Mercurial > hg > graal-jvmci-8
annotate src/share/vm/gc_implementation/parNew/parNewGeneration.cpp @ 679:cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
Summary: When using compressed oops, rather than chaining the overflowed grey objects' pre-images through their klass words, we use GC-worker thread-local overflow stacks.
Reviewed-by: jcoomes, jmasa
author | ysr |
---|---|
date | Sat, 28 Mar 2009 15:47:29 -0700 |
parents | 0fbdb4381b99 |
children | becb17ad5e51 |
rev | line source |
---|---|
0 | 1 /* |
579 | 2 * Copyright 2001-2009 Sun Microsystems, Inc. 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 * | |
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
21 * have any questions. | |
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_, | |
37 size_t desired_plab_sz_, | |
38 ParallelTaskTerminator& term_) : | |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
39 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_), |
0 | 40 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false), |
41 _ageTable(false), // false ==> not the global age table, no perf data. | |
42 _to_space_alloc_buffer(desired_plab_sz_), | |
43 _to_space_closure(gen_, this), _old_gen_closure(gen_, this), | |
44 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this), | |
45 _older_gen_closure(gen_, this), | |
46 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure, | |
47 &_to_space_root_closure, gen_, &_old_gen_root_closure, | |
48 work_queue_set_, &term_), | |
49 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this), | |
50 _keep_alive_closure(&_scan_weak_ref_closure), | |
51 _pushes(0), _pops(0), _steals(0), _steal_attempts(0), _term_attempts(0), | |
52 _strong_roots_time(0.0), _term_time(0.0) | |
53 { | |
54 _survivor_chunk_array = | |
55 (ChunkArray*) old_gen()->get_data_recorder(thread_num()); | |
56 _hash_seed = 17; // Might want to take time-based random value. | |
57 _start = os::elapsedTime(); | |
58 _old_gen_closure.set_generation(old_gen_); | |
59 _old_gen_root_closure.set_generation(old_gen_); | |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
60 if (UseCompressedOops) { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
61 _overflow_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(512, true); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
62 } else { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
63 _overflow_stack = NULL; |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
64 } |
0 | 65 } |
66 #ifdef _MSC_VER | |
67 #pragma warning( pop ) | |
68 #endif | |
69 | |
70 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start, | |
71 size_t plab_word_size) { | |
72 ChunkArray* sca = survivor_chunk_array(); | |
73 if (sca != NULL) { | |
74 // A non-null SCA implies that we want the PLAB data recorded. | |
75 sca->record_sample(plab_start, plab_word_size); | |
76 } | |
77 } | |
78 | |
79 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const { | |
80 return new_obj->is_objArray() && | |
81 arrayOop(new_obj)->length() > ParGCArrayScanChunk && | |
82 new_obj != old_obj; | |
83 } | |
84 | |
85 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) { | |
86 assert(old->is_objArray(), "must be obj array"); | |
87 assert(old->is_forwarded(), "must be forwarded"); | |
88 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
|
89 assert(!old_gen()->is_in(old), "must be in young generation."); |
0 | 90 |
91 objArrayOop obj = objArrayOop(old->forwardee()); | |
92 // Process ParGCArrayScanChunk elements now | |
93 // and push the remainder back onto queue | |
94 int start = arrayOop(old)->length(); | |
95 int end = obj->length(); | |
96 int remainder = end - start; | |
97 assert(start <= end, "just checking"); | |
98 if (remainder > 2 * ParGCArrayScanChunk) { | |
99 // Test above combines last partial chunk with a full chunk | |
100 end = start + ParGCArrayScanChunk; | |
101 arrayOop(old)->set_length(end); | |
102 // Push remainder. | |
103 bool ok = work_queue()->push(old); | |
104 assert(ok, "just popped, push must be okay"); | |
105 note_push(); | |
106 } else { | |
107 // Restore length so that it can be used if there | |
108 // is a promotion failure and forwarding pointers | |
109 // must be removed. | |
110 arrayOop(old)->set_length(end); | |
111 } | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
112 |
0 | 113 // 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
|
114 // should make sure end is even (aligned to HeapWord in case of compressed oops) |
0 | 115 if ((HeapWord *)obj < young_old_boundary()) { |
116 // 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
|
117 obj->oop_iterate_range(&_to_space_closure, start, end); |
0 | 118 } else { |
119 // 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
|
120 obj->oop_iterate_range(&_old_gen_closure, start, end); |
0 | 121 } |
122 } | |
123 | |
124 | |
125 void ParScanThreadState::trim_queues(int max_size) { | |
126 ObjToScanQueue* queue = work_queue(); | |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
127 do { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
128 while (queue->size() > (juint)max_size) { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
129 oop obj_to_scan; |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
130 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
|
131 note_pop(); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
132 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
|
133 if (obj_to_scan->is_objArray() && |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
134 obj_to_scan->is_forwarded() && |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
135 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
|
136 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
|
137 } else { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
138 // object is in to_space |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
139 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
|
140 } |
0 | 141 } else { |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
142 // object is in old generation |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
143 obj_to_scan->oop_iterate(&_old_gen_closure); |
0 | 144 } |
145 } | |
146 } | |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
147 // 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
|
148 // 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
|
149 // 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
|
150 // 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
|
151 // 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
|
152 // 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
|
153 // private overflow stack. |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
154 } 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
|
155 } |
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 bool ParScanThreadState::take_from_overflow_stack() { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
158 assert(UseCompressedOops, "Else should not call"); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
159 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
|
160 ObjToScanQueue* queue = work_queue(); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
161 GrowableArray<oop>* of_stack = overflow_stack(); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
162 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
|
163 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
|
164 (juint)ParGCDesiredObjsFromOverflowList), |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
165 num_overflow_elems); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
166 // 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
|
167 // stack to our work queue. |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
168 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
|
169 oop cur = of_stack->pop(); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
170 oop obj_to_push = cur->forwardee(); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
171 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
|
172 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
|
173 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
|
174 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
|
175 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
|
176 obj_to_push = cur; |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
177 } |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
178 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
|
179 assert(ok, "Should have succeeded"); |
0 | 180 } |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
181 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
|
182 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
|
183 } |
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 void ParScanThreadState::push_on_overflow_stack(oop p) { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
186 assert(UseCompressedOops, "Else should not call"); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
187 overflow_stack()->push(p); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
188 assert(young_gen()->overflow_list() == NULL, "Error"); |
0 | 189 } |
190 | |
191 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) { | |
192 | |
193 // Otherwise, if the object is small enough, try to reallocate the | |
194 // buffer. | |
195 HeapWord* obj = NULL; | |
196 if (!_to_space_full) { | |
197 ParGCAllocBuffer* const plab = to_space_alloc_buffer(); | |
198 Space* const sp = to_space(); | |
199 if (word_sz * 100 < | |
200 ParallelGCBufferWastePct * plab->word_sz()) { | |
201 // Is small enough; abandon this buffer and start a new one. | |
202 plab->retire(false, false); | |
203 size_t buf_size = plab->word_sz(); | |
204 HeapWord* buf_space = sp->par_allocate(buf_size); | |
205 if (buf_space == NULL) { | |
206 const size_t min_bytes = | |
207 ParGCAllocBuffer::min_size() << LogHeapWordSize; | |
208 size_t free_bytes = sp->free(); | |
209 while(buf_space == NULL && free_bytes >= min_bytes) { | |
210 buf_size = free_bytes >> LogHeapWordSize; | |
211 assert(buf_size == (size_t)align_object_size(buf_size), | |
212 "Invariant"); | |
213 buf_space = sp->par_allocate(buf_size); | |
214 free_bytes = sp->free(); | |
215 } | |
216 } | |
217 if (buf_space != NULL) { | |
218 plab->set_word_size(buf_size); | |
219 plab->set_buf(buf_space); | |
220 record_survivor_plab(buf_space, buf_size); | |
221 obj = plab->allocate(word_sz); | |
222 // Note that we cannot compare buf_size < word_sz below | |
223 // because of AlignmentReserve (see ParGCAllocBuffer::allocate()). | |
224 assert(obj != NULL || plab->words_remaining() < word_sz, | |
225 "Else should have been able to allocate"); | |
226 // It's conceivable that we may be able to use the | |
227 // buffer we just grabbed for subsequent small requests | |
228 // even if not for this one. | |
229 } else { | |
230 // We're used up. | |
231 _to_space_full = true; | |
232 } | |
233 | |
234 } else { | |
235 // Too large; allocate the object individually. | |
236 obj = sp->par_allocate(word_sz); | |
237 } | |
238 } | |
239 return obj; | |
240 } | |
241 | |
242 | |
243 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, | |
244 size_t word_sz) { | |
245 // Is the alloc in the current alloc buffer? | |
246 if (to_space_alloc_buffer()->contains(obj)) { | |
247 assert(to_space_alloc_buffer()->contains(obj + word_sz - 1), | |
248 "Should contain whole object."); | |
249 to_space_alloc_buffer()->undo_allocation(obj, word_sz); | |
250 } else { | |
481
7d7a7c599c17
6578152: fill_region_with_object has usability and safety issues
jcoomes
parents:
457
diff
changeset
|
251 CollectedHeap::fill_with_object(obj, word_sz); |
0 | 252 } |
253 } | |
254 | |
255 class ParScanThreadStateSet: private ResourceArray { | |
256 public: | |
257 // Initializes states for the specified number of threads; | |
258 ParScanThreadStateSet(int num_threads, | |
259 Space& to_space, | |
260 ParNewGeneration& gen, | |
261 Generation& old_gen, | |
262 ObjToScanQueueSet& queue_set, | |
263 size_t desired_plab_sz, | |
264 ParallelTaskTerminator& term); | |
265 inline ParScanThreadState& thread_sate(int i); | |
266 int pushes() { return _pushes; } | |
267 int pops() { return _pops; } | |
268 int steals() { return _steals; } | |
269 void reset(); | |
270 void flush(); | |
271 private: | |
272 ParallelTaskTerminator& _term; | |
273 ParNewGeneration& _gen; | |
274 Generation& _next_gen; | |
275 // staticstics | |
276 int _pushes; | |
277 int _pops; | |
278 int _steals; | |
279 }; | |
280 | |
281 | |
282 ParScanThreadStateSet::ParScanThreadStateSet( | |
283 int num_threads, Space& to_space, ParNewGeneration& gen, | |
284 Generation& old_gen, ObjToScanQueueSet& queue_set, | |
285 size_t desired_plab_sz, ParallelTaskTerminator& term) | |
286 : ResourceArray(sizeof(ParScanThreadState), num_threads), | |
287 _gen(gen), _next_gen(old_gen), _term(term), | |
288 _pushes(0), _pops(0), _steals(0) | |
289 { | |
290 assert(num_threads > 0, "sanity check!"); | |
291 // Initialize states. | |
292 for (int i = 0; i < num_threads; ++i) { | |
293 new ((ParScanThreadState*)_data + i) | |
294 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set, | |
295 desired_plab_sz, term); | |
296 } | |
297 } | |
298 | |
299 inline ParScanThreadState& ParScanThreadStateSet::thread_sate(int i) | |
300 { | |
301 assert(i >= 0 && i < length(), "sanity check!"); | |
302 return ((ParScanThreadState*)_data)[i]; | |
303 } | |
304 | |
305 | |
306 void ParScanThreadStateSet::reset() | |
307 { | |
308 _term.reset_for_reuse(); | |
309 } | |
310 | |
311 void ParScanThreadStateSet::flush() | |
312 { | |
313 for (int i = 0; i < length(); ++i) { | |
314 ParScanThreadState& par_scan_state = thread_sate(i); | |
315 | |
316 // Flush stats related to To-space PLAB activity and | |
317 // retire the last buffer. | |
318 par_scan_state.to_space_alloc_buffer()-> | |
319 flush_stats_and_retire(_gen.plab_stats(), | |
320 false /* !retain */); | |
321 | |
322 // Every thread has its own age table. We need to merge | |
323 // them all into one. | |
324 ageTable *local_table = par_scan_state.age_table(); | |
325 _gen.age_table()->merge(local_table); | |
326 | |
327 // Inform old gen that we're done. | |
328 _next_gen.par_promote_alloc_done(i); | |
329 _next_gen.par_oop_since_save_marks_iterate_done(i); | |
330 | |
331 // Flush stats related to work queue activity (push/pop/steal) | |
332 // This could conceivably become a bottleneck; if so, we'll put the | |
333 // stat's gathering under the flag. | |
334 if (PAR_STATS_ENABLED) { | |
335 _pushes += par_scan_state.pushes(); | |
336 _pops += par_scan_state.pops(); | |
337 _steals += par_scan_state.steals(); | |
338 if (ParallelGCVerbose) { | |
339 gclog_or_tty->print("Thread %d complete:\n" | |
340 " Pushes: %7d Pops: %7d Steals %7d (in %d attempts)\n", | |
341 i, par_scan_state.pushes(), par_scan_state.pops(), | |
342 par_scan_state.steals(), par_scan_state.steal_attempts()); | |
343 if (par_scan_state.overflow_pushes() > 0 || | |
344 par_scan_state.overflow_refills() > 0) { | |
345 gclog_or_tty->print(" Overflow pushes: %7d " | |
346 "Overflow refills: %7d for %d objs.\n", | |
347 par_scan_state.overflow_pushes(), | |
348 par_scan_state.overflow_refills(), | |
349 par_scan_state.overflow_refill_objs()); | |
350 } | |
351 | |
352 double elapsed = par_scan_state.elapsed(); | |
353 double strong_roots = par_scan_state.strong_roots_time(); | |
354 double term = par_scan_state.term_time(); | |
355 gclog_or_tty->print( | |
356 " Elapsed: %7.2f ms.\n" | |
357 " Strong roots: %7.2f ms (%6.2f%%)\n" | |
358 " Termination: %7.2f ms (%6.2f%%) (in %d entries)\n", | |
359 elapsed * 1000.0, | |
360 strong_roots * 1000.0, (strong_roots*100.0/elapsed), | |
361 term * 1000.0, (term*100.0/elapsed), | |
362 par_scan_state.term_attempts()); | |
363 } | |
364 } | |
365 } | |
366 } | |
367 | |
368 ParScanClosure::ParScanClosure(ParNewGeneration* g, | |
369 ParScanThreadState* par_scan_state) : | |
370 OopsInGenClosure(g), _par_scan_state(par_scan_state), _g(g) | |
371 { | |
372 assert(_g->level() == 0, "Optimized for youngest generation"); | |
373 _boundary = _g->reserved().end(); | |
374 } | |
375 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
376 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
|
377 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
|
378 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
379 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
|
380 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
|
381 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
382 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
|
383 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
|
384 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
385 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
|
386 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
|
387 |
0 | 388 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g, |
389 ParScanThreadState* par_scan_state) | |
390 : 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
|
391 {} |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
392 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
393 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
|
394 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); } |
0 | 395 |
396 #ifdef WIN32 | |
397 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */ | |
398 #endif | |
399 | |
400 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure( | |
401 ParScanThreadState* par_scan_state_, | |
402 ParScanWithoutBarrierClosure* to_space_closure_, | |
403 ParScanWithBarrierClosure* old_gen_closure_, | |
404 ParRootScanWithoutBarrierClosure* to_space_root_closure_, | |
405 ParNewGeneration* par_gen_, | |
406 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_, | |
407 ObjToScanQueueSet* task_queues_, | |
408 ParallelTaskTerminator* terminator_) : | |
409 | |
410 _par_scan_state(par_scan_state_), | |
411 _to_space_closure(to_space_closure_), | |
412 _old_gen_closure(old_gen_closure_), | |
413 _to_space_root_closure(to_space_root_closure_), | |
414 _old_gen_root_closure(old_gen_root_closure_), | |
415 _par_gen(par_gen_), | |
416 _task_queues(task_queues_), | |
417 _terminator(terminator_) | |
418 {} | |
419 | |
420 void ParEvacuateFollowersClosure::do_void() { | |
421 ObjToScanQueue* work_q = par_scan_state()->work_queue(); | |
422 | |
423 while (true) { | |
424 | |
425 // Scan to-space and old-gen objs until we run out of both. | |
426 oop obj_to_scan; | |
427 par_scan_state()->trim_queues(0); | |
428 | |
429 // We have no local work, attempt to steal from other threads. | |
430 | |
431 // attempt to steal work from promoted. | |
432 par_scan_state()->note_steal_attempt(); | |
433 if (task_queues()->steal(par_scan_state()->thread_num(), | |
434 par_scan_state()->hash_seed(), | |
435 obj_to_scan)) { | |
436 par_scan_state()->note_steal(); | |
437 bool res = work_q->push(obj_to_scan); | |
438 assert(res, "Empty queue should have room for a push."); | |
439 | |
440 par_scan_state()->note_push(); | |
441 // if successful, goto Start. | |
442 continue; | |
443 | |
444 // try global overflow list. | |
445 } else if (par_gen()->take_from_overflow_list(par_scan_state())) { | |
446 continue; | |
447 } | |
448 | |
449 // Otherwise, offer termination. | |
450 par_scan_state()->start_term_time(); | |
451 if (terminator()->offer_termination()) break; | |
452 par_scan_state()->end_term_time(); | |
453 } | |
534 | 454 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0, |
455 "Broken overflow list?"); | |
0 | 456 // Finish the last termination pause. |
457 par_scan_state()->end_term_time(); | |
458 } | |
459 | |
460 ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen, | |
461 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) : | |
462 AbstractGangTask("ParNewGeneration collection"), | |
463 _gen(gen), _next_gen(next_gen), | |
464 _young_old_boundary(young_old_boundary), | |
465 _state_set(state_set) | |
466 {} | |
467 | |
468 void ParNewGenTask::work(int i) { | |
469 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
470 // Since this is being done in a separate thread, need new resource | |
471 // and handle marks. | |
472 ResourceMark rm; | |
473 HandleMark hm; | |
474 // 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
|
475 assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen."); |
0 | 476 |
477 Generation* old_gen = gch->next_gen(_gen); | |
478 | |
479 ParScanThreadState& par_scan_state = _state_set->thread_sate(i); | |
480 par_scan_state.set_young_old_boundary(_young_old_boundary); | |
481 | |
482 par_scan_state.start_strong_roots(); | |
483 gch->gen_process_strong_roots(_gen->level(), | |
484 true, // Process younger gens, if any, | |
485 // as strong roots. | |
486 false,// not collecting perm generation. | |
487 SharedHeap::SO_AllClasses, | |
488 &par_scan_state.older_gen_closure(), | |
489 &par_scan_state.to_space_root_closure()); | |
490 par_scan_state.end_strong_roots(); | |
491 | |
492 // "evacuate followers". | |
493 par_scan_state.evacuate_followers_closure().do_void(); | |
494 } | |
495 | |
496 #ifdef _MSC_VER | |
497 #pragma warning( push ) | |
498 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list | |
499 #endif | |
500 ParNewGeneration:: | |
501 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level) | |
502 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"), | |
503 _overflow_list(NULL), | |
504 _is_alive_closure(this), | |
505 _plab_stats(YoungPLABSize, PLABWeight) | |
506 { | |
534 | 507 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;) |
508 NOT_PRODUCT(_num_par_pushes = 0;) | |
0 | 509 _task_queues = new ObjToScanQueueSet(ParallelGCThreads); |
510 guarantee(_task_queues != NULL, "task_queues allocation failure."); | |
511 | |
512 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) { | |
513 ObjToScanQueuePadded *q_padded = new ObjToScanQueuePadded(); | |
514 guarantee(q_padded != NULL, "work_queue Allocation failure."); | |
515 | |
516 _task_queues->register_queue(i1, &q_padded->work_queue); | |
517 } | |
518 | |
519 for (uint i2 = 0; i2 < ParallelGCThreads; i2++) | |
520 _task_queues->queue(i2)->initialize(); | |
521 | |
522 if (UsePerfData) { | |
523 EXCEPTION_MARK; | |
524 ResourceMark rm; | |
525 | |
526 const char* cname = | |
527 PerfDataManager::counter_name(_gen_counters->name_space(), "threads"); | |
528 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None, | |
529 ParallelGCThreads, CHECK); | |
530 } | |
531 } | |
532 #ifdef _MSC_VER | |
533 #pragma warning( pop ) | |
534 #endif | |
535 | |
536 // ParNewGeneration:: | |
537 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) : | |
538 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {} | |
539 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
540 template <class T> |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
541 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
|
542 #ifdef ASSERT |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
543 { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
544 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
|
545 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
|
546 // 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
|
547 // as a weak reference. |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
548 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
|
549 } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
550 #endif // ASSERT |
0 | 551 |
552 _par_cl->do_oop_nv(p); | |
553 | |
554 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
|
555 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
|
556 _rs->write_ref_field_gc_par(p, obj); |
0 | 557 } |
558 } | |
559 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
560 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
|
561 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
|
562 |
0 | 563 // ParNewGeneration:: |
564 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) : | |
565 DefNewGeneration::KeepAliveClosure(cl) {} | |
566 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
567 template <class T> |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
568 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
|
569 #ifdef ASSERT |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
570 { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
571 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
|
572 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
|
573 // 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
|
574 // as a weak reference. |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
575 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
|
576 } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
577 #endif // ASSERT |
0 | 578 |
579 _cl->do_oop_nv(p); | |
580 | |
581 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
|
582 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
|
583 _rs->write_ref_field_gc_par(p, obj); |
0 | 584 } |
585 } | |
586 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
587 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
|
588 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
|
589 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
590 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
|
591 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
|
592 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
|
593 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); |
0 | 594 if ((HeapWord*)obj < _boundary) { |
595 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
|
596 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
|
597 ? obj->forwardee() |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
598 : _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
|
599 oopDesc::encode_store_heap_oop_not_null(p, new_obj); |
0 | 600 } |
601 if (_gc_barrier) { | |
602 // If p points to a younger generation, mark the card. | |
603 if ((HeapWord*)obj < _gen_boundary) { | |
604 _rs->write_ref_field_gc_par(p, obj); | |
605 } | |
606 } | |
607 } | |
608 } | |
609 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
6
diff
changeset
|
610 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
|
611 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
|
612 |
0 | 613 class ParNewRefProcTaskProxy: public AbstractGangTask { |
614 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; | |
615 public: | |
616 ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen, | |
617 Generation& next_gen, | |
618 HeapWord* young_old_boundary, | |
619 ParScanThreadStateSet& state_set); | |
620 | |
621 private: | |
622 virtual void work(int i); | |
623 | |
624 private: | |
625 ParNewGeneration& _gen; | |
626 ProcessTask& _task; | |
627 Generation& _next_gen; | |
628 HeapWord* _young_old_boundary; | |
629 ParScanThreadStateSet& _state_set; | |
630 }; | |
631 | |
632 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy( | |
633 ProcessTask& task, ParNewGeneration& gen, | |
634 Generation& next_gen, | |
635 HeapWord* young_old_boundary, | |
636 ParScanThreadStateSet& state_set) | |
637 : AbstractGangTask("ParNewGeneration parallel reference processing"), | |
638 _gen(gen), | |
639 _task(task), | |
640 _next_gen(next_gen), | |
641 _young_old_boundary(young_old_boundary), | |
642 _state_set(state_set) | |
643 { | |
644 } | |
645 | |
646 void ParNewRefProcTaskProxy::work(int i) | |
647 { | |
648 ResourceMark rm; | |
649 HandleMark hm; | |
650 ParScanThreadState& par_scan_state = _state_set.thread_sate(i); | |
651 par_scan_state.set_young_old_boundary(_young_old_boundary); | |
652 _task.work(i, par_scan_state.is_alive_closure(), | |
653 par_scan_state.keep_alive_closure(), | |
654 par_scan_state.evacuate_followers_closure()); | |
655 } | |
656 | |
657 class ParNewRefEnqueueTaskProxy: public AbstractGangTask { | |
658 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; | |
659 EnqueueTask& _task; | |
660 | |
661 public: | |
662 ParNewRefEnqueueTaskProxy(EnqueueTask& task) | |
663 : AbstractGangTask("ParNewGeneration parallel reference enqueue"), | |
664 _task(task) | |
665 { } | |
666 | |
667 virtual void work(int i) | |
668 { | |
669 _task.work(i); | |
670 } | |
671 }; | |
672 | |
673 | |
674 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) | |
675 { | |
676 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
677 assert(gch->kind() == CollectedHeap::GenCollectedHeap, | |
678 "not a generational heap"); | |
679 WorkGang* workers = gch->workers(); | |
680 assert(workers != NULL, "Need parallel worker threads."); | |
681 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(), | |
682 _generation.reserved().end(), _state_set); | |
683 workers->run_task(&rp_task); | |
684 _state_set.reset(); | |
685 } | |
686 | |
687 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) | |
688 { | |
689 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
690 WorkGang* workers = gch->workers(); | |
691 assert(workers != NULL, "Need parallel worker threads."); | |
692 ParNewRefEnqueueTaskProxy enq_task(task); | |
693 workers->run_task(&enq_task); | |
694 } | |
695 | |
696 void ParNewRefProcTaskExecutor::set_single_threaded_mode() | |
697 { | |
698 _state_set.flush(); | |
699 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
700 gch->set_par_threads(0); // 0 ==> non-parallel. | |
701 gch->save_marks(); | |
702 } | |
703 | |
704 ScanClosureWithParBarrier:: | |
705 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) : | |
706 ScanClosure(g, gc_barrier) {} | |
707 | |
708 EvacuateFollowersClosureGeneral:: | |
709 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level, | |
710 OopsInGenClosure* cur, | |
711 OopsInGenClosure* older) : | |
712 _gch(gch), _level(level), | |
713 _scan_cur_or_nonheap(cur), _scan_older(older) | |
714 {} | |
715 | |
716 void EvacuateFollowersClosureGeneral::do_void() { | |
717 do { | |
718 // Beware: this call will lead to closure applications via virtual | |
719 // calls. | |
720 _gch->oop_since_save_marks_iterate(_level, | |
721 _scan_cur_or_nonheap, | |
722 _scan_older); | |
723 } while (!_gch->no_allocs_since_save_marks(_level)); | |
724 } | |
725 | |
726 | |
727 bool ParNewGeneration::_avoid_promotion_undo = false; | |
728 | |
729 void ParNewGeneration::adjust_desired_tenuring_threshold() { | |
730 // Set the desired survivor size to half the real survivor space | |
731 _tenuring_threshold = | |
732 age_table()->compute_tenuring_threshold(to()->capacity()/HeapWordSize); | |
733 } | |
734 | |
735 // A Generation that does parallel young-gen collection. | |
736 | |
737 void ParNewGeneration::collect(bool full, | |
738 bool clear_all_soft_refs, | |
739 size_t size, | |
740 bool is_tlab) { | |
741 assert(full || size > 0, "otherwise we don't want to collect"); | |
742 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
743 assert(gch->kind() == CollectedHeap::GenCollectedHeap, | |
744 "not a CMS generational heap"); | |
745 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy(); | |
746 WorkGang* workers = gch->workers(); | |
747 _next_gen = gch->next_gen(this); | |
748 assert(_next_gen != NULL, | |
749 "This must be the youngest gen, and not the only gen"); | |
750 assert(gch->n_gens() == 2, | |
751 "Par collection currently only works with single older gen."); | |
752 // Do we have to avoid promotion_undo? | |
753 if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) { | |
754 set_avoid_promotion_undo(true); | |
755 } | |
756 | |
757 // If the next generation is too full to accomodate worst-case promotion | |
758 // from this generation, pass on collection; let the next generation | |
759 // do it. | |
760 if (!collection_attempt_is_safe()) { | |
761 gch->set_incremental_collection_will_fail(); | |
762 return; | |
763 } | |
764 assert(to()->is_empty(), "Else not collection_attempt_is_safe"); | |
765 | |
766 init_assuming_no_promotion_failure(); | |
767 | |
768 if (UseAdaptiveSizePolicy) { | |
769 set_survivor_overflow(false); | |
770 size_policy->minor_collection_begin(); | |
771 } | |
772 | |
773 TraceTime t1("GC", PrintGC && !PrintGCDetails, true, gclog_or_tty); | |
774 // Capture heap used before collection (for printing). | |
775 size_t gch_prev_used = gch->used(); | |
776 | |
777 SpecializationStats::clear(); | |
778 | |
779 age_table()->clear(); | |
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
780 to()->clear(SpaceDecorator::Mangle); |
0 | 781 |
782 gch->save_marks(); | |
783 assert(workers != NULL, "Need parallel worker threads."); | |
784 ParallelTaskTerminator _term(workers->total_workers(), task_queues()); | |
785 ParScanThreadStateSet thread_state_set(workers->total_workers(), | |
786 *to(), *this, *_next_gen, *task_queues(), | |
787 desired_plab_sz(), _term); | |
788 | |
789 ParNewGenTask tsk(this, _next_gen, reserved().end(), &thread_state_set); | |
790 int n_workers = workers->total_workers(); | |
791 gch->set_par_threads(n_workers); | |
792 gch->change_strong_roots_parity(); | |
793 gch->rem_set()->prepare_for_younger_refs_iterate(true); | |
794 // It turns out that even when we're using 1 thread, doing the work in a | |
795 // separate thread causes wide variance in run times. We can't help this | |
796 // in the multi-threaded case, but we special-case n=1 here to get | |
797 // repeatable measurements of the 1-thread overhead of the parallel code. | |
798 if (n_workers > 1) { | |
799 workers->run_task(&tsk); | |
800 } else { | |
801 tsk.work(0); | |
802 } | |
803 thread_state_set.reset(); | |
804 | |
805 if (PAR_STATS_ENABLED && ParallelGCVerbose) { | |
806 gclog_or_tty->print("Thread totals:\n" | |
807 " Pushes: %7d Pops: %7d Steals %7d (sum = %7d).\n", | |
808 thread_state_set.pushes(), thread_state_set.pops(), | |
809 thread_state_set.steals(), | |
810 thread_state_set.pops()+thread_state_set.steals()); | |
811 } | |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
812 assert(thread_state_set.pushes() == thread_state_set.pops() |
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
813 + thread_state_set.steals(), |
0 | 814 "Or else the queues are leaky."); |
815 | |
816 // Process (weak) reference objects found during scavenge. | |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
817 ReferenceProcessor* rp = ref_processor(); |
0 | 818 IsAliveClosure is_alive(this); |
819 ScanWeakRefClosure scan_weak_ref(this); | |
820 KeepAliveClosure keep_alive(&scan_weak_ref); | |
821 ScanClosure scan_without_gc_barrier(this, false); | |
822 ScanClosureWithParBarrier scan_with_gc_barrier(this, true); | |
823 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier); | |
824 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level, | |
825 &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
|
826 rp->setup_policy(clear_all_soft_refs); |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
827 if (rp->processing_is_mt()) { |
0 | 828 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
829 rp->process_discovered_references(&is_alive, &keep_alive, |
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
830 &evacuate_followers, &task_executor); |
0 | 831 } else { |
832 thread_state_set.flush(); | |
833 gch->set_par_threads(0); // 0 ==> non-parallel. | |
834 gch->save_marks(); | |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
835 rp->process_discovered_references(&is_alive, &keep_alive, |
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
836 &evacuate_followers, NULL); |
0 | 837 } |
838 if (!promotion_failed()) { | |
839 // Swap the survivor spaces. | |
263
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
840 eden()->clear(SpaceDecorator::Mangle); |
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
841 from()->clear(SpaceDecorator::Mangle); |
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
842 if (ZapUnusedHeapArea) { |
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
843 // 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
|
844 // 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
|
845 // 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
|
846 // 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
|
847 // 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
|
848 // 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
|
849 // other spaces. |
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
850 to()->mangle_unused_area(); |
12eea04c8b06
6672698: mangle_unused_area() should not remangle the entire heap at each collection.
jmasa
parents:
167
diff
changeset
|
851 } |
0 | 852 swap_spaces(); |
853 | |
854 assert(to()->is_empty(), "to space should be empty now"); | |
855 } else { | |
856 assert(HandlePromotionFailure, | |
857 "Should only be here if promotion failure handling is on"); | |
858 if (_promo_failure_scan_stack != NULL) { | |
859 // Can be non-null because of reference processing. | |
860 // Free stack with its elements. | |
861 delete _promo_failure_scan_stack; | |
862 _promo_failure_scan_stack = NULL; | |
863 } | |
864 remove_forwarding_pointers(); | |
865 if (PrintGCDetails) { | |
866 gclog_or_tty->print(" (promotion failed)"); | |
867 } | |
868 // All the spaces are in play for mark-sweep. | |
869 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690. | |
870 from()->set_next_compaction_space(to()); | |
871 gch->set_incremental_collection_will_fail(); | |
6
73e96e5c30df
6624765: Guarantee failure "Unexpected dirty card found"
jmasa
parents:
0
diff
changeset
|
872 |
73e96e5c30df
6624765: Guarantee failure "Unexpected dirty card found"
jmasa
parents:
0
diff
changeset
|
873 // Reset the PromotionFailureALot counters. |
73e96e5c30df
6624765: Guarantee failure "Unexpected dirty card found"
jmasa
parents:
0
diff
changeset
|
874 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();) |
0 | 875 } |
876 // set new iteration safe limit for the survivor spaces | |
877 from()->set_concurrent_iteration_safe_limit(from()->top()); | |
878 to()->set_concurrent_iteration_safe_limit(to()->top()); | |
879 | |
880 adjust_desired_tenuring_threshold(); | |
881 if (ResizePLAB) { | |
882 plab_stats()->adjust_desired_plab_sz(); | |
883 } | |
884 | |
885 if (PrintGC && !PrintGCDetails) { | |
886 gch->print_heap_change(gch_prev_used); | |
887 } | |
888 | |
889 if (UseAdaptiveSizePolicy) { | |
890 size_policy->minor_collection_end(gch->gc_cause()); | |
891 size_policy->avg_survived()->sample(from()->used()); | |
892 } | |
893 | |
894 update_time_of_last_gc(os::javaTimeMillis()); | |
895 | |
896 SpecializationStats::print(); | |
897 | |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
898 rp->set_enqueuing_is_done(true); |
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
899 if (rp->processing_is_mt()) { |
0 | 900 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set); |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
901 rp->enqueue_discovered_references(&task_executor); |
0 | 902 } else { |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
903 rp->enqueue_discovered_references(NULL); |
0 | 904 } |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
269
diff
changeset
|
905 rp->verify_no_references_recorded(); |
0 | 906 } |
907 | |
908 static int sum; | |
909 void ParNewGeneration::waste_some_time() { | |
910 for (int i = 0; i < 100; i++) { | |
911 sum += i; | |
912 } | |
913 } | |
914 | |
915 static const oop ClaimedForwardPtr = oop(0x4); | |
916 | |
917 // Because of concurrency, there are times where an object for which | |
918 // "is_forwarded()" is true contains an "interim" forwarding pointer | |
919 // value. Such a value will soon be overwritten with a real value. | |
920 // This method requires "obj" to have a forwarding pointer, and waits, if | |
921 // necessary for a real one to be inserted, and returns it. | |
922 | |
923 oop ParNewGeneration::real_forwardee(oop obj) { | |
924 oop forward_ptr = obj->forwardee(); | |
925 if (forward_ptr != ClaimedForwardPtr) { | |
926 return forward_ptr; | |
927 } else { | |
928 return real_forwardee_slow(obj); | |
929 } | |
930 } | |
931 | |
932 oop ParNewGeneration::real_forwardee_slow(oop obj) { | |
933 // Spin-read if it is claimed but not yet written by another thread. | |
934 oop forward_ptr = obj->forwardee(); | |
935 while (forward_ptr == ClaimedForwardPtr) { | |
936 waste_some_time(); | |
937 assert(obj->is_forwarded(), "precondition"); | |
938 forward_ptr = obj->forwardee(); | |
939 } | |
940 return forward_ptr; | |
941 } | |
942 | |
943 #ifdef ASSERT | |
944 bool ParNewGeneration::is_legal_forward_ptr(oop p) { | |
945 return | |
946 (_avoid_promotion_undo && p == ClaimedForwardPtr) | |
947 || Universe::heap()->is_in_reserved(p); | |
948 } | |
949 #endif | |
950 | |
951 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) { | |
952 if ((m != markOopDesc::prototype()) && | |
953 (!UseBiasedLocking || (m != markOopDesc::biased_locking_prototype()))) { | |
954 MutexLocker ml(ParGCRareEvent_lock); | |
955 DefNewGeneration::preserve_mark_if_necessary(obj, m); | |
956 } | |
957 } | |
958 | |
959 // Multiple GC threads may try to promote an object. If the object | |
960 // is successfully promoted, a forwarding pointer will be installed in | |
961 // the object in the young generation. This method claims the right | |
962 // to install the forwarding pointer before it copies the object, | |
963 // thus avoiding the need to undo the copy as in | |
964 // copy_to_survivor_space_avoiding_with_undo. | |
965 | |
966 oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo( | |
967 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { | |
968 // In the sequential version, this assert also says that the object is | |
969 // not forwarded. That might not be the case here. It is the case that | |
970 // the caller observed it to be not forwarded at some time in the past. | |
971 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); | |
972 | |
973 // The sequential code read "old->age()" below. That doesn't work here, | |
974 // since the age is in the mark word, and that might be overwritten with | |
975 // a forwarding pointer by a parallel thread. So we must save the mark | |
976 // word in a local and then analyze it. | |
977 oopDesc dummyOld; | |
978 dummyOld.set_mark(m); | |
979 assert(!dummyOld.is_forwarded(), | |
980 "should not be called with forwarding pointer mark word."); | |
981 | |
982 oop new_obj = NULL; | |
983 oop forward_ptr; | |
984 | |
985 // Try allocating obj in to-space (unless too old) | |
986 if (dummyOld.age() < tenuring_threshold()) { | |
987 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); | |
988 if (new_obj == NULL) { | |
989 set_survivor_overflow(true); | |
990 } | |
991 } | |
992 | |
993 if (new_obj == NULL) { | |
994 // Either to-space is full or we decided to promote | |
995 // try allocating obj tenured | |
996 | |
997 // Attempt to install a null forwarding pointer (atomically), | |
998 // to claim the right to install the real forwarding pointer. | |
999 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr); | |
1000 if (forward_ptr != NULL) { | |
1001 // someone else beat us to it. | |
1002 return real_forwardee(old); | |
1003 } | |
1004 | |
1005 new_obj = _next_gen->par_promote(par_scan_state->thread_num(), | |
1006 old, m, sz); | |
1007 | |
1008 if (new_obj == NULL) { | |
1009 if (!HandlePromotionFailure) { | |
1010 // A failed promotion likely means the MaxLiveObjectEvacuationRatio flag | |
1011 // is incorrectly set. In any case, its seriously wrong to be here! | |
1012 vm_exit_out_of_memory(sz*wordSize, "promotion"); | |
1013 } | |
1014 // promotion failed, forward to self | |
1015 _promotion_failed = true; | |
1016 new_obj = old; | |
1017 | |
1018 preserve_mark_if_necessary(old, m); | |
1019 } | |
1020 | |
1021 old->forward_to(new_obj); | |
1022 forward_ptr = NULL; | |
1023 } else { | |
1024 // Is in to-space; do copying ourselves. | |
1025 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); | |
1026 forward_ptr = old->forward_to_atomic(new_obj); | |
1027 // Restore the mark word copied above. | |
1028 new_obj->set_mark(m); | |
1029 // Increment age if obj still in new generation | |
1030 new_obj->incr_age(); | |
1031 par_scan_state->age_table()->add(new_obj, sz); | |
1032 } | |
1033 assert(new_obj != NULL, "just checking"); | |
1034 | |
1035 if (forward_ptr == NULL) { | |
1036 oop obj_to_push = new_obj; | |
1037 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { | |
1038 // Length field used as index of next element to be scanned. | |
1039 // Real length can be obtained from real_forwardee() | |
1040 arrayOop(old)->set_length(0); | |
1041 obj_to_push = old; | |
1042 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, | |
1043 "push forwarded object"); | |
1044 } | |
1045 // Push it on one of the queues of to-be-scanned objects. | |
534 | 1046 bool simulate_overflow = false; |
1047 NOT_PRODUCT( | |
1048 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { | |
1049 // simulate a stack overflow | |
1050 simulate_overflow = true; | |
1051 } | |
1052 ) | |
1053 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { | |
0 | 1054 // Add stats for overflow pushes. |
1055 if (Verbose && PrintGCDetails) { | |
1056 gclog_or_tty->print("queue overflow!\n"); | |
1057 } | |
534 | 1058 push_on_overflow_list(old, par_scan_state); |
0 | 1059 par_scan_state->note_overflow_push(); |
1060 } | |
1061 par_scan_state->note_push(); | |
1062 | |
1063 return new_obj; | |
1064 } | |
1065 | |
1066 // Oops. Someone beat us to it. Undo the allocation. Where did we | |
1067 // allocate it? | |
1068 if (is_in_reserved(new_obj)) { | |
1069 // Must be in to_space. | |
1070 assert(to()->is_in_reserved(new_obj), "Checking"); | |
1071 if (forward_ptr == ClaimedForwardPtr) { | |
1072 // Wait to get the real forwarding pointer value. | |
1073 forward_ptr = real_forwardee(old); | |
1074 } | |
1075 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); | |
1076 } | |
1077 | |
1078 return forward_ptr; | |
1079 } | |
1080 | |
1081 | |
1082 // Multiple GC threads may try to promote the same object. If two | |
1083 // or more GC threads copy the object, only one wins the race to install | |
1084 // the forwarding pointer. The other threads have to undo their copy. | |
1085 | |
1086 oop ParNewGeneration::copy_to_survivor_space_with_undo( | |
1087 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) { | |
1088 | |
1089 // In the sequential version, this assert also says that the object is | |
1090 // not forwarded. That might not be the case here. It is the case that | |
1091 // the caller observed it to be not forwarded at some time in the past. | |
1092 assert(is_in_reserved(old), "shouldn't be scavenging this oop"); | |
1093 | |
1094 // The sequential code read "old->age()" below. That doesn't work here, | |
1095 // since the age is in the mark word, and that might be overwritten with | |
1096 // a forwarding pointer by a parallel thread. So we must save the mark | |
1097 // word here, install it in a local oopDesc, and then analyze it. | |
1098 oopDesc dummyOld; | |
1099 dummyOld.set_mark(m); | |
1100 assert(!dummyOld.is_forwarded(), | |
1101 "should not be called with forwarding pointer mark word."); | |
1102 | |
1103 bool failed_to_promote = false; | |
1104 oop new_obj = NULL; | |
1105 oop forward_ptr; | |
1106 | |
1107 // Try allocating obj in to-space (unless too old) | |
1108 if (dummyOld.age() < tenuring_threshold()) { | |
1109 new_obj = (oop)par_scan_state->alloc_in_to_space(sz); | |
1110 if (new_obj == NULL) { | |
1111 set_survivor_overflow(true); | |
1112 } | |
1113 } | |
1114 | |
1115 if (new_obj == NULL) { | |
1116 // Either to-space is full or we decided to promote | |
1117 // try allocating obj tenured | |
1118 new_obj = _next_gen->par_promote(par_scan_state->thread_num(), | |
1119 old, m, sz); | |
1120 | |
1121 if (new_obj == NULL) { | |
1122 if (!HandlePromotionFailure) { | |
1123 // A failed promotion likely means the MaxLiveObjectEvacuationRatio | |
1124 // flag is incorrectly set. In any case, its seriously wrong to be | |
1125 // here! | |
1126 vm_exit_out_of_memory(sz*wordSize, "promotion"); | |
1127 } | |
1128 // promotion failed, forward to self | |
1129 forward_ptr = old->forward_to_atomic(old); | |
1130 new_obj = old; | |
1131 | |
1132 if (forward_ptr != NULL) { | |
1133 return forward_ptr; // someone else succeeded | |
1134 } | |
1135 | |
1136 _promotion_failed = true; | |
1137 failed_to_promote = true; | |
1138 | |
1139 preserve_mark_if_necessary(old, m); | |
1140 } | |
1141 } else { | |
1142 // Is in to-space; do copying ourselves. | |
1143 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz); | |
1144 // Restore the mark word copied above. | |
1145 new_obj->set_mark(m); | |
1146 // Increment age if new_obj still in new generation | |
1147 new_obj->incr_age(); | |
1148 par_scan_state->age_table()->add(new_obj, sz); | |
1149 } | |
1150 assert(new_obj != NULL, "just checking"); | |
1151 | |
1152 // Now attempt to install the forwarding pointer (atomically). | |
1153 // We have to copy the mark word before overwriting with forwarding | |
1154 // ptr, so we can restore it below in the copy. | |
1155 if (!failed_to_promote) { | |
1156 forward_ptr = old->forward_to_atomic(new_obj); | |
1157 } | |
1158 | |
1159 if (forward_ptr == NULL) { | |
1160 oop obj_to_push = new_obj; | |
1161 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) { | |
1162 // Length field used as index of next element to be scanned. | |
1163 // Real length can be obtained from real_forwardee() | |
1164 arrayOop(old)->set_length(0); | |
1165 obj_to_push = old; | |
1166 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push, | |
1167 "push forwarded object"); | |
1168 } | |
1169 // Push it on one of the queues of to-be-scanned objects. | |
534 | 1170 bool simulate_overflow = false; |
1171 NOT_PRODUCT( | |
1172 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) { | |
1173 // simulate a stack overflow | |
1174 simulate_overflow = true; | |
1175 } | |
1176 ) | |
1177 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) { | |
0 | 1178 // Add stats for overflow pushes. |
534 | 1179 push_on_overflow_list(old, par_scan_state); |
0 | 1180 par_scan_state->note_overflow_push(); |
1181 } | |
1182 par_scan_state->note_push(); | |
1183 | |
1184 return new_obj; | |
1185 } | |
1186 | |
1187 // Oops. Someone beat us to it. Undo the allocation. Where did we | |
1188 // allocate it? | |
1189 if (is_in_reserved(new_obj)) { | |
1190 // Must be in to_space. | |
1191 assert(to()->is_in_reserved(new_obj), "Checking"); | |
1192 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz); | |
1193 } else { | |
1194 assert(!_avoid_promotion_undo, "Should not be here if avoiding."); | |
1195 _next_gen->par_promote_alloc_undo(par_scan_state->thread_num(), | |
1196 (HeapWord*)new_obj, sz); | |
1197 } | |
1198 | |
1199 return forward_ptr; | |
1200 } | |
1201 | |
534 | 1202 #ifndef PRODUCT |
1203 // It's OK to call this multi-threaded; the worst thing | |
1204 // that can happen is that we'll get a bunch of closely | |
1205 // spaced simulated oveflows, but that's OK, in fact | |
1206 // probably good as it would exercise the overflow code | |
1207 // under contention. | |
1208 bool ParNewGeneration::should_simulate_overflow() { | |
1209 if (_overflow_counter-- <= 0) { // just being defensive | |
1210 _overflow_counter = ParGCWorkQueueOverflowInterval; | |
1211 return true; | |
1212 } else { | |
1213 return false; | |
1214 } | |
1215 } | |
1216 #endif | |
1217 | |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1218 // 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
|
1219 // 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
|
1220 // 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
|
1221 // 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
|
1222 // 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
|
1223 // 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
|
1224 // 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
|
1225 // 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
|
1226 // 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
|
1227 // 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
|
1228 // 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
|
1229 // 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
|
1230 // 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
|
1231 // 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
|
1232 // 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
|
1233 // 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
|
1234 // 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
|
1235 // (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
|
1236 // 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
|
1237 // as we were made painfully aware not long ago, see 6786503). |
534 | 1238 #define BUSY (oop(0x1aff1aff)) |
1239 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
|
1240 assert(is_in_reserved(from_space_obj), "Should be from this generation"); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1241 if (UseCompressedOops) { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1242 // 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
|
1243 // overflow stack. |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1244 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
|
1245 } else { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1246 // 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
|
1247 // 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
|
1248 // 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
|
1249 // 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
|
1250 // and should be fixed. XXX FIX ME !!! |
534 | 1251 #ifndef PRODUCT |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1252 Atomic::inc_ptr(&_num_par_pushes); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1253 assert(_num_par_pushes > 0, "Tautology"); |
534 | 1254 #endif |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1255 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
|
1256 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
|
1257 listhead->forward_to(from_space_obj); |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1258 from_space_obj = listhead; |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1259 } |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1260 oop observed_overflow_list = _overflow_list; |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1261 oop cur_overflow_list; |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1262 do { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1263 cur_overflow_list = observed_overflow_list; |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1264 if (cur_overflow_list != BUSY) { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1265 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
|
1266 } else { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1267 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
|
1268 } |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1269 observed_overflow_list = |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1270 (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
|
1271 } while (cur_overflow_list != observed_overflow_list); |
0 | 1272 } |
1273 } | |
1274 | |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1275 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
|
1276 bool res; |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1277 |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1278 if (UseCompressedOops) { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1279 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
|
1280 } else { |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1281 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
|
1282 } |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1283 return res; |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1284 } |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1285 |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1286 |
534 | 1287 // *NOTE*: The overflow list manipulation code here and |
1288 // in CMSCollector:: are very similar in shape, | |
1289 // except that in the CMS case we thread the objects | |
1290 // directly into the list via their mark word, and do | |
1291 // not need to deal with special cases below related | |
1292 // to chunking of object arrays and promotion failure | |
1293 // handling. | |
1294 // CR 6797058 has been filed to attempt consolidation of | |
1295 // the common code. | |
1296 // Because of the common code, if you make any changes in | |
1297 // the code below, please check the CMS version to see if | |
1298 // similar changes might be needed. | |
1299 // See CMSCollector::par_take_from_overflow_list() for | |
1300 // more extensive documentation comments. | |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1301 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) { |
0 | 1302 ObjToScanQueue* work_q = par_scan_state->work_queue(); |
1303 // How many to take? | |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1304 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, |
534 | 1305 (size_t)ParGCDesiredObjsFromOverflowList); |
0 | 1306 |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
579
diff
changeset
|
1307 assert(par_scan_state->overflow_stack() == NULL, "Error"); |
0 | 1308 if (_overflow_list == NULL) return false; |
1309 | |
1310 // Otherwise, there was something there; try claiming the list. | |
534 | 1311 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); |
1312 // Trim off a prefix of at most objsFromOverflow items | |
1313 Thread* tid = Thread::current(); | |
1314 size_t spin_count = (size_t)ParallelGCThreads; | |
1315 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100); | |
1316 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) { | |
1317 // someone grabbed it before we did ... | |
1318 // ... we spin for a short while... | |
1319 os::sleep(tid, sleep_time_millis, false); | |
1320 if (_overflow_list == NULL) { | |
1321 // nothing left to take | |
1322 return false; | |
1323 } else if (_overflow_list != BUSY) { | |
1324 // try and grab the prefix | |
1325 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); | |
1326 } | |
0 | 1327 } |
534 | 1328 if (prefix == NULL || prefix == BUSY) { |
1329 // Nothing to take or waited long enough | |
1330 if (prefix == NULL) { | |
1331 // Write back the NULL in case we overwrote it with BUSY above | |
1332 // and it is still the same value. | |
1333 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); | |
1334 } | |
1335 return false; | |
1336 } | |
1337 assert(prefix != NULL && prefix != BUSY, "Error"); | |
1338 size_t i = 1; | |
0 | 1339 oop cur = prefix; |
167
feeb96a45707
6696264: assert("narrow oop can never be zero") for GCBasher & ParNewGC
coleenp
parents:
113
diff
changeset
|
1340 while (i < objsFromOverflow && cur->klass_or_null() != NULL) { |
0 | 1341 i++; cur = oop(cur->klass()); |
1342 } | |
1343 | |
1344 // Reattach remaining (suffix) to overflow list | |
534 | 1345 if (cur->klass_or_null() == NULL) { |
1346 // Write back the NULL in lieu of the BUSY we wrote | |
1347 // above and it is still the same value. | |
1348 if (_overflow_list == BUSY) { | |
1349 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); | |
0 | 1350 } |
534 | 1351 } else { |
1352 assert(cur->klass_or_null() != BUSY, "Error"); | |
1353 oop suffix = oop(cur->klass()); // suffix will be put back on global list | |
1354 cur->set_klass_to_list_ptr(NULL); // break off suffix | |
1355 // It's possible that the list is still in the empty(busy) state | |
1356 // we left it in a short while ago; in that case we may be | |
1357 // able to place back the suffix. | |
1358 oop observed_overflow_list = _overflow_list; | |
1359 oop cur_overflow_list = observed_overflow_list; | |
1360 bool attached = false; | |
1361 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { | |
1362 observed_overflow_list = | |
1363 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); | |
1364 if (cur_overflow_list == observed_overflow_list) { | |
1365 attached = true; | |
1366 break; | |
1367 } else cur_overflow_list = observed_overflow_list; | |
1368 } | |
1369 if (!attached) { | |
1370 // Too bad, someone else got in in between; we'll need to do a splice. | |
1371 // Find the last item of suffix list | |
1372 oop last = suffix; | |
1373 while (last->klass_or_null() != NULL) { | |
1374 last = oop(last->klass()); | |
1375 } | |
1376 // Atomically prepend suffix to current overflow list | |
1377 observed_overflow_list = _overflow_list; | |
1378 do { | |
1379 cur_overflow_list = observed_overflow_list; | |
1380 if (cur_overflow_list != BUSY) { | |
1381 // Do the splice ... | |
1382 last->set_klass_to_list_ptr(cur_overflow_list); | |
1383 } else { // cur_overflow_list == BUSY | |
1384 last->set_klass_to_list_ptr(NULL); | |
1385 } | |
1386 observed_overflow_list = | |
1387 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list); | |
1388 } while (cur_overflow_list != observed_overflow_list); | |
0 | 1389 } |
1390 } | |
1391 | |
1392 // Push objects on prefix list onto this thread's work queue | |
534 | 1393 assert(prefix != NULL && prefix != BUSY, "program logic"); |
0 | 1394 cur = prefix; |
534 | 1395 ssize_t n = 0; |
0 | 1396 while (cur != NULL) { |
1397 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
|
1398 oop next = oop(cur->klass_or_null()); |
0 | 1399 cur->set_klass(obj_to_push->klass()); |
534 | 1400 // This may be an array object that is self-forwarded. In that case, the list pointer |
1401 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed. | |
1402 if (!is_in_reserved(cur)) { | |
1403 // This can become a scaling bottleneck when there is work queue overflow coincident | |
1404 // with promotion failure. | |
1405 oopDesc* f = cur; | |
1406 FREE_C_HEAP_ARRAY(oopDesc, f); | |
1407 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) { | |
0 | 1408 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned"); |
534 | 1409 obj_to_push = cur; |
0 | 1410 } |
534 | 1411 bool ok = work_q->push(obj_to_push); |
1412 assert(ok, "Should have succeeded"); | |
0 | 1413 cur = next; |
1414 n++; | |
1415 } | |
1416 par_scan_state->note_overflow_refill(n); | |
534 | 1417 #ifndef PRODUCT |
1418 assert(_num_par_pushes >= n, "Too many pops?"); | |
1419 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); | |
1420 #endif | |
0 | 1421 return true; |
1422 } | |
534 | 1423 #undef BUSY |
0 | 1424 |
1425 void ParNewGeneration::ref_processor_init() | |
1426 { | |
1427 if (_ref_processor == NULL) { | |
1428 // Allocate and initialize a reference processor | |
1429 _ref_processor = ReferenceProcessor::create_ref_processor( | |
1430 _reserved, // span | |
1431 refs_discovery_is_atomic(), // atomic_discovery | |
1432 refs_discovery_is_mt(), // mt_discovery | |
1433 NULL, // is_alive_non_header | |
1434 ParallelGCThreads, | |
1435 ParallelRefProcEnabled); | |
1436 } | |
1437 } | |
1438 | |
1439 const char* ParNewGeneration::name() const { | |
1440 return "par new generation"; | |
1441 } |