Mercurial > hg > graal-compiler
annotate src/share/vm/gc_implementation/parNew/parNewGeneration.cpp @ 1710:94251661de76
6970376: ParNew: shared TaskQueue statistics
Reviewed-by: ysr
| author | jcoomes |
|---|---|
| date | Mon, 09 Aug 2010 18:03:50 -0700 |
| parents | a93a9eda13f7 |
| children | 2d6b74c9a797 |
| rev | line source |
|---|---|
| 0 | 1 /* |
|
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1387
diff
changeset
|
2 * Copyright (c) 2001, 2010, 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 | |
| 25 # include "incls/_precompiled.incl" | |
| 26 # include "incls/_parNewGeneration.cpp.incl" | |
| 27 | |
| 28 #ifdef _MSC_VER | |
| 29 #pragma warning( push ) | |
| 30 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list | |
| 31 #endif | |
| 32 ParScanThreadState::ParScanThreadState(Space* to_space_, | |
| 33 ParNewGeneration* gen_, | |
| 34 Generation* old_gen_, | |
| 35 int thread_num_, | |
| 36 ObjToScanQueueSet* work_queue_set_, | |
| 695 | 37 GrowableArray<oop>** overflow_stack_set_, |
| 0 | 38 size_t desired_plab_sz_, |
| 39 ParallelTaskTerminator& term_) : | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
40 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_), |
| 0 | 41 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false), |
| 695 | 42 _overflow_stack(overflow_stack_set_[thread_num_]), |
| 0 | 43 _ageTable(false), // false ==> not the global age table, no perf data. |
| 44 _to_space_alloc_buffer(desired_plab_sz_), | |
| 45 _to_space_closure(gen_, this), _old_gen_closure(gen_, this), | |
| 46 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this), | |
| 47 _older_gen_closure(gen_, this), | |
| 48 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure, | |
| 49 &_to_space_root_closure, gen_, &_old_gen_root_closure, | |
| 50 work_queue_set_, &term_), | |
| 51 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this), | |
| 52 _keep_alive_closure(&_scan_weak_ref_closure), | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
53 _promotion_failure_size(0), |
| 0 | 54 _strong_roots_time(0.0), _term_time(0.0) |
| 55 { | |
| 1710 | 56 #if TASKQUEUE_STATS |
| 57 _term_attempts = 0; | |
| 58 _overflow_refills = 0; | |
| 59 _overflow_refill_objs = 0; | |
| 60 #endif // TASKQUEUE_STATS | |
| 61 | |
| 0 | 62 _survivor_chunk_array = |
| 63 (ChunkArray*) old_gen()->get_data_recorder(thread_num()); | |
| 64 _hash_seed = 17; // Might want to take time-based random value. | |
| 65 _start = os::elapsedTime(); | |
| 66 _old_gen_closure.set_generation(old_gen_); | |
| 67 _old_gen_root_closure.set_generation(old_gen_); | |
| 68 } | |
| 69 #ifdef _MSC_VER | |
| 70 #pragma warning( pop ) | |
| 71 #endif | |
| 72 | |
| 73 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start, | |
| 74 size_t plab_word_size) { | |
| 75 ChunkArray* sca = survivor_chunk_array(); | |
| 76 if (sca != NULL) { | |
| 77 // A non-null SCA implies that we want the PLAB data recorded. | |
| 78 sca->record_sample(plab_start, plab_word_size); | |
| 79 } | |
| 80 } | |
| 81 | |
| 82 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const { | |
| 83 return new_obj->is_objArray() && | |
| 84 arrayOop(new_obj)->length() > ParGCArrayScanChunk && | |
| 85 new_obj != old_obj; | |
| 86 } | |
| 87 | |
| 88 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) { | |
| 89 assert(old->is_objArray(), "must be obj array"); | |
| 90 assert(old->is_forwarded(), "must be forwarded"); | |
| 91 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
|
92 assert(!old_gen()->is_in(old), "must be in young generation."); |
| 0 | 93 |
| 94 objArrayOop obj = objArrayOop(old->forwardee()); | |
| 95 // Process ParGCArrayScanChunk elements now | |
| 96 // and push the remainder back onto queue | |
| 97 int start = arrayOop(old)->length(); | |
| 98 int end = obj->length(); | |
| 99 int remainder = end - start; | |
| 100 assert(start <= end, "just checking"); | |
| 101 if (remainder > 2 * ParGCArrayScanChunk) { | |
| 102 // Test above combines last partial chunk with a full chunk | |
| 103 end = start + ParGCArrayScanChunk; | |
| 104 arrayOop(old)->set_length(end); | |
| 105 // Push remainder. | |
| 106 bool ok = work_queue()->push(old); | |
| 107 assert(ok, "just popped, push must be okay"); | |
| 108 } else { | |
| 109 // Restore length so that it can be used if there | |
| 110 // is a promotion failure and forwarding pointers | |
| 111 // must be removed. | |
| 112 arrayOop(old)->set_length(end); | |
| 113 } | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
114 |
| 0 | 115 // 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
|
116 // should make sure end is even (aligned to HeapWord in case of compressed oops) |
| 0 | 117 if ((HeapWord *)obj < young_old_boundary()) { |
| 118 // 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
|
119 obj->oop_iterate_range(&_to_space_closure, start, end); |
| 0 | 120 } else { |
| 121 // 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
|
122 obj->oop_iterate_range(&_old_gen_closure, start, end); |
| 0 | 123 } |
| 124 } | |
| 125 | |
| 126 | |
| 127 void ParScanThreadState::trim_queues(int max_size) { | |
| 128 ObjToScanQueue* queue = work_queue(); | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
129 do { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
130 while (queue->size() > (juint)max_size) { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
131 oop obj_to_scan; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
132 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
|
133 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
|
134 if (obj_to_scan->is_objArray() && |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
135 obj_to_scan->is_forwarded() && |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
136 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
|
137 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
|
138 } else { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
139 // object is in to_space |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
140 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
|
141 } |
| 0 | 142 } else { |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
143 // object is in old generation |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
144 obj_to_scan->oop_iterate(&_old_gen_closure); |
| 0 | 145 } |
| 146 } | |
| 147 } | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
148 // 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
|
149 // 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
|
150 // 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
|
151 // 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
|
152 // 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
|
153 // 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
|
154 // private overflow stack. |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
155 } 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
|
156 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
157 |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
158 bool ParScanThreadState::take_from_overflow_stack() { |
| 695 | 159 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
|
160 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
|
161 ObjToScanQueue* queue = work_queue(); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
162 GrowableArray<oop>* of_stack = overflow_stack(); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
163 uint num_overflow_elems = of_stack->length(); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
164 uint num_take_elems = MIN2(MIN2((queue->max_elems() - queue->size())/4, |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
165 (juint)ParGCDesiredObjsFromOverflowList), |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
166 num_overflow_elems); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
167 // 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
|
168 // stack to our work queue. |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
169 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
|
170 oop cur = of_stack->pop(); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
171 oop obj_to_push = cur->forwardee(); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
172 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
|
173 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
|
174 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
|
175 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
|
176 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
|
177 obj_to_push = cur; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
178 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
179 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
|
180 assert(ok, "Should have succeeded"); |
| 0 | 181 } |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
182 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
|
183 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
|
184 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
185 |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
186 void ParScanThreadState::push_on_overflow_stack(oop p) { |
| 695 | 187 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
|
188 overflow_stack()->push(p); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
189 assert(young_gen()->overflow_list() == NULL, "Error"); |
| 0 | 190 } |
| 191 | |
| 192 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) { | |
| 193 | |
| 194 // Otherwise, if the object is small enough, try to reallocate the | |
| 195 // buffer. | |
| 196 HeapWord* obj = NULL; | |
| 197 if (!_to_space_full) { | |
| 198 ParGCAllocBuffer* const plab = to_space_alloc_buffer(); | |
| 199 Space* const sp = to_space(); | |
| 200 if (word_sz * 100 < | |
| 201 ParallelGCBufferWastePct * plab->word_sz()) { | |
| 202 // Is small enough; abandon this buffer and start a new one. | |
| 203 plab->retire(false, false); | |
| 204 size_t buf_size = plab->word_sz(); | |
| 205 HeapWord* buf_space = sp->par_allocate(buf_size); | |
| 206 if (buf_space == NULL) { | |
| 207 const size_t min_bytes = | |
| 208 ParGCAllocBuffer::min_size() << LogHeapWordSize; | |
| 209 size_t free_bytes = sp->free(); | |
| 210 while(buf_space == NULL && free_bytes >= min_bytes) { | |
| 211 buf_size = free_bytes >> LogHeapWordSize; | |
| 212 assert(buf_size == (size_t)align_object_size(buf_size), | |
| 213 "Invariant"); | |
| 214 buf_space = sp->par_allocate(buf_size); | |
| 215 free_bytes = sp->free(); | |
| 216 } | |
| 217 } | |
| 218 if (buf_space != NULL) { | |
| 219 plab->set_word_size(buf_size); | |
| 220 plab->set_buf(buf_space); | |
| 221 record_survivor_plab(buf_space, buf_size); | |
| 222 obj = plab->allocate(word_sz); | |
| 223 // Note that we cannot compare buf_size < word_sz below | |
| 224 // because of AlignmentReserve (see ParGCAllocBuffer::allocate()). | |
| 225 assert(obj != NULL || plab->words_remaining() < word_sz, | |
| 226 "Else should have been able to allocate"); | |
| 227 // It's conceivable that we may be able to use the | |
| 228 // buffer we just grabbed for subsequent small requests | |
| 229 // even if not for this one. | |
| 230 } else { | |
| 231 // We're used up. | |
| 232 _to_space_full = true; | |
| 233 } | |
| 234 | |
| 235 } else { | |
| 236 // Too large; allocate the object individually. | |
| 237 obj = sp->par_allocate(word_sz); | |
| 238 } | |
| 239 } | |
| 240 return obj; | |
| 241 } | |
| 242 | |
| 243 | |
| 244 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, | |
| 245 size_t word_sz) { | |
| 246 // Is the alloc in the current alloc buffer? | |
| 247 if (to_space_alloc_buffer()->contains(obj)) { | |
| 248 assert(to_space_alloc_buffer()->contains(obj + word_sz - 1), | |
| 249 "Should contain whole object."); | |
| 250 to_space_alloc_buffer()->undo_allocation(obj, word_sz); | |
| 251 } else { | |
|
481
7d7a7c599c17
6578152: fill_region_with_object has usability and safety issues
jcoomes
parents:
457
diff
changeset
|
252 CollectedHeap::fill_with_object(obj, word_sz); |
| 0 | 253 } |
| 254 } | |
| 255 | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
256 void ParScanThreadState::print_and_clear_promotion_failure_size() { |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
257 if (_promotion_failure_size != 0) { |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
258 if (PrintPromotionFailure) { |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
259 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ", |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
260 _thread_num, _promotion_failure_size); |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
261 } |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
262 _promotion_failure_size = 0; |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
263 } |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
264 } |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
265 |
| 0 | 266 class ParScanThreadStateSet: private ResourceArray { |
| 267 public: | |
| 268 // Initializes states for the specified number of threads; | |
| 269 ParScanThreadStateSet(int num_threads, | |
| 270 Space& to_space, | |
| 271 ParNewGeneration& gen, | |
| 272 Generation& old_gen, | |
| 273 ObjToScanQueueSet& queue_set, | |
| 695 | 274 GrowableArray<oop>** overflow_stacks_, |
| 0 | 275 size_t desired_plab_sz, |
| 276 ParallelTaskTerminator& term); | |
| 1710 | 277 |
| 278 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); } | |
| 279 | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
280 inline ParScanThreadState& thread_state(int i); |
| 1710 | 281 |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
282 void reset(bool promotion_failed); |
| 0 | 283 void flush(); |
| 1710 | 284 |
| 285 #if TASKQUEUE_STATS | |
| 286 static void | |
| 287 print_termination_stats_hdr(outputStream* const st = gclog_or_tty); | |
| 288 void print_termination_stats(outputStream* const st = gclog_or_tty); | |
| 289 static void | |
| 290 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty); | |
| 291 void print_taskqueue_stats(outputStream* const st = gclog_or_tty); | |
| 292 void reset_stats(); | |
| 293 #endif // TASKQUEUE_STATS | |
| 294 | |
| 0 | 295 private: |
| 296 ParallelTaskTerminator& _term; | |
| 297 ParNewGeneration& _gen; | |
| 298 Generation& _next_gen; | |
| 299 }; | |
| 300 | |
| 301 | |
| 302 ParScanThreadStateSet::ParScanThreadStateSet( | |
| 303 int num_threads, Space& to_space, ParNewGeneration& gen, | |
| 304 Generation& old_gen, ObjToScanQueueSet& queue_set, | |
| 695 | 305 GrowableArray<oop>** overflow_stack_set_, |
| 0 | 306 size_t desired_plab_sz, ParallelTaskTerminator& term) |
| 307 : ResourceArray(sizeof(ParScanThreadState), num_threads), | |
| 1710 | 308 _gen(gen), _next_gen(old_gen), _term(term) |
| 0 | 309 { |
| 310 assert(num_threads > 0, "sanity check!"); | |
| 311 // Initialize states. | |
| 312 for (int i = 0; i < num_threads; ++i) { | |
| 313 new ((ParScanThreadState*)_data + i) | |
| 314 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set, | |
| 695 | 315 overflow_stack_set_, desired_plab_sz, term); |
| 0 | 316 } |
| 317 } | |
| 318 | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
319 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) |
| 0 | 320 { |
| 321 assert(i >= 0 && i < length(), "sanity check!"); | |
| 322 return ((ParScanThreadState*)_data)[i]; | |
| 323 } | |
| 324 | |
| 325 | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
326 void ParScanThreadStateSet::reset(bool promotion_failed) |
| 0 | 327 { |
| 328 _term.reset_for_reuse(); | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
329 if (promotion_failed) { |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
330 for (int i = 0; i < length(); ++i) { |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
331 thread_state(i).print_and_clear_promotion_failure_size(); |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
332 } |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
333 } |
| 0 | 334 } |
| 335 | |
| 1710 | 336 #if TASKQUEUE_STATS |
| 337 void | |
| 338 ParScanThreadState::reset_stats() | |
| 339 { | |
| 340 taskqueue_stats().reset(); | |
| 341 _term_attempts = 0; | |
| 342 _overflow_refills = 0; | |
| 343 _overflow_refill_objs = 0; | |
| 344 } | |
| 345 | |
| 346 void ParScanThreadStateSet::reset_stats() | |
| 347 { | |
| 348 for (int i = 0; i < length(); ++i) { | |
| 349 thread_state(i).reset_stats(); | |
| 350 } | |
| 351 } | |
| 352 | |
| 353 void | |
| 354 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) | |
| 355 { | |
| 356 st->print_raw_cr("GC Termination Stats"); | |
| 357 st->print_raw_cr(" elapsed --strong roots-- " | |
| 358 "-------termination-------"); | |
| 359 st->print_raw_cr("thr ms ms % " | |
| 360 " ms % attempts"); | |
| 361 st->print_raw_cr("--- --------- --------- ------ " | |
| 362 "--------- ------ --------"); | |
| 363 } | |
| 364 | |
| 365 void ParScanThreadStateSet::print_termination_stats(outputStream* const st) | |
| 366 { | |
| 367 print_termination_stats_hdr(st); | |
| 368 | |
| 369 for (int i = 0; i < length(); ++i) { | |
| 370 const ParScanThreadState & pss = thread_state(i); | |
| 371 const double elapsed_ms = pss.elapsed_time() * 1000.0; | |
| 372 const double s_roots_ms = pss.strong_roots_time() * 1000.0; | |
| 373 const double term_ms = pss.term_time() * 1000.0; | |
| 374 st->print_cr("%3d %9.2f %9.2f %6.2f " | |
| 375 "%9.2f %6.2f " SIZE_FORMAT_W(8), | |
| 376 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, | |
| 377 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts()); | |
| 378 } | |
| 379 } | |
| 380 | |
| 381 // Print stats related to work queue activity. | |
| 382 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) | |
| 383 { | |
| 384 st->print_raw_cr("GC Task Stats"); | |
| 385 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); | |
| 386 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); | |
| 387 } | |
| 388 | |
| 389 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st) | |
| 390 { | |
| 391 print_taskqueue_stats_hdr(st); | |
| 392 | |
| 393 TaskQueueStats totals; | |
| 394 for (int i = 0; i < length(); ++i) { | |
| 395 const ParScanThreadState & pss = thread_state(i); | |
| 396 const TaskQueueStats & stats = pss.taskqueue_stats(); | |
| 397 st->print("%3d ", i); stats.print(st); st->cr(); | |
| 398 totals += stats; | |
| 399 | |
| 400 if (pss.overflow_refills() > 0) { | |
| 401 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills " | |
| 402 SIZE_FORMAT_W(10) " overflow objects", | |
| 403 pss.overflow_refills(), pss.overflow_refill_objs()); | |
| 404 } | |
| 405 } | |
| 406 st->print("tot "); totals.print(st); st->cr(); | |
| 407 | |
| 408 DEBUG_ONLY(totals.verify()); | |
| 409 } | |
| 410 #endif // TASKQUEUE_STATS | |
| 411 | |
| 0 | 412 void ParScanThreadStateSet::flush() |
| 413 { | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
414 // Work in this loop should be kept as lightweight as |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
415 // possible since this might otherwise become a bottleneck |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
416 // to scaling. Should we add heavy-weight work into this |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
417 // loop, consider parallelizing the loop into the worker threads. |
| 0 | 418 for (int i = 0; i < length(); ++i) { |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
419 ParScanThreadState& par_scan_state = thread_state(i); |
| 0 | 420 |
| 421 // Flush stats related to To-space PLAB activity and | |
| 422 // retire the last buffer. | |
| 423 par_scan_state.to_space_alloc_buffer()-> | |
| 424 flush_stats_and_retire(_gen.plab_stats(), | |
| 425 false /* !retain */); | |
| 426 | |
| 427 // Every thread has its own age table. We need to merge | |
| 428 // them all into one. | |
| 429 ageTable *local_table = par_scan_state.age_table(); | |
| 430 _gen.age_table()->merge(local_table); | |
| 431 | |
| 432 // Inform old gen that we're done. | |
| 433 _next_gen.par_promote_alloc_done(i); | |
| 434 _next_gen.par_oop_since_save_marks_iterate_done(i); | |
| 1710 | 435 } |
| 0 | 436 |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
437 if (UseConcMarkSweepGC && ParallelGCThreads > 0) { |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
438 // We need to call this even when ResizeOldPLAB is disabled |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
439 // 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
|
440 // 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
|
441 // 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
|
442 // performance. |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
443 CFLS_LAB::compute_desired_plab_size(); |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
444 } |
| 0 | 445 } |
| 446 | |
| 447 ParScanClosure::ParScanClosure(ParNewGeneration* g, | |
| 448 ParScanThreadState* par_scan_state) : | |
| 449 OopsInGenClosure(g), _par_scan_state(par_scan_state), _g(g) | |
| 450 { | |
| 451 assert(_g->level() == 0, "Optimized for youngest generation"); | |
| 452 _boundary = _g->reserved().end(); | |
| 453 } | |
| 454 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
455 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
|
456 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
|
457 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
458 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
|
459 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
|
460 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
461 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
|
462 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
|
463 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
464 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
|
465 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
|
466 |
| 0 | 467 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g, |
| 468 ParScanThreadState* par_scan_state) | |
| 469 : 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
|
470 {} |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
471 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
472 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
|
473 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); } |
| 0 | 474 |
| 475 #ifdef WIN32 | |
| 476 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */ | |
| 477 #endif | |
| 478 | |
| 479 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure( | |
| 480 ParScanThreadState* par_scan_state_, | |
| 481 ParScanWithoutBarrierClosure* to_space_closure_, | |
| 482 ParScanWithBarrierClosure* old_gen_closure_, | |
| 483 ParRootScanWithoutBarrierClosure* to_space_root_closure_, | |
| 484 ParNewGeneration* par_gen_, | |
| 485 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_, | |
| 486 ObjToScanQueueSet* task_queues_, | |
| 487 ParallelTaskTerminator* terminator_) : | |
| 488 | |
| 489 _par_scan_state(par_scan_state_), | |
| 490 _to_space_closure(to_space_closure_), | |
| 491 _old_gen_closure(old_gen_closure_), | |
| 492 _to_space_root_closure(to_space_root_closure_), | |
| 493 _old_gen_root_closure(old_gen_root_closure_), | |
| 494 _par_gen(par_gen_), | |
| 495 _task_queues(task_queues_), | |
| 496 _terminator(terminator_) | |
| 497 {} | |
| 498 | |
| 499 void ParEvacuateFollowersClosure::do_void() { | |
| 500 ObjToScanQueue* work_q = par_scan_state()->work_queue(); | |
| 501 | |
| 502 while (true) { | |
| 503 | |
| 504 // Scan to-space and old-gen objs until we run out of both. | |
| 505 oop obj_to_scan; | |
| 506 par_scan_state()->trim_queues(0); | |
| 507 | |
| 508 // We have no local work, attempt to steal from other threads. | |
| 509 | |
| 510 // attempt to steal work from promoted. | |
| 511 if (task_queues()->steal(par_scan_state()->thread_num(), | |
| 512 par_scan_state()->hash_seed(), | |
| 513 obj_to_scan)) { | |
| 514 bool res = work_q->push(obj_to_scan); | |
| 515 assert(res, "Empty queue should have room for a push."); | |
| 516 | |
| 517 // if successful, goto Start. | |
| 518 continue; | |
| 519 | |
| 520 // try global overflow list. | |
| 521 } else if (par_gen()->take_from_overflow_list(par_scan_state())) { | |
| 522 continue; | |
| 523 } | |
| 524 | |
| 525 // Otherwise, offer termination. | |
| 526 par_scan_state()->start_term_time(); | |
| 527 if (terminator()->offer_termination()) break; | |
| 528 par_scan_state()->end_term_time(); | |
| 529 } | |
| 534 | 530 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0, |
| 531 "Broken overflow list?"); | |
| 0 | 532 // Finish the last termination pause. |
| 533 par_scan_state()->end_term_time(); | |
| 534 } | |
| 535 | |
| 536 ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen, | |
| 537 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) : | |
| 538 AbstractGangTask("ParNewGeneration collection"), | |
| 539 _gen(gen), _next_gen(next_gen), | |
| 540 _young_old_boundary(young_old_boundary), | |
| 541 _state_set(state_set) | |
| 542 {} | |
| 543 | |
| 544 void ParNewGenTask::work(int i) { | |
| 545 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
| 546 // Since this is being done in a separate thread, need new resource | |
| 547 // and handle marks. | |
| 548 ResourceMark rm; | |
| 549 HandleMark hm; | |
| 550 // 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
|
551 assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen."); |
| 0 | 552 |
| 553 Generation* old_gen = gch->next_gen(_gen); | |
| 554 | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
555 ParScanThreadState& par_scan_state = _state_set->thread_state(i); |
| 0 | 556 par_scan_state.set_young_old_boundary(_young_old_boundary); |
| 557 | |
| 558 par_scan_state.start_strong_roots(); | |
| 559 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
|
560 true, // Process younger gens, if any, |
|
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
561 // as strong roots. |
|
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
562 false, // no scope; this is parallel code |
|
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
563 false, // not collecting perm generation. |
| 0 | 564 SharedHeap::SO_AllClasses, |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
565 &par_scan_state.to_space_root_closure(), |
|
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
566 true, // walk *all* scavengable nmethods |
|
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
567 &par_scan_state.older_gen_closure()); |
| 0 | 568 par_scan_state.end_strong_roots(); |
| 569 | |
| 570 // "evacuate followers". | |
| 571 par_scan_state.evacuate_followers_closure().do_void(); | |
| 572 } | |
| 573 | |
| 574 #ifdef _MSC_VER | |
| 575 #pragma warning( push ) | |
| 576 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list | |
| 577 #endif | |
| 578 ParNewGeneration:: | |
| 579 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level) | |
| 580 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"), | |
| 581 _overflow_list(NULL), | |
| 582 _is_alive_closure(this), | |
| 583 _plab_stats(YoungPLABSize, PLABWeight) | |
| 584 { | |
| 534 | 585 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;) |
| 586 NOT_PRODUCT(_num_par_pushes = 0;) | |
| 0 | 587 _task_queues = new ObjToScanQueueSet(ParallelGCThreads); |
| 588 guarantee(_task_queues != NULL, "task_queues allocation failure."); | |
| 589 | |
| 590 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) { | |
| 1665 | 591 ObjToScanQueue *q = new ObjToScanQueue(); |
| 592 guarantee(q != NULL, "work_queue Allocation failure."); | |
| 593 _task_queues->register_queue(i1, q); | |
| 0 | 594 } |
| 595 | |
| 596 for (uint i2 = 0; i2 < ParallelGCThreads; i2++) | |
| 597 _task_queues->queue(i2)->initialize(); | |
| 598 | |
| 695 | 599 _overflow_stacks = NEW_C_HEAP_ARRAY(GrowableArray<oop>*, ParallelGCThreads); |
| 600 guarantee(_overflow_stacks != NULL, "Overflow stack set allocation failure"); | |
| 601 for (uint i = 0; i < ParallelGCThreads; i++) { | |
| 602 if (ParGCUseLocalOverflow) { | |
| 603 _overflow_stacks[i] = new (ResourceObj::C_HEAP) GrowableArray<oop>(512, true); | |
| 604 guarantee(_overflow_stacks[i] != NULL, "Overflow Stack allocation failure."); | |
| 605 } else { | |
| 606 _overflow_stacks[i] = NULL; | |
| 607 } | |
| 608 } | |
| 609 | |
| 0 | 610 if (UsePerfData) { |
| 611 EXCEPTION_MARK; | |
| 612 ResourceMark rm; | |
| 613 | |
| 614 const char* cname = | |
| 615 PerfDataManager::counter_name(_gen_counters->name_space(), "threads"); | |
| 616 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None, | |
| 617 ParallelGCThreads, CHECK); | |
| 618 } | |
| 619 } | |
| 620 #ifdef _MSC_VER | |
| 621 #pragma warning( pop ) | |
| 622 #endif | |
| 623 | |
| 624 // ParNewGeneration:: | |
| 625 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) : | |
| 626 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {} | |
| 627 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
628 template <class T> |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
629 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
|
630 #ifdef ASSERT |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
631 { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
632 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
|
633 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
|
634 // 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
|
635 // as a weak reference. |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
636 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
|
637 } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
638 #endif // ASSERT |
| 0 | 639 |
| 640 _par_cl->do_oop_nv(p); | |
| 641 | |
| 642 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
|
643 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
|
644 _rs->write_ref_field_gc_par(p, obj); |
| 0 | 645 } |
| 646 } | |
| 647 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
648 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
|
649 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
|
650 |
| 0 | 651 // ParNewGeneration:: |
| 652 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) : | |
| 653 DefNewGeneration::KeepAliveClosure(cl) {} | |
| 654 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
655 template <class T> |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
656 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
|
657 #ifdef ASSERT |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
658 { |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
659 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
|
660 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
|
661 // 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
|
662 // as a weak reference. |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
663 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
|
664 } |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
665 #endif // ASSERT |
| 0 | 666 |
| 667 _cl->do_oop_nv(p); | |
| 668 | |
| 669 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
|
670 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
|
671 _rs->write_ref_field_gc_par(p, obj); |
| 0 | 672 } |
| 673 } | |
| 674 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
675 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
|
676 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
|
677 |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
678 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
|
679 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
|
680 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
|
681 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); |
| 0 | 682 if ((HeapWord*)obj < _boundary) { |
| 683 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
|
684 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
|
685 ? obj->forwardee() |
|
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
686 : _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
|
687 oopDesc::encode_store_heap_oop_not_null(p, new_obj); |
| 0 | 688 } |
| 689 if (_gc_barrier) { | |
| 690 // If p points to a younger generation, mark the card. | |
| 691 if ((HeapWord*)obj < _gen_boundary) { | |
| 692 _rs->write_ref_field_gc_par(p, obj); | |
| 693 } | |
| 694 } | |
| 695 } | |
| 696 } | |
| 697 | |
|
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
698 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
|
699 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
|
700 |
| 0 | 701 class ParNewRefProcTaskProxy: public AbstractGangTask { |
| 702 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; | |
| 703 public: | |
| 704 ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen, | |
| 705 Generation& next_gen, | |
| 706 HeapWord* young_old_boundary, | |
| 707 ParScanThreadStateSet& state_set); | |
| 708 | |
| 709 private: | |
| 710 virtual void work(int i); | |
| 711 | |
| 712 private: | |
| 713 ParNewGeneration& _gen; | |
| 714 ProcessTask& _task; | |
| 715 Generation& _next_gen; | |
| 716 HeapWord* _young_old_boundary; | |
| 717 ParScanThreadStateSet& _state_set; | |
| 718 }; | |
| 719 | |
| 720 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy( | |
| 721 ProcessTask& task, ParNewGeneration& gen, | |
| 722 Generation& next_gen, | |
| 723 HeapWord* young_old_boundary, | |
| 724 ParScanThreadStateSet& state_set) | |
| 725 : AbstractGangTask("ParNewGeneration parallel reference processing"), | |
| 726 _gen(gen), | |
| 727 _task(task), | |
| 728 _next_gen(next_gen), | |
| 729 _young_old_boundary(young_old_boundary), | |
| 730 _state_set(state_set) | |
| 731 { | |
| 732 } | |
| 733 | |
| 734 void ParNewRefProcTaskProxy::work(int i) | |
| 735 { | |
| 736 ResourceMark rm; | |
| 737 HandleMark hm; | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
738 ParScanThreadState& par_scan_state = _state_set.thread_state(i); |
| 0 | 739 par_scan_state.set_young_old_boundary(_young_old_boundary); |
| 740 _task.work(i, par_scan_state.is_alive_closure(), | |
| 741 par_scan_state.keep_alive_closure(), | |
| 742 par_scan_state.evacuate_followers_closure()); | |
| 743 } | |
| 744 | |
| 745 class ParNewRefEnqueueTaskProxy: public AbstractGangTask { | |
| 746 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; | |
| 747 EnqueueTask& _task; | |
| 748 | |
| 749 public: | |
| 750 ParNewRefEnqueueTaskProxy(EnqueueTask& task) | |
| 751 : AbstractGangTask("ParNewGeneration parallel reference enqueue"), | |
| 752 _task(task) | |
| 753 { } | |
| 754 | |
| 755 virtual void work(int i) | |
| 756 { | |
| 757 _task.work(i); | |
| 758 } | |
| 759 }; | |
| 760 | |
| 761 | |
| 762 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) | |
| 763 { | |
| 764 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
| 765 assert(gch->kind() == CollectedHeap::GenCollectedHeap, | |
| 766 "not a generational heap"); | |
| 767 WorkGang* workers = gch->workers(); | |
| 768 assert(workers != NULL, "Need parallel worker threads."); | |
| 769 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(), | |
| 770 _generation.reserved().end(), _state_set); | |
| 771 workers->run_task(&rp_task); | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
772 _state_set.reset(_generation.promotion_failed()); |
| 0 | 773 } |
| 774 | |
| 775 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) | |
| 776 { | |
| 777 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
| 778 WorkGang* workers = gch->workers(); | |
| 779 assert(workers != NULL, "Need parallel worker threads."); | |
| 780 ParNewRefEnqueueTaskProxy enq_task(task); | |
| 781 workers->run_task(&enq_task); | |
| 782 } | |
| 783 | |
| 784 void ParNewRefProcTaskExecutor::set_single_threaded_mode() | |
| 785 { | |
| 786 _state_set.flush(); | |
| 787 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
| 788 gch->set_par_threads(0); // 0 ==> non-parallel. | |
| 789 gch->save_marks(); | |
| 790 } | |
| 791 | |
| 792 ScanClosureWithParBarrier:: | |
| 793 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) : | |
| 794 ScanClosure(g, gc_barrier) {} | |
| 795 | |
| 796 EvacuateFollowersClosureGeneral:: | |
| 797 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level, | |
| 798 OopsInGenClosure* cur, | |
| 799 OopsInGenClosure* older) : | |
| 800 _gch(gch), _level(level), | |
| 801 _scan_cur_or_nonheap(cur), _scan_older(older) | |
| 802 {} | |
| 803 | |
| 804 void EvacuateFollowersClosureGeneral::do_void() { | |
| 805 do { | |
| 806 // Beware: this call will lead to closure applications via virtual | |
| 807 // calls. | |
| 808 _gch->oop_since_save_marks_iterate(_level, | |
| 809 _scan_cur_or_nonheap, | |
| 810 _scan_older); | |
| 811 } while (!_gch->no_allocs_since_save_marks(_level)); | |
| 812 } | |
| 813 | |
| 814 | |
| 815 bool ParNewGeneration::_avoid_promotion_undo = false; | |
| 816 | |
| 817 void ParNewGeneration::adjust_desired_tenuring_threshold() { | |
| 818 // Set the desired survivor size to half the real survivor space | |
| 819 _tenuring_threshold = | |
| 820 age_table()->compute_tenuring_threshold(to()->capacity()/HeapWordSize); | |
| 821 } | |
| 822 | |
| 823 // A Generation that does parallel young-gen collection. | |
| 824 | |
| 825 void ParNewGeneration::collect(bool full, | |
| 826 bool clear_all_soft_refs, | |
| 827 size_t size, | |
| 828 bool is_tlab) { | |
| 829 assert(full || size > 0, "otherwise we don't want to collect"); | |
| 830 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
| 831 assert(gch->kind() == CollectedHeap::GenCollectedHeap, | |
| 832 "not a CMS generational heap"); | |
| 833 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy(); | |
| 834 WorkGang* workers = gch->workers(); | |
| 835 _next_gen = gch->next_gen(this); | |
| 836 assert(_next_gen != NULL, | |
| 837 "This must be the youngest gen, and not the only gen"); | |
| 838 assert(gch->n_gens() == 2, | |
| 839 "Par collection currently only works with single older gen."); | |
| 840 // Do we have to avoid promotion_undo? | |
| 841 if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) { | |
| 842 set_avoid_promotion_undo(true); | |
| 843 } | |
| 844 | |
| 845 // If the next generation is too full to accomodate worst-case promotion | |
| 846 // from this generation, pass on collection; let the next generation | |
| 847 // do it. | |
| 848 if (!collection_attempt_is_safe()) { | |
| 849 gch->set_incremental_collection_will_fail(); | |
| 850 return; | |
| 851 } | |
| 852 assert(to()->is_empty(), "Else not collection_attempt_is_safe"); | |
| 853 | |
| 854 init_assuming_no_promotion_failure(); | |
| 855 | |
| 856 if (UseAdaptiveSizePolicy) { | |
| 857 set_survivor_overflow(false); | |
| 858 size_policy->minor_collection_begin(); | |
| 859 } | |
| 860 | |
| 861 TraceTime t1("GC", PrintGC && !PrintGCDetails, true, gclog_or_tty); | |
| 862 // Capture heap used before collection (for printing). | |
| 863 size_t gch_prev_used = gch->used(); | |
| 864 | |
| 865 SpecializationStats::clear(); | |
| 866 | |
| 867 age_table()->clear(); | |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
868 to()->clear(SpaceDecorator::Mangle); |
| 0 | 869 |
| 870 gch->save_marks(); | |
| 871 assert(workers != NULL, "Need parallel worker threads."); | |
| 872 ParallelTaskTerminator _term(workers->total_workers(), task_queues()); | |
| 873 ParScanThreadStateSet thread_state_set(workers->total_workers(), | |
| 874 *to(), *this, *_next_gen, *task_queues(), | |
| 695 | 875 _overflow_stacks, desired_plab_sz(), _term); |
| 0 | 876 |
| 877 ParNewGenTask tsk(this, _next_gen, reserved().end(), &thread_state_set); | |
| 878 int n_workers = workers->total_workers(); | |
| 879 gch->set_par_threads(n_workers); | |
| 880 gch->rem_set()->prepare_for_younger_refs_iterate(true); | |
| 881 // It turns out that even when we're using 1 thread, doing the work in a | |
| 882 // separate thread causes wide variance in run times. We can't help this | |
| 883 // in the multi-threaded case, but we special-case n=1 here to get | |
| 884 // repeatable measurements of the 1-thread overhead of the parallel code. | |
| 885 if (n_workers > 1) { | |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
886 GenCollectedHeap::StrongRootsScope srs(gch); |
| 0 | 887 workers->run_task(&tsk); |
| 888 } else { | |
|
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
695
diff
changeset
|
889 GenCollectedHeap::StrongRootsScope srs(gch); |
| 0 | 890 tsk.work(0); |
| 891 } | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
892 thread_state_set.reset(promotion_failed()); |
| 0 | 893 |
| 894 // Process (weak) reference objects found during scavenge. | |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
895 ReferenceProcessor* rp = ref_processor(); |
| 0 | 896 IsAliveClosure is_alive(this); |
| 897 ScanWeakRefClosure scan_weak_ref(this); | |
| 898 KeepAliveClosure keep_alive(&scan_weak_ref); | |
| 899 ScanClosure scan_without_gc_barrier(this, false); | |
| 900 ScanClosureWithParBarrier scan_with_gc_barrier(this, true); | |
| 901 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier); | |
| 902 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level, | |
| 903 &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
|
904 rp->setup_policy(clear_all_soft_refs); |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
905 if (rp->processing_is_mt()) { |
| 0 | 906 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
907 rp->process_discovered_references(&is_alive, &keep_alive, |
|
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
908 &evacuate_followers, &task_executor); |
| 0 | 909 } else { |
| 910 thread_state_set.flush(); | |
| 911 gch->set_par_threads(0); // 0 ==> non-parallel. | |
| 912 gch->save_marks(); | |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
913 rp->process_discovered_references(&is_alive, &keep_alive, |
|
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
914 &evacuate_followers, NULL); |
| 0 | 915 } |
| 916 if (!promotion_failed()) { | |
| 917 // Swap the survivor spaces. | |
|
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
918 eden()->clear(SpaceDecorator::Mangle); |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
919 from()->clear(SpaceDecorator::Mangle); |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
920 if (ZapUnusedHeapArea) { |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
921 // 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
|
922 // 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
|
923 // 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
|
924 // 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
|
925 // 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
|
926 // 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
|
927 // other spaces. |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
928 to()->mangle_unused_area(); |
|
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
929 } |
| 0 | 930 swap_spaces(); |
| 931 | |
|
1387
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1145
diff
changeset
|
932 // 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
|
933 // for full GC's. |
|
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1145
diff
changeset
|
934 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
|
935 |
| 0 | 936 assert(to()->is_empty(), "to space should be empty now"); |
| 937 } else { | |
| 938 assert(HandlePromotionFailure, | |
| 939 "Should only be here if promotion failure handling is on"); | |
| 940 if (_promo_failure_scan_stack != NULL) { | |
| 941 // Can be non-null because of reference processing. | |
| 942 // Free stack with its elements. | |
| 943 delete _promo_failure_scan_stack; | |
| 944 _promo_failure_scan_stack = NULL; | |
| 945 } | |
| 946 remove_forwarding_pointers(); | |
| 947 if (PrintGCDetails) { | |
| 948 gclog_or_tty->print(" (promotion failed)"); | |
| 949 } | |
| 950 // All the spaces are in play for mark-sweep. | |
| 951 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690. | |
| 952 from()->set_next_compaction_space(to()); | |
| 953 gch->set_incremental_collection_will_fail(); | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
954 // Inform the next generation that a promotion failure occurred. |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
955 _next_gen->promotion_failure_occurred(); |
|
6
73e96e5c30df
6624765: Guarantee failure "Unexpected dirty card found"
jmasa
parents:
0
diff
changeset
|
956 |
|
73e96e5c30df
6624765: Guarantee failure "Unexpected dirty card found"
jmasa
parents:
0
diff
changeset
|
957 // Reset the PromotionFailureALot counters. |
|
73e96e5c30df
6624765: Guarantee failure "Unexpected dirty card found"
jmasa
parents:
0
diff
changeset
|
958 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();) |
| 0 | 959 } |
| 960 // set new iteration safe limit for the survivor spaces | |
| 961 from()->set_concurrent_iteration_safe_limit(from()->top()); | |
| 962 to()->set_concurrent_iteration_safe_limit(to()->top()); | |
| 963 | |
| 964 adjust_desired_tenuring_threshold(); | |
| 965 if (ResizePLAB) { | |
| 966 plab_stats()->adjust_desired_plab_sz(); | |
| 967 } | |
| 968 | |
| 969 if (PrintGC && !PrintGCDetails) { | |
| 970 gch->print_heap_change(gch_prev_used); | |
| 971 } | |
| 972 | |
| 1710 | 973 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats()); |
| 974 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats()); | |
| 975 | |
| 0 | 976 if (UseAdaptiveSizePolicy) { |
| 977 size_policy->minor_collection_end(gch->gc_cause()); | |
| 978 size_policy->avg_survived()->sample(from()->used()); | |
| 979 } | |
| 980 | |
| 981 update_time_of_last_gc(os::javaTimeMillis()); | |
| 982 | |
| 983 SpecializationStats::print(); | |
| 984 | |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
985 rp->set_enqueuing_is_done(true); |
|
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
986 if (rp->processing_is_mt()) { |
| 0 | 987 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
988 rp->enqueue_discovered_references(&task_executor); |
| 0 | 989 } else { |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
990 rp->enqueue_discovered_references(NULL); |
| 0 | 991 } |
|
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
992 rp->verify_no_references_recorded(); |
| 0 | 993 } |
| 994 | |
| 995 static int sum; | |
| 996 void ParNewGeneration::waste_some_time() { | |
| 997 for (int i = 0; i < 100; i++) { | |
| 998 sum += i; | |
| 999 } | |
| 1000 } | |
| 1001 | |
| 1002 static const oop ClaimedForwardPtr = oop(0x4); | |
| 1003 | |
| 1004 // Because of concurrency, there are times where an object for which | |
| 1005 // "is_forwarded()" is true contains an "interim" forwarding pointer | |
| 1006 // value. Such a value will soon be overwritten with a real value. | |
| 1007 // This method requires "obj" to have a forwarding pointer, and waits, if | |
| 1008 // necessary for a real one to be inserted, and returns it. | |
| 1009 | |
| 1010 oop ParNewGeneration::real_forwardee(oop obj) { | |
| 1011 oop forward_ptr = obj->forwardee(); | |
| 1012 if (forward_ptr != ClaimedForwardPtr) { | |
| 1013 return forward_ptr; | |
| 1014 } else { | |
| 1015 return real_forwardee_slow(obj); | |
| 1016 } | |
| 1017 } | |
| 1018 | |
| 1019 oop ParNewGeneration::real_forwardee_slow(oop obj) { | |
| 1020 // Spin-read if it is claimed but not yet written by another thread. | |
| 1021 oop forward_ptr = obj->forwardee(); | |
| 1022 while (forward_ptr == ClaimedForwardPtr) { | |
| 1023 waste_some_time(); | |
| 1024 assert(obj->is_forwarded(), "precondition"); | |
| 1025 forward_ptr = obj->forwardee(); | |
| 1026 } | |
| 1027 return forward_ptr; | |
| 1028 } | |
| 1029 | |
| 1030 #ifdef ASSERT | |
| 1031 bool ParNewGeneration::is_legal_forward_ptr(oop p) { | |
| 1032 return | |
| 1033 (_avoid_promotion_undo && p == ClaimedForwardPtr) | |
| 1034 || Universe::heap()->is_in_reserved(p); | |
| 1035 } | |
| 1036 #endif | |
| 1037 | |
| 1038 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) { | |
| 1039 if ((m != markOopDesc::prototype()) && | |
| 1040 (!UseBiasedLocking || (m != markOopDesc::biased_locking_prototype()))) { | |
| 1041 MutexLocker ml(ParGCRareEvent_lock); | |
| 1042 DefNewGeneration::preserve_mark_if_necessary(obj, m); | |
| 1043 } | |
| 1044 } | |
| 1045 | |
| 1046 // Multiple GC threads may try to promote an object. If the object | |
| 1047 // is successfully promoted, a forwarding pointer will be installed in | |
| 1048 // the object in the young generation. This method claims the right | |
| 1049 // to install the forwarding pointer before it copies the object, | |
| 1050 // thus avoiding the need to undo the copy as in | |
| 1051 // copy_to_survivor_space_avoiding_with_undo. | |
| 1052 | |
| 1053 oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo( | |
| 1054 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { | |
| 1055 // In the sequential version, this assert also says that the object is | |
| 1056 // not forwarded. That might not be the case here. It is the case that | |
| 1057 // the caller observed it to be not forwarded at some time in the past. | |
| 1058 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); | |
| 1059 | |
| 1060 // The sequential code read "old->age()" below. That doesn't work here, | |
| 1061 // since the age is in the mark word, and that might be overwritten with | |
| 1062 // a forwarding pointer by a parallel thread. So we must save the mark | |
| 1063 // word in a local and then analyze it. | |
| 1064 oopDesc dummyOld; | |
| 1065 dummyOld.set_mark(m); | |
| 1066 assert(!dummyOld.is_forwarded(), | |
| 1067 "should not be called with forwarding pointer mark word."); | |
| 1068 | |
| 1069 oop new_obj = NULL; | |
| 1070 oop forward_ptr; | |
| 1071 | |
| 1072 // Try allocating obj in to-space (unless too old) | |
| 1073 if (dummyOld.age() < tenuring_threshold()) { | |
| 1074 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); | |
| 1075 if (new_obj == NULL) { | |
| 1076 set_survivor_overflow(true); | |
| 1077 } | |
| 1078 } | |
| 1079 | |
| 1080 if (new_obj == NULL) { | |
| 1081 // Either to-space is full or we decided to promote | |
| 1082 // try allocating obj tenured | |
| 1083 | |
| 1084 // Attempt to install a null forwarding pointer (atomically), | |
| 1085 // to claim the right to install the real forwarding pointer. | |
| 1086 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr); | |
| 1087 if (forward_ptr != NULL) { | |
| 1088 // someone else beat us to it. | |
| 1089 return real_forwardee(old); | |
| 1090 } | |
| 1091 | |
| 1092 new_obj = _next_gen->par_promote(par_scan_state->thread_num(), | |
| 1093 old, m, sz); | |
| 1094 | |
| 1095 if (new_obj == NULL) { | |
| 1096 if (!HandlePromotionFailure) { | |
| 1097 // A failed promotion likely means the MaxLiveObjectEvacuationRatio flag | |
| 1098 // is incorrectly set. In any case, its seriously wrong to be here! | |
| 1099 vm_exit_out_of_memory(sz*wordSize, "promotion"); | |
| 1100 } | |
| 1101 // promotion failed, forward to self | |
| 1102 _promotion_failed = true; | |
| 1103 new_obj = old; | |
| 1104 | |
| 1105 preserve_mark_if_necessary(old, m); | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
1106 // Log the size of the maiden promotion failure |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
1107 par_scan_state->log_promotion_failure(sz); |
| 0 | 1108 } |
| 1109 | |
| 1110 old->forward_to(new_obj); | |
| 1111 forward_ptr = NULL; | |
| 1112 } else { | |
| 1113 // Is in to-space; do copying ourselves. | |
| 1114 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); | |
| 1115 forward_ptr = old->forward_to_atomic(new_obj); | |
| 1116 // Restore the mark word copied above. | |
| 1117 new_obj->set_mark(m); | |
| 1118 // Increment age if obj still in new generation | |
| 1119 new_obj->incr_age(); | |
| 1120 par_scan_state->age_table()->add(new_obj, sz); | |
| 1121 } | |
| 1122 assert(new_obj != NULL, "just checking"); | |
| 1123 | |
| 1124 if (forward_ptr == NULL) { | |
| 1125 oop obj_to_push = new_obj; | |
| 1126 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { | |
| 1127 // Length field used as index of next element to be scanned. | |
| 1128 // Real length can be obtained from real_forwardee() | |
| 1129 arrayOop(old)->set_length(0); | |
| 1130 obj_to_push = old; | |
| 1131 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, | |
| 1132 "push forwarded object"); | |
| 1133 } | |
| 1134 // Push it on one of the queues of to-be-scanned objects. | |
| 534 | 1135 bool simulate_overflow = false; |
| 1136 NOT_PRODUCT( | |
| 1137 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { | |
| 1138 // simulate a stack overflow | |
| 1139 simulate_overflow = true; | |
| 1140 } | |
| 1141 ) | |
| 1142 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { | |
| 0 | 1143 // Add stats for overflow pushes. |
| 1144 if (Verbose && PrintGCDetails) { | |
| 1145 gclog_or_tty->print("queue overflow!\n"); | |
| 1146 } | |
| 534 | 1147 push_on_overflow_list(old, par_scan_state); |
| 1710 | 1148 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); |
| 0 | 1149 } |
| 1150 | |
| 1151 return new_obj; | |
| 1152 } | |
| 1153 | |
| 1154 // Oops. Someone beat us to it. Undo the allocation. Where did we | |
| 1155 // allocate it? | |
| 1156 if (is_in_reserved(new_obj)) { | |
| 1157 // Must be in to_space. | |
| 1158 assert(to()->is_in_reserved(new_obj), "Checking"); | |
| 1159 if (forward_ptr == ClaimedForwardPtr) { | |
| 1160 // Wait to get the real forwarding pointer value. | |
| 1161 forward_ptr = real_forwardee(old); | |
| 1162 } | |
| 1163 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); | |
| 1164 } | |
| 1165 | |
| 1166 return forward_ptr; | |
| 1167 } | |
| 1168 | |
| 1169 | |
| 1170 // Multiple GC threads may try to promote the same object. If two | |
| 1171 // or more GC threads copy the object, only one wins the race to install | |
| 1172 // the forwarding pointer. The other threads have to undo their copy. | |
| 1173 | |
| 1174 oop ParNewGeneration::copy_to_survivor_space_with_undo( | |
| 1175 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { | |
| 1176 | |
| 1177 // In the sequential version, this assert also says that the object is | |
| 1178 // not forwarded. That might not be the case here. It is the case that | |
| 1179 // the caller observed it to be not forwarded at some time in the past. | |
| 1180 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); | |
| 1181 | |
| 1182 // The sequential code read "old->age()" below. That doesn't work here, | |
| 1183 // since the age is in the mark word, and that might be overwritten with | |
| 1184 // a forwarding pointer by a parallel thread. So we must save the mark | |
| 1185 // word here, install it in a local oopDesc, and then analyze it. | |
| 1186 oopDesc dummyOld; | |
| 1187 dummyOld.set_mark(m); | |
| 1188 assert(!dummyOld.is_forwarded(), | |
| 1189 "should not be called with forwarding pointer mark word."); | |
| 1190 | |
| 1191 bool failed_to_promote = false; | |
| 1192 oop new_obj = NULL; | |
| 1193 oop forward_ptr; | |
| 1194 | |
| 1195 // Try allocating obj in to-space (unless too old) | |
| 1196 if (dummyOld.age() < tenuring_threshold()) { | |
| 1197 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); | |
| 1198 if (new_obj == NULL) { | |
| 1199 set_survivor_overflow(true); | |
| 1200 } | |
| 1201 } | |
| 1202 | |
| 1203 if (new_obj == NULL) { | |
| 1204 // Either to-space is full or we decided to promote | |
| 1205 // try allocating obj tenured | |
| 1206 new_obj = _next_gen->par_promote(par_scan_state->thread_num(), | |
| 1207 old, m, sz); | |
| 1208 | |
| 1209 if (new_obj == NULL) { | |
| 1210 if (!HandlePromotionFailure) { | |
| 1211 // A failed promotion likely means the MaxLiveObjectEvacuationRatio | |
| 1212 // flag is incorrectly set. In any case, its seriously wrong to be | |
| 1213 // here! | |
| 1214 vm_exit_out_of_memory(sz*wordSize, "promotion"); | |
| 1215 } | |
| 1216 // promotion failed, forward to self | |
| 1217 forward_ptr = old->forward_to_atomic(old); | |
| 1218 new_obj = old; | |
| 1219 | |
| 1220 if (forward_ptr != NULL) { | |
| 1221 return forward_ptr; // someone else succeeded | |
| 1222 } | |
| 1223 | |
| 1224 _promotion_failed = true; | |
| 1225 failed_to_promote = true; | |
| 1226 | |
| 1227 preserve_mark_if_necessary(old, m); | |
|
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
1228 // Log the size of the maiden promotion failure |
|
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
989
diff
changeset
|
1229 par_scan_state->log_promotion_failure(sz); |
| 0 | 1230 } |
| 1231 } else { | |
| 1232 // Is in to-space; do copying ourselves. | |
| 1233 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); | |
| 1234 // Restore the mark word copied above. | |
| 1235 new_obj->set_mark(m); | |
| 1236 // Increment age if new_obj still in new generation | |
| 1237 new_obj->incr_age(); | |
| 1238 par_scan_state->age_table()->add(new_obj, sz); | |
| 1239 } | |
| 1240 assert(new_obj != NULL, "just checking"); | |
| 1241 | |
| 1242 // Now attempt to install the forwarding pointer (atomically). | |
| 1243 // We have to copy the mark word before overwriting with forwarding | |
| 1244 // ptr, so we can restore it below in the copy. | |
| 1245 if (!failed_to_promote) { | |
| 1246 forward_ptr = old->forward_to_atomic(new_obj); | |
| 1247 } | |
| 1248 | |
| 1249 if (forward_ptr == NULL) { | |
| 1250 oop obj_to_push = new_obj; | |
| 1251 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { | |
| 1252 // Length field used as index of next element to be scanned. | |
| 1253 // Real length can be obtained from real_forwardee() | |
| 1254 arrayOop(old)->set_length(0); | |
| 1255 obj_to_push = old; | |
| 1256 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, | |
| 1257 "push forwarded object"); | |
| 1258 } | |
| 1259 // Push it on one of the queues of to-be-scanned objects. | |
| 534 | 1260 bool simulate_overflow = false; |
| 1261 NOT_PRODUCT( | |
| 1262 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { | |
| 1263 // simulate a stack overflow | |
| 1264 simulate_overflow = true; | |
| 1265 } | |
| 1266 ) | |
| 1267 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { | |
| 0 | 1268 // Add stats for overflow pushes. |
| 534 | 1269 push_on_overflow_list(old, par_scan_state); |
| 1710 | 1270 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0)); |
| 0 | 1271 } |
| 1272 | |
| 1273 return new_obj; | |
| 1274 } | |
| 1275 | |
| 1276 // Oops. Someone beat us to it. Undo the allocation. Where did we | |
| 1277 // allocate it? | |
| 1278 if (is_in_reserved(new_obj)) { | |
| 1279 // Must be in to_space. | |
| 1280 assert(to()->is_in_reserved(new_obj), "Checking"); | |
| 1281 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); | |
| 1282 } else { | |
| 1283 assert(!_avoid_promotion_undo, "Should not be here if avoiding."); | |
| 1284 _next_gen->par_promote_alloc_undo(par_scan_state->thread_num(), | |
| 1285 (HeapWord*)new_obj, sz); | |
| 1286 } | |
| 1287 | |
| 1288 return forward_ptr; | |
| 1289 } | |
| 1290 | |
| 534 | 1291 #ifndef PRODUCT |
| 1292 // It's OK to call this multi-threaded; the worst thing | |
| 1293 // that can happen is that we'll get a bunch of closely | |
| 1294 // spaced simulated oveflows, but that's OK, in fact | |
| 1295 // probably good as it would exercise the overflow code | |
| 1296 // under contention. | |
| 1297 bool ParNewGeneration::should_simulate_overflow() { | |
| 1298 if (_overflow_counter-- <= 0) { // just being defensive | |
| 1299 _overflow_counter = ParGCWorkQueueOverflowInterval; | |
| 1300 return true; | |
| 1301 } else { | |
| 1302 return false; | |
| 1303 } | |
| 1304 } | |
| 1305 #endif | |
| 1306 | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1307 // 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
|
1308 // 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
|
1309 // 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
|
1310 // 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
|
1311 // 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
|
1312 // 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
|
1313 // 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
|
1314 // 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
|
1315 // 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
|
1316 // 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
|
1317 // 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
|
1318 // 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
|
1319 // 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
|
1320 // 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
|
1321 // 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
|
1322 // 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
|
1323 // 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
|
1324 // (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
|
1325 // 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
|
1326 // as we were made painfully aware not long ago, see 6786503). |
| 534 | 1327 #define BUSY (oop(0x1aff1aff)) |
| 1328 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
|
1329 assert(is_in_reserved(from_space_obj), "Should be from this generation"); |
| 695 | 1330 if (ParGCUseLocalOverflow) { |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1331 // 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
|
1332 // overflow stack. |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1333 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
|
1334 } else { |
| 695 | 1335 assert(!UseCompressedOops, "Error"); |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1336 // 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
|
1337 // 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
|
1338 // 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
|
1339 // 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
|
1340 // and should be fixed. XXX FIX ME !!! |
| 534 | 1341 #ifndef PRODUCT |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1342 Atomic::inc_ptr(&_num_par_pushes); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1343 assert(_num_par_pushes > 0, "Tautology"); |
| 534 | 1344 #endif |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1345 if (from_space_obj->forwardee() == from_space_obj) { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1346 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1347 listhead->forward_to(from_space_obj); |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1348 from_space_obj = listhead; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1349 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1350 oop observed_overflow_list = _overflow_list; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1351 oop cur_overflow_list; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1352 do { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1353 cur_overflow_list = observed_overflow_list; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1354 if (cur_overflow_list != BUSY) { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1355 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
|
1356 } else { |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1357 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
|
1358 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1359 observed_overflow_list = |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1360 (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
|
1361 } while (cur_overflow_list != observed_overflow_list); |
| 0 | 1362 } |
| 1363 } | |
| 1364 | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1365 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
|
1366 bool res; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1367 |
| 695 | 1368 if (ParGCUseLocalOverflow) { |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1369 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
|
1370 } else { |
| 695 | 1371 assert(!UseCompressedOops, "Error"); |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1372 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
|
1373 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1374 return res; |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1375 } |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1376 |
|
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1377 |
| 534 | 1378 // *NOTE*: The overflow list manipulation code here and |
| 1379 // in CMSCollector:: are very similar in shape, | |
| 1380 // except that in the CMS case we thread the objects | |
| 1381 // directly into the list via their mark word, and do | |
| 1382 // not need to deal with special cases below related | |
| 1383 // to chunking of object arrays and promotion failure | |
| 1384 // handling. | |
| 1385 // CR 6797058 has been filed to attempt consolidation of | |
| 1386 // the common code. | |
| 1387 // Because of the common code, if you make any changes in | |
| 1388 // the code below, please check the CMS version to see if | |
| 1389 // similar changes might be needed. | |
| 1390 // See CMSCollector::par_take_from_overflow_list() for | |
| 1391 // more extensive documentation comments. | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1392 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) { |
| 0 | 1393 ObjToScanQueue* work_q = par_scan_state->work_queue(); |
| 1394 // How many to take? | |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1395 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, |
| 534 | 1396 (size_t)ParGCDesiredObjsFromOverflowList); |
| 0 | 1397 |
|
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1398 assert(par_scan_state->overflow_stack() == NULL, "Error"); |
| 695 | 1399 assert(!UseCompressedOops, "Error"); |
| 0 | 1400 if (_overflow_list == NULL) return false; |
| 1401 | |
| 1402 // Otherwise, there was something there; try claiming the list. | |
| 534 | 1403 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); |
| 1404 // Trim off a prefix of at most objsFromOverflow items | |
| 1405 Thread* tid = Thread::current(); | |
| 1406 size_t spin_count = (size_t)ParallelGCThreads; | |
| 1407 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100); | |
| 1408 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) { | |
| 1409 // someone grabbed it before we did ... | |
| 1410 // ... we spin for a short while... | |
| 1411 os::sleep(tid, sleep_time_millis, false); | |
| 1412 if (_overflow_list == NULL) { | |
| 1413 // nothing left to take | |
| 1414 return false; | |
| 1415 } else if (_overflow_list != BUSY) { | |
| 1416 // try and grab the prefix | |
| 1417 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); | |
| 1418 } | |
| 0 | 1419 } |
| 534 | 1420 if (prefix == NULL || prefix == BUSY) { |
| 1421 // Nothing to take or waited long enough | |
| 1422 if (prefix == NULL) { | |
| 1423 // Write back the NULL in case we overwrote it with BUSY above | |
| 1424 // and it is still the same value. | |
| 1425 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); | |
| 1426 } | |
| 1427 return false; | |
| 1428 } | |
| 1429 assert(prefix != NULL && prefix != BUSY, "Error"); | |
| 1430 size_t i = 1; | |
| 0 | 1431 oop cur = prefix; |
|
167
feeb96a45707
6696264: assert("narrow oop can never be zero") for GCBasher & ParNewGC
coleenp
parents:
113
diff
changeset
|
1432 while (i < objsFromOverflow && cur->klass_or_null() != NULL) { |
| 0 | 1433 i++; cur = oop(cur->klass()); |
| 1434 } | |
| 1435 | |
| 1436 // Reattach remaining (suffix) to overflow list | |
| 534 | 1437 if (cur->klass_or_null() == NULL) { |
| 1438 // Write back the NULL in lieu of the BUSY we wrote | |
| 1439 // above and it is still the same value. | |
| 1440 if (_overflow_list == BUSY) { | |
| 1441 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); | |
| 0 | 1442 } |
| 534 | 1443 } else { |
| 1444 assert(cur->klass_or_null() != BUSY, "Error"); | |
| 1445 oop suffix = oop(cur->klass()); // suffix will be put back on global list | |
| 1446 cur->set_klass_to_list_ptr(NULL); // break off suffix | |
| 1447 // It's possible that the list is still in the empty(busy) state | |
| 1448 // we left it in a short while ago; in that case we may be | |
| 1449 // able to place back the suffix. | |
| 1450 oop observed_overflow_list = _overflow_list; | |
| 1451 oop cur_overflow_list = observed_overflow_list; | |
| 1452 bool attached = false; | |
| 1453 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { | |
| 1454 observed_overflow_list = | |
| 1455 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); | |
| 1456 if (cur_overflow_list == observed_overflow_list) { | |
| 1457 attached = true; | |
| 1458 break; | |
| 1459 } else cur_overflow_list = observed_overflow_list; | |
| 1460 } | |
| 1461 if (!attached) { | |
| 1462 // Too bad, someone else got in in between; we'll need to do a splice. | |
| 1463 // Find the last item of suffix list | |
| 1464 oop last = suffix; | |
| 1465 while (last->klass_or_null() != NULL) { | |
| 1466 last = oop(last->klass()); | |
| 1467 } | |
| 1468 // Atomically prepend suffix to current overflow list | |
| 1469 observed_overflow_list = _overflow_list; | |
| 1470 do { | |
| 1471 cur_overflow_list = observed_overflow_list; | |
| 1472 if (cur_overflow_list != BUSY) { | |
| 1473 // Do the splice ... | |
| 1474 last->set_klass_to_list_ptr(cur_overflow_list); | |
| 1475 } else { // cur_overflow_list == BUSY | |
| 1476 last->set_klass_to_list_ptr(NULL); | |
| 1477 } | |
| 1478 observed_overflow_list = | |
| 1479 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); | |
| 1480 } while (cur_overflow_list != observed_overflow_list); | |
| 0 | 1481 } |
| 1482 } | |
| 1483 | |
| 1484 // Push objects on prefix list onto this thread's work queue | |
| 534 | 1485 assert(prefix != NULL && prefix != BUSY, "program logic"); |
| 0 | 1486 cur = prefix; |
| 534 | 1487 ssize_t n = 0; |
| 0 | 1488 while (cur != NULL) { |
| 1489 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
|
1490 oop next = oop(cur->klass_or_null()); |
| 0 | 1491 cur->set_klass(obj_to_push->klass()); |
| 534 | 1492 // This may be an array object that is self-forwarded. In that case, the list pointer |
| 1493 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed. | |
| 1494 if (!is_in_reserved(cur)) { | |
| 1495 // This can become a scaling bottleneck when there is work queue overflow coincident | |
| 1496 // with promotion failure. | |
| 1497 oopDesc* f = cur; | |
| 1498 FREE_C_HEAP_ARRAY(oopDesc, f); | |
| 1499 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) { | |
| 0 | 1500 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); |
| 534 | 1501 obj_to_push = cur; |
| 0 | 1502 } |
| 534 | 1503 bool ok = work_q->push(obj_to_push); |
| 1504 assert(ok, "Should have succeeded"); | |
| 0 | 1505 cur = next; |
| 1506 n++; | |
| 1507 } | |
| 1710 | 1508 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n)); |
| 534 | 1509 #ifndef PRODUCT |
| 1510 assert(_num_par_pushes >= n, "Too many pops?"); | |
| 1511 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); | |
| 1512 #endif | |
| 0 | 1513 return true; |
| 1514 } | |
| 534 | 1515 #undef BUSY |
| 0 | 1516 |
| 1517 void ParNewGeneration::ref_processor_init() | |
| 1518 { | |
| 1519 if (_ref_processor == NULL) { | |
| 1520 // Allocate and initialize a reference processor | |
| 1521 _ref_processor = ReferenceProcessor::create_ref_processor( | |
| 1522 _reserved, // span | |
| 1523 refs_discovery_is_atomic(), // atomic_discovery | |
| 1524 refs_discovery_is_mt(), // mt_discovery | |
| 1525 NULL, // is_alive_non_header | |
| 1526 ParallelGCThreads, | |
| 1527 ParallelRefProcEnabled); | |
| 1528 } | |
| 1529 } | |
| 1530 | |
| 1531 const char* ParNewGeneration::name() const { | |
| 1532 return "par new generation"; | |
| 1533 } |