Mercurial > hg > graal-compiler
annotate src/share/vm/gc_implementation/parNew/parNewGeneration.cpp @ 6197:d2a62e0f25eb
6995781: Native Memory Tracking (Phase 1)
7151532: DCmd for hotspot native memory tracking
Summary: Implementation of native memory tracking phase 1, which tracks VM native memory usage, and related DCmd
Reviewed-by: acorn, coleenp, fparain
| author | zgu |
|---|---|
| date | Thu, 28 Jun 2012 17:03:16 -0400 |
| parents | 9d679effd28c |
| children | aaf61e68b255 |
| rev | line source |
|---|---|
| 0 | 1 /* |
| 4911 | 2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved. |
| 0 | 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 * | |
|
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1387
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
|
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1387
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
|
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1387
diff
changeset
|
21 * questions. |
| 0 | 22 * |
| 23 */ | |
| 24 | |
| 1972 | 25 #include "precompiled.hpp" |
| 26 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp" | |
| 27 #include "gc_implementation/parNew/parGCAllocBuffer.hpp" | |
| 28 #include "gc_implementation/parNew/parNewGeneration.hpp" | |
| 29 #include "gc_implementation/parNew/parOopClosures.inline.hpp" | |
| 30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp" | |
| 31 #include "gc_implementation/shared/ageTable.hpp" | |
| 32 #include "gc_implementation/shared/spaceDecorator.hpp" | |
| 33 #include "memory/defNewGeneration.inline.hpp" | |
| 34 #include "memory/genCollectedHeap.hpp" | |
| 35 #include "memory/genOopClosures.inline.hpp" | |
| 36 #include "memory/generation.hpp" | |
| 37 #include "memory/generation.inline.hpp" | |
| 38 #include "memory/referencePolicy.hpp" | |
| 39 #include "memory/resourceArea.hpp" | |
| 40 #include "memory/sharedHeap.hpp" | |
| 41 #include "memory/space.hpp" | |
| 42 #include "oops/objArrayOop.hpp" | |
| 43 #include "oops/oop.inline.hpp" | |
| 44 #include "oops/oop.pcgc.inline.hpp" | |
| 45 #include "runtime/handles.hpp" | |
| 46 #include "runtime/handles.inline.hpp" | |
| 47 #include "runtime/java.hpp" | |
| 48 #include "runtime/thread.hpp" | |
| 49 #include "utilities/copy.hpp" | |
| 50 #include "utilities/globalDefinitions.hpp" | |
| 51 #include "utilities/workgroup.hpp" | |
| 0 | 52 |
| 53 #ifdef _MSC_VER | |
| 54 #pragma warning( push ) | |
| 55 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list | |
| 56 #endif | |
| 57 ParScanThreadState::ParScanThreadState(Space* to_space_, | |
| 58 ParNewGeneration* gen_, | |
| 59 Generation* old_gen_, | |
| 60 int thread_num_, | |
| 61 ObjToScanQueueSet* work_queue_set_, | |
| 6197 | 62 Stack<oop, mtGC>* overflow_stacks_, |
| 0 | 63 size_t desired_plab_sz_, |
| 64 ParallelTaskTerminator& term_) : | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
65 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_), |
| 0 | 66 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false), |
|
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
67 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL), |
| 0 | 68 _ageTable(false), // false ==> not the global age table, no perf data. |
| 69 _to_space_alloc_buffer(desired_plab_sz_), | |
| 70 _to_space_closure(gen_, this), _old_gen_closure(gen_, this), | |
| 71 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this), | |
| 72 _older_gen_closure(gen_, this), | |
| 73 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure, | |
| 74 &_to_space_root_closure, gen_, &_old_gen_root_closure, | |
| 75 work_queue_set_, &term_), | |
| 76 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this), | |
| 77 _keep_alive_closure(&_scan_weak_ref_closure), | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
78 _promotion_failure_size(0), |
| 0 | 79 _strong_roots_time(0.0), _term_time(0.0) |
| 80 { | |
| 1710 | 81 #if TASKQUEUE_STATS |
| 82 _term_attempts = 0; | |
| 83 _overflow_refills = 0; | |
| 84 _overflow_refill_objs = 0; | |
| 85 #endif // TASKQUEUE_STATS | |
| 86 | |
| 0 | 87 _survivor_chunk_array = |
| 88 (ChunkArray*) old_gen()->get_data_recorder(thread_num()); | |
| 89 _hash_seed = 17; // Might want to take time-based random value. | |
| 90 _start = os::elapsedTime(); | |
| 91 _old_gen_closure.set_generation(old_gen_); | |
| 92 _old_gen_root_closure.set_generation(old_gen_); | |
| 93 } | |
| 94 #ifdef _MSC_VER | |
| 95 #pragma warning( pop ) | |
| 96 #endif | |
| 97 | |
| 98 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start, | |
| 99 size_t plab_word_size) { | |
| 100 ChunkArray* sca = survivor_chunk_array(); | |
| 101 if (sca != NULL) { | |
| 102 // A non-null SCA implies that we want the PLAB data recorded. | |
| 103 sca->record_sample(plab_start, plab_word_size); | |
| 104 } | |
| 105 } | |
| 106 | |
| 107 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const { | |
| 108 return new_obj->is_objArray() && | |
| 109 arrayOop(new_obj)->length() > ParGCArrayScanChunk && | |
| 110 new_obj != old_obj; | |
| 111 } | |
| 112 | |
| 113 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) { | |
| 114 assert(old->is_objArray(), "must be obj array"); | |
| 115 assert(old->is_forwarded(), "must be forwarded"); | |
| 116 assert(Universe::heap()->is_in_reserved(old), "must be in heap."); | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
117 assert(!old_gen()->is_in(old), "must be in young generation."); |
| 0 | 118 |
| 119 objArrayOop obj = objArrayOop(old->forwardee()); | |
| 120 // Process ParGCArrayScanChunk elements now | |
| 121 // and push the remainder back onto queue | |
| 122 int start = arrayOop(old)->length(); | |
| 123 int end = obj->length(); | |
| 124 int remainder = end - start; | |
| 125 assert(start <= end, "just checking"); | |
| 126 if (remainder > 2 * ParGCArrayScanChunk) { | |
| 127 // Test above combines last partial chunk with a full chunk | |
| 128 end = start + ParGCArrayScanChunk; | |
| 129 arrayOop(old)->set_length(end); | |
| 130 // Push remainder. | |
| 131 bool ok = work_queue()->push(old); | |
| 132 assert(ok, "just popped, push must be okay"); | |
| 133 } else { | |
| 134 // Restore length so that it can be used if there | |
| 135 // is a promotion failure and forwarding pointers | |
| 136 // must be removed. | |
| 137 arrayOop(old)->set_length(end); | |
| 138 } | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
139 |
| 0 | 140 // process our set of indices (include header in first chunk) |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
141 // should make sure end is even (aligned to HeapWord in case of compressed oops) |
| 0 | 142 if ((HeapWord *)obj < young_old_boundary()) { |
| 143 // object is in to_space | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
144 obj->oop_iterate_range(&_to_space_closure, start, end); |
| 0 | 145 } else { |
| 146 // object is in old generation | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
147 obj->oop_iterate_range(&_old_gen_closure, start, end); |
| 0 | 148 } |
| 149 } | |
| 150 | |
| 151 | |
| 152 void ParScanThreadState::trim_queues(int max_size) { | |
| 153 ObjToScanQueue* queue = work_queue(); | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
154 do { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
155 while (queue->size() > (juint)max_size) { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
156 oop obj_to_scan; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
157 if (queue->pop_local(obj_to_scan)) { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
158 if ((HeapWord *)obj_to_scan < young_old_boundary()) { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
159 if (obj_to_scan->is_objArray() && |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
160 obj_to_scan->is_forwarded() && |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
161 obj_to_scan->forwardee() != obj_to_scan) { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
162 scan_partial_array_and_push_remainder(obj_to_scan); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
163 } else { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
164 // object is in to_space |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
165 obj_to_scan->oop_iterate(&_to_space_closure); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
166 } |
| 0 | 167 } else { |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
168 // object is in old generation |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
169 obj_to_scan->oop_iterate(&_old_gen_closure); |
| 0 | 170 } |
| 171 } | |
| 172 } | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
173 // For the case of compressed oops, we have a private, non-shared |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
174 // overflow stack, so we eagerly drain it so as to more evenly |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
175 // distribute load early. Note: this may be good to do in |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
176 // general rather than delay for the final stealing phase. |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
177 // If applicable, we'll transfer a set of objects over to our |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
178 // work queue, allowing them to be stolen and draining our |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
179 // private overflow stack. |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
180 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this)); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
181 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
182 |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
183 bool ParScanThreadState::take_from_overflow_stack() { |
| 695 | 184 assert(ParGCUseLocalOverflow, "Else should not call"); |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
185 assert(young_gen()->overflow_list() == NULL, "Error"); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
186 ObjToScanQueue* queue = work_queue(); |
| 6197 | 187 Stack<oop, mtGC>* const of_stack = overflow_stack(); |
|
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
188 const size_t num_overflow_elems = of_stack->size(); |
|
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
189 const size_t space_available = queue->max_elems() - queue->size(); |
|
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
190 const size_t num_take_elems = MIN3(space_available / 4, |
|
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
191 ParGCDesiredObjsFromOverflowList, |
|
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
192 num_overflow_elems); |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
193 // Transfer the most recent num_take_elems from the overflow |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
194 // stack to our work queue. |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
195 for (size_t i = 0; i != num_take_elems; i++) { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
196 oop cur = of_stack->pop(); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
197 oop obj_to_push = cur->forwardee(); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
198 assert(Universe::heap()->is_in_reserved(cur), "Should be in heap"); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
199 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen"); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
200 assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap"); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
201 if (should_be_partially_scanned(obj_to_push, cur)) { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
202 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
203 obj_to_push = cur; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
204 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
205 bool ok = queue->push(obj_to_push); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
206 assert(ok, "Should have succeeded"); |
| 0 | 207 } |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
208 assert(young_gen()->overflow_list() == NULL, "Error"); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
209 return num_take_elems > 0; // was something transferred? |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
210 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
211 |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
212 void ParScanThreadState::push_on_overflow_stack(oop p) { |
| 695 | 213 assert(ParGCUseLocalOverflow, "Else should not call"); |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
214 overflow_stack()->push(p); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
215 assert(young_gen()->overflow_list() == NULL, "Error"); |
| 0 | 216 } |
| 217 | |
| 218 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) { | |
| 219 | |
| 220 // Otherwise, if the object is small enough, try to reallocate the | |
| 221 // buffer. | |
| 222 HeapWord* obj = NULL; | |
| 223 if (!_to_space_full) { | |
| 224 ParGCAllocBuffer* const plab = to_space_alloc_buffer(); | |
| 225 Space* const sp = to_space(); | |
| 226 if (word_sz * 100 < | |
| 227 ParallelGCBufferWastePct * plab->word_sz()) { | |
| 228 // Is small enough; abandon this buffer and start a new one. | |
| 229 plab->retire(false, false); | |
| 230 size_t buf_size = plab->word_sz(); | |
| 231 HeapWord* buf_space = sp->par_allocate(buf_size); | |
| 232 if (buf_space == NULL) { | |
| 233 const size_t min_bytes = | |
| 234 ParGCAllocBuffer::min_size() << LogHeapWordSize; | |
| 235 size_t free_bytes = sp->free(); | |
| 236 while(buf_space == NULL && free_bytes >= min_bytes) { | |
| 237 buf_size = free_bytes >> LogHeapWordSize; | |
| 238 assert(buf_size == (size_t)align_object_size(buf_size), | |
| 239 "Invariant"); | |
| 240 buf_space = sp->par_allocate(buf_size); | |
| 241 free_bytes = sp->free(); | |
| 242 } | |
| 243 } | |
| 244 if (buf_space != NULL) { | |
| 245 plab->set_word_size(buf_size); | |
| 246 plab->set_buf(buf_space); | |
| 247 record_survivor_plab(buf_space, buf_size); | |
| 248 obj = plab->allocate(word_sz); | |
| 249 // Note that we cannot compare buf_size < word_sz below | |
| 250 // because of AlignmentReserve (see ParGCAllocBuffer::allocate()). | |
| 251 assert(obj != NULL || plab->words_remaining() < word_sz, | |
| 252 "Else should have been able to allocate"); | |
| 253 // It's conceivable that we may be able to use the | |
| 254 // buffer we just grabbed for subsequent small requests | |
| 255 // even if not for this one. | |
| 256 } else { | |
| 257 // We're used up. | |
| 258 _to_space_full = true; | |
| 259 } | |
| 260 | |
| 261 } else { | |
| 262 // Too large; allocate the object individually. | |
| 263 obj = sp->par_allocate(word_sz); | |
| 264 } | |
| 265 } | |
| 266 return obj; | |
| 267 } | |
| 268 | |
| 269 | |
| 270 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, | |
| 271 size_t word_sz) { | |
| 272 // Is the alloc in the current alloc buffer? | |
| 273 if (to_space_alloc_buffer()->contains(obj)) { | |
| 274 assert(to_space_alloc_buffer()->contains(obj + word_sz - 1), | |
| 275 "Should contain whole object."); | |
| 276 to_space_alloc_buffer()->undo_allocation(obj, word_sz); | |
| 277 } else { | |
|
481
7d7a7c599c17
6578152: fill_region_with_object has usability and safety issues
jcoomes
parents:
457
diff
changeset
|
278 CollectedHeap::fill_with_object(obj, word_sz); |
| 0 | 279 } |
| 280 } | |
| 281 | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
282 void ParScanThreadState::print_and_clear_promotion_failure_size() { |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
283 if (_promotion_failure_size != 0) { |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
284 if (PrintPromotionFailure) { |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
285 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ", |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
286 _thread_num, _promotion_failure_size); |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
287 } |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
288 _promotion_failure_size = 0; |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
289 } |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
290 } |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
291 |
| 0 | 292 class ParScanThreadStateSet: private ResourceArray { |
| 293 public: | |
| 294 // Initializes states for the specified number of threads; | |
| 295 ParScanThreadStateSet(int num_threads, | |
| 296 Space& to_space, | |
| 297 ParNewGeneration& gen, | |
| 298 Generation& old_gen, | |
| 299 ObjToScanQueueSet& queue_set, | |
| 6197 | 300 Stack<oop, mtGC>* overflow_stacks_, |
| 0 | 301 size_t desired_plab_sz, |
| 302 ParallelTaskTerminator& term); | |
| 1710 | 303 |
| 304 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); } | |
| 305 | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
306 inline ParScanThreadState& thread_state(int i); |
| 1710 | 307 |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
308 void reset(int active_workers, bool promotion_failed); |
| 0 | 309 void flush(); |
| 1710 | 310 |
| 311 #if TASKQUEUE_STATS | |
| 312 static void | |
| 313 print_termination_stats_hdr(outputStream* const st = gclog_or_tty); | |
| 314 void print_termination_stats(outputStream* const st = gclog_or_tty); | |
| 315 static void | |
| 316 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty); | |
| 317 void print_taskqueue_stats(outputStream* const st = gclog_or_tty); | |
| 318 void reset_stats(); | |
| 319 #endif // TASKQUEUE_STATS | |
| 320 | |
| 0 | 321 private: |
| 322 ParallelTaskTerminator& _term; | |
| 323 ParNewGeneration& _gen; | |
| 324 Generation& _next_gen; | |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
325 public: |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
326 bool is_valid(int id) const { return id < length(); } |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
327 ParallelTaskTerminator* terminator() { return &_term; } |
| 0 | 328 }; |
| 329 | |
| 330 | |
| 331 ParScanThreadStateSet::ParScanThreadStateSet( | |
| 332 int num_threads, Space& to_space, ParNewGeneration& gen, | |
| 333 Generation& old_gen, ObjToScanQueueSet& queue_set, | |
| 6197 | 334 Stack<oop, mtGC>* overflow_stacks, |
| 0 | 335 size_t desired_plab_sz, ParallelTaskTerminator& term) |
| 336 : ResourceArray(sizeof(ParScanThreadState), num_threads), | |
| 1710 | 337 _gen(gen), _next_gen(old_gen), _term(term) |
| 0 | 338 { |
| 339 assert(num_threads > 0, "sanity check!"); | |
|
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
340 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL), |
|
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
341 "overflow_stack allocation mismatch"); |
| 0 | 342 // Initialize states. |
| 343 for (int i = 0; i < num_threads; ++i) { | |
| 344 new ((ParScanThreadState*)_data + i) | |
| 345 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set, | |
|
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
346 overflow_stacks, desired_plab_sz, term); |
| 0 | 347 } |
| 348 } | |
| 349 | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
350 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) |
| 0 | 351 { |
| 352 assert(i >= 0 && i < length(), "sanity check!"); | |
| 353 return ((ParScanThreadState*)_data)[i]; | |
| 354 } | |
| 355 | |
| 356 | |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
357 void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed) |
| 0 | 358 { |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
359 _term.reset_for_reuse(active_threads); |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
360 if (promotion_failed) { |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
361 for (int i = 0; i < length(); ++i) { |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
362 thread_state(i).print_and_clear_promotion_failure_size(); |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
363 } |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
364 } |
| 0 | 365 } |
| 366 | |
| 1710 | 367 #if TASKQUEUE_STATS |
| 368 void | |
| 369 ParScanThreadState::reset_stats() | |
| 370 { | |
| 371 taskqueue_stats().reset(); | |
| 372 _term_attempts = 0; | |
| 373 _overflow_refills = 0; | |
| 374 _overflow_refill_objs = 0; | |
| 375 } | |
| 376 | |
| 377 void ParScanThreadStateSet::reset_stats() | |
| 378 { | |
| 379 for (int i = 0; i < length(); ++i) { | |
| 380 thread_state(i).reset_stats(); | |
| 381 } | |
| 382 } | |
| 383 | |
| 384 void | |
| 385 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) | |
| 386 { | |
| 387 st->print_raw_cr("GC Termination Stats"); | |
| 388 st->print_raw_cr(" elapsed --strong roots-- " | |
| 389 "-------termination-------"); | |
| 390 st->print_raw_cr("thr ms ms % " | |
| 391 " ms % attempts"); | |
| 392 st->print_raw_cr("--- --------- --------- ------ " | |
| 393 "--------- ------ --------"); | |
| 394 } | |
| 395 | |
| 396 void ParScanThreadStateSet::print_termination_stats(outputStream* const st) | |
| 397 { | |
| 398 print_termination_stats_hdr(st); | |
| 399 | |
| 400 for (int i = 0; i < length(); ++i) { | |
| 401 const ParScanThreadState & pss = thread_state(i); | |
| 402 const double elapsed_ms = pss.elapsed_time() * 1000.0; | |
| 403 const double s_roots_ms = pss.strong_roots_time() * 1000.0; | |
| 404 const double term_ms = pss.term_time() * 1000.0; | |
| 405 st->print_cr("%3d %9.2f %9.2f %6.2f " | |
| 406 "%9.2f %6.2f " SIZE_FORMAT_W(8), | |
| 407 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, | |
| 408 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts()); | |
| 409 } | |
| 410 } | |
| 411 | |
| 412 // Print stats related to work queue activity. | |
| 413 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) | |
| 414 { | |
| 415 st->print_raw_cr("GC Task Stats"); | |
| 416 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); | |
| 417 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); | |
| 418 } | |
| 419 | |
| 420 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st) | |
| 421 { | |
| 422 print_taskqueue_stats_hdr(st); | |
| 423 | |
| 424 TaskQueueStats totals; | |
| 425 for (int i = 0; i < length(); ++i) { | |
| 426 const ParScanThreadState & pss = thread_state(i); | |
| 427 const TaskQueueStats & stats = pss.taskqueue_stats(); | |
| 428 st->print("%3d ", i); stats.print(st); st->cr(); | |
| 429 totals += stats; | |
| 430 | |
| 431 if (pss.overflow_refills() > 0) { | |
| 432 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills " | |
| 433 SIZE_FORMAT_W(10) " overflow objects", | |
| 434 pss.overflow_refills(), pss.overflow_refill_objs()); | |
| 435 } | |
| 436 } | |
| 437 st->print("tot "); totals.print(st); st->cr(); | |
| 438 | |
| 439 DEBUG_ONLY(totals.verify()); | |
| 440 } | |
| 441 #endif // TASKQUEUE_STATS | |
| 442 | |
| 0 | 443 void ParScanThreadStateSet::flush() |
| 444 { | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
445 // Work in this loop should be kept as lightweight as |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
446 // possible since this might otherwise become a bottleneck |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
447 // to scaling. Should we add heavy-weight work into this |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
448 // loop, consider parallelizing the loop into the worker threads. |
| 0 | 449 for (int i = 0; i < length(); ++i) { |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
450 ParScanThreadState& par_scan_state = thread_state(i); |
| 0 | 451 |
| 452 // Flush stats related to To-space PLAB activity and | |
| 453 // retire the last buffer. | |
| 454 par_scan_state.to_space_alloc_buffer()-> | |
| 455 flush_stats_and_retire(_gen.plab_stats(), | |
| 456 false /* !retain */); | |
| 457 | |
| 458 // Every thread has its own age table. We need to merge | |
| 459 // them all into one. | |
| 460 ageTable *local_table = par_scan_state.age_table(); | |
| 461 _gen.age_table()->merge(local_table); | |
| 462 | |
| 463 // Inform old gen that we're done. | |
| 464 _next_gen.par_promote_alloc_done(i); | |
| 465 _next_gen.par_oop_since_save_marks_iterate_done(i); | |
| 1710 | 466 } |
| 0 | 467 |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
468 if (UseConcMarkSweepGC && ParallelGCThreads > 0) { |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
469 // We need to call this even when ResizeOldPLAB is disabled |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
470 // so as to avoid breaking some asserts. While we may be able |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
471 // to avoid this by reorganizing the code a bit, I am loathe |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
472 // to do that unless we find cases where ergo leads to bad |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
473 // performance. |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
474 CFLS_LAB::compute_desired_plab_size(); |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
475 } |
| 0 | 476 } |
| 477 | |
| 478 ParScanClosure::ParScanClosure(ParNewGeneration* g, | |
| 479 ParScanThreadState* par_scan_state) : | |
| 480 OopsInGenClosure(g), _par_scan_state(par_scan_state), _g(g) | |
| 481 { | |
| 482 assert(_g->level() == 0, "Optimized for youngest generation"); | |
| 483 _boundary = _g->reserved().end(); | |
| 484 } | |
| 485 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
486 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
487 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
488 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
489 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
490 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
491 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
492 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
493 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
494 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
495 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
496 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
497 |
| 0 | 498 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g, |
| 499 ParScanThreadState* par_scan_state) | |
| 500 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state) | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
501 {} |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
502 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
503 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
504 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); } |
| 0 | 505 |
| 506 #ifdef WIN32 | |
| 507 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */ | |
| 508 #endif | |
| 509 | |
| 510 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure( | |
| 511 ParScanThreadState* par_scan_state_, | |
| 512 ParScanWithoutBarrierClosure* to_space_closure_, | |
| 513 ParScanWithBarrierClosure* old_gen_closure_, | |
| 514 ParRootScanWithoutBarrierClosure* to_space_root_closure_, | |
| 515 ParNewGeneration* par_gen_, | |
| 516 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_, | |
| 517 ObjToScanQueueSet* task_queues_, | |
| 518 ParallelTaskTerminator* terminator_) : | |
| 519 | |
| 520 _par_scan_state(par_scan_state_), | |
| 521 _to_space_closure(to_space_closure_), | |
| 522 _old_gen_closure(old_gen_closure_), | |
| 523 _to_space_root_closure(to_space_root_closure_), | |
| 524 _old_gen_root_closure(old_gen_root_closure_), | |
| 525 _par_gen(par_gen_), | |
| 526 _task_queues(task_queues_), | |
| 527 _terminator(terminator_) | |
| 528 {} | |
| 529 | |
| 530 void ParEvacuateFollowersClosure::do_void() { | |
| 531 ObjToScanQueue* work_q = par_scan_state()->work_queue(); | |
| 532 | |
| 533 while (true) { | |
| 534 | |
| 535 // Scan to-space and old-gen objs until we run out of both. | |
| 536 oop obj_to_scan; | |
| 537 par_scan_state()->trim_queues(0); | |
| 538 | |
| 539 // We have no local work, attempt to steal from other threads. | |
| 540 | |
| 541 // attempt to steal work from promoted. | |
| 542 if (task_queues()->steal(par_scan_state()->thread_num(), | |
| 543 par_scan_state()->hash_seed(), | |
| 544 obj_to_scan)) { | |
| 545 bool res = work_q->push(obj_to_scan); | |
| 546 assert(res, "Empty queue should have room for a push."); | |
| 547 | |
| 548 // if successful, goto Start. | |
| 549 continue; | |
| 550 | |
| 551 // try global overflow list. | |
| 552 } else if (par_gen()->take_from_overflow_list(par_scan_state())) { | |
| 553 continue; | |
| 554 } | |
| 555 | |
| 556 // Otherwise, offer termination. | |
| 557 par_scan_state()->start_term_time(); | |
| 558 if (terminator()->offer_termination()) break; | |
| 559 par_scan_state()->end_term_time(); | |
| 560 } | |
| 534 | 561 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0, |
| 562 "Broken overflow list?"); | |
| 0 | 563 // Finish the last termination pause. |
| 564 par_scan_state()->end_term_time(); | |
| 565 } | |
| 566 | |
| 567 ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen, | |
| 568 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) : | |
| 569 AbstractGangTask("ParNewGeneration collection"), | |
| 570 _gen(gen), _next_gen(next_gen), | |
| 571 _young_old_boundary(young_old_boundary), | |
| 572 _state_set(state_set) | |
| 573 {} | |
| 574 | |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
575 // Reset the terminator for the given number of |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
576 // active threads. |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
577 void ParNewGenTask::set_for_termination(int active_workers) { |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
578 _state_set->reset(active_workers, _gen->promotion_failed()); |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
579 // Should the heap be passed in? There's only 1 for now so |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
580 // grab it instead. |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
581 GenCollectedHeap* gch = GenCollectedHeap::heap(); |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
582 gch->set_n_termination(active_workers); |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
583 } |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
584 |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
585 // The "i" passed to this method is the part of the work for |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
586 // this thread. It is not the worker ID. The "i" is derived |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
587 // from _started_workers which is incremented in internal_note_start() |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
588 // called in GangWorker loop() and which is called under the |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
589 // which is called under the protection of the gang monitor and is |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
590 // called after a task is started. So "i" is based on |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
591 // first-come-first-served. |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
592 |
|
4728
441e946dc1af
7121618: Change type of number of GC workers to unsigned int.
jmasa
parents:
4095
diff
changeset
|
593 void ParNewGenTask::work(uint worker_id) { |
| 0 | 594 GenCollectedHeap* gch = GenCollectedHeap::heap(); |
| 595 // Since this is being done in a separate thread, need new resource | |
| 596 // and handle marks. | |
| 597 ResourceMark rm; | |
| 598 HandleMark hm; | |
| 599 // We would need multiple old-gen queues otherwise. | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
600 assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen."); |
| 0 | 601 |
| 602 Generation* old_gen = gch->next_gen(_gen); | |
| 603 | |
|
4728
441e946dc1af
7121618: Change type of number of GC workers to unsigned int.
jmasa
parents:
4095
diff
changeset
|
604 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id); |
|
441e946dc1af
7121618: Change type of number of GC workers to unsigned int.
jmasa
parents:
4095
diff
changeset
|
605 assert(_state_set->is_valid(worker_id), "Should not have been called"); |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
606 |
| 0 | 607 par_scan_state.set_young_old_boundary(_young_old_boundary); |
| 608 | |
| 609 par_scan_state.start_strong_roots(); | |
| 610 gch->gen_process_strong_roots(_gen->level(), | |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
611 true, // Process younger gens, if any, |
|
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
612 // as strong roots. |
|
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
613 false, // no scope; this is parallel code |
|
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
614 false, // not collecting perm generation. |
| 0 | 615 SharedHeap::SO_AllClasses, |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
616 &par_scan_state.to_space_root_closure(), |
|
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
617 true, // walk *all* scavengable nmethods |
|
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
618 &par_scan_state.older_gen_closure()); |
| 0 | 619 par_scan_state.end_strong_roots(); |
| 620 | |
| 621 // "evacuate followers". | |
| 622 par_scan_state.evacuate_followers_closure().do_void(); | |
| 623 } | |
| 624 | |
| 625 #ifdef _MSC_VER | |
| 626 #pragma warning( push ) | |
| 627 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list | |
| 628 #endif | |
| 629 ParNewGeneration:: | |
| 630 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level) | |
| 631 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"), | |
| 632 _overflow_list(NULL), | |
| 633 _is_alive_closure(this), | |
| 634 _plab_stats(YoungPLABSize, PLABWeight) | |
| 635 { | |
| 534 | 636 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;) |
| 637 NOT_PRODUCT(_num_par_pushes = 0;) | |
| 0 | 638 _task_queues = new ObjToScanQueueSet(ParallelGCThreads); |
| 639 guarantee(_task_queues != NULL, "task_queues allocation failure."); | |
| 640 | |
| 641 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) { | |
| 1665 | 642 ObjToScanQueue *q = new ObjToScanQueue(); |
| 643 guarantee(q != NULL, "work_queue Allocation failure."); | |
| 644 _task_queues->register_queue(i1, q); | |
| 0 | 645 } |
| 646 | |
| 647 for (uint i2 = 0; i2 < ParallelGCThreads; i2++) | |
| 648 _task_queues->queue(i2)->initialize(); | |
| 649 | |
|
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
650 _overflow_stacks = NULL; |
|
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
651 if (ParGCUseLocalOverflow) { |
| 6197 | 652 |
| 653 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal | |
| 654 // with ',' | |
| 655 typedef Stack<oop, mtGC> GCOopStack; | |
| 656 | |
| 657 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC); | |
|
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
658 for (size_t i = 0; i < ParallelGCThreads; ++i) { |
| 6197 | 659 new (_overflow_stacks + i) Stack<oop, mtGC>(); |
| 695 | 660 } |
| 661 } | |
| 662 | |
| 0 | 663 if (UsePerfData) { |
| 664 EXCEPTION_MARK; | |
| 665 ResourceMark rm; | |
| 666 | |
| 667 const char* cname = | |
| 668 PerfDataManager::counter_name(_gen_counters->name_space(), "threads"); | |
| 669 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None, | |
| 670 ParallelGCThreads, CHECK); | |
| 671 } | |
| 672 } | |
| 673 #ifdef _MSC_VER | |
| 674 #pragma warning( pop ) | |
| 675 #endif | |
| 676 | |
| 677 // ParNewGeneration:: | |
| 678 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) : | |
| 679 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {} | |
| 680 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
681 template <class T> |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
682 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
683 #ifdef ASSERT |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
684 { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
685 assert(!oopDesc::is_null(*p), "expected non-null ref"); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
686 oop obj = oopDesc::load_decode_heap_oop_not_null(p); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
687 // We never expect to see a null reference being processed |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
688 // as a weak reference. |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
689 assert(obj->is_oop(), "expected an oop while scanning weak refs"); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
690 } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
691 #endif // ASSERT |
| 0 | 692 |
| 693 _par_cl->do_oop_nv(p); | |
| 694 | |
| 695 if (Universe::heap()->is_in_reserved(p)) { | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
696 oop obj = oopDesc::load_decode_heap_oop_not_null(p); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
697 _rs->write_ref_field_gc_par(p, obj); |
| 0 | 698 } |
| 699 } | |
| 700 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
701 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
702 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
703 |
| 0 | 704 // ParNewGeneration:: |
| 705 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) : | |
| 706 DefNewGeneration::KeepAliveClosure(cl) {} | |
| 707 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
708 template <class T> |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
709 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
710 #ifdef ASSERT |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
711 { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
712 assert(!oopDesc::is_null(*p), "expected non-null ref"); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
713 oop obj = oopDesc::load_decode_heap_oop_not_null(p); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
714 // We never expect to see a null reference being processed |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
715 // as a weak reference. |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
716 assert(obj->is_oop(), "expected an oop while scanning weak refs"); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
717 } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
718 #endif // ASSERT |
| 0 | 719 |
| 720 _cl->do_oop_nv(p); | |
| 721 | |
| 722 if (Universe::heap()->is_in_reserved(p)) { | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
723 oop obj = oopDesc::load_decode_heap_oop_not_null(p); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
724 _rs->write_ref_field_gc_par(p, obj); |
| 0 | 725 } |
| 726 } | |
| 727 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
728 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
729 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
730 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
731 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
732 T heap_oop = oopDesc::load_heap_oop(p); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
733 if (!oopDesc::is_null(heap_oop)) { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
734 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); |
| 0 | 735 if ((HeapWord*)obj < _boundary) { |
| 736 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?"); | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
737 oop new_obj = obj->is_forwarded() |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
738 ? obj->forwardee() |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
739 : _g->DefNewGeneration::copy_to_survivor_space(obj); |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
740 oopDesc::encode_store_heap_oop_not_null(p, new_obj); |
| 0 | 741 } |
| 742 if (_gc_barrier) { | |
| 743 // If p points to a younger generation, mark the card. | |
| 744 if ((HeapWord*)obj < _gen_boundary) { | |
| 745 _rs->write_ref_field_gc_par(p, obj); | |
| 746 } | |
| 747 } | |
| 748 } | |
| 749 } | |
| 750 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
751 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
752 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
753 |
| 0 | 754 class ParNewRefProcTaskProxy: public AbstractGangTask { |
| 755 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; | |
| 756 public: | |
| 757 ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen, | |
| 758 Generation& next_gen, | |
| 759 HeapWord* young_old_boundary, | |
| 760 ParScanThreadStateSet& state_set); | |
| 761 | |
| 762 private: | |
|
4728
441e946dc1af
7121618: Change type of number of GC workers to unsigned int.
jmasa
parents:
4095
diff
changeset
|
763 virtual void work(uint worker_id); |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
764 virtual void set_for_termination(int active_workers) { |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
765 _state_set.terminator()->reset_for_reuse(active_workers); |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
766 } |
| 0 | 767 private: |
| 768 ParNewGeneration& _gen; | |
| 769 ProcessTask& _task; | |
| 770 Generation& _next_gen; | |
| 771 HeapWord* _young_old_boundary; | |
| 772 ParScanThreadStateSet& _state_set; | |
| 773 }; | |
| 774 | |
| 775 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy( | |
| 776 ProcessTask& task, ParNewGeneration& gen, | |
| 777 Generation& next_gen, | |
| 778 HeapWord* young_old_boundary, | |
| 779 ParScanThreadStateSet& state_set) | |
| 780 : AbstractGangTask("ParNewGeneration parallel reference processing"), | |
| 781 _gen(gen), | |
| 782 _task(task), | |
| 783 _next_gen(next_gen), | |
| 784 _young_old_boundary(young_old_boundary), | |
| 785 _state_set(state_set) | |
| 786 { | |
| 787 } | |
| 788 | |
|
4728
441e946dc1af
7121618: Change type of number of GC workers to unsigned int.
jmasa
parents:
4095
diff
changeset
|
789 void ParNewRefProcTaskProxy::work(uint worker_id) |
| 0 | 790 { |
| 791 ResourceMark rm; | |
| 792 HandleMark hm; | |
|
4728
441e946dc1af
7121618: Change type of number of GC workers to unsigned int.
jmasa
parents:
4095
diff
changeset
|
793 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id); |
| 0 | 794 par_scan_state.set_young_old_boundary(_young_old_boundary); |
|
4728
441e946dc1af
7121618: Change type of number of GC workers to unsigned int.
jmasa
parents:
4095
diff
changeset
|
795 _task.work(worker_id, par_scan_state.is_alive_closure(), |
| 0 | 796 par_scan_state.keep_alive_closure(), |
| 797 par_scan_state.evacuate_followers_closure()); | |
| 798 } | |
| 799 | |
| 800 class ParNewRefEnqueueTaskProxy: public AbstractGangTask { | |
| 801 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; | |
| 802 EnqueueTask& _task; | |
| 803 | |
| 804 public: | |
| 805 ParNewRefEnqueueTaskProxy(EnqueueTask& task) | |
| 806 : AbstractGangTask("ParNewGeneration parallel reference enqueue"), | |
| 807 _task(task) | |
| 808 { } | |
| 809 | |
|
4728
441e946dc1af
7121618: Change type of number of GC workers to unsigned int.
jmasa
parents:
4095
diff
changeset
|
810 virtual void work(uint worker_id) |
| 0 | 811 { |
|
4728
441e946dc1af
7121618: Change type of number of GC workers to unsigned int.
jmasa
parents:
4095
diff
changeset
|
812 _task.work(worker_id); |
| 0 | 813 } |
| 814 }; | |
| 815 | |
| 816 | |
| 817 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) | |
| 818 { | |
| 819 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
| 820 assert(gch->kind() == CollectedHeap::GenCollectedHeap, | |
| 821 "not a generational heap"); | |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
822 FlexibleWorkGang* workers = gch->workers(); |
| 0 | 823 assert(workers != NULL, "Need parallel worker threads."); |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
824 _state_set.reset(workers->active_workers(), _generation.promotion_failed()); |
| 0 | 825 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(), |
| 826 _generation.reserved().end(), _state_set); | |
| 827 workers->run_task(&rp_task); | |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
828 _state_set.reset(0 /* bad value in debug if not reset */, |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
829 _generation.promotion_failed()); |
| 0 | 830 } |
| 831 | |
| 832 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) | |
| 833 { | |
| 834 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
835 FlexibleWorkGang* workers = gch->workers(); |
| 0 | 836 assert(workers != NULL, "Need parallel worker threads."); |
| 837 ParNewRefEnqueueTaskProxy enq_task(task); | |
| 838 workers->run_task(&enq_task); | |
| 839 } | |
| 840 | |
| 841 void ParNewRefProcTaskExecutor::set_single_threaded_mode() | |
| 842 { | |
| 843 _state_set.flush(); | |
| 844 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
| 845 gch->set_par_threads(0); // 0 ==> non-parallel. | |
| 846 gch->save_marks(); | |
| 847 } | |
| 848 | |
| 849 ScanClosureWithParBarrier:: | |
| 850 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) : | |
| 851 ScanClosure(g, gc_barrier) {} | |
| 852 | |
| 853 EvacuateFollowersClosureGeneral:: | |
| 854 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level, | |
| 855 OopsInGenClosure* cur, | |
| 856 OopsInGenClosure* older) : | |
| 857 _gch(gch), _level(level), | |
| 858 _scan_cur_or_nonheap(cur), _scan_older(older) | |
| 859 {} | |
| 860 | |
| 861 void EvacuateFollowersClosureGeneral::do_void() { | |
| 862 do { | |
| 863 // Beware: this call will lead to closure applications via virtual | |
| 864 // calls. | |
| 865 _gch->oop_since_save_marks_iterate(_level, | |
| 866 _scan_cur_or_nonheap, | |
| 867 _scan_older); | |
| 868 } while (!_gch->no_allocs_since_save_marks(_level)); | |
| 869 } | |
| 870 | |
| 871 | |
| 872 bool ParNewGeneration::_avoid_promotion_undo = false; | |
| 873 | |
| 874 void ParNewGeneration::adjust_desired_tenuring_threshold() { | |
| 875 // Set the desired survivor size to half the real survivor space | |
| 876 _tenuring_threshold = | |
| 877 age_table()->compute_tenuring_threshold(to()->capacity()/HeapWordSize); | |
| 878 } | |
| 879 | |
| 880 // A Generation that does parallel young-gen collection. | |
| 881 | |
| 882 void ParNewGeneration::collect(bool full, | |
| 883 bool clear_all_soft_refs, | |
| 884 size_t size, | |
| 885 bool is_tlab) { | |
| 886 assert(full || size > 0, "otherwise we don't want to collect"); | |
| 887 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
| 888 assert(gch->kind() == CollectedHeap::GenCollectedHeap, | |
| 889 "not a CMS generational heap"); | |
| 890 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy(); | |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
891 FlexibleWorkGang* workers = gch->workers(); |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
892 assert(workers != NULL, "Need workgang for parallel work"); |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
893 int active_workers = |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
894 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(), |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
895 workers->active_workers(), |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
896 Threads::number_of_non_daemon_threads()); |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
897 workers->set_active_workers(active_workers); |
| 0 | 898 _next_gen = gch->next_gen(this); |
| 899 assert(_next_gen != NULL, | |
| 900 "This must be the youngest gen, and not the only gen"); | |
| 901 assert(gch->n_gens() == 2, | |
| 902 "Par collection currently only works with single older gen."); | |
| 903 // Do we have to avoid promotion_undo? | |
| 904 if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) { | |
| 905 set_avoid_promotion_undo(true); | |
| 906 } | |
| 907 | |
| 908 // If the next generation is too full to accomodate worst-case promotion | |
| 909 // from this generation, pass on collection; let the next generation | |
| 910 // do it. | |
| 911 if (!collection_attempt_is_safe()) { | |
|
1888
a7214d79fcf1
6896603: CMS/GCH: collection_attempt_is_safe() ergo should use more recent data
ysr
parents:
1836
diff
changeset
|
912 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one |
| 0 | 913 return; |
| 914 } | |
| 915 assert(to()->is_empty(), "Else not collection_attempt_is_safe"); | |
| 916 | |
| 917 init_assuming_no_promotion_failure(); | |
| 918 | |
| 919 if (UseAdaptiveSizePolicy) { | |
| 920 set_survivor_overflow(false); | |
| 921 size_policy->minor_collection_begin(); | |
| 922 } | |
| 923 | |
|
6064
9d679effd28c
7166894: Add gc cause to GC logging for all collectors
brutisso
parents:
4911
diff
changeset
|
924 TraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, gclog_or_tty); |
| 0 | 925 // Capture heap used before collection (for printing). |
| 926 size_t gch_prev_used = gch->used(); | |
| 927 | |
| 928 SpecializationStats::clear(); | |
| 929 | |
| 930 age_table()->clear(); | |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
931 to()->clear(SpaceDecorator::Mangle); |
| 0 | 932 |
| 933 gch->save_marks(); | |
| 934 assert(workers != NULL, "Need parallel worker threads."); | |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
935 int n_workers = active_workers; |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
936 |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
937 // Set the correct parallelism (number of queues) in the reference processor |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
938 ref_processor()->set_active_mt_degree(n_workers); |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
939 |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
940 // Always set the terminator for the active number of workers |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
941 // because only those workers go through the termination protocol. |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
942 ParallelTaskTerminator _term(n_workers, task_queues()); |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
943 ParScanThreadStateSet thread_state_set(workers->active_workers(), |
| 0 | 944 *to(), *this, *_next_gen, *task_queues(), |
| 695 | 945 _overflow_stacks, desired_plab_sz(), _term); |
| 0 | 946 |
| 947 ParNewGenTask tsk(this, _next_gen, reserved().end(), &thread_state_set); | |
| 948 gch->set_par_threads(n_workers); | |
| 949 gch->rem_set()->prepare_for_younger_refs_iterate(true); | |
| 950 // It turns out that even when we're using 1 thread, doing the work in a | |
| 951 // separate thread causes wide variance in run times. We can't help this | |
| 952 // in the multi-threaded case, but we special-case n=1 here to get | |
| 953 // repeatable measurements of the 1-thread overhead of the parallel code. | |
| 954 if (n_workers > 1) { | |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
955 GenCollectedHeap::StrongRootsScope srs(gch); |
| 0 | 956 workers->run_task(&tsk); |
| 957 } else { | |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
958 GenCollectedHeap::StrongRootsScope srs(gch); |
| 0 | 959 tsk.work(0); |
| 960 } | |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
961 thread_state_set.reset(0 /* Bad value in debug if not reset */, |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
962 promotion_failed()); |
| 0 | 963 |
| 964 // Process (weak) reference objects found during scavenge. | |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
965 ReferenceProcessor* rp = ref_processor(); |
| 0 | 966 IsAliveClosure is_alive(this); |
| 967 ScanWeakRefClosure scan_weak_ref(this); | |
| 968 KeepAliveClosure keep_alive(&scan_weak_ref); | |
| 969 ScanClosure scan_without_gc_barrier(this, false); | |
| 970 ScanClosureWithParBarrier scan_with_gc_barrier(this, true); | |
| 971 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier); | |
| 972 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level, | |
| 973 &scan_without_gc_barrier, &scan_with_gc_barrier); | |
|
457
27a80744a83b
6778647: snap(), snap_policy() should be renamed setup(), setup_policy()
ysr
parents:
454
diff
changeset
|
974 rp->setup_policy(clear_all_soft_refs); |
|
4095
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
975 // Can the mt_degree be set later (at run_task() time would be best)? |
|
bca17e38de00
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
jmasa
parents:
2369
diff
changeset
|
976 rp->set_active_mt_degree(active_workers); |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
977 if (rp->processing_is_mt()) { |
| 0 | 978 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
979 rp->process_discovered_references(&is_alive, &keep_alive, |
|
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
980 &evacuate_followers, &task_executor); |
| 0 | 981 } else { |
| 982 thread_state_set.flush(); | |
| 983 gch->set_par_threads(0); // 0 ==> non-parallel. | |
| 984 gch->save_marks(); | |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
985 rp->process_discovered_references(&is_alive, &keep_alive, |
|
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
986 &evacuate_followers, NULL); |
| 0 | 987 } |
| 988 if (!promotion_failed()) { | |
| 989 // Swap the survivor spaces. | |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
990 eden()->clear(SpaceDecorator::Mangle); |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
991 from()->clear(SpaceDecorator::Mangle); |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
992 if (ZapUnusedHeapArea) { |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
993 // This is now done here because of the piece-meal mangling which |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
994 // can check for valid mangling at intermediate points in the |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
995 // collection(s). When a minor collection fails to collect |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
996 // sufficient space resizing of the young generation can occur |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
997 // an redistribute the spaces in the young generation. Mangle |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
998 // here so that unzapped regions don't get distributed to |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
999 // other spaces. |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
1000 to()->mangle_unused_area(); |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
1001 } |
| 0 | 1002 swap_spaces(); |
| 1003 | |
|
1387
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1145
diff
changeset
|
1004 // A successful scavenge should restart the GC time limit count which is |
|
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1145
diff
changeset
|
1005 // for full GC's. |
|
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1145
diff
changeset
|
1006 size_policy->reset_gc_overhead_limit_count(); |
|
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1145
diff
changeset
|
1007 |
| 0 | 1008 assert(to()->is_empty(), "to space should be empty now"); |
| 1009 } else { | |
|
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
1010 assert(_promo_failure_scan_stack.is_empty(), "post condition"); |
|
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
1011 _promo_failure_scan_stack.clear(true); // Clear cached segments. |
|
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
1012 |
| 0 | 1013 remove_forwarding_pointers(); |
| 1014 if (PrintGCDetails) { | |
| 1015 gclog_or_tty->print(" (promotion failed)"); | |
| 1016 } | |
| 1017 // All the spaces are in play for mark-sweep. | |
| 1018 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690. | |
| 1019 from()->set_next_compaction_space(to()); | |
|
1888
a7214d79fcf1
6896603: CMS/GCH: collection_attempt_is_safe() ergo should use more recent data
ysr
parents:
1836
diff
changeset
|
1020 gch->set_incremental_collection_failed(); |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
1021 // Inform the next generation that a promotion failure occurred. |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
1022 _next_gen->promotion_failure_occurred(); |
|
6
73e96e5c30df
6624765: Guarantee failure "Unexpected dirty card found"
jmasa
parents:
0
diff
changeset
|
1023 |
|
73e96e5c30df
6624765: Guarantee failure "Unexpected dirty card found"
jmasa
parents:
0
diff
changeset
|
1024 // Reset the PromotionFailureALot counters. |
|
73e96e5c30df
6624765: Guarantee failure "Unexpected dirty card found"
jmasa
parents:
0
diff
changeset
|
1025 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();) |
| 0 | 1026 } |
| 1027 // set new iteration safe limit for the survivor spaces | |
| 1028 from()->set_concurrent_iteration_safe_limit(from()->top()); | |
| 1029 to()->set_concurrent_iteration_safe_limit(to()->top()); | |
| 1030 | |
| 1031 adjust_desired_tenuring_threshold(); | |
| 1032 if (ResizePLAB) { | |
| 1033 plab_stats()->adjust_desired_plab_sz(); | |
| 1034 } | |
| 1035 | |
| 1036 if (PrintGC && !PrintGCDetails) { | |
| 1037 gch->print_heap_change(gch_prev_used); | |
| 1038 } | |
| 1039 | |
|
1712
2d6b74c9a797
6976378: ParNew: stats are printed unconditionally in debug builds
jcoomes
parents:
1710
diff
changeset
|
1040 if (PrintGCDetails && ParallelGCVerbose) { |
|
2d6b74c9a797
6976378: ParNew: stats are printed unconditionally in debug builds
jcoomes
parents:
1710
diff
changeset
|
1041 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats()); |
|
2d6b74c9a797
6976378: ParNew: stats are printed unconditionally in debug builds
jcoomes
parents:
1710
diff
changeset
|
1042 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats()); |
|
2d6b74c9a797
6976378: ParNew: stats are printed unconditionally in debug builds
jcoomes
parents:
1710
diff
changeset
|
1043 } |
| 1710 | 1044 |
| 0 | 1045 if (UseAdaptiveSizePolicy) { |
| 1046 size_policy->minor_collection_end(gch->gc_cause()); | |
| 1047 size_policy->avg_survived()->sample(from()->used()); | |
| 1048 } | |
| 1049 | |
| 4911 | 1050 // We need to use a monotonically non-deccreasing time in ms |
| 1051 // or we will see time-warp warnings and os::javaTimeMillis() | |
| 1052 // does not guarantee monotonicity. | |
| 1053 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; | |
| 1054 update_time_of_last_gc(now); | |
| 0 | 1055 |
| 1056 SpecializationStats::print(); | |
| 1057 | |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
1058 rp->set_enqueuing_is_done(true); |
|
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
1059 if (rp->processing_is_mt()) { |
| 0 | 1060 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
1061 rp->enqueue_discovered_references(&task_executor); |
| 0 | 1062 } else { |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
1063 rp->enqueue_discovered_references(NULL); |
| 0 | 1064 } |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
1065 rp->verify_no_references_recorded(); |
| 0 | 1066 } |
| 1067 | |
| 1068 static int sum; | |
| 1069 void ParNewGeneration::waste_some_time() { | |
| 1070 for (int i = 0; i < 100; i++) { | |
| 1071 sum += i; | |
| 1072 } | |
| 1073 } | |
| 1074 | |
| 1075 static const oop ClaimedForwardPtr = oop(0x4); | |
| 1076 | |
| 1077 // Because of concurrency, there are times where an object for which | |
| 1078 // "is_forwarded()" is true contains an "interim" forwarding pointer | |
| 1079 // value. Such a value will soon be overwritten with a real value. | |
| 1080 // This method requires "obj" to have a forwarding pointer, and waits, if | |
| 1081 // necessary for a real one to be inserted, and returns it. | |
| 1082 | |
| 1083 oop ParNewGeneration::real_forwardee(oop obj) { | |
| 1084 oop forward_ptr = obj->forwardee(); | |
| 1085 if (forward_ptr != ClaimedForwardPtr) { | |
| 1086 return forward_ptr; | |
| 1087 } else { | |
| 1088 return real_forwardee_slow(obj); | |
| 1089 } | |
| 1090 } | |
| 1091 | |
| 1092 oop ParNewGeneration::real_forwardee_slow(oop obj) { | |
| 1093 // Spin-read if it is claimed but not yet written by another thread. | |
| 1094 oop forward_ptr = obj->forwardee(); | |
| 1095 while (forward_ptr == ClaimedForwardPtr) { | |
| 1096 waste_some_time(); | |
| 1097 assert(obj->is_forwarded(), "precondition"); | |
| 1098 forward_ptr = obj->forwardee(); | |
| 1099 } | |
| 1100 return forward_ptr; | |
| 1101 } | |
| 1102 | |
| 1103 #ifdef ASSERT | |
| 1104 bool ParNewGeneration::is_legal_forward_ptr(oop p) { | |
| 1105 return | |
| 1106 (_avoid_promotion_undo && p == ClaimedForwardPtr) | |
| 1107 || Universe::heap()->is_in_reserved(p); | |
| 1108 } | |
| 1109 #endif | |
| 1110 | |
| 1111 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) { | |
|
2038
74ee0db180fa
6807801: CMS: could save/restore fewer header words during scavenge
ysr
parents:
1972
diff
changeset
|
1112 if (m->must_be_preserved_for_promotion_failure(obj)) { |
|
74ee0db180fa
6807801: CMS: could save/restore fewer header words during scavenge
ysr
parents:
1972
diff
changeset
|
1113 // We should really have separate per-worker stacks, rather |
|
74ee0db180fa
6807801: CMS: could save/restore fewer header words during scavenge
ysr
parents:
1972
diff
changeset
|
1114 // than use locking of a common pair of stacks. |
| 0 | 1115 MutexLocker ml(ParGCRareEvent_lock); |
|
2038
74ee0db180fa
6807801: CMS: could save/restore fewer header words during scavenge
ysr
parents:
1972
diff
changeset
|
1116 preserve_mark(obj, m); |
| 0 | 1117 } |
| 1118 } | |
| 1119 | |
| 1120 // Multiple GC threads may try to promote an object. If the object | |
| 1121 // is successfully promoted, a forwarding pointer will be installed in | |
| 1122 // the object in the young generation. This method claims the right | |
| 1123 // to install the forwarding pointer before it copies the object, | |
| 1124 // thus avoiding the need to undo the copy as in | |
| 1125 // copy_to_survivor_space_avoiding_with_undo. | |
| 1126 | |
| 1127 oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo( | |
| 1128 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { | |
| 1129 // In the sequential version, this assert also says that the object is | |
| 1130 // not forwarded. That might not be the case here. It is the case that | |
| 1131 // the caller observed it to be not forwarded at some time in the past. | |
| 1132 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); | |
| 1133 | |
| 1134 // The sequential code read "old->age()" below. That doesn't work here, | |
| 1135 // since the age is in the mark word, and that might be overwritten with | |
| 1136 // a forwarding pointer by a parallel thread. So we must save the mark | |
| 1137 // word in a local and then analyze it. | |
| 1138 oopDesc dummyOld; | |
| 1139 dummyOld.set_mark(m); | |
| 1140 assert(!dummyOld.is_forwarded(), | |
| 1141 "should not be called with forwarding pointer mark word."); | |
| 1142 | |
| 1143 oop new_obj = NULL; | |
| 1144 oop forward_ptr; | |
| 1145 | |
| 1146 // Try allocating obj in to-space (unless too old) | |
| 1147 if (dummyOld.age() < tenuring_threshold()) { | |
| 1148 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); | |
| 1149 if (new_obj == NULL) { | |
| 1150 set_survivor_overflow(true); | |
| 1151 } | |
| 1152 } | |
| 1153 | |
| 1154 if (new_obj == NULL) { | |
| 1155 // Either to-space is full or we decided to promote | |
| 1156 // try allocating obj tenured | |
| 1157 | |
| 1158 // Attempt to install a null forwarding pointer (atomically), | |
| 1159 // to claim the right to install the real forwarding pointer. | |
| 1160 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr); | |
| 1161 if (forward_ptr != NULL) { | |
| 1162 // someone else beat us to it. | |
| 1163 return real_forwardee(old); | |
| 1164 } | |
| 1165 | |
| 1166 new_obj = _next_gen->par_promote(par_scan_state->thread_num(), | |
| 1167 old, m, sz); | |
| 1168 | |
| 1169 if (new_obj == NULL) { | |
| 1170 // promotion failed, forward to self | |
| 1171 _promotion_failed = true; | |
| 1172 new_obj = old; | |
| 1173 | |
| 1174 preserve_mark_if_necessary(old, m); | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
1175 // Log the size of the maiden promotion failure |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
1176 par_scan_state->log_promotion_failure(sz); |
| 0 | 1177 } |
| 1178 | |
| 1179 old->forward_to(new_obj); | |
| 1180 forward_ptr = NULL; | |
| 1181 } else { | |
| 1182 // Is in to-space; do copying ourselves. | |
| 1183 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); | |
| 1184 forward_ptr = old->forward_to_atomic(new_obj); | |
| 1185 // Restore the mark word copied above. | |
| 1186 new_obj->set_mark(m); | |
| 1187 // Increment age if obj still in new generation | |
| 1188 new_obj->incr_age(); | |
| 1189 par_scan_state->age_table()->add(new_obj, sz); | |
| 1190 } | |
| 1191 assert(new_obj != NULL, "just checking"); | |
| 1192 | |
| 1193 if (forward_ptr == NULL) { | |
| 1194 oop obj_to_push = new_obj; | |
| 1195 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { | |
| 1196 // Length field used as index of next element to be scanned. | |
| 1197 // Real length can be obtained from real_forwardee() | |
| 1198 arrayOop(old)->set_length(0); | |
| 1199 obj_to_push = old; | |
| 1200 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, | |
| 1201 "push forwarded object"); | |
| 1202 } | |
| 1203 // Push it on one of the queues of to-be-scanned objects. | |
| 534 | 1204 bool simulate_overflow = false; |
| 1205 NOT_PRODUCT( | |
| 1206 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { | |
| 1207 // simulate a stack overflow | |
| 1208 simulate_overflow = true; | |
| 1209 } | |
| 1210 ) | |
| 1211 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { | |
| 0 | 1212 // Add stats for overflow pushes. |
| 1213 if (Verbose && PrintGCDetails) { | |
| 1214 gclog_or_tty->print("queue overflow!\n"); | |
| 1215 } | |
| 534 | 1216 push_on_overflow_list(old, par_scan_state); |
| 1710 | 1217 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); |
| 0 | 1218 } |
| 1219 | |
| 1220 return new_obj; | |
| 1221 } | |
| 1222 | |
| 1223 // Oops. Someone beat us to it. Undo the allocation. Where did we | |
| 1224 // allocate it? | |
| 1225 if (is_in_reserved(new_obj)) { | |
| 1226 // Must be in to_space. | |
| 1227 assert(to()->is_in_reserved(new_obj), "Checking"); | |
| 1228 if (forward_ptr == ClaimedForwardPtr) { | |
| 1229 // Wait to get the real forwarding pointer value. | |
| 1230 forward_ptr = real_forwardee(old); | |
| 1231 } | |
| 1232 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); | |
| 1233 } | |
| 1234 | |
| 1235 return forward_ptr; | |
| 1236 } | |
| 1237 | |
| 1238 | |
| 1239 // Multiple GC threads may try to promote the same object. If two | |
| 1240 // or more GC threads copy the object, only one wins the race to install | |
| 1241 // the forwarding pointer. The other threads have to undo their copy. | |
| 1242 | |
| 1243 oop ParNewGeneration::copy_to_survivor_space_with_undo( | |
| 1244 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { | |
| 1245 | |
| 1246 // In the sequential version, this assert also says that the object is | |
| 1247 // not forwarded. That might not be the case here. It is the case that | |
| 1248 // the caller observed it to be not forwarded at some time in the past. | |
| 1249 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); | |
| 1250 | |
| 1251 // The sequential code read "old->age()" below. That doesn't work here, | |
| 1252 // since the age is in the mark word, and that might be overwritten with | |
| 1253 // a forwarding pointer by a parallel thread. So we must save the mark | |
| 1254 // word here, install it in a local oopDesc, and then analyze it. | |
| 1255 oopDesc dummyOld; | |
| 1256 dummyOld.set_mark(m); | |
| 1257 assert(!dummyOld.is_forwarded(), | |
| 1258 "should not be called with forwarding pointer mark word."); | |
| 1259 | |
| 1260 bool failed_to_promote = false; | |
| 1261 oop new_obj = NULL; | |
| 1262 oop forward_ptr; | |
| 1263 | |
| 1264 // Try allocating obj in to-space (unless too old) | |
| 1265 if (dummyOld.age() < tenuring_threshold()) { | |
| 1266 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); | |
| 1267 if (new_obj == NULL) { | |
| 1268 set_survivor_overflow(true); | |
| 1269 } | |
| 1270 } | |
| 1271 | |
| 1272 if (new_obj == NULL) { | |
| 1273 // Either to-space is full or we decided to promote | |
| 1274 // try allocating obj tenured | |
| 1275 new_obj = _next_gen->par_promote(par_scan_state->thread_num(), | |
| 1276 old, m, sz); | |
| 1277 | |
| 1278 if (new_obj == NULL) { | |
| 1279 // promotion failed, forward to self | |
| 1280 forward_ptr = old->forward_to_atomic(old); | |
| 1281 new_obj = old; | |
| 1282 | |
| 1283 if (forward_ptr != NULL) { | |
| 1284 return forward_ptr; // someone else succeeded | |
| 1285 } | |
| 1286 | |
| 1287 _promotion_failed = true; | |
| 1288 failed_to_promote = true; | |
| 1289 | |
| 1290 preserve_mark_if_necessary(old, m); | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
1291 // Log the size of the maiden promotion failure |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
1292 par_scan_state->log_promotion_failure(sz); |
| 0 | 1293 } |
| 1294 } else { | |
| 1295 // Is in to-space; do copying ourselves. | |
| 1296 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); | |
| 1297 // Restore the mark word copied above. | |
| 1298 new_obj->set_mark(m); | |
| 1299 // Increment age if new_obj still in new generation | |
| 1300 new_obj->incr_age(); | |
| 1301 par_scan_state->age_table()->add(new_obj, sz); | |
| 1302 } | |
| 1303 assert(new_obj != NULL, "just checking"); | |
| 1304 | |
| 1305 // Now attempt to install the forwarding pointer (atomically). | |
| 1306 // We have to copy the mark word before overwriting with forwarding | |
| 1307 // ptr, so we can restore it below in the copy. | |
| 1308 if (!failed_to_promote) { | |
| 1309 forward_ptr = old->forward_to_atomic(new_obj); | |
| 1310 } | |
| 1311 | |
| 1312 if (forward_ptr == NULL) { | |
| 1313 oop obj_to_push = new_obj; | |
| 1314 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { | |
| 1315 // Length field used as index of next element to be scanned. | |
| 1316 // Real length can be obtained from real_forwardee() | |
| 1317 arrayOop(old)->set_length(0); | |
| 1318 obj_to_push = old; | |
| 1319 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, | |
| 1320 "push forwarded object"); | |
| 1321 } | |
| 1322 // Push it on one of the queues of to-be-scanned objects. | |
| 534 | 1323 bool simulate_overflow = false; |
| 1324 NOT_PRODUCT( | |
| 1325 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { | |
| 1326 // simulate a stack overflow | |
| 1327 simulate_overflow = true; | |
| 1328 } | |
| 1329 ) | |
| 1330 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { | |
| 0 | 1331 // Add stats for overflow pushes. |
| 534 | 1332 push_on_overflow_list(old, par_scan_state); |
| 1710 | 1333 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); |
| 0 | 1334 } |
| 1335 | |
| 1336 return new_obj; | |
| 1337 } | |
| 1338 | |
| 1339 // Oops. Someone beat us to it. Undo the allocation. Where did we | |
| 1340 // allocate it? | |
| 1341 if (is_in_reserved(new_obj)) { | |
| 1342 // Must be in to_space. | |
| 1343 assert(to()->is_in_reserved(new_obj), "Checking"); | |
| 1344 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); | |
| 1345 } else { | |
| 1346 assert(!_avoid_promotion_undo, "Should not be here if avoiding."); | |
| 1347 _next_gen->par_promote_alloc_undo(par_scan_state->thread_num(), | |
| 1348 (HeapWord*)new_obj, sz); | |
| 1349 } | |
| 1350 | |
| 1351 return forward_ptr; | |
| 1352 } | |
| 1353 | |
| 534 | 1354 #ifndef PRODUCT |
| 1355 // It's OK to call this multi-threaded; the worst thing | |
| 1356 // that can happen is that we'll get a bunch of closely | |
| 1357 // spaced simulated oveflows, but that's OK, in fact | |
| 1358 // probably good as it would exercise the overflow code | |
| 1359 // under contention. | |
| 1360 bool ParNewGeneration::should_simulate_overflow() { | |
| 1361 if (_overflow_counter-- <= 0) { // just being defensive | |
| 1362 _overflow_counter = ParGCWorkQueueOverflowInterval; | |
| 1363 return true; | |
| 1364 } else { | |
| 1365 return false; | |
| 1366 } | |
| 1367 } | |
| 1368 #endif | |
| 1369 | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1370 // In case we are using compressed oops, we need to be careful. |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1371 // If the object being pushed is an object array, then its length |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1372 // field keeps track of the "grey boundary" at which the next |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1373 // incremental scan will be done (see ParGCArrayScanChunk). |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1374 // When using compressed oops, this length field is kept in the |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1375 // lower 32 bits of the erstwhile klass word and cannot be used |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1376 // for the overflow chaining pointer (OCP below). As such the OCP |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1377 // would itself need to be compressed into the top 32-bits in this |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1378 // case. Unfortunately, see below, in the event that we have a |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1379 // promotion failure, the node to be pushed on the list can be |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1380 // outside of the Java heap, so the heap-based pointer compression |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1381 // would not work (we would have potential aliasing between C-heap |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1382 // and Java-heap pointers). For this reason, when using compressed |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1383 // oops, we simply use a worker-thread-local, non-shared overflow |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1384 // list in the form of a growable array, with a slightly different |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1385 // overflow stack draining strategy. If/when we start using fat |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1386 // stacks here, we can go back to using (fat) pointer chains |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1387 // (although some performance comparisons would be useful since |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1388 // single global lists have their own performance disadvantages |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1389 // as we were made painfully aware not long ago, see 6786503). |
| 534 | 1390 #define BUSY (oop(0x1aff1aff)) |
| 1391 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) { | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1392 assert(is_in_reserved(from_space_obj), "Should be from this generation"); |
| 695 | 1393 if (ParGCUseLocalOverflow) { |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1394 // In the case of compressed oops, we use a private, not-shared |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1395 // overflow stack. |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1396 par_scan_state->push_on_overflow_stack(from_space_obj); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1397 } else { |
| 695 | 1398 assert(!UseCompressedOops, "Error"); |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1399 // if the object has been forwarded to itself, then we cannot |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1400 // use the klass pointer for the linked list. Instead we have |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1401 // to allocate an oopDesc in the C-Heap and use that for the linked list. |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1402 // XXX This is horribly inefficient when a promotion failure occurs |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1403 // and should be fixed. XXX FIX ME !!! |
| 534 | 1404 #ifndef PRODUCT |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1405 Atomic::inc_ptr(&_num_par_pushes); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1406 assert(_num_par_pushes > 0, "Tautology"); |
| 534 | 1407 #endif |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1408 if (from_space_obj->forwardee() == from_space_obj) { |
| 6197 | 1409 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC); |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1410 listhead->forward_to(from_space_obj); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1411 from_space_obj = listhead; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1412 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1413 oop observed_overflow_list = _overflow_list; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1414 oop cur_overflow_list; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1415 do { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1416 cur_overflow_list = observed_overflow_list; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1417 if (cur_overflow_list != BUSY) { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1418 from_space_obj->set_klass_to_list_ptr(cur_overflow_list); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1419 } else { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1420 from_space_obj->set_klass_to_list_ptr(NULL); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1421 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1422 observed_overflow_list = |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1423 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1424 } while (cur_overflow_list != observed_overflow_list); |
| 0 | 1425 } |
| 1426 } | |
| 1427 | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1428 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1429 bool res; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1430 |
| 695 | 1431 if (ParGCUseLocalOverflow) { |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1432 res = par_scan_state->take_from_overflow_stack(); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1433 } else { |
| 695 | 1434 assert(!UseCompressedOops, "Error"); |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1435 res = take_from_overflow_list_work(par_scan_state); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1436 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1437 return res; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1438 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1439 |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1440 |
| 534 | 1441 // *NOTE*: The overflow list manipulation code here and |
| 1442 // in CMSCollector:: are very similar in shape, | |
| 1443 // except that in the CMS case we thread the objects | |
| 1444 // directly into the list via their mark word, and do | |
| 1445 // not need to deal with special cases below related | |
| 1446 // to chunking of object arrays and promotion failure | |
| 1447 // handling. | |
| 1448 // CR 6797058 has been filed to attempt consolidation of | |
| 1449 // the common code. | |
| 1450 // Because of the common code, if you make any changes in | |
| 1451 // the code below, please check the CMS version to see if | |
| 1452 // similar changes might be needed. | |
| 1453 // See CMSCollector::par_take_from_overflow_list() for | |
| 1454 // more extensive documentation comments. | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1455 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) { |
| 0 | 1456 ObjToScanQueue* work_q = par_scan_state->work_queue(); |
| 1457 // How many to take? | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1458 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, |
| 534 | 1459 (size_t)ParGCDesiredObjsFromOverflowList); |
| 0 | 1460 |
|
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
1461 assert(!UseCompressedOops, "Error"); |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1462 assert(par_scan_state->overflow_stack() == NULL, "Error"); |
| 0 | 1463 if (_overflow_list == NULL) return false; |
| 1464 | |
| 1465 // Otherwise, there was something there; try claiming the list. | |
| 534 | 1466 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); |
| 1467 // Trim off a prefix of at most objsFromOverflow items | |
| 1468 Thread* tid = Thread::current(); | |
| 1469 size_t spin_count = (size_t)ParallelGCThreads; | |
| 1470 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100); | |
| 1471 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) { | |
| 1472 // someone grabbed it before we did ... | |
| 1473 // ... we spin for a short while... | |
| 1474 os::sleep(tid, sleep_time_millis, false); | |
| 1475 if (_overflow_list == NULL) { | |
| 1476 // nothing left to take | |
| 1477 return false; | |
| 1478 } else if (_overflow_list != BUSY) { | |
| 1479 // try and grab the prefix | |
| 1480 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); | |
| 1481 } | |
| 0 | 1482 } |
| 534 | 1483 if (prefix == NULL || prefix == BUSY) { |
| 1484 // Nothing to take or waited long enough | |
| 1485 if (prefix == NULL) { | |
| 1486 // Write back the NULL in case we overwrote it with BUSY above | |
| 1487 // and it is still the same value. | |
| 1488 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); | |
| 1489 } | |
| 1490 return false; | |
| 1491 } | |
| 1492 assert(prefix != NULL && prefix != BUSY, "Error"); | |
| 1493 size_t i = 1; | |
| 0 | 1494 oop cur = prefix; |
|
167
feeb96a45707
6696264: assert("narrow oop can never be zero") for GCBasher & ParNewGC
coleenp
parents:
113
diff
changeset
|
1495 while (i < objsFromOverflow && cur->klass_or_null() != NULL) { |
| 0 | 1496 i++; cur = oop(cur->klass()); |
| 1497 } | |
| 1498 | |
| 1499 // Reattach remaining (suffix) to overflow list | |
| 534 | 1500 if (cur->klass_or_null() == NULL) { |
| 1501 // Write back the NULL in lieu of the BUSY we wrote | |
| 1502 // above and it is still the same value. | |
| 1503 if (_overflow_list == BUSY) { | |
| 1504 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); | |
| 0 | 1505 } |
| 534 | 1506 } else { |
| 1507 assert(cur->klass_or_null() != BUSY, "Error"); | |
| 1508 oop suffix = oop(cur->klass()); // suffix will be put back on global list | |
| 1509 cur->set_klass_to_list_ptr(NULL); // break off suffix | |
| 1510 // It's possible that the list is still in the empty(busy) state | |
| 1511 // we left it in a short while ago; in that case we may be | |
| 1512 // able to place back the suffix. | |
| 1513 oop observed_overflow_list = _overflow_list; | |
| 1514 oop cur_overflow_list = observed_overflow_list; | |
| 1515 bool attached = false; | |
| 1516 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { | |
| 1517 observed_overflow_list = | |
| 1518 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); | |
| 1519 if (cur_overflow_list == observed_overflow_list) { | |
| 1520 attached = true; | |
| 1521 break; | |
| 1522 } else cur_overflow_list = observed_overflow_list; | |
| 1523 } | |
| 1524 if (!attached) { | |
| 1525 // Too bad, someone else got in in between; we'll need to do a splice. | |
| 1526 // Find the last item of suffix list | |
| 1527 oop last = suffix; | |
| 1528 while (last->klass_or_null() != NULL) { | |
| 1529 last = oop(last->klass()); | |
| 1530 } | |
| 1531 // Atomically prepend suffix to current overflow list | |
| 1532 observed_overflow_list = _overflow_list; | |
| 1533 do { | |
| 1534 cur_overflow_list = observed_overflow_list; | |
| 1535 if (cur_overflow_list != BUSY) { | |
| 1536 // Do the splice ... | |
| 1537 last->set_klass_to_list_ptr(cur_overflow_list); | |
| 1538 } else { // cur_overflow_list == BUSY | |
| 1539 last->set_klass_to_list_ptr(NULL); | |
| 1540 } | |
| 1541 observed_overflow_list = | |
| 1542 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); | |
| 1543 } while (cur_overflow_list != observed_overflow_list); | |
| 0 | 1544 } |
| 1545 } | |
| 1546 | |
| 1547 // Push objects on prefix list onto this thread's work queue | |
| 534 | 1548 assert(prefix != NULL && prefix != BUSY, "program logic"); |
| 0 | 1549 cur = prefix; |
| 534 | 1550 ssize_t n = 0; |
| 0 | 1551 while (cur != NULL) { |
| 1552 oop obj_to_push = cur->forwardee(); | |
|
454
df4305d4c1a1
6774607: SIGSEGV or (!is_null(v),"oop value can never be zero") assertion when running with CMS and COOPs
ysr
parents:
453
diff
changeset
|
1553 oop next = oop(cur->klass_or_null()); |
| 0 | 1554 cur->set_klass(obj_to_push->klass()); |
| 534 | 1555 // This may be an array object that is self-forwarded. In that case, the list pointer |
| 1556 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed. | |
| 1557 if (!is_in_reserved(cur)) { | |
| 1558 // This can become a scaling bottleneck when there is work queue overflow coincident | |
| 1559 // with promotion failure. | |
| 1560 oopDesc* f = cur; | |
| 6197 | 1561 FREE_C_HEAP_ARRAY(oopDesc, f, mtGC); |
| 534 | 1562 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) { |
| 0 | 1563 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); |
| 534 | 1564 obj_to_push = cur; |
| 0 | 1565 } |
| 534 | 1566 bool ok = work_q->push(obj_to_push); |
| 1567 assert(ok, "Should have succeeded"); | |
| 0 | 1568 cur = next; |
| 1569 n++; | |
| 1570 } | |
| 1710 | 1571 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n)); |
| 534 | 1572 #ifndef PRODUCT |
| 1573 assert(_num_par_pushes >= n, "Too many pops?"); | |
| 1574 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); | |
| 1575 #endif | |
| 0 | 1576 return true; |
| 1577 } | |
| 534 | 1578 #undef BUSY |
| 0 | 1579 |
| 1580 void ParNewGeneration::ref_processor_init() | |
| 1581 { | |
| 1582 if (_ref_processor == NULL) { | |
| 1583 // Allocate and initialize a reference processor | |
|
2369
92da084fefc9
6668573: CMS: reference processing crash if ParallelCMSThreads > ParallelGCThreads
ysr
parents:
2038
diff
changeset
|
1584 _ref_processor = |
|
92da084fefc9
6668573: CMS: reference processing crash if ParallelCMSThreads > ParallelGCThreads
ysr
parents:
2038
diff
changeset
|
1585 new ReferenceProcessor(_reserved, // span |
|
92da084fefc9
6668573: CMS: reference processing crash if ParallelCMSThreads > ParallelGCThreads
ysr
parents:
2038
diff
changeset
|
1586 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing |
|
92da084fefc9
6668573: CMS: reference processing crash if ParallelCMSThreads > ParallelGCThreads
ysr
parents:
2038
diff
changeset
|
1587 (int) ParallelGCThreads, // mt processing degree |
|
92da084fefc9
6668573: CMS: reference processing crash if ParallelCMSThreads > ParallelGCThreads
ysr
parents:
2038
diff
changeset
|
1588 refs_discovery_is_mt(), // mt discovery |
|
92da084fefc9
6668573: CMS: reference processing crash if ParallelCMSThreads > ParallelGCThreads
ysr
parents:
2038
diff
changeset
|
1589 (int) ParallelGCThreads, // mt discovery degree |
|
92da084fefc9
6668573: CMS: reference processing crash if ParallelCMSThreads > ParallelGCThreads
ysr
parents:
2038
diff
changeset
|
1590 refs_discovery_is_atomic(), // atomic_discovery |
|
92da084fefc9
6668573: CMS: reference processing crash if ParallelCMSThreads > ParallelGCThreads
ysr
parents:
2038
diff
changeset
|
1591 NULL, // is_alive_non_header |
|
92da084fefc9
6668573: CMS: reference processing crash if ParallelCMSThreads > ParallelGCThreads
ysr
parents:
2038
diff
changeset
|
1592 false); // write barrier for next field updates |
| 0 | 1593 } |
| 1594 } | |
| 1595 | |
| 1596 const char* ParNewGeneration::name() const { | |
| 1597 return "par new generation"; | |
| 1598 } | |
|
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1712
diff
changeset
|
1599 |
|
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1712
diff
changeset
|
1600 bool ParNewGeneration::in_use() { |
|
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1712
diff
changeset
|
1601 return UseParNewGC && ParallelGCThreads > 0; |
|
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1712
diff
changeset
|
1602 } |