1 /*
2 * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/cms/compactibleFreeListSpace.hpp"
27 #include "gc/cms/concurrentMarkSweepGeneration.hpp"
28 #include "gc/cms/parNewGeneration.inline.hpp"
29 #include "gc/cms/parOopClosures.inline.hpp"
30 #include "gc/serial/defNewGeneration.inline.hpp"
31 #include "gc/shared/adaptiveSizePolicy.hpp"
32 #include "gc/shared/ageTable.inline.hpp"
33 #include "gc/shared/copyFailedInfo.hpp"
34 #include "gc/shared/gcHeapSummary.hpp"
35 #include "gc/shared/gcTimer.hpp"
36 #include "gc/shared/gcTrace.hpp"
37 #include "gc/shared/gcTraceTime.inline.hpp"
38 #include "gc/shared/genCollectedHeap.hpp"
39 #include "gc/shared/genOopClosures.inline.hpp"
40 #include "gc/shared/generation.hpp"
41 #include "gc/shared/plab.inline.hpp"
42 #include "gc/shared/preservedMarks.inline.hpp"
43 #include "gc/shared/referencePolicy.hpp"
44 #include "gc/shared/space.hpp"
45 #include "gc/shared/spaceDecorator.hpp"
46 #include "gc/shared/strongRootsScope.hpp"
47 #include "gc/shared/taskqueue.inline.hpp"
48 #include "gc/shared/workgroup.hpp"
49 #include "logging/log.hpp"
50 #include "logging/logStream.hpp"
51 #include "memory/resourceArea.hpp"
52 #include "oops/objArrayOop.hpp"
53 #include "oops/oop.inline.hpp"
54 #include "runtime/atomic.hpp"
55 #include "runtime/handles.hpp"
56 #include "runtime/handles.inline.hpp"
57 #include "runtime/java.hpp"
58 #include "runtime/thread.inline.hpp"
59 #include "utilities/copy.hpp"
60 #include "utilities/globalDefinitions.hpp"
61 #include "utilities/stack.inline.hpp"
62
63 ParScanThreadState::ParScanThreadState(Space* to_space_,
64 ParNewGeneration* young_gen_,
65 Generation* old_gen_,
66 int thread_num_,
67 ObjToScanQueueSet* work_queue_set_,
68 Stack<oop, mtGC>* overflow_stacks_,
69 PreservedMarks* preserved_marks_,
70 size_t desired_plab_sz_,
71 ParallelTaskTerminator& term_) :
72 _to_space(to_space_),
73 _old_gen(old_gen_),
74 _young_gen(young_gen_),
75 _thread_num(thread_num_),
76 _work_queue(work_queue_set_->queue(thread_num_)),
77 _to_space_full(false),
78 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
79 _preserved_marks(preserved_marks_),
80 _ageTable(false), // false ==> not the global age table, no perf data.
81 _to_space_alloc_buffer(desired_plab_sz_),
82 _to_space_closure(young_gen_, this),
83 _old_gen_closure(young_gen_, this),
84 _to_space_root_closure(young_gen_, this),
85 _old_gen_root_closure(young_gen_, this),
86 _older_gen_closure(young_gen_, this),
87 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
88 &_to_space_root_closure, young_gen_, &_old_gen_root_closure,
89 work_queue_set_, &term_),
90 _is_alive_closure(young_gen_),
91 _scan_weak_ref_closure(young_gen_, this),
92 _keep_alive_closure(&_scan_weak_ref_closure),
93 _strong_roots_time(0.0),
94 _term_time(0.0)
95 {
96 #if TASKQUEUE_STATS
97 _term_attempts = 0;
98 _overflow_refills = 0;
99 _overflow_refill_objs = 0;
100 #endif // TASKQUEUE_STATS
101
102 _survivor_chunk_array = (ChunkArray*) old_gen()->get_data_recorder(thread_num());
103 _hash_seed = 17; // Might want to take time-based random value.
104 _start = os::elapsedTime();
105 _old_gen_closure.set_generation(old_gen_);
106 _old_gen_root_closure.set_generation(old_gen_);
107 }
108
109 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
110 size_t plab_word_size) {
111 ChunkArray* sca = survivor_chunk_array();
112 if (sca != NULL) {
113 // A non-null SCA implies that we want the PLAB data recorded.
114 sca->record_sample(plab_start, plab_word_size);
115 }
116 }
117
118 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
119 return new_obj->is_objArray() &&
120 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
121 new_obj != old_obj;
122 }
123
124 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
125 assert(old->is_objArray(), "must be obj array");
126 assert(old->is_forwarded(), "must be forwarded");
127 assert(GenCollectedHeap::heap()->is_in_reserved(old), "must be in heap.");
128 assert(!old_gen()->is_in(old), "must be in young generation.");
129
130 objArrayOop obj = objArrayOop(old->forwardee());
131 // Process ParGCArrayScanChunk elements now
132 // and push the remainder back onto queue
133 int start = arrayOop(old)->length();
134 int end = obj->length();
135 int remainder = end - start;
136 assert(start <= end, "just checking");
137 if (remainder > 2 * ParGCArrayScanChunk) {
138 // Test above combines last partial chunk with a full chunk
139 end = start + ParGCArrayScanChunk;
140 arrayOop(old)->set_length(end);
141 // Push remainder.
142 bool ok = work_queue()->push(old);
143 assert(ok, "just popped, push must be okay");
144 } else {
145 // Restore length so that it can be used if there
146 // is a promotion failure and forwarding pointers
147 // must be removed.
148 arrayOop(old)->set_length(end);
149 }
150
151 // process our set of indices (include header in first chunk)
152 // should make sure end is even (aligned to HeapWord in case of compressed oops)
153 if ((HeapWord *)obj < young_old_boundary()) {
154 // object is in to_space
155 obj->oop_iterate_range(&_to_space_closure, start, end);
156 } else {
157 // object is in old generation
158 obj->oop_iterate_range(&_old_gen_closure, start, end);
159 }
160 }
161
162 void ParScanThreadState::trim_queues(int max_size) {
163 ObjToScanQueue* queue = work_queue();
164 do {
165 while (queue->size() > (juint)max_size) {
166 oop obj_to_scan;
167 if (queue->pop_local(obj_to_scan)) {
168 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
169 if (obj_to_scan->is_objArray() &&
170 obj_to_scan->is_forwarded() &&
171 obj_to_scan->forwardee() != obj_to_scan) {
172 scan_partial_array_and_push_remainder(obj_to_scan);
173 } else {
174 // object is in to_space
175 obj_to_scan->oop_iterate(&_to_space_closure);
176 }
177 } else {
178 // object is in old generation
179 obj_to_scan->oop_iterate(&_old_gen_closure);
180 }
181 }
182 }
183 // For the case of compressed oops, we have a private, non-shared
184 // overflow stack, so we eagerly drain it so as to more evenly
185 // distribute load early. Note: this may be good to do in
186 // general rather than delay for the final stealing phase.
187 // If applicable, we'll transfer a set of objects over to our
188 // work queue, allowing them to be stolen and draining our
189 // private overflow stack.
190 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
191 }
192
193 bool ParScanThreadState::take_from_overflow_stack() {
194 assert(ParGCUseLocalOverflow, "Else should not call");
195 assert(young_gen()->overflow_list() == NULL, "Error");
196 ObjToScanQueue* queue = work_queue();
197 Stack<oop, mtGC>* const of_stack = overflow_stack();
198 const size_t num_overflow_elems = of_stack->size();
199 const size_t space_available = queue->max_elems() - queue->size();
200 const size_t num_take_elems = MIN3(space_available / 4,
201 ParGCDesiredObjsFromOverflowList,
202 num_overflow_elems);
203 // Transfer the most recent num_take_elems from the overflow
204 // stack to our work queue.
205 for (size_t i = 0; i != num_take_elems; i++) {
206 oop cur = of_stack->pop();
207 oop obj_to_push = cur->forwardee();
208 assert(GenCollectedHeap::heap()->is_in_reserved(cur), "Should be in heap");
209 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
210 assert(GenCollectedHeap::heap()->is_in_reserved(obj_to_push), "Should be in heap");
211 if (should_be_partially_scanned(obj_to_push, cur)) {
212 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
213 obj_to_push = cur;
214 }
215 bool ok = queue->push(obj_to_push);
216 assert(ok, "Should have succeeded");
217 }
218 assert(young_gen()->overflow_list() == NULL, "Error");
219 return num_take_elems > 0; // was something transferred?
220 }
221
222 void ParScanThreadState::push_on_overflow_stack(oop p) {
223 assert(ParGCUseLocalOverflow, "Else should not call");
224 overflow_stack()->push(p);
225 assert(young_gen()->overflow_list() == NULL, "Error");
226 }
227
228 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
229 // If the object is small enough, try to reallocate the buffer.
230 HeapWord* obj = NULL;
231 if (!_to_space_full) {
232 PLAB* const plab = to_space_alloc_buffer();
233 Space* const sp = to_space();
234 if (word_sz * 100 < ParallelGCBufferWastePct * plab->word_sz()) {
235 // Is small enough; abandon this buffer and start a new one.
236 plab->retire();
237 // The minimum size has to be twice SurvivorAlignmentInBytes to
238 // allow for padding used in the alignment of 1 word. A padding
239 // of 1 is too small for a filler word so the padding size will
240 // be increased by SurvivorAlignmentInBytes.
241 size_t min_usable_size = 2 * static_cast<size_t>(SurvivorAlignmentInBytes >> LogHeapWordSize);
242 size_t buf_size = MAX2(plab->word_sz(), min_usable_size);
243 HeapWord* buf_space = sp->par_allocate(buf_size);
244 if (buf_space == NULL) {
245 const size_t min_bytes = MAX2(PLAB::min_size(), min_usable_size) << LogHeapWordSize;
246 size_t free_bytes = sp->free();
247 while(buf_space == NULL && free_bytes >= min_bytes) {
248 buf_size = free_bytes >> LogHeapWordSize;
249 assert(buf_size == (size_t)align_object_size(buf_size), "Invariant");
250 buf_space = sp->par_allocate(buf_size);
251 free_bytes = sp->free();
252 }
253 }
254 if (buf_space != NULL) {
255 plab->set_buf(buf_space, buf_size);
256 record_survivor_plab(buf_space, buf_size);
257 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
258 // Note that we cannot compare buf_size < word_sz below
259 // because of AlignmentReserve (see PLAB::allocate()).
260 assert(obj != NULL || plab->words_remaining() < word_sz,
261 "Else should have been able to allocate requested object size "
262 SIZE_FORMAT ", PLAB size " SIZE_FORMAT ", SurvivorAlignmentInBytes "
263 SIZE_FORMAT ", words_remaining " SIZE_FORMAT,
264 word_sz, buf_size, SurvivorAlignmentInBytes, plab->words_remaining());
265 // It's conceivable that we may be able to use the
266 // buffer we just grabbed for subsequent small requests
267 // even if not for this one.
268 } else {
269 // We're used up.
270 _to_space_full = true;
271 }
272 } else {
273 // Too large; allocate the object individually.
274 obj = sp->par_allocate(word_sz);
275 }
276 }
277 return obj;
278 }
279
280 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) {
281 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
282 }
283
284 void ParScanThreadState::print_promotion_failure_size() {
285 if (_promotion_failed_info.has_failed()) {
286 log_trace(gc, promotion)(" (%d: promotion failure size = " SIZE_FORMAT ") ",
287 _thread_num, _promotion_failed_info.first_size());
288 }
289 }
290
291 class ParScanThreadStateSet: StackObj {
292 public:
293 // Initializes states for the specified number of threads;
294 ParScanThreadStateSet(int num_threads,
295 Space& to_space,
296 ParNewGeneration& young_gen,
297 Generation& old_gen,
298 ObjToScanQueueSet& queue_set,
299 Stack<oop, mtGC>* overflow_stacks_,
300 PreservedMarksSet& preserved_marks_set,
301 size_t desired_plab_sz,
302 ParallelTaskTerminator& term);
303
304 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
305
306 inline ParScanThreadState& thread_state(int i);
307
308 void trace_promotion_failed(const YoungGCTracer* gc_tracer);
309 void reset(uint active_workers, bool promotion_failed);
310 void flush();
311
312 #if TASKQUEUE_STATS
313 static void
314 print_termination_stats_hdr(outputStream* const st);
315 void print_termination_stats();
316 static void
317 print_taskqueue_stats_hdr(outputStream* const st);
318 void print_taskqueue_stats();
319 void reset_stats();
320 #endif // TASKQUEUE_STATS
321
322 private:
323 ParallelTaskTerminator& _term;
324 ParNewGeneration& _young_gen;
325 Generation& _old_gen;
326 ParScanThreadState* _per_thread_states;
327 const int _num_threads;
328 public:
329 bool is_valid(int id) const { return id < _num_threads; }
330 ParallelTaskTerminator* terminator() { return &_term; }
331 };
332
333 ParScanThreadStateSet::ParScanThreadStateSet(int num_threads,
334 Space& to_space,
335 ParNewGeneration& young_gen,
336 Generation& old_gen,
337 ObjToScanQueueSet& queue_set,
338 Stack<oop, mtGC>* overflow_stacks,
339 PreservedMarksSet& preserved_marks_set,
340 size_t desired_plab_sz,
341 ParallelTaskTerminator& term)
342 : _young_gen(young_gen),
343 _old_gen(old_gen),
344 _term(term),
345 _per_thread_states(NEW_RESOURCE_ARRAY(ParScanThreadState, num_threads)),
346 _num_threads(num_threads)
347 {
348 assert(num_threads > 0, "sanity check!");
349 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
350 "overflow_stack allocation mismatch");
351 // Initialize states.
352 for (int i = 0; i < num_threads; ++i) {
353 new(_per_thread_states + i)
354 ParScanThreadState(&to_space, &young_gen, &old_gen, i, &queue_set,
355 overflow_stacks, preserved_marks_set.get(i),
356 desired_plab_sz, term);
357 }
358 }
359
360 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) {
361 assert(i >= 0 && i < _num_threads, "sanity check!");
362 return _per_thread_states[i];
363 }
364
365 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) {
366 for (int i = 0; i < _num_threads; ++i) {
367 if (thread_state(i).promotion_failed()) {
368 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info());
369 thread_state(i).promotion_failed_info().reset();
370 }
371 }
372 }
373
374 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed) {
375 _term.reset_for_reuse(active_threads);
376 if (promotion_failed) {
377 for (int i = 0; i < _num_threads; ++i) {
378 thread_state(i).print_promotion_failure_size();
379 }
380 }
381 }
382
383 #if TASKQUEUE_STATS
384 void ParScanThreadState::reset_stats() {
385 taskqueue_stats().reset();
386 _term_attempts = 0;
387 _overflow_refills = 0;
388 _overflow_refill_objs = 0;
389 }
390
391 void ParScanThreadStateSet::reset_stats() {
392 for (int i = 0; i < _num_threads; ++i) {
393 thread_state(i).reset_stats();
394 }
395 }
396
397 void ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) {
398 st->print_raw_cr("GC Termination Stats");
399 st->print_raw_cr(" elapsed --strong roots-- -------termination-------");
400 st->print_raw_cr("thr ms ms % ms % attempts");
401 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------");
402 }
403
404 void ParScanThreadStateSet::print_termination_stats() {
405 Log(gc, task, stats) log;
406 if (!log.is_debug()) {
407 return;
408 }
409
410 ResourceMark rm;
411 LogStream ls(log.debug());
412 outputStream* st = &ls;
413
414 print_termination_stats_hdr(st);
415
416 for (int i = 0; i < _num_threads; ++i) {
417 const ParScanThreadState & pss = thread_state(i);
418 const double elapsed_ms = pss.elapsed_time() * 1000.0;
419 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
420 const double term_ms = pss.term_time() * 1000.0;
421 st->print_cr("%3d %9.2f %9.2f %6.2f %9.2f %6.2f " SIZE_FORMAT_W(8),
422 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
423 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
424 }
425 }
426
427 // Print stats related to work queue activity.
428 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) {
429 st->print_raw_cr("GC Task Stats");
430 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
431 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
432 }
433
434 void ParScanThreadStateSet::print_taskqueue_stats() {
435 if (!log_develop_is_enabled(Trace, gc, task, stats)) {
436 return;
437 }
438 Log(gc, task, stats) log;
439 ResourceMark rm;
440 LogStream ls(log.trace());
441 outputStream* st = &ls;
442 print_taskqueue_stats_hdr(st);
443
444 TaskQueueStats totals;
445 for (int i = 0; i < _num_threads; ++i) {
446 const ParScanThreadState & pss = thread_state(i);
447 const TaskQueueStats & stats = pss.taskqueue_stats();
448 st->print("%3d ", i); stats.print(st); st->cr();
449 totals += stats;
450
451 if (pss.overflow_refills() > 0) {
452 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
453 SIZE_FORMAT_W(10) " overflow objects",
454 pss.overflow_refills(), pss.overflow_refill_objs());
455 }
456 }
457 st->print("tot "); totals.print(st); st->cr();
458
459 DEBUG_ONLY(totals.verify());
460 }
461 #endif // TASKQUEUE_STATS
462
463 void ParScanThreadStateSet::flush() {
464 // Work in this loop should be kept as lightweight as
465 // possible since this might otherwise become a bottleneck
466 // to scaling. Should we add heavy-weight work into this
467 // loop, consider parallelizing the loop into the worker threads.
468 for (int i = 0; i < _num_threads; ++i) {
469 ParScanThreadState& par_scan_state = thread_state(i);
470
471 // Flush stats related to To-space PLAB activity and
472 // retire the last buffer.
473 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_young_gen.plab_stats());
474
475 // Every thread has its own age table. We need to merge
476 // them all into one.
477 AgeTable *local_table = par_scan_state.age_table();
478 _young_gen.age_table()->merge(local_table);
479
480 // Inform old gen that we're done.
481 _old_gen.par_promote_alloc_done(i);
482 }
483
484 if (UseConcMarkSweepGC) {
485 // We need to call this even when ResizeOldPLAB is disabled
486 // so as to avoid breaking some asserts. While we may be able
487 // to avoid this by reorganizing the code a bit, I am loathe
488 // to do that unless we find cases where ergo leads to bad
489 // performance.
490 CompactibleFreeListSpaceLAB::compute_desired_plab_size();
491 }
492 }
493
494 ParScanClosure::ParScanClosure(ParNewGeneration* g,
495 ParScanThreadState* par_scan_state) :
496 OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) {
497 _boundary = _g->reserved().end();
498 }
499
500 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
501 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
502
503 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
504 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
505
506 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
507 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
508
509 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
510 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
511
512 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
513 ParScanThreadState* par_scan_state)
514 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
515 {}
516
517 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
518 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
519
520 #ifdef WIN32
521 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
522 #endif
523
524 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
525 ParScanThreadState* par_scan_state_,
526 ParScanWithoutBarrierClosure* to_space_closure_,
527 ParScanWithBarrierClosure* old_gen_closure_,
528 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
529 ParNewGeneration* par_gen_,
530 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
531 ObjToScanQueueSet* task_queues_,
532 ParallelTaskTerminator* terminator_) :
533
534 _par_scan_state(par_scan_state_),
535 _to_space_closure(to_space_closure_),
536 _old_gen_closure(old_gen_closure_),
537 _to_space_root_closure(to_space_root_closure_),
538 _old_gen_root_closure(old_gen_root_closure_),
539 _par_gen(par_gen_),
540 _task_queues(task_queues_),
541 _terminator(terminator_)
542 {}
543
544 void ParEvacuateFollowersClosure::do_void() {
545 ObjToScanQueue* work_q = par_scan_state()->work_queue();
546
547 while (true) {
548 // Scan to-space and old-gen objs until we run out of both.
549 oop obj_to_scan;
550 par_scan_state()->trim_queues(0);
551
552 // We have no local work, attempt to steal from other threads.
553
554 // Attempt to steal work from promoted.
555 if (task_queues()->steal(par_scan_state()->thread_num(),
556 par_scan_state()->hash_seed(),
557 obj_to_scan)) {
558 bool res = work_q->push(obj_to_scan);
559 assert(res, "Empty queue should have room for a push.");
560
561 // If successful, goto Start.
562 continue;
563
564 // Try global overflow list.
565 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
566 continue;
567 }
568
569 // Otherwise, offer termination.
570 par_scan_state()->start_term_time();
571 if (terminator()->offer_termination()) break;
572 par_scan_state()->end_term_time();
573 }
574 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
575 "Broken overflow list?");
576 // Finish the last termination pause.
577 par_scan_state()->end_term_time();
578 }
579
580 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen,
581 Generation* old_gen,
582 HeapWord* young_old_boundary,
583 ParScanThreadStateSet* state_set,
584 StrongRootsScope* strong_roots_scope) :
585 AbstractGangTask("ParNewGeneration collection"),
586 _young_gen(young_gen), _old_gen(old_gen),
587 _young_old_boundary(young_old_boundary),
588 _state_set(state_set),
589 _strong_roots_scope(strong_roots_scope)
590 {}
591
592 void ParNewGenTask::work(uint worker_id) {
593 GenCollectedHeap* gch = GenCollectedHeap::heap();
594 // Since this is being done in a separate thread, need new resource
595 // and handle marks.
596 ResourceMark rm;
597 HandleMark hm;
598
599 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
600 assert(_state_set->is_valid(worker_id), "Should not have been called");
601
602 par_scan_state.set_young_old_boundary(_young_old_boundary);
603
604 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(),
605 gch->rem_set()->klass_rem_set());
606 CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure,
607 &par_scan_state.to_space_root_closure(),
608 false);
609
610 par_scan_state.start_strong_roots();
611 gch->young_process_roots(_strong_roots_scope,
612 &par_scan_state.to_space_root_closure(),
613 &par_scan_state.older_gen_closure(),
614 &cld_scan_closure);
615
616 par_scan_state.end_strong_roots();
617
618 // "evacuate followers".
619 par_scan_state.evacuate_followers_closure().do_void();
620
621 // This will collapse this worker's promoted object list that's
622 // created during the main ParNew parallel phase of ParNew. This has
623 // to be called after all workers have finished promoting objects
624 // and scanning promoted objects. It should be safe calling it from
625 // here, given that we can only reach here after all thread have
626 // offered termination, i.e., after there is no more work to be
627 // done. It will also disable promotion tracking for the rest of
628 // this GC as it's not necessary to be on during reference processing.
629 _old_gen->par_oop_since_save_marks_iterate_done((int) worker_id);
630 }
631
632 ParNewGeneration::ParNewGeneration(ReservedSpace rs, size_t initial_byte_size)
633 : DefNewGeneration(rs, initial_byte_size, "PCopy"),
634 _overflow_list(NULL),
635 _is_alive_closure(this),
636 _plab_stats("Young", YoungPLABSize, PLABWeight)
637 {
638 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
639 NOT_PRODUCT(_num_par_pushes = 0;)
640 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
641 guarantee(_task_queues != NULL, "task_queues allocation failure.");
642
643 for (uint i = 0; i < ParallelGCThreads; i++) {
644 ObjToScanQueue *q = new ObjToScanQueue();
645 guarantee(q != NULL, "work_queue Allocation failure.");
646 _task_queues->register_queue(i, q);
647 }
648
649 for (uint i = 0; i < ParallelGCThreads; i++) {
650 _task_queues->queue(i)->initialize();
651 }
652
653 _overflow_stacks = NULL;
654 if (ParGCUseLocalOverflow) {
655 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal with ','
656 typedef Stack<oop, mtGC> GCOopStack;
657
658 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
659 for (size_t i = 0; i < ParallelGCThreads; ++i) {
660 new (_overflow_stacks + i) Stack<oop, mtGC>();
661 }
662 }
663
664 if (UsePerfData) {
665 EXCEPTION_MARK;
666 ResourceMark rm;
667
668 const char* cname =
669 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
670 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
671 ParallelGCThreads, CHECK);
672 }
673 }
674
675 // ParNewGeneration::
676 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
677 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
678
679 template <class T>
680 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
681 #ifdef ASSERT
682 {
683 assert(!oopDesc::is_null(*p), "expected non-null ref");
684 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
685 // We never expect to see a null reference being processed
686 // as a weak reference.
687 assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs");
688 }
689 #endif // ASSERT
690
691 _par_cl->do_oop_nv(p);
692
693 if (GenCollectedHeap::heap()->is_in_reserved(p)) {
694 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
695 _rs->write_ref_field_gc_par(p, obj);
696 }
697 }
698
699 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
700 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
701
702 // ParNewGeneration::
703 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
704 DefNewGeneration::KeepAliveClosure(cl) {}
705
706 template <class T>
707 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
708 #ifdef ASSERT
709 {
710 assert(!oopDesc::is_null(*p), "expected non-null ref");
711 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
712 // We never expect to see a null reference being processed
713 // as a weak reference.
714 assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs");
715 }
716 #endif // ASSERT
717
718 _cl->do_oop_nv(p);
719
720 if (GenCollectedHeap::heap()->is_in_reserved(p)) {
721 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
722 _rs->write_ref_field_gc_par(p, obj);
723 }
724 }
725
726 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
727 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
728
729 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
730 T heap_oop = oopDesc::load_heap_oop(p);
731 if (!oopDesc::is_null(heap_oop)) {
732 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
733 if ((HeapWord*)obj < _boundary) {
734 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
735 oop new_obj = obj->is_forwarded()
736 ? obj->forwardee()
737 : _g->DefNewGeneration::copy_to_survivor_space(obj);
738 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
739 }
740 if (_gc_barrier) {
741 // If p points to a younger generation, mark the card.
742 if ((HeapWord*)obj < _gen_boundary) {
743 _rs->write_ref_field_gc_par(p, obj);
744 }
745 }
746 }
747 }
748
749 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
750 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
751
752 class ParNewRefProcTaskProxy: public AbstractGangTask {
753 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
754 public:
755 ParNewRefProcTaskProxy(ProcessTask& task,
756 ParNewGeneration& young_gen,
757 Generation& old_gen,
758 HeapWord* young_old_boundary,
759 ParScanThreadStateSet& state_set);
760
761 private:
762 virtual void work(uint worker_id);
763 private:
764 ParNewGeneration& _young_gen;
765 ProcessTask& _task;
766 Generation& _old_gen;
767 HeapWord* _young_old_boundary;
768 ParScanThreadStateSet& _state_set;
769 };
770
771 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
772 ParNewGeneration& young_gen,
773 Generation& old_gen,
774 HeapWord* young_old_boundary,
775 ParScanThreadStateSet& state_set)
776 : AbstractGangTask("ParNewGeneration parallel reference processing"),
777 _young_gen(young_gen),
778 _task(task),
779 _old_gen(old_gen),
780 _young_old_boundary(young_old_boundary),
781 _state_set(state_set)
782 { }
783
784 void ParNewRefProcTaskProxy::work(uint worker_id) {
785 ResourceMark rm;
786 HandleMark hm;
787 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
788 par_scan_state.set_young_old_boundary(_young_old_boundary);
789 _task.work(worker_id, par_scan_state.is_alive_closure(),
790 par_scan_state.keep_alive_closure(),
791 par_scan_state.evacuate_followers_closure());
792 }
793
794 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
795 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
796 EnqueueTask& _task;
797
798 public:
799 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
800 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
801 _task(task)
802 { }
803
804 virtual void work(uint worker_id) {
805 _task.work(worker_id);
806 }
807 };
808
809 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) {
810 GenCollectedHeap* gch = GenCollectedHeap::heap();
811 WorkGang* workers = gch->workers();
812 assert(workers != NULL, "Need parallel worker threads.");
813 _state_set.reset(workers->active_workers(), _young_gen.promotion_failed());
814 ParNewRefProcTaskProxy rp_task(task, _young_gen, _old_gen,
815 _young_gen.reserved().end(), _state_set);
816 workers->run_task(&rp_task);
817 _state_set.reset(0 /* bad value in debug if not reset */,
818 _young_gen.promotion_failed());
819 }
820
821 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) {
822 GenCollectedHeap* gch = GenCollectedHeap::heap();
823 WorkGang* workers = gch->workers();
824 assert(workers != NULL, "Need parallel worker threads.");
825 ParNewRefEnqueueTaskProxy enq_task(task);
826 workers->run_task(&enq_task);
827 }
828
829 void ParNewRefProcTaskExecutor::set_single_threaded_mode() {
830 _state_set.flush();
831 GenCollectedHeap* gch = GenCollectedHeap::heap();
832 gch->save_marks();
833 }
834
835 ScanClosureWithParBarrier::
836 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
837 ScanClosure(g, gc_barrier)
838 { }
839
840 EvacuateFollowersClosureGeneral::
841 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch,
842 OopsInGenClosure* cur,
843 OopsInGenClosure* older) :
844 _gch(gch),
845 _scan_cur_or_nonheap(cur), _scan_older(older)
846 { }
847
848 void EvacuateFollowersClosureGeneral::do_void() {
849 do {
850 // Beware: this call will lead to closure applications via virtual
851 // calls.
852 _gch->oop_since_save_marks_iterate(GenCollectedHeap::YoungGen,
853 _scan_cur_or_nonheap,
854 _scan_older);
855 } while (!_gch->no_allocs_since_save_marks());
856 }
857
858 // A Generation that does parallel young-gen collection.
859
860 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set) {
861 assert(_promo_failure_scan_stack.is_empty(), "post condition");
862 _promo_failure_scan_stack.clear(true); // Clear cached segments.
863
864 remove_forwarding_pointers();
865 log_info(gc, promotion)("Promotion failed");
866 // All the spaces are in play for mark-sweep.
867 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
868 from()->set_next_compaction_space(to());
869 gch->set_incremental_collection_failed();
870 // Inform the next generation that a promotion failure occurred.
871 _old_gen->promotion_failure_occurred();
872
873 // Trace promotion failure in the parallel GC threads
874 thread_state_set.trace_promotion_failed(gc_tracer());
875 // Single threaded code may have reported promotion failure to the global state
876 if (_promotion_failed_info.has_failed()) {
877 _gc_tracer.report_promotion_failed(_promotion_failed_info);
878 }
879 // Reset the PromotionFailureALot counters.
880 NOT_PRODUCT(gch->reset_promotion_should_fail();)
881 }
882
883 void ParNewGeneration::collect(bool full,
884 bool clear_all_soft_refs,
885 size_t size,
886 bool is_tlab) {
887 assert(full || size > 0, "otherwise we don't want to collect");
888
889 GenCollectedHeap* gch = GenCollectedHeap::heap();
890
891 _gc_timer->register_gc_start();
892
893 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
894 WorkGang* workers = gch->workers();
895 assert(workers != NULL, "Need workgang for parallel work");
896 uint active_workers =
897 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
898 workers->active_workers(),
899 Threads::number_of_non_daemon_threads());
900 active_workers = workers->update_active_workers(active_workers);
901 log_info(gc,task)("Using %u workers of %u for evacuation", active_workers, workers->total_workers());
902
903 _old_gen = gch->old_gen();
904
905 // If the next generation is too full to accommodate worst-case promotion
906 // from this generation, pass on collection; let the next generation
907 // do it.
908 if (!collection_attempt_is_safe()) {
909 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
910 return;
911 }
912 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
913
914 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
915 gch->trace_heap_before_gc(gc_tracer());
916
917 init_assuming_no_promotion_failure();
918
919 if (UseAdaptiveSizePolicy) {
920 set_survivor_overflow(false);
921 size_policy->minor_collection_begin();
922 }
923
924 GCTraceTime(Trace, gc, phases) t1("ParNew", NULL, gch->gc_cause());
925
926 age_table()->clear();
927 to()->clear(SpaceDecorator::Mangle);
928
929 gch->save_marks();
930
931 // Set the correct parallelism (number of queues) in the reference processor
932 ref_processor()->set_active_mt_degree(active_workers);
933
934 // Need to initialize the preserved marks before the ThreadStateSet c'tor.
935 _preserved_marks_set.init(active_workers);
936
937 // Always set the terminator for the active number of workers
938 // because only those workers go through the termination protocol.
939 ParallelTaskTerminator _term(active_workers, task_queues());
940 ParScanThreadStateSet thread_state_set(active_workers,
941 *to(), *this, *_old_gen, *task_queues(),
942 _overflow_stacks, _preserved_marks_set,
943 desired_plab_sz(), _term);
944
945 thread_state_set.reset(active_workers, promotion_failed());
946
947 {
948 StrongRootsScope srs(active_workers);
949
950 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs);
951 gch->rem_set()->prepare_for_younger_refs_iterate(true);
952 // It turns out that even when we're using 1 thread, doing the work in a
953 // separate thread causes wide variance in run times. We can't help this
954 // in the multi-threaded case, but we special-case n=1 here to get
955 // repeatable measurements of the 1-thread overhead of the parallel code.
956 // Might multiple workers ever be used? If yes, initialization
957 // has been done such that the single threaded path should not be used.
958 if (workers->total_workers() > 1) {
959 workers->run_task(&tsk);
960 } else {
961 tsk.work(0);
962 }
963 }
964
965 thread_state_set.reset(0 /* Bad value in debug if not reset */,
966 promotion_failed());
967
968 // Trace and reset failed promotion info.
969 if (promotion_failed()) {
970 thread_state_set.trace_promotion_failed(gc_tracer());
971 }
972
973 // Process (weak) reference objects found during scavenge.
974 ReferenceProcessor* rp = ref_processor();
975 IsAliveClosure is_alive(this);
976 ScanWeakRefClosure scan_weak_ref(this);
977 KeepAliveClosure keep_alive(&scan_weak_ref);
978 ScanClosure scan_without_gc_barrier(this, false);
979 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
980 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
981 EvacuateFollowersClosureGeneral evacuate_followers(gch,
982 &scan_without_gc_barrier, &scan_with_gc_barrier);
983 rp->setup_policy(clear_all_soft_refs);
984 // Can the mt_degree be set later (at run_task() time would be best)?
985 rp->set_active_mt_degree(active_workers);
986 ReferenceProcessorStats stats;
987 ReferenceProcessorPhaseTimes pt(_gc_timer, rp->num_q());
988 if (rp->processing_is_mt()) {
989 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set);
990 stats = rp->process_discovered_references(&is_alive, &keep_alive,
991 &evacuate_followers, &task_executor,
992 &pt);
993 } else {
994 thread_state_set.flush();
995 gch->save_marks();
996 stats = rp->process_discovered_references(&is_alive, &keep_alive,
997 &evacuate_followers, NULL,
998 &pt);
999 }
1000 _gc_tracer.report_gc_reference_stats(stats);
1001 _gc_tracer.report_tenuring_threshold(tenuring_threshold());
1002 pt.print_all_references();
1003
1004 if (!promotion_failed()) {
1005 // Swap the survivor spaces.
1006 eden()->clear(SpaceDecorator::Mangle);
1007 from()->clear(SpaceDecorator::Mangle);
1008 if (ZapUnusedHeapArea) {
1009 // This is now done here because of the piece-meal mangling which
1010 // can check for valid mangling at intermediate points in the
1011 // collection(s). When a young collection fails to collect
1012 // sufficient space resizing of the young generation can occur
1013 // and redistribute the spaces in the young generation. Mangle
1014 // here so that unzapped regions don't get distributed to
1015 // other spaces.
1016 to()->mangle_unused_area();
1017 }
1018 swap_spaces();
1019
1020 // A successful scavenge should restart the GC time limit count which is
1021 // for full GC's.
1022 size_policy->reset_gc_overhead_limit_count();
1023
1024 assert(to()->is_empty(), "to space should be empty now");
1025
1026 adjust_desired_tenuring_threshold();
1027 } else {
1028 handle_promotion_failed(gch, thread_state_set);
1029 }
1030 _preserved_marks_set.reclaim();
1031 // set new iteration safe limit for the survivor spaces
1032 from()->set_concurrent_iteration_safe_limit(from()->top());
1033 to()->set_concurrent_iteration_safe_limit(to()->top());
1034
1035 plab_stats()->adjust_desired_plab_sz();
1036
1037 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
1038 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
1039
1040 if (UseAdaptiveSizePolicy) {
1041 size_policy->minor_collection_end(gch->gc_cause());
1042 size_policy->avg_survived()->sample(from()->used());
1043 }
1044
1045 // We need to use a monotonically non-decreasing time in ms
1046 // or we will see time-warp warnings and os::javaTimeMillis()
1047 // does not guarantee monotonicity.
1048 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1049 update_time_of_last_gc(now);
1050
1051 rp->set_enqueuing_is_done(true);
1052 if (rp->processing_is_mt()) {
1053 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set);
1054 rp->enqueue_discovered_references(&task_executor, &pt);
1055 } else {
1056 rp->enqueue_discovered_references(NULL, &pt);
1057 }
1058 rp->verify_no_references_recorded();
1059
1060 gch->trace_heap_after_gc(gc_tracer());
1061
1062 pt.print_enqueue_phase();
1063
1064 _gc_timer->register_gc_end();
1065
1066 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1067 }
1068
1069 size_t ParNewGeneration::desired_plab_sz() {
1070 return _plab_stats.desired_plab_sz(GenCollectedHeap::heap()->workers()->active_workers());
1071 }
1072
1073 static int sum;
1074 void ParNewGeneration::waste_some_time() {
1075 for (int i = 0; i < 100; i++) {
1076 sum += i;
1077 }
1078 }
1079
1080 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1081
1082 // Because of concurrency, there are times where an object for which
1083 // "is_forwarded()" is true contains an "interim" forwarding pointer
1084 // value. Such a value will soon be overwritten with a real value.
1085 // This method requires "obj" to have a forwarding pointer, and waits, if
1086 // necessary for a real one to be inserted, and returns it.
1087
1088 oop ParNewGeneration::real_forwardee(oop obj) {
1089 oop forward_ptr = obj->forwardee();
1090 if (forward_ptr != ClaimedForwardPtr) {
1091 return forward_ptr;
1092 } else {
1093 return real_forwardee_slow(obj);
1094 }
1095 }
1096
1097 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1098 // Spin-read if it is claimed but not yet written by another thread.
1099 oop forward_ptr = obj->forwardee();
1100 while (forward_ptr == ClaimedForwardPtr) {
1101 waste_some_time();
1102 assert(obj->is_forwarded(), "precondition");
1103 forward_ptr = obj->forwardee();
1104 }
1105 return forward_ptr;
1106 }
1107
1108 // Multiple GC threads may try to promote an object. If the object
1109 // is successfully promoted, a forwarding pointer will be installed in
1110 // the object in the young generation. This method claims the right
1111 // to install the forwarding pointer before it copies the object,
1112 // thus avoiding the need to undo the copy as in
1113 // copy_to_survivor_space_avoiding_with_undo.
1114
1115 oop ParNewGeneration::copy_to_survivor_space(ParScanThreadState* par_scan_state,
1116 oop old,
1117 size_t sz,
1118 markOop m) {
1119 // In the sequential version, this assert also says that the object is
1120 // not forwarded. That might not be the case here. It is the case that
1121 // the caller observed it to be not forwarded at some time in the past.
1122 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1123
1124 // The sequential code read "old->age()" below. That doesn't work here,
1125 // since the age is in the mark word, and that might be overwritten with
1126 // a forwarding pointer by a parallel thread. So we must save the mark
1127 // word in a local and then analyze it.
1128 oopDesc dummyOld;
1129 dummyOld.set_mark(m);
1130 assert(!dummyOld.is_forwarded(),
1131 "should not be called with forwarding pointer mark word.");
1132
1133 oop new_obj = NULL;
1134 oop forward_ptr;
1135
1136 // Try allocating obj in to-space (unless too old)
1137 if (dummyOld.age() < tenuring_threshold()) {
1138 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1139 if (new_obj == NULL) {
1140 set_survivor_overflow(true);
1141 }
1142 }
1143
1144 if (new_obj == NULL) {
1145 // Either to-space is full or we decided to promote try allocating obj tenured
1146
1147 // Attempt to install a null forwarding pointer (atomically),
1148 // to claim the right to install the real forwarding pointer.
1149 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1150 if (forward_ptr != NULL) {
1151 // someone else beat us to it.
1152 return real_forwardee(old);
1153 }
1154
1155 if (!_promotion_failed) {
1156 new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1157 old, m, sz);
1158 }
1159
1160 if (new_obj == NULL) {
1161 // promotion failed, forward to self
1162 _promotion_failed = true;
1163 new_obj = old;
1164
1165 par_scan_state->preserved_marks()->push_if_necessary(old, m);
1166 par_scan_state->register_promotion_failure(sz);
1167 }
1168
1169 old->forward_to(new_obj);
1170 forward_ptr = NULL;
1171 } else {
1172 // Is in to-space; do copying ourselves.
1173 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1174 assert(GenCollectedHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1175 forward_ptr = old->forward_to_atomic(new_obj);
1176 // Restore the mark word copied above.
1177 new_obj->set_mark(m);
1178 // Increment age if obj still in new generation
1179 new_obj->incr_age();
1180 par_scan_state->age_table()->add(new_obj, sz);
1181 }
1182 assert(new_obj != NULL, "just checking");
1183
1184 // This code must come after the CAS test, or it will print incorrect
1185 // information.
1186 log_develop_trace(gc, scavenge)("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1187 is_in_reserved(new_obj) ? "copying" : "tenuring",
1188 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size());
1189
1190 if (forward_ptr == NULL) {
1191 oop obj_to_push = new_obj;
1192 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1193 // Length field used as index of next element to be scanned.
1194 // Real length can be obtained from real_forwardee()
1195 arrayOop(old)->set_length(0);
1196 obj_to_push = old;
1197 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1198 "push forwarded object");
1199 }
1200 // Push it on one of the queues of to-be-scanned objects.
1201 bool simulate_overflow = false;
1202 NOT_PRODUCT(
1203 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1204 // simulate a stack overflow
1205 simulate_overflow = true;
1206 }
1207 )
1208 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1209 // Add stats for overflow pushes.
1210 log_develop_trace(gc)("Queue Overflow");
1211 push_on_overflow_list(old, par_scan_state);
1212 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1213 }
1214
1215 return new_obj;
1216 }
1217
1218 // Oops. Someone beat us to it. Undo the allocation. Where did we
1219 // allocate it?
1220 if (is_in_reserved(new_obj)) {
1221 // Must be in to_space.
1222 assert(to()->is_in_reserved(new_obj), "Checking");
1223 if (forward_ptr == ClaimedForwardPtr) {
1224 // Wait to get the real forwarding pointer value.
1225 forward_ptr = real_forwardee(old);
1226 }
1227 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1228 }
1229
1230 return forward_ptr;
1231 }
1232
1233 #ifndef PRODUCT
1234 // It's OK to call this multi-threaded; the worst thing
1235 // that can happen is that we'll get a bunch of closely
1236 // spaced simulated overflows, but that's OK, in fact
1237 // probably good as it would exercise the overflow code
1238 // under contention.
1239 bool ParNewGeneration::should_simulate_overflow() {
1240 if (_overflow_counter-- <= 0) { // just being defensive
1241 _overflow_counter = ParGCWorkQueueOverflowInterval;
1242 return true;
1243 } else {
1244 return false;
1245 }
1246 }
1247 #endif
1248
1249 // In case we are using compressed oops, we need to be careful.
1250 // If the object being pushed is an object array, then its length
1251 // field keeps track of the "grey boundary" at which the next
1252 // incremental scan will be done (see ParGCArrayScanChunk).
1253 // When using compressed oops, this length field is kept in the
1254 // lower 32 bits of the erstwhile klass word and cannot be used
1255 // for the overflow chaining pointer (OCP below). As such the OCP
1256 // would itself need to be compressed into the top 32-bits in this
1257 // case. Unfortunately, see below, in the event that we have a
1258 // promotion failure, the node to be pushed on the list can be
1259 // outside of the Java heap, so the heap-based pointer compression
1260 // would not work (we would have potential aliasing between C-heap
1261 // and Java-heap pointers). For this reason, when using compressed
1262 // oops, we simply use a worker-thread-local, non-shared overflow
1263 // list in the form of a growable array, with a slightly different
1264 // overflow stack draining strategy. If/when we start using fat
1265 // stacks here, we can go back to using (fat) pointer chains
1266 // (although some performance comparisons would be useful since
1267 // single global lists have their own performance disadvantages
1268 // as we were made painfully aware not long ago, see 6786503).
1269 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1270 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1271 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1272 if (ParGCUseLocalOverflow) {
1273 // In the case of compressed oops, we use a private, not-shared
1274 // overflow stack.
1275 par_scan_state->push_on_overflow_stack(from_space_obj);
1276 } else {
1277 assert(!UseCompressedOops, "Error");
1278 // if the object has been forwarded to itself, then we cannot
1279 // use the klass pointer for the linked list. Instead we have
1280 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1281 // XXX This is horribly inefficient when a promotion failure occurs
1282 // and should be fixed. XXX FIX ME !!!
1283 #ifndef PRODUCT
1284 Atomic::inc_ptr(&_num_par_pushes);
1285 assert(_num_par_pushes > 0, "Tautology");
1286 #endif
1287 if (from_space_obj->forwardee() == from_space_obj) {
1288 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1289 listhead->forward_to(from_space_obj);
1290 from_space_obj = listhead;
1291 }
1292 oop observed_overflow_list = _overflow_list;
1293 oop cur_overflow_list;
1294 do {
1295 cur_overflow_list = observed_overflow_list;
1296 if (cur_overflow_list != BUSY) {
1297 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1298 } else {
1299 from_space_obj->set_klass_to_list_ptr(NULL);
1300 }
1301 observed_overflow_list =
1302 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1303 } while (cur_overflow_list != observed_overflow_list);
1304 }
1305 }
1306
1307 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1308 bool res;
1309
1310 if (ParGCUseLocalOverflow) {
1311 res = par_scan_state->take_from_overflow_stack();
1312 } else {
1313 assert(!UseCompressedOops, "Error");
1314 res = take_from_overflow_list_work(par_scan_state);
1315 }
1316 return res;
1317 }
1318
1319
1320 // *NOTE*: The overflow list manipulation code here and
1321 // in CMSCollector:: are very similar in shape,
1322 // except that in the CMS case we thread the objects
1323 // directly into the list via their mark word, and do
1324 // not need to deal with special cases below related
1325 // to chunking of object arrays and promotion failure
1326 // handling.
1327 // CR 6797058 has been filed to attempt consolidation of
1328 // the common code.
1329 // Because of the common code, if you make any changes in
1330 // the code below, please check the CMS version to see if
1331 // similar changes might be needed.
1332 // See CMSCollector::par_take_from_overflow_list() for
1333 // more extensive documentation comments.
1334 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1335 ObjToScanQueue* work_q = par_scan_state->work_queue();
1336 // How many to take?
1337 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1338 (size_t)ParGCDesiredObjsFromOverflowList);
1339
1340 assert(!UseCompressedOops, "Error");
1341 assert(par_scan_state->overflow_stack() == NULL, "Error");
1342 if (_overflow_list == NULL) return false;
1343
1344 // Otherwise, there was something there; try claiming the list.
1345 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1346 // Trim off a prefix of at most objsFromOverflow items
1347 Thread* tid = Thread::current();
1348 size_t spin_count = ParallelGCThreads;
1349 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1350 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1351 // someone grabbed it before we did ...
1352 // ... we spin for a short while...
1353 os::sleep(tid, sleep_time_millis, false);
1354 if (_overflow_list == NULL) {
1355 // nothing left to take
1356 return false;
1357 } else if (_overflow_list != BUSY) {
1358 // try and grab the prefix
1359 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1360 }
1361 }
1362 if (prefix == NULL || prefix == BUSY) {
1363 // Nothing to take or waited long enough
1364 if (prefix == NULL) {
1365 // Write back the NULL in case we overwrote it with BUSY above
1366 // and it is still the same value.
1367 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1368 }
1369 return false;
1370 }
1371 assert(prefix != NULL && prefix != BUSY, "Error");
1372 oop cur = prefix;
1373 for (size_t i = 1; i < objsFromOverflow; ++i) {
1374 oop next = cur->list_ptr_from_klass();
1375 if (next == NULL) break;
1376 cur = next;
1377 }
1378 assert(cur != NULL, "Loop postcondition");
1379
1380 // Reattach remaining (suffix) to overflow list
1381 oop suffix = cur->list_ptr_from_klass();
1382 if (suffix == NULL) {
1383 // Write back the NULL in lieu of the BUSY we wrote
1384 // above and it is still the same value.
1385 if (_overflow_list == BUSY) {
1386 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1387 }
1388 } else {
1389 assert(suffix != BUSY, "Error");
1390 // suffix will be put back on global list
1391 cur->set_klass_to_list_ptr(NULL); // break off suffix
1392 // It's possible that the list is still in the empty(busy) state
1393 // we left it in a short while ago; in that case we may be
1394 // able to place back the suffix.
1395 oop observed_overflow_list = _overflow_list;
1396 oop cur_overflow_list = observed_overflow_list;
1397 bool attached = false;
1398 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1399 observed_overflow_list =
1400 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1401 if (cur_overflow_list == observed_overflow_list) {
1402 attached = true;
1403 break;
1404 } else cur_overflow_list = observed_overflow_list;
1405 }
1406 if (!attached) {
1407 // Too bad, someone else got in in between; we'll need to do a splice.
1408 // Find the last item of suffix list
1409 oop last = suffix;
1410 while (true) {
1411 oop next = last->list_ptr_from_klass();
1412 if (next == NULL) break;
1413 last = next;
1414 }
1415 // Atomically prepend suffix to current overflow list
1416 observed_overflow_list = _overflow_list;
1417 do {
1418 cur_overflow_list = observed_overflow_list;
1419 if (cur_overflow_list != BUSY) {
1420 // Do the splice ...
1421 last->set_klass_to_list_ptr(cur_overflow_list);
1422 } else { // cur_overflow_list == BUSY
1423 last->set_klass_to_list_ptr(NULL);
1424 }
1425 observed_overflow_list =
1426 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1427 } while (cur_overflow_list != observed_overflow_list);
1428 }
1429 }
1430
1431 // Push objects on prefix list onto this thread's work queue
1432 assert(prefix != NULL && prefix != BUSY, "program logic");
1433 cur = prefix;
1434 ssize_t n = 0;
1435 while (cur != NULL) {
1436 oop obj_to_push = cur->forwardee();
1437 oop next = cur->list_ptr_from_klass();
1438 cur->set_klass(obj_to_push->klass());
1439 // This may be an array object that is self-forwarded. In that case, the list pointer
1440 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1441 if (!is_in_reserved(cur)) {
1442 // This can become a scaling bottleneck when there is work queue overflow coincident
1443 // with promotion failure.
1444 oopDesc* f = cur;
1445 FREE_C_HEAP_ARRAY(oopDesc, f);
1446 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1447 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1448 obj_to_push = cur;
1449 }
1450 bool ok = work_q->push(obj_to_push);
1451 assert(ok, "Should have succeeded");
1452 cur = next;
1453 n++;
1454 }
1455 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1456 #ifndef PRODUCT
1457 assert(_num_par_pushes >= n, "Too many pops?");
1458 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1459 #endif
1460 return true;
1461 }
1462 #undef BUSY
1463
1464 void ParNewGeneration::ref_processor_init() {
1465 if (_ref_processor == NULL) {
1466 // Allocate and initialize a reference processor
1467 _ref_processor =
1468 new ReferenceProcessor(_reserved, // span
1469 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1470 ParallelGCThreads, // mt processing degree
1471 refs_discovery_is_mt(), // mt discovery
1472 ParallelGCThreads, // mt discovery degree
1473 refs_discovery_is_atomic(), // atomic_discovery
1474 NULL); // is_alive_non_header
1475 }
1476 }
1477
1478 const char* ParNewGeneration::name() const {
1479 return "par new generation";
1480 }