@5237 | 905 | } |
sla@5237 | 906 | // Reset the PromotionFailureALot counters. |
sla@5237 | 907 | NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();) |
sla@5237 | 908 | } |
duke@435 | 909 | |
duke@435 | 910 | void ParNewGeneration::collect(bool full, |
duke@435 | 911 | bool clear_all_soft_refs, |
duke@435 | 912 | size_t size, |
duke@435 | 913 | bool is_tlab) { |
duke@435 | 914 | assert(full || size > 0, "otherwise we don't want to collect"); |
sla@5237 | 915 | |
duke@435 | 916 | GenCollectedHeap* gch = GenCollectedHeap::heap(); |
sla@5237 | 917 | |
sla@5237 | 918 | _gc_timer->register_gc_start(os::elapsed_counter()); |
sla@5237 | 919 | |
duke@435 | 920 | assert(gch->kind() == CollectedHeap::GenCollectedHeap, |
duke@435 | 921 | "not a CMS generational heap"); |
duke@435 | 922 | AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy(); |
jmasa@3294 | 923 | FlexibleWorkGang* workers = gch->workers(); |
jmasa@3294 | 924 | assert(workers != NULL, "Need workgang for parallel work"); |
jmasa@3294 | 925 | int active_workers = |
jmasa@3294 | 926 | AdaptiveSizePolicy::calc_active_workers(workers->total_workers(), |
jmasa@3294 | 927 | workers->active_workers(), |
jmasa@3294 | 928 | Threads::number_of_non_daemon_threads()); |
jmasa@3294 | 929 | workers->set_active_workers(active_workers); |
duke@435 | 930 | assert(gch->n_gens() == 2, |
duke@435 | 931 | "Par collection currently only works with single older gen."); |
brutisso@5516 | 932 | _next_gen = gch->next_gen(this); |
duke@435 | 933 | // Do we have to avoid promotion_undo? |
duke@435 | 934 | if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) { |
duke@435 | 935 | set_avoid_promotion_undo(true); |
duke@435 | 936 | } |
duke@435 | 937 | |
sla@5237 | 938 | // If the next generation is too full to accommodate worst-case promotion |
duke@435 | 939 | // from this generation, pass on collection; let the next generation |
duke@435 | 940 | // do it. |
duke@435 | 941 | if (!collection_attempt_is_safe()) { |
ysr@2243 | 942 | gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one |
duke@435 | 943 | return; |
duke@435 | 944 | } |
duke@435 | 945 | assert(to()->is_empty(), "Else not collection_attempt_is_safe"); |
duke@435 | 946 | |
sla@5237 | 947 | ParNewTracer gc_tracer; |
sla@5237 | 948 | gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start()); |
sla@5237 | 949 | gch->trace_heap_before_gc(&gc_tracer); |
sla@5237 | 950 | |
duke@435 | 951 | init_assuming_no_promotion_failure(); |
duke@435 | 952 | |
duke@435 | 953 | if (UseAdaptiveSizePolicy) { |
duke@435 | 954 | set_survivor_overflow(false); |
duke@435 | 955 | size_policy->minor_collection_begin(); |
duke@435 | 956 | } |
duke@435 | 957 | |
sla@5237 | 958 | GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL); |
duke@435 | 959 | // Capture heap used before collection (for printing). |
duke@435 | 960 | size_t gch_prev_used = gch->used(); |
duke@435 | 961 | |
duke@435 | 962 | SpecializationStats::clear(); |
duke@435 | 963 | |
duke@435 | 964 | age_table()->clear(); |
jmasa@698 | 965 | to()->clear(SpaceDecorator::Mangle); |
duke@435 | 966 | |
duke@435 | 967 | gch->save_marks(); |
duke@435 | 968 | assert(workers != NULL, "Need parallel worker threads."); |
jmasa@3294 | 969 | int n_workers = active_workers; |
jmasa@3294 | 970 | |
jmasa@3294 | 971 | // Set the correct parallelism (number of queues) in the reference processor |
jmasa@3294 | 972 | ref_processor()->set_active_mt_degree(n_workers); |
jmasa@3294 | 973 | |
jmasa@3294 | 974 | // Always set the terminator for the active number of workers |
jmasa@3294 | 975 | // because only those workers go through the termination protocol. |
jmasa@3294 | 976 | ParallelTaskTerminator _term(n_workers, task_queues()); |
jmasa@3294 | 977 | ParScanThreadStateSet thread_state_set(workers->active_workers(), |
duke@435 | 978 | *to(), *this, *_next_gen, *task_queues(), |
ysr@1130 | 979 | _overflow_stacks, desired_plab_sz(), _term); |
duke@435 | 980 | |
duke@435 | 981 | ParNewGenTask tsk(this, _next_gen, reserved().end(), &thread_state_set); |
duke@435 | 982 | gch->set_par_threads(n_workers); |
duke@435 | 983 | gch->rem_set()->prepare_for_younger_refs_iterate(true); |
duke@435 | 984 | // It turns out that even when we're using 1 thread, doing the work in a |
duke@435 | 985 | // separate thread causes wide variance in run times. We can't help this |
duke@435 | 986 | // in the multi-threaded case, but we special-case n=1 here to get |
duke@435 | 987 | // repeatable measurements of the 1-thread overhead of the parallel code. |
duke@435 | 988 | if (n_workers > 1) { |
jrose@1424 | 989 | GenCollectedHeap::StrongRootsScope srs(gch); |
duke@435 | 990 | workers->run_task(&tsk); |
duke@435 | 991 | } else { |
jrose@1424 | 992 | GenCollectedHeap::StrongRootsScope srs(gch); |
duke@435 | 993 | tsk.work(0); |
duke@435 | 994 | } |
jmasa@3294 | 995 | thread_state_set.reset(0 /* Bad value in debug if not reset */, |
jmasa@3294 | 996 | promotion_failed()); |
duke@435 | 997 | |
duke@435 | 998 | // Process (weak) reference objects found during scavenge. |
ysr@888 | 999 | ReferenceProcessor* rp = ref_processor(); |
duke@435 | 1000 | IsAliveClosure is_alive(this); |
duke@435 | 1001 | ScanWeakRefClosure scan_weak_ref(this); |
duke@435 | 1002 | KeepAliveClosure keep_alive(&scan_weak_ref); |
duke@435 | 1003 | ScanClosure scan_without_gc_barrier(this, false); |
duke@435 | 1004 | ScanClosureWithParBarrier scan_with_gc_barrier(this, true); |
duke@435 | 1005 | set_promo_failure_scan_stack_closure(&scan_without_gc_barrier); |
duke@435 | 1006 | EvacuateFollowersClosureGeneral evacuate_followers(gch, _level, |
duke@435 | 1007 | &scan_without_gc_barrier, &scan_with_gc_barrier); |
ysr@892 | 1008 | rp->setup_policy(clear_all_soft_refs); |
jmasa@3294 | 1009 | // Can the mt_degree be set later (at run_task() time would be best)? |
jmasa@3294 | 1010 | rp->set_active_mt_degree(active_workers); |
sla@5237 | 1011 | ReferenceProcessorStats stats; |
ysr@888 | 1012 | if (rp->processing_is_mt()) { |
duke@435 | 1013 | ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); |
sla@5237 | 1014 | stats = rp->process_discovered_references(&is_alive, &keep_alive, |
sla@5237 | 1015 | &evacuate_followers, &task_executor, |
sla@5237 | 1016 | _gc_timer); |
duke@435 | 1017 | } else { |
duke@435 | 1018 | thread_state_set.flush(); |
duke@435 | 1019 | gch->set_par_threads(0); // 0 ==> non-parallel. |
duke@435 | 1020 | gch->save_marks(); |
sla@5237 | 1021 | stats = rp->process_discovered_references(&is_alive, &keep_alive, |
sla@5237 | 1022 | &evacuate_followers, NULL, |
sla@5237 | 1023 | _gc_timer); |
duke@435 | 1024 | } |
sla@5237 | 1025 | gc_tracer.report_gc_reference_stats(stats); |
duke@435 | 1026 | if (!promotion_failed()) { |
duke@435 | 1027 | // Swap the survivor spaces. |
jmasa@698 | 1028 | eden()->clear(SpaceDecorator::Mangle); |
jmasa@698 | 1029 | from()->clear(SpaceDecorator::Mangle); |
jmasa@698 | 1030 | if (ZapUnusedHeapArea) { |
jmasa@698 | 1031 | // This is now done here because of the piece-meal mangling which |
jmasa@698 | 1032 | // can check for valid mangling at intermediate points in the |
jmasa@698 | 1033 | // collection(s). When a minor collection fails to collect |
jmasa@698 | 1034 | // sufficient space resizing of the young generation can occur |
jmasa@698 | 1035 | // an redistribute the spaces in the young generation. Mangle |
jmasa@698 | 1036 | // here so that unzapped regions don't get distributed to |
jmasa@698 | 1037 | // other spaces. |
jmasa@698 | 1038 | to()->mangle_unused_area(); |
jmasa@698 | 1039 | } |
duke@435 | 1040 | swap_spaces(); |
duke@435 | 1041 | |
jmasa@1822 | 1042 | // A successful scavenge should restart the GC time limit count which is |
jmasa@1822 | 1043 | // for full GC's. |
jmasa@1822 | 1044 | size_policy->reset_gc_overhead_limit_count(); |
jmasa@1822 | 1045 | |
duke@435 | 1046 | assert(to()->is_empty(), "to space should be empty now"); |
brutisso@4452 | 1047 | |
brutisso@4452 | 1048 | adjust_desired_tenuring_threshold(); |
duke@435 | 1049 | } else { |
sla@5237 | 1050 | handle_promotion_failed(gch, thread_state_set, gc_tracer); |
duke@435 | 1051 | } |
duke@435 | 1052 | // set new iteration safe limit for the survivor spaces |
duke@435 | 1053 | from()->set_concurrent_iteration_safe_limit(from()->top()); |
duke@435 | 1054 | to()->set_concurrent_iteration_safe_limit(to()->top()); |
duke@435 | 1055 | |
duke@435 | 1056 | if (ResizePLAB) { |
johnc@4130 | 1057 | plab_stats()->adjust_desired_plab_sz(n_workers); |
duke@435 | 1058 | } |
duke@435 | 1059 | |
duke@435 | 1060 | if (PrintGC && !PrintGCDetails) { |
duke@435 | 1061 | gch->print_heap_change(gch_prev_used); |
duke@435 | 1062 | } |
duke@435 | 1063 | |
jcoomes@2067 | 1064 | if (PrintGCDetails && ParallelGCVerbose) { |
jcoomes@2067 | 1065 | TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats()); |
jcoomes@2067 | 1066 | TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats()); |
jcoomes@2067 | 1067 | } |
jcoomes@2065 | 1068 | |
duke@435 | 1069 | if (UseAdaptiveSizePolicy) { |
duke@435 | 1070 | size_policy->minor_collection_end(gch->gc_cause()); |
duke@435 | 1071 | size_policy->avg_survived()->sample(from()->used()); |
duke@435 | 1072 | } |
duke@435 | 1073 | |
johnc@3538 | 1074 | // We need to use a monotonically non-deccreasing time in ms |
johnc@3538 | 1075 | // or we will see time-warp warnings and os::javaTimeMillis() |
johnc@3538 | 1076 | // does not guarantee monotonicity. |
johnc@3538 | 1077 | jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; |
johnc@3538 | 1078 | update_time_of_last_gc(now); |
duke@435 | 1079 | |
duke@435 | 1080 | SpecializationStats::print(); |
duke@435 | 1081 | |
ysr@888 | 1082 | rp->set_enqueuing_is_done(true); |
ysr@888 | 1083 | if (rp->processing_is_mt()) { |
duke@435 | 1084 | ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); |
ysr@888 | 1085 | rp->enqueue_discovered_references(&task_executor); |
duke@435 | 1086 | } else { |
ysr@888 | 1087 | rp->enqueue_discovered_references(NULL); |
duke@435 | 1088 | } |
ysr@888 | 1089 | rp->verify_no_references_recorded(); |
sla@5237 | 1090 | |
sla@5237 | 1091 | gch->trace_heap_after_gc(&gc_tracer); |
sla@5237 | 1092 | gc_tracer.report_tenuring_threshold(tenuring_threshold()); |
sla@5237 | 1093 | |
sla@5237 | 1094 | _gc_timer->register_gc_end(os::elapsed_counter()); |
sla@5237 | 1095 | |
sla@5237 | 1096 | gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions()); |
duke@435 | 1097 | } |
duke@435 | 1098 | |
duke@435 | 1099 | static int sum; |
duke@435 | 1100 | void ParNewGeneration::waste_some_time() { |
duke@435 | 1101 | for (int i = 0; i < 100; i++) { |
duke@435 | 1102 | sum += i; |
duke@435 | 1103 | } |
duke@435 | 1104 | } |
duke@435 | 1105 | |
hseigel@5784 | 1106 | static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4); |
duke@435 | 1107 | |
duke@435 | 1108 | // Because of concurrency, there are times where an object for which |
duke@435 | 1109 | // "is_forwarded()" is true contains an "interim" forwarding pointer |
duke@435 | 1110 | // value. Such a value will soon be overwritten with a real value. |
duke@435 | 1111 | // This method requires "obj" to have a forwarding pointer, and waits, if |
duke@435 | 1112 | // necessary for a real one to be inserted, and returns it. |
duke@435 | 1113 | |
duke@435 | 1114 | oop ParNewGeneration::real_forwardee(oop obj) { |
duke@435 | 1115 | oop forward_ptr = obj->forwardee(); |
duke@435 | 1116 | if (forward_ptr != ClaimedForwardPtr) { |
duke@435 | 1117 | return forward_ptr; |
duke@435 | 1118 | } else { |
duke@435 | 1119 | return real_forwardee_slow(obj); |
duke@435 | 1120 | } |
duke@435 | 1121 | } |
duke@435 | 1122 | |
duke@435 | 1123 | oop ParNewGeneration::real_forwardee_slow(oop obj) { |
duke@435 | 1124 | // Spin-read if it is claimed but not yet written by another thread. |
duke@435 | 1125 | oop forward_ptr = obj->forwardee(); |
duke@435 | 1126 | while (forward_ptr == ClaimedForwardPtr) { |
duke@435 | 1127 | waste_some_time(); |
duke@435 | 1128 | assert(obj->is_forwarded(), "precondition"); |
duke@435 | 1129 | forward_ptr = obj->forwardee(); |
duke@435 | 1130 | } |
duke@435 | 1131 | return forward_ptr; |
duke@435 | 1132 | } |
duke@435 | 1133 | |
duke@435 | 1134 | #ifdef ASSERT |
duke@435 | 1135 | bool ParNewGeneration::is_legal_forward_ptr(oop p) { |
duke@435 | 1136 | return |
duke@435 | 1137 | (_avoid_promotion_undo && p == ClaimedForwardPtr) |
duke@435 | 1138 | || Universe::heap()->is_in_reserved(p); |
duke@435 | 1139 | } |
duke@435 | 1140 | #endif |
duke@435 | 1141 | |
duke@435 | 1142 | void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) { |
ysr@2380 | 1143 | if (m->must_be_preserved_for_promotion_failure(obj)) { |
ysr@2380 | 1144 | // We should really have separate per-worker stacks, rather |
ysr@2380 | 1145 | // than use locking of a common pair of stacks. |
duke@435 | 1146 | MutexLocker ml(ParGCRareEvent_lock); |
ysr@2380 | 1147 | preserve_mark(obj, m); |
duke@435 | 1148 | } |
duke@435 | 1149 | } |
duke@435 | 1150 | |
duke@435 | 1151 | // Multiple GC threads may try to promote an object. If the object |
duke@435 | 1152 | // is successfully promoted, a forwarding pointer will be installed in |
duke@435 | 1153 | // the object in the young generation. This method claims the right |
duke@435 | 1154 | // to install the forwarding pointer before it copies the object, |
duke@435 | 1155 | // thus avoiding the need to undo the copy as in |
duke@435 | 1156 | // copy_to_survivor_space_avoiding_with_undo. |
duke@435 | 1157 | |
duke@435 | 1158 | oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo( |
duke@435 | 1159 | ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { |
duke@435 | 1160 | // In the sequential version, this assert also says that the object is |
duke@435 | 1161 | // not forwarded. That might not be the case here. It is the case that |
duke@435 | 1162 | // the caller observed it to be not forwarded at some time in the past. |
duke@435 | 1163 | assert(is_in_reserved(old), "shouldn't be scavenging this oop"); |
duke@435 | 1164 | |
duke@435 | 1165 | // The sequential code read "old->age()" below. That doesn't work here, |
duke@435 | 1166 | // since the age is in the mark word, and that might be overwritten with |
duke@435 | 1167 | // a forwarding pointer by a parallel thread. So we must save the mark |
duke@435 | 1168 | // word in a local and then analyze it. |
duke@435 | 1169 | oopDesc dummyOld; |
duke@435 | 1170 | dummyOld.set_mark(m); |
duke@435 | 1171 | assert(!dummyOld.is_forwarded(), |
duke@435 | 1172 | "should not be called with forwarding pointer mark word."); |
duke@435 | 1173 | |
duke@435 | 1174 | oop new_obj = NULL; |
duke@435 | 1175 | oop forward_ptr; |
duke@435 | 1176 | |
duke@435 | 1177 | // Try allocating obj in to-space (unless too old) |
duke@435 | 1178 | if (dummyOld.age() < tenuring_threshold()) { |
duke@435 | 1179 | new_obj = (oop)par_scan_state->alloc_in_to_space(sz); |
duke@435 | 1180 | if (new_obj == NULL) { |
duke@435 | 1181 | set_survivor_overflow(true); |
duke@435 | 1182 | } |
duke@435 | 1183 | } |
duke@435 | 1184 | |
duke@435 | 1185 | if (new_obj == NULL) { |
duke@435 | 1186 | // Either to-space is full or we decided to promote |
duke@435 | 1187 | // try allocating obj tenured |
duke@435 | 1188 | |
duke@435 | 1189 | // Attempt to install a null forwarding pointer (atomically), |
duke@435 | 1190 | // to claim the right to install the real forwarding pointer. |
duke@435 | 1191 | forward_ptr = old->forward_to_atomic(ClaimedForwardPtr); |
duke@435 | 1192 | if (forward_ptr != NULL) { |
duke@435 | 1193 | // someone else beat us to it. |
duke@435 | 1194 | return real_forwardee(old); |
duke@435 | 1195 | } |
duke@435 | 1196 | |
duke@435 | 1197 | new_obj = _next_gen->par_promote(par_scan_state->thread_num(), |
duke@435 | 1198 | old, m, sz); |
duke@435 | 1199 | |
duke@435 | 1200 | if (new_obj == NULL) { |
duke@435 | 1201 | // promotion failed, forward to self |
duke@435 | 1202 | _promotion_failed = true; |
duke@435 | 1203 | new_obj = old; |
duke@435 | 1204 | |
duke@435 | 1205 | preserve_mark_if_necessary(old, m); |
sla@5237 | 1206 | par_scan_state->register_promotion_failure(sz); |
duke@435 | 1207 | } |
duke@435 | 1208 | |
duke@435 | 1209 | old->forward_to(new_obj); |
duke@435 | 1210 | forward_ptr = NULL; |
duke@435 | 1211 | } else { |
duke@435 | 1212 | // Is in to-space; do copying ourselves. |
duke@435 | 1213 | Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); |
duke@435 | 1214 | forward_ptr = old->forward_to_atomic(new_obj); |
duke@435 | 1215 | // Restore the mark word copied above. |
duke@435 | 1216 | new_obj->set_mark(m); |
duke@435 | 1217 | // Increment age if obj still in new generation |
duke@435 | 1218 | new_obj->incr_age(); |
duke@435 | 1219 | par_scan_state->age_table()->add(new_obj, sz); |
duke@435 | 1220 | } |
duke@435 | 1221 | assert(new_obj != NULL, "just checking"); |
duke@435 | 1222 | |
coleenp@4037 | 1223 | #ifndef PRODUCT |
coleenp@4037 | 1224 | // This code must come after the CAS test, or it will print incorrect |
coleenp@4037 | 1225 | // information. |
coleenp@4037 | 1226 | if (TraceScavenge) { |
coleenp@4037 | 1227 | gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}", |
coleenp@4037 | 1228 | is_in_reserved(new_obj) ? "copying" : "tenuring", |
hseigel@5784 | 1229 | new_obj->klass()->internal_name(), (void *)old, (void *)new_obj, new_obj->size()); |
coleenp@4037 | 1230 | } |
coleenp@4037 | 1231 | #endif |
coleenp@4037 | 1232 | |
duke@435 | 1233 | if (forward_ptr == NULL) { |
duke@435 | 1234 | oop obj_to_push = new_obj; |
duke@435 | 1235 | if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { |
duke@435 | 1236 | // Length field used as index of next element to be scanned. |
duke@435 | 1237 | // Real length can be obtained from real_forwardee() |
duke@435 | 1238 | arrayOop(old)->set_length(0); |
duke@435 | 1239 | obj_to_push = old; |
duke@435 | 1240 | assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, |
duke@435 | 1241 | "push forwarded object"); |
duke@435 | 1242 | } |
duke@435 | 1243 | // Push it on one of the queues of to-be-scanned objects. |
ysr@969 | 1244 | bool simulate_overflow = false; |
ysr@969 | 1245 | NOT_PRODUCT( |
ysr@969 | 1246 | if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { |
ysr@969 | 1247 | // simulate a stack overflow |
ysr@969 | 1248 | simulate_overflow = true; |
ysr@969 | 1249 | } |
ysr@969 | 1250 | ) |
ysr@969 | 1251 | if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { |
duke@435 | 1252 | // Add stats for overflow pushes. |
duke@435 | 1253 | if (Verbose && PrintGCDetails) { |
duke@435 | 1254 | gclog_or_tty->print("queue overflow!\n"); |
duke@435 | 1255 | } |
ysr@969 | 1256 | push_on_overflow_list(old, par_scan_state); |
jcoomes@2065 | 1257 | TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); |
duke@435 | 1258 | } |
duke@435 | 1259 | |
duke@435 | 1260 | return new_obj; |
duke@435 | 1261 | } |
duke@435 | 1262 | |
duke@435 | 1263 | // Oops. Someone beat us to it. Undo the allocation. Where did we |
duke@435 | 1264 | // allocate it? |
duke@435 | 1265 | if (is_in_reserved(new_obj)) { |
duke@435 | 1266 | // Must be in to_space. |
duke@435 | 1267 | assert(to()->is_in_reserved(new_obj), "Checking"); |
duke@435 | 1268 | if (forward_ptr == ClaimedForwardPtr) { |
duke@435 | 1269 | // Wait to get the real forwarding pointer value. |
duke@435 | 1270 | forward_ptr = real_forwardee(old); |
duke@435 | 1271 | } |
duke@435 | 1272 | par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); |
duke@435 | 1273 | } |
duke@435 | 1274 | |
duke@435 | 1275 | return forward_ptr; |
duke@435 | 1276 | } |
duke@435 | 1277 | |
duke@435 | 1278 | |
duke@435 | 1279 | // Multiple GC threads may try to promote the same object. If two |
duke@435 | 1280 | // or more GC threads copy the object, only one wins the race to install |
duke@435 | 1281 | // the forwarding pointer. The other threads have to undo their copy. |
duke@435 | 1282 | |
duke@435 | 1283 | oop ParNewGeneration::copy_to_survivor_space_with_undo( |
duke@435 | 1284 | ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { |
duke@435 | 1285 | |
duke@435 | 1286 | // In the sequential version, this assert also says that the object is |
duke@435 | 1287 | // not forwarded. That might not be the case here. It is the case that |
duke@435 | 1288 | // the caller observed it to be not forwarded at some time in the past. |
duke@435 | 1289 | assert(is_in_reserved(old), "shouldn't be scavenging this oop"); |
duke@435 | 1290 | |
duke@435 | 1291 | // The sequential code read "old->age()" below. That doesn't work here, |
duke@435 | 1292 | // since the age is in the mark word, and that might be overwritten with |
duke@435 | 1293 | // a forwarding pointer by a parallel thread. So we must save the mark |
duke@435 | 1294 | // word here, install it in a local oopDesc, and then analyze it. |
duke@435 | 1295 | oopDesc dummyOld; |
duke@435 | 1296 | dummyOld.set_mark(m); |
duke@435 | 1297 | assert(!dummyOld.is_forwarded(), |
duke@435 | 1298 | "should not be called with forwarding pointer mark word."); |
duke@435 | 1299 | |
duke@435 | 1300 | bool failed_to_promote = false; |
duke@435 | 1301 | oop new_obj = NULL; |
duke@435 | 1302 | oop forward_ptr; |
duke@435 | 1303 | |
duke@435 | 1304 | // Try allocating obj in to-space (unless too old) |
duke@435 | 1305 | if (dummyOld.age() < tenuring_threshold()) { |
duke@435 | 1306 | new_obj = (oop)par_scan_state->alloc_in_to_space(sz); |
duke@435 | 1307 | if (new_obj == NULL) { |
duke@435 | 1308 | set_survivor_overflow(true); |
duke@435 | 1309 | } |
duke@435 | 1310 | } |
duke@435 | 1311 | |
duke@435 | 1312 | if (new_obj == NULL) { |
duke@435 | 1313 | // Either to-space is full or we decided to promote |
duke@435 | 1314 | // try allocating obj tenured |
duke@435 | 1315 | new_obj = _next_gen->par_promote(par_scan_state->thread_num(), |
duke@435 | 1316 | old, m, sz); |
duke@435 | 1317 | |
duke@435 | 1318 | if (new_obj == NULL) { |
duke@435 | 1319 | // promotion failed, forward to self |
duke@435 | 1320 | forward_ptr = old->forward_to_atomic(old); |
duke@435 | 1321 | new_obj = old; |
duke@435 | 1322 | |
duke@435 | 1323 | if (forward_ptr != NULL) { |
duke@435 | 1324 | return forward_ptr; // someone else succeeded |
duke@435 | 1325 | } |
duke@435 | 1326 | |
duke@435 | 1327 | _promotion_failed = true; |
duke@435 | 1328 | failed_to_promote = true; |
duke@435 | 1329 | |
duke@435 | 1330 | preserve_mark_if_necessary(old, m); |
sla@5237 | 1331 | par_scan_state->register_promotion_failure(sz); |
duke@435 | 1332 | } |
duke@435 | 1333 | } else { |
duke@435 | 1334 | // Is in to-space; do copying ourselves. |
duke@435 | 1335 | Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); |
duke@435 | 1336 | // Restore the mark word copied above. |
duke@435 | 1337 | new_obj->set_mark(m); |
duke@435 | 1338 | // Increment age if new_obj still in new generation |
duke@435 | 1339 | new_obj->incr_age(); |
duke@435 | 1340 | par_scan_state->age_table()->add(new_obj, sz); |
duke@435 | 1341 | } |
duke@435 | 1342 | assert(new_obj != NULL, "just checking"); |
duke@435 | 1343 | |
coleenp@4037 | 1344 | #ifndef PRODUCT |
coleenp@4037 | 1345 | // This code must come after the CAS test, or it will print incorrect |
coleenp@4037 | 1346 | // information. |
coleenp@4037 | 1347 | if (TraceScavenge) { |
coleenp@4037 | 1348 | gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}", |
coleenp@4037 | 1349 | is_in_reserved(new_obj) ? "copying" : "tenuring", |
hseigel@5784 | 1350 | new_obj->klass()->internal_name(), (void *)old, (void *)new_obj, new_obj->size()); |
coleenp@4037 | 1351 | } |
coleenp@4037 | 1352 | #endif |
coleenp@4037 | 1353 | |
duke@435 | 1354 | // Now attempt to install the forwarding pointer (atomically). |
duke@435 | 1355 | // We have to copy the mark word before overwriting with forwarding |
duke@435 | 1356 | // ptr, so we can restore it below in the copy. |
duke@435 | 1357 | if (!failed_to_promote) { |
duke@435 | 1358 | forward_ptr = old->forward_to_atomic(new_obj); |
duke@435 | 1359 | } |
duke@435 | 1360 | |
duke@435 | 1361 | if (forward_ptr == NULL) { |
duke@435 | 1362 | oop obj_to_push = new_obj; |
duke@435 | 1363 | if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { |
duke@435 | 1364 | // Length field used as index of next element to be scanned. |
duke@435 | 1365 | // Real length can be obtained from real_forwardee() |
duke@435 | 1366 | arrayOop(old)->set_length(0); |
duke@435 | 1367 | obj_to_push = old; |
duke@435 | 1368 | assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, |
duke@435 | 1369 | "push forwarded object"); |
duke@435 | 1370 | } |
duke@435 | 1371 | // Push it on one of the queues of to-be-scanned objects. |
ysr@969 | 1372 | bool simulate_overflow = false; |
ysr@969 | 1373 | NOT_PRODUCT( |
ysr@969 | 1374 | if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { |
ysr@969 | 1375 | // simulate a stack overflow |
ysr@969 | 1376 | simulate_overflow = true; |
ysr@969 | 1377 | } |
ysr@969 | 1378 | ) |
ysr@969 | 1379 | if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { |
duke@435 | 1380 | // Add stats for overflow pushes. |
ysr@969 | 1381 | push_on_overflow_list(old, par_scan_state); |
jcoomes@2065 | 1382 | TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); |
duke@435 | 1383 | } |
duke@435 | 1384 | |
duke@435 | 1385 | return new_obj; |
duke@435 | 1386 | } |
duke@435 | 1387 | |
duke@435 | 1388 | // Oops. Someone beat us to it. Undo the allocation. Where did we |
duke@435 | 1389 | // allocate it? |
duke@435 | 1390 | if (is_in_reserved(new_obj)) { |
duke@435 | 1391 | // Must be in to_space. |
duke@435 | 1392 | assert(to()->is_in_reserved(new_obj), "Checking"); |
duke@435 | 1393 | par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); |
duke@435 | 1394 | } else { |
duke@435 | 1395 | assert(!_avoid_promotion_undo, "Should not be here if avoiding."); |
duke@435 | 1396 | _next_gen->par_promote_alloc_undo(par_scan_state->thread_num(), |
duke@435 | 1397 | (HeapWord*)new_obj, sz); |
duke@435 | 1398 | } |
duke@435 | 1399 | |
duke@435 | 1400 | return forward_ptr; |
duke@435 | 1401 | } |
duke@435 | 1402 | |
ysr@969 | 1403 | #ifndef PRODUCT |
ysr@969 | 1404 | // It's OK to call this multi-threaded; the worst thing |
ysr@969 | 1405 | // that can happen is that we'll get a bunch of closely |
ysr@969 | 1406 | // spaced simulated oveflows, but that's OK, in fact |
ysr@969 | 1407 | // probably good as it would exercise the overflow code |
ysr@969 | 1408 | // under contention. |
ysr@969 | 1409 | bool ParNewGeneration::should_simulate_overflow() { |
ysr@969 | 1410 | if (_overflow_counter-- <= 0) { // just being defensive |
ysr@969 | 1411 | _overflow_counter = ParGCWorkQueueOverflowInterval; |
ysr@969 | 1412 | return true; |
ysr@969 | 1413 | } else { |
ysr@969 | 1414 | return false; |
ysr@969 | 1415 | } |
ysr@969 | 1416 | } |
ysr@969 | 1417 | #endif |
ysr@969 | 1418 | |
ysr@1114 | 1419 | // In case we are using compressed oops, we need to be careful. |
ysr@1114 | 1420 | // If the object being pushed is an object array, then its length |
ysr@1114 | 1421 | // field keeps track of the "grey boundary" at which the next |
ysr@1114 | 1422 | // incremental scan will be done (see ParGCArrayScanChunk). |
ysr@1114 | 1423 | // When using compressed oops, this length field is kept in the |
ysr@1114 | 1424 | // lower 32 bits of the erstwhile klass word and cannot be used |
ysr@1114 | 1425 | // for the overflow chaining pointer (OCP below). As such the OCP |
ysr@1114 | 1426 | // would itself need to be compressed into the top 32-bits in this |
ysr@1114 | 1427 | // case. Unfortunately, see below, in the event that we have a |
ysr@1114 | 1428 | // promotion failure, the node to be pushed on the list can be |
ysr@1114 | 1429 | // outside of the Java heap, so the heap-based pointer compression |
ysr@1114 | 1430 | // would not work (we would have potential aliasing between C-heap |
ysr@1114 | 1431 | // and Java-heap pointers). For this reason, when using compressed |
ysr@1114 | 1432 | // oops, we simply use a worker-thread-local, non-shared overflow |
ysr@1114 | 1433 | // list in the form of a growable array, with a slightly different |
ysr@1114 | 1434 | // overflow stack draining strategy. If/when we start using fat |
ysr@1114 | 1435 | // stacks here, we can go back to using (fat) pointer chains |
ysr@1114 | 1436 | // (although some performance comparisons would be useful since |
ysr@1114 | 1437 | // single global lists have their own performance disadvantages |
ysr@1114 | 1438 | // as we were made painfully aware not long ago, see 6786503). |
hseigel@5784 | 1439 | #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) |
ysr@969 | 1440 | void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) { |
ysr@1114 | 1441 | assert(is_in_reserved(from_space_obj), "Should be from this generation"); |
ysr@1130 | 1442 | if (ParGCUseLocalOverflow) { |
ysr@1114 | 1443 | // In the case of compressed oops, we use a private, not-shared |
ysr@1114 | 1444 | // overflow stack. |
ysr@1114 | 1445 | par_scan_state->push_on_overflow_stack(from_space_obj); |
ysr@1114 | 1446 | } else { |
ysr@1130 | 1447 | assert(!UseCompressedOops, "Error"); |
ysr@1114 | 1448 | // if the object has been forwarded to itself, then we cannot |
ysr@1114 | 1449 | // use the klass pointer for the linked list. Instead we have |
ysr@1114 | 1450 | // to allocate an oopDesc in the C-Heap and use that for the linked list. |
ysr@1114 | 1451 | // XXX This is horribly inefficient when a promotion failure occurs |
ysr@1114 | 1452 | // and should be fixed. XXX FIX ME !!! |
ysr@969 | 1453 | #ifndef PRODUCT |
ysr@1114 | 1454 | Atomic::inc_ptr(&_num_par_pushes); |
ysr@1114 | 1455 | assert(_num_par_pushes > 0, "Tautology"); |
ysr@969 | 1456 | #endif |
ysr@1114 | 1457 | if (from_space_obj->forwardee() == from_space_obj) { |
zgu@3900 | 1458 | oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC); |
ysr@1114 | 1459 | listhead->forward_to(from_space_obj); |
ysr@1114 | 1460 | from_space_obj = listhead; |
ysr@1114 | 1461 | } |
ysr@1114 | 1462 | oop observed_overflow_list = _overflow_list; |
ysr@1114 | 1463 | oop cur_overflow_list; |
ysr@1114 | 1464 | do { |
ysr@1114 | 1465 | cur_overflow_list = observed_overflow_list; |
ysr@1114 | 1466 | if (cur_overflow_list != BUSY) { |
ysr@1114 | 1467 | from_space_obj->set_klass_to_list_ptr(cur_overflow_list); |
ysr@1114 | 1468 | } else { |
ysr@1114 | 1469 | from_space_obj->set_klass_to_list_ptr(NULL); |
ysr@1114 | 1470 | } |
ysr@1114 | 1471 | observed_overflow_list = |
ysr@1114 | 1472 | (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list); |
ysr@1114 | 1473 | } while (cur_overflow_list != observed_overflow_list); |
duke@435 | 1474 | } |
duke@435 | 1475 | } |
duke@435 | 1476 | |
ysr@1114 | 1477 | bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) { |
ysr@1114 | 1478 | bool res; |
ysr@1114 | 1479 | |
ysr@1130 | 1480 | if (ParGCUseLocalOverflow) { |
ysr@1114 | 1481 | res = par_scan_state->take_from_overflow_stack(); |
ysr@1114 | 1482 | } else { |
ysr@1130 | 1483 | assert(!UseCompressedOops, "Error"); |
ysr@1114 | 1484 | res = take_from_overflow_list_work(par_scan_state); |
ysr@1114 | 1485 | } |
ysr@1114 | 1486 | return res; |
ysr@1114 | 1487 | } |
ysr@1114 | 1488 | |
ysr@1114 | 1489 | |
ysr@969 | 1490 | // *NOTE*: The overflow list manipulation code here and |
ysr@969 | 1491 | // in CMSCollector:: are very similar in shape, |
ysr@969 | 1492 | // except that in the CMS case we thread the objects |
ysr@969 | 1493 | // directly into the list via their mark word, and do |
ysr@969 | 1494 | // not need to deal with special cases below related |
ysr@969 | 1495 | // to chunking of object arrays and promotion failure |
ysr@969 | 1496 | // handling. |
ysr@969 | 1497 | // CR 6797058 has been filed to attempt consolidation of |
ysr@969 | 1498 | // the common code. |
ysr@969 | 1499 | // Because of the common code, if you make any changes in |
ysr@969 | 1500 | // the code below, please check the CMS version to see if |
ysr@969 | 1501 | // similar changes might be needed. |
ysr@969 | 1502 | // See CMSCollector::par_take_from_overflow_list() for |
ysr@969 | 1503 | // more extensive documentation comments. |
ysr@1114 | 1504 | bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) { |
duke@435 | 1505 | ObjToScanQueue* work_q = par_scan_state->work_queue(); |
duke@435 | 1506 | // How many to take? |
ysr@1114 | 1507 | size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, |
ysr@969 | 1508 | (size_t)ParGCDesiredObjsFromOverflowList); |
duke@435 | 1509 | |
jcoomes@2191 | 1510 | assert(!UseCompressedOops, "Error"); |
ysr@1114 | 1511 | assert(par_scan_state->overflow_stack() == NULL, "Error"); |
duke@435 | 1512 | if (_overflow_list == NULL) return false; |
duke@435 | 1513 | |
duke@435 | 1514 | // Otherwise, there was something there; try claiming the list. |
hseigel@5784 | 1515 | oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); |
ysr@969 | 1516 | // Trim off a prefix of at most objsFromOverflow items |
ysr@969 | 1517 | Thread* tid = Thread::current(); |
ysr@969 | 1518 | size_t spin_count = (size_t)ParallelGCThreads; |
ysr@969 | 1519 | size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100); |
ysr@969 | 1520 | for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) { |
ysr@969 | 1521 | // someone grabbed it before we did ... |
ysr@969 | 1522 | // ... we spin for a short while... |
ysr@969 | 1523 | os::sleep(tid, sleep_time_millis, false); |
ysr@969 | 1524 | if (_overflow_list == NULL) { |
ysr@969 | 1525 | // nothing left to take |
ysr@969 | 1526 | return false; |
ysr@969 | 1527 | } else if (_overflow_list != BUSY) { |
ysr@969 | 1528 | // try and grab the prefix |
hseigel@5784 | 1529 | prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); |
ysr@969 | 1530 | } |
duke@435 | 1531 | } |
ysr@969 | 1532 | if (prefix == NULL || prefix == BUSY) { |
ysr@969 | 1533 | // Nothing to take or waited long enough |
ysr@969 | 1534 | if (prefix == NULL) { |
ysr@969 | 1535 | // Write back the NULL in case we overwrote it with BUSY above |
ysr@969 | 1536 | // and it is still the same value. |
ysr@969 | 1537 | (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); |
ysr@969 | 1538 | } |
ysr@969 | 1539 | return false; |
ysr@969 | 1540 | } |
ysr@969 | 1541 | assert(prefix != NULL && prefix != BUSY, "Error"); |
ysr@969 | 1542 | size_t i = 1; |
duke@435 | 1543 | oop cur = prefix; |
coleenp@602 | 1544 | while (i < objsFromOverflow && cur->klass_or_null() != NULL) { |
coleenp@4037 | 1545 | i++; cur = cur->list_ptr_from_klass(); |
duke@435 | 1546 | } |
duke@435 | 1547 | |
duke@435 | 1548 | // Reattach remaining (suffix) to overflow list |
ysr@969 | 1549 | if (cur->klass_or_null() == NULL) { |
ysr@969 | 1550 | // Write back the NULL in lieu of the BUSY we wrote |
ysr@969 | 1551 | // above and it is still the same value. |
ysr@969 | 1552 | if (_overflow_list == BUSY) { |
ysr@969 | 1553 | (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); |
duke@435 | 1554 | } |
ysr@969 | 1555 | } else { |
coleenp@4037 | 1556 | assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error"); |
coleenp@4037 | 1557 | oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list |
ysr@969 | 1558 | cur->set_klass_to_list_ptr(NULL); // break off suffix |
ysr@969 | 1559 | // It's possible that the list is still in the empty(busy) state |
ysr@969 | 1560 | // we left it in a short while ago; in that case we may be |
ysr@969 | 1561 | // able to place back the suffix. |
ysr@969 | 1562 | oop observed_overflow_list = _overflow_list; |
ysr@969 | 1563 | oop cur_overflow_list = observed_overflow_list; |
ysr@969 | 1564 | bool attached = false; |
ysr@969 | 1565 | while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { |
ysr@969 | 1566 | observed_overflow_list = |
ysr@969 | 1567 | (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); |
ysr@969 | 1568 | if (cur_overflow_list == observed_overflow_list) { |
ysr@969 | 1569 | attached = true; |
ysr@969 | 1570 | break; |
ysr@969 | 1571 | } else cur_overflow_list = observed_overflow_list; |
ysr@969 | 1572 | } |
ysr@969 | 1573 | if (!attached) { |
ysr@969 | 1574 | // Too bad, someone else got in in between; we'll need to do a splice. |
ysr@969 | 1575 | // Find the last item of suffix list |
ysr@969 | 1576 | oop last = suffix; |
ysr@969 | 1577 | while (last->klass_or_null() != NULL) { |
coleenp@4037 | 1578 | last = last->list_ptr_from_klass(); |
ysr@969 | 1579 | } |
ysr@969 | 1580 | // Atomically prepend suffix to current overflow list |
ysr@969 | 1581 | observed_overflow_list = _overflow_list; |
ysr@969 | 1582 | do { |
ysr@969 | 1583 | cur_overflow_list = observed_overflow_list; |
ysr@969 | 1584 | if (cur_overflow_list != BUSY) { |
ysr@969 | 1585 | // Do the splice ... |
ysr@969 | 1586 | last->set_klass_to_list_ptr(cur_overflow_list); |
ysr@969 | 1587 | } else { // cur_overflow_list == BUSY |
ysr@969 | 1588 | last->set_klass_to_list_ptr(NULL); |
ysr@969 | 1589 | } |
ysr@969 | 1590 | observed_overflow_list = |
ysr@969 | 1591 | (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); |
ysr@969 | 1592 | } while (cur_overflow_list != observed_overflow_list); |
duke@435 | 1593 | } |
duke@435 | 1594 | } |
duke@435 | 1595 | |
duke@435 | 1596 | // Push objects on prefix list onto this thread's work queue |
ysr@969 | 1597 | assert(prefix != NULL && prefix != BUSY, "program logic"); |
duke@435 | 1598 | cur = prefix; |
ysr@969 | 1599 | ssize_t n = 0; |
duke@435 | 1600 | while (cur != NULL) { |
duke@435 | 1601 | oop obj_to_push = cur->forwardee(); |
coleenp@4037 | 1602 | oop next = cur->list_ptr_from_klass(); |
duke@435 | 1603 | cur->set_klass(obj_to_push->klass()); |
ysr@969 | 1604 | // This may be an array object that is self-forwarded. In that case, the list pointer |
ysr@969 | 1605 | // space, cur, is not in the Java heap, but rather in the C-heap and should be freed. |
ysr@969 | 1606 | if (!is_in_reserved(cur)) { |
ysr@969 | 1607 | // This can become a scaling bottleneck when there is work queue overflow coincident |
ysr@969 | 1608 | // with promotion failure. |
ysr@969 | 1609 | oopDesc* f = cur; |
zgu@3900 | 1610 | FREE_C_HEAP_ARRAY(oopDesc, f, mtGC); |
ysr@969 | 1611 | } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) { |
ysr@969 | 1612 | assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); |
duke@435 | 1613 | obj_to_push = cur; |
duke@435 | 1614 | } |
ysr@969 | 1615 | bool ok = work_q->push(obj_to_push); |
ysr@969 | 1616 | assert(ok, "Should have succeeded"); |
duke@435 | 1617 | cur = next; |
duke@435 | 1618 | n++; |
duke@435 | 1619 | } |
jcoomes@2065 | 1620 | TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n)); |
ysr@969 | 1621 | #ifndef PRODUCT |
ysr@969 | 1622 | assert(_num_par_pushes >= n, "Too many pops?"); |
ysr@969 | 1623 | Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); |
ysr@969 | 1624 | #endif |
duke@435 | 1625 | return true; |
duke@435 | 1626 | } |
ysr@969 | 1627 | #undef BUSY |
duke@435 | 1628 | |
sla@5237 | 1629 | void ParNewGeneration::ref_processor_init() { |
duke@435 | 1630 | if (_ref_processor == NULL) { |
duke@435 | 1631 | // Allocate and initialize a reference processor |
ysr@2651 | 1632 | _ref_processor = |
ysr@2651 | 1633 | new ReferenceProcessor(_reserved, // span |
ysr@2651 | 1634 | ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing |
ysr@2651 | 1635 | (int) ParallelGCThreads, // mt processing degree |
ysr@2651 | 1636 | refs_discovery_is_mt(), // mt discovery |
ysr@2651 | 1637 | (int) ParallelGCThreads, // mt discovery degree |
ysr@2651 | 1638 | refs_discovery_is_atomic(), // atomic_discovery |
ysr@2651 | 1639 | NULL, // is_alive_non_header |
ysr@2651 | 1640 | false); // write barrier for next field updates |
duke@435 | 1641 | } |
duke@435 | 1642 | } |
duke@435 | 1643 | |
duke@435 | 1644 | const char* ParNewGeneration::name() const { |
duke@435 | 1645 | return "par new generation"; |
duke@435 | 1646 | } |