1 /*
2 * Copyright (c) 2001, 2018, 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 "classfile/stringTable.hpp"
27 #include "gc/cms/cmsHeap.inline.hpp"
28 #include "gc/cms/compactibleFreeListSpace.hpp"
29 #include "gc/cms/concurrentMarkSweepGeneration.hpp"
30 #include "gc/cms/parNewGeneration.inline.hpp"
31 #include "gc/cms/parOopClosures.inline.hpp"
32 #include "gc/serial/defNewGeneration.inline.hpp"
33 #include "gc/shared/adaptiveSizePolicy.hpp"
34 #include "gc/shared/ageTable.inline.hpp"
35 #include "gc/shared/copyFailedInfo.hpp"
36 #include "gc/shared/gcHeapSummary.hpp"
37 #include "gc/shared/gcTimer.hpp"
38 #include "gc/shared/gcTrace.hpp"
39 #include "gc/shared/gcTraceTime.inline.hpp"
40 #include "gc/shared/genOopClosures.inline.hpp"
41 #include "gc/shared/generation.hpp"
42 #include "gc/shared/plab.inline.hpp"
43 #include "gc/shared/preservedMarks.inline.hpp"
44 #include "gc/shared/referencePolicy.hpp"
45 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
46 #include "gc/shared/space.hpp"
47 #include "gc/shared/spaceDecorator.hpp"
48 #include "gc/shared/strongRootsScope.hpp"
49 #include "gc/shared/taskqueue.inline.hpp"
50 #include "gc/shared/weakProcessor.hpp"
51 #include "gc/shared/workgroup.hpp"
52 #include "logging/log.hpp"
53 #include "logging/logStream.hpp"
54 #include "memory/iterator.inline.hpp"
55 #include "memory/resourceArea.hpp"
56 #include "oops/access.inline.hpp"
57 #include "oops/compressedOops.inline.hpp"
58 #include "oops/objArrayOop.hpp"
59 #include "oops/oop.inline.hpp"
60 #include "runtime/atomic.hpp"
61 #include "runtime/handles.hpp"
62 #include "runtime/handles.inline.hpp"
63 #include "runtime/java.hpp"
64 #include "runtime/thread.inline.hpp"
65 #include "utilities/copy.hpp"
66 #include "utilities/globalDefinitions.hpp"
67 #include "utilities/stack.inline.hpp"
68
69 ParScanThreadState::ParScanThreadState(Space* to_space_,
70 ParNewGeneration* young_gen_,
71 Generation* old_gen_,
72 int thread_num_,
73 ObjToScanQueueSet* work_queue_set_,
74 Stack<oop, mtGC>* overflow_stacks_,
75 PreservedMarks* preserved_marks_,
76 size_t desired_plab_sz_,
77 ParallelTaskTerminator& term_) :
78 _work_queue(work_queue_set_->queue(thread_num_)),
79 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
80 _preserved_marks(preserved_marks_),
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 _older_gen_closure(young_gen_, this),
86 _old_gen_root_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 _to_space(to_space_),
94 _young_gen(young_gen_),
95 _old_gen(old_gen_),
96 _young_old_boundary(NULL),
97 _thread_num(thread_num_),
98 _ageTable(false), // false ==> not the global age table, no perf data.
99 _to_space_full(false),
100 _strong_roots_time(0.0),
101 _term_time(0.0)
102 {
103 #if TASKQUEUE_STATS
104 _term_attempts = 0;
105 _overflow_refills = 0;
106 _overflow_refill_objs = 0;
107 #endif // TASKQUEUE_STATS
108
109 _survivor_chunk_array = (ChunkArray*) old_gen()->get_data_recorder(thread_num());
110 _start = os::elapsedTime();
111 _old_gen_closure.set_generation(old_gen_);
112 _old_gen_root_closure.set_generation(old_gen_);
113 }
114
115 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
116 size_t plab_word_size) {
117 ChunkArray* sca = survivor_chunk_array();
118 if (sca != NULL) {
119 // A non-null SCA implies that we want the PLAB data recorded.
120 sca->record_sample(plab_start, plab_word_size);
121 }
122 }
123
124 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
125 return new_obj->is_objArray() &&
126 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
127 new_obj != old_obj;
128 }
129
130 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
131 assert(old->is_objArray(), "must be obj array");
132 assert(old->is_forwarded(), "must be forwarded");
133 assert(CMSHeap::heap()->is_in_reserved(old), "must be in heap.");
134 assert(!old_gen()->is_in(old), "must be in young generation.");
135
136 objArrayOop obj = objArrayOop(old->forwardee());
137 // Process ParGCArrayScanChunk elements now
138 // and push the remainder back onto queue
139 int start = arrayOop(old)->length();
140 int end = obj->length();
141 int remainder = end - start;
142 assert(start <= end, "just checking");
143 if (remainder > 2 * ParGCArrayScanChunk) {
144 // Test above combines last partial chunk with a full chunk
145 end = start + ParGCArrayScanChunk;
146 arrayOop(old)->set_length(end);
147 // Push remainder.
148 bool ok = work_queue()->push(old);
149 assert(ok, "just popped, push must be okay");
150 } else {
151 // Restore length so that it can be used if there
152 // is a promotion failure and forwarding pointers
153 // must be removed.
154 arrayOop(old)->set_length(end);
155 }
156
157 // process our set of indices (include header in first chunk)
158 // should make sure end is even (aligned to HeapWord in case of compressed oops)
159 if ((HeapWord *)obj < young_old_boundary()) {
160 // object is in to_space
161 obj->oop_iterate_range(&_to_space_closure, start, end);
162 } else {
163 // object is in old generation
164 obj->oop_iterate_range(&_old_gen_closure, start, end);
165 }
166 }
167
168 void ParScanThreadState::trim_queues(int max_size) {
169 ObjToScanQueue* queue = work_queue();
170 do {
171 while (queue->size() > (juint)max_size) {
172 oop obj_to_scan;
173 if (queue->pop_local(obj_to_scan)) {
174 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
175 if (obj_to_scan->is_objArray() &&
176 obj_to_scan->is_forwarded() &&
177 obj_to_scan->forwardee() != obj_to_scan) {
178 scan_partial_array_and_push_remainder(obj_to_scan);
179 } else {
180 // object is in to_space
181 obj_to_scan->oop_iterate(&_to_space_closure);
182 }
183 } else {
184 // object is in old generation
185 obj_to_scan->oop_iterate(&_old_gen_closure);
186 }
187 }
188 }
189 // For the case of compressed oops, we have a private, non-shared
190 // overflow stack, so we eagerly drain it so as to more evenly
191 // distribute load early. Note: this may be good to do in
192 // general rather than delay for the final stealing phase.
193 // If applicable, we'll transfer a set of objects over to our
194 // work queue, allowing them to be stolen and draining our
195 // private overflow stack.
196 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
197 }
198
199 bool ParScanThreadState::take_from_overflow_stack() {
200 assert(ParGCUseLocalOverflow, "Else should not call");
201 assert(young_gen()->overflow_list() == NULL, "Error");
202 ObjToScanQueue* queue = work_queue();
203 Stack<oop, mtGC>* const of_stack = overflow_stack();
204 const size_t num_overflow_elems = of_stack->size();
205 const size_t space_available = queue->max_elems() - queue->size();
206 const size_t num_take_elems = MIN3(space_available / 4,
207 (size_t)ParGCDesiredObjsFromOverflowList,
208 num_overflow_elems);
209 // Transfer the most recent num_take_elems from the overflow
210 // stack to our work queue.
211 for (size_t i = 0; i != num_take_elems; i++) {
212 oop cur = of_stack->pop();
213 oop obj_to_push = cur->forwardee();
214 assert(CMSHeap::heap()->is_in_reserved(cur), "Should be in heap");
215 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
216 assert(CMSHeap::heap()->is_in_reserved(obj_to_push), "Should be in heap");
217 if (should_be_partially_scanned(obj_to_push, cur)) {
218 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
219 obj_to_push = cur;
220 }
221 bool ok = queue->push(obj_to_push);
222 assert(ok, "Should have succeeded");
223 }
224 assert(young_gen()->overflow_list() == NULL, "Error");
225 return num_take_elems > 0; // was something transferred?
226 }
227
228 void ParScanThreadState::push_on_overflow_stack(oop p) {
229 assert(ParGCUseLocalOverflow, "Else should not call");
230 overflow_stack()->push(p);
231 assert(young_gen()->overflow_list() == NULL, "Error");
232 }
233
234 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
235 // If the object is small enough, try to reallocate the buffer.
236 HeapWord* obj = NULL;
237 if (!_to_space_full) {
238 PLAB* const plab = to_space_alloc_buffer();
239 Space* const sp = to_space();
240 if (word_sz * 100 < ParallelGCBufferWastePct * plab->word_sz()) {
241 // Is small enough; abandon this buffer and start a new one.
242 plab->retire();
243 // The minimum size has to be twice SurvivorAlignmentInBytes to
244 // allow for padding used in the alignment of 1 word. A padding
245 // of 1 is too small for a filler word so the padding size will
246 // be increased by SurvivorAlignmentInBytes.
247 size_t min_usable_size = 2 * static_cast<size_t>(SurvivorAlignmentInBytes >> LogHeapWordSize);
248 size_t buf_size = MAX2(plab->word_sz(), min_usable_size);
249 HeapWord* buf_space = sp->par_allocate(buf_size);
250 if (buf_space == NULL) {
251 const size_t min_bytes = MAX2(PLAB::min_size(), min_usable_size) << LogHeapWordSize;
252 size_t free_bytes = sp->free();
253 while(buf_space == NULL && free_bytes >= min_bytes) {
254 buf_size = free_bytes >> LogHeapWordSize;
255 assert(buf_size == (size_t)align_object_size(buf_size), "Invariant");
256 buf_space = sp->par_allocate(buf_size);
257 free_bytes = sp->free();
258 }
259 }
260 if (buf_space != NULL) {
261 plab->set_buf(buf_space, buf_size);
262 record_survivor_plab(buf_space, buf_size);
263 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
264 // Note that we cannot compare buf_size < word_sz below
265 // because of AlignmentReserve (see PLAB::allocate()).
266 assert(obj != NULL || plab->words_remaining() < word_sz,
267 "Else should have been able to allocate requested object size "
268 SIZE_FORMAT ", PLAB size " SIZE_FORMAT ", SurvivorAlignmentInBytes "
269 SIZE_FORMAT ", words_remaining " SIZE_FORMAT,
270 word_sz, buf_size, SurvivorAlignmentInBytes, plab->words_remaining());
271 // It's conceivable that we may be able to use the
272 // buffer we just grabbed for subsequent small requests
273 // even if not for this one.
274 } else {
275 // We're used up.
276 _to_space_full = true;
277 }
278 } else {
279 // Too large; allocate the object individually.
280 obj = sp->par_allocate(word_sz);
281 }
282 }
283 return obj;
284 }
285
286 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) {
287 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
288 }
289
290 void ParScanThreadState::print_promotion_failure_size() {
291 if (_promotion_failed_info.has_failed()) {
292 log_trace(gc, promotion)(" (%d: promotion failure size = " SIZE_FORMAT ") ",
293 _thread_num, _promotion_failed_info.first_size());
294 }
295 }
296
297 class ParScanThreadStateSet: StackObj {
298 public:
299 // Initializes states for the specified number of threads;
300 ParScanThreadStateSet(int num_threads,
301 Space& to_space,
302 ParNewGeneration& young_gen,
303 Generation& old_gen,
304 ObjToScanQueueSet& queue_set,
305 Stack<oop, mtGC>* overflow_stacks_,
306 PreservedMarksSet& preserved_marks_set,
307 size_t desired_plab_sz,
308 ParallelTaskTerminator& term);
309
310 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
311
312 inline ParScanThreadState& thread_state(int i);
313
314 void trace_promotion_failed(const YoungGCTracer* gc_tracer);
315 void reset(uint active_workers, bool promotion_failed);
316 void flush();
317
318 #if TASKQUEUE_STATS
319 static void
320 print_termination_stats_hdr(outputStream* const st);
321 void print_termination_stats();
322 static void
323 print_taskqueue_stats_hdr(outputStream* const st);
324 void print_taskqueue_stats();
325 void reset_stats();
326 #endif // TASKQUEUE_STATS
327
328 private:
329 ParallelTaskTerminator& _term;
330 ParNewGeneration& _young_gen;
331 Generation& _old_gen;
332 ParScanThreadState* _per_thread_states;
333 const int _num_threads;
334 public:
335 bool is_valid(int id) const { return id < _num_threads; }
336 ParallelTaskTerminator* terminator() { return &_term; }
337 };
338
339 ParScanThreadStateSet::ParScanThreadStateSet(int num_threads,
340 Space& to_space,
341 ParNewGeneration& young_gen,
342 Generation& old_gen,
343 ObjToScanQueueSet& queue_set,
344 Stack<oop, mtGC>* overflow_stacks,
345 PreservedMarksSet& preserved_marks_set,
346 size_t desired_plab_sz,
347 ParallelTaskTerminator& term)
348 : _term(term),
349 _young_gen(young_gen),
350 _old_gen(old_gen),
351 _per_thread_states(NEW_RESOURCE_ARRAY(ParScanThreadState, num_threads)),
352 _num_threads(num_threads)
353 {
354 assert(num_threads > 0, "sanity check!");
355 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
356 "overflow_stack allocation mismatch");
357 // Initialize states.
358 for (int i = 0; i < num_threads; ++i) {
359 new(_per_thread_states + i)
360 ParScanThreadState(&to_space, &young_gen, &old_gen, i, &queue_set,
361 overflow_stacks, preserved_marks_set.get(i),
362 desired_plab_sz, term);
363 }
364 }
365
366 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) {
367 assert(i >= 0 && i < _num_threads, "sanity check!");
368 return _per_thread_states[i];
369 }
370
371 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) {
372 for (int i = 0; i < _num_threads; ++i) {
373 if (thread_state(i).promotion_failed()) {
374 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info());
375 thread_state(i).promotion_failed_info().reset();
376 }
377 }
378 }
379
380 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed) {
381 _term.reset_for_reuse(active_threads);
382 if (promotion_failed) {
383 for (int i = 0; i < _num_threads; ++i) {
384 thread_state(i).print_promotion_failure_size();
385 }
386 }
387 }
388
389 #if TASKQUEUE_STATS
390 void ParScanThreadState::reset_stats() {
391 taskqueue_stats().reset();
392 _term_attempts = 0;
393 _overflow_refills = 0;
394 _overflow_refill_objs = 0;
395 }
396
397 void ParScanThreadStateSet::reset_stats() {
398 for (int i = 0; i < _num_threads; ++i) {
399 thread_state(i).reset_stats();
400 }
401 }
402
403 void ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) {
404 st->print_raw_cr("GC Termination Stats");
405 st->print_raw_cr(" elapsed --strong roots-- -------termination-------");
406 st->print_raw_cr("thr ms ms % ms % attempts");
407 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------");
408 }
409
410 void ParScanThreadStateSet::print_termination_stats() {
411 Log(gc, task, stats) log;
412 if (!log.is_debug()) {
413 return;
414 }
415
416 ResourceMark rm;
417 LogStream ls(log.debug());
418 outputStream* st = &ls;
419
420 print_termination_stats_hdr(st);
421
422 for (int i = 0; i < _num_threads; ++i) {
423 const ParScanThreadState & pss = thread_state(i);
424 const double elapsed_ms = pss.elapsed_time() * 1000.0;
425 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
426 const double term_ms = pss.term_time() * 1000.0;
427 st->print_cr("%3d %9.2f %9.2f %6.2f %9.2f %6.2f " SIZE_FORMAT_W(8),
428 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
429 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
430 }
431 }
432
433 // Print stats related to work queue activity.
434 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) {
435 st->print_raw_cr("GC Task Stats");
436 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
437 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
438 }
439
440 void ParScanThreadStateSet::print_taskqueue_stats() {
441 if (!log_is_enabled(Trace, gc, task, stats)) {
442 return;
443 }
444 Log(gc, task, stats) log;
445 ResourceMark rm;
446 LogStream ls(log.trace());
447 outputStream* st = &ls;
448 print_taskqueue_stats_hdr(st);
449
450 TaskQueueStats totals;
451 for (int i = 0; i < _num_threads; ++i) {
452 const ParScanThreadState & pss = thread_state(i);
453 const TaskQueueStats & stats = pss.taskqueue_stats();
454 st->print("%3d ", i); stats.print(st); st->cr();
455 totals += stats;
456
457 if (pss.overflow_refills() > 0) {
458 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
459 SIZE_FORMAT_W(10) " overflow objects",
460 pss.overflow_refills(), pss.overflow_refill_objs());
461 }
462 }
463 st->print("tot "); totals.print(st); st->cr();
464
465 DEBUG_ONLY(totals.verify());
466 }
467 #endif // TASKQUEUE_STATS
468
469 void ParScanThreadStateSet::flush() {
470 // Work in this loop should be kept as lightweight as
471 // possible since this might otherwise become a bottleneck
472 // to scaling. Should we add heavy-weight work into this
473 // loop, consider parallelizing the loop into the worker threads.
474 for (int i = 0; i < _num_threads; ++i) {
475 ParScanThreadState& par_scan_state = thread_state(i);
476
477 // Flush stats related to To-space PLAB activity and
478 // retire the last buffer.
479 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_young_gen.plab_stats());
480
481 // Every thread has its own age table. We need to merge
482 // them all into one.
483 AgeTable *local_table = par_scan_state.age_table();
484 _young_gen.age_table()->merge(local_table);
485
486 // Inform old gen that we're done.
487 _old_gen.par_promote_alloc_done(i);
488 }
489
490 if (UseConcMarkSweepGC) {
491 // We need to call this even when ResizeOldPLAB is disabled
492 // so as to avoid breaking some asserts. While we may be able
493 // to avoid this by reorganizing the code a bit, I am loathe
494 // to do that unless we find cases where ergo leads to bad
495 // performance.
496 CompactibleFreeListSpaceLAB::compute_desired_plab_size();
497 }
498 }
499
500 ParScanClosure::ParScanClosure(ParNewGeneration* g,
501 ParScanThreadState* par_scan_state) :
502 OopsInClassLoaderDataOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) {
503 _boundary = _g->reserved().end();
504 }
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 #ifdef WIN32
518 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
519 #endif
520
521 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
522 ParScanThreadState* par_scan_state_,
523 ParScanWithoutBarrierClosure* to_space_closure_,
524 ParScanWithBarrierClosure* old_gen_closure_,
525 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
526 ParNewGeneration* par_gen_,
527 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
528 ObjToScanQueueSet* task_queues_,
529 ParallelTaskTerminator* terminator_) :
530
531 _par_scan_state(par_scan_state_),
532 _to_space_closure(to_space_closure_),
533 _to_space_root_closure(to_space_root_closure_),
534 _old_gen_closure(old_gen_closure_),
535 _old_gen_root_closure(old_gen_root_closure_),
536 _par_gen(par_gen_),
537 _task_queues(task_queues_),
538 _terminator(terminator_)
539 {}
540
541 void ParEvacuateFollowersClosure::do_void() {
542 ObjToScanQueue* work_q = par_scan_state()->work_queue();
543
544 while (true) {
545 // Scan to-space and old-gen objs until we run out of both.
546 oop obj_to_scan;
547 par_scan_state()->trim_queues(0);
548
549 // We have no local work, attempt to steal from other threads.
550
551 // Attempt to steal work from promoted.
552 if (task_queues()->steal(par_scan_state()->thread_num(),
553 obj_to_scan)) {
554 bool res = work_q->push(obj_to_scan);
555 assert(res, "Empty queue should have room for a push.");
556
557 // If successful, goto Start.
558 continue;
559
560 // Try global overflow list.
561 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
562 continue;
563 }
564
565 // Otherwise, offer termination.
566 par_scan_state()->start_term_time();
567 if (terminator()->offer_termination()) break;
568 par_scan_state()->end_term_time();
569 }
570 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
571 "Broken overflow list?");
572 // Finish the last termination pause.
573 par_scan_state()->end_term_time();
574 }
575
576 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen,
577 Generation* old_gen,
578 HeapWord* young_old_boundary,
579 ParScanThreadStateSet* state_set,
580 StrongRootsScope* strong_roots_scope) :
581 AbstractGangTask("ParNewGeneration collection"),
582 _young_gen(young_gen), _old_gen(old_gen),
583 _young_old_boundary(young_old_boundary),
584 _state_set(state_set),
585 _strong_roots_scope(strong_roots_scope),
586 _par_state_string(StringTable::weak_storage())
587 {}
588
589 void ParNewGenTask::work(uint worker_id) {
590 CMSHeap* heap = CMSHeap::heap();
591 // Since this is being done in a separate thread, need new resource
592 // and handle marks.
593 ResourceMark rm;
594 HandleMark hm;
595
596 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
597 assert(_state_set->is_valid(worker_id), "Should not have been called");
598
599 par_scan_state.set_young_old_boundary(_young_old_boundary);
600
601 CLDScanClosure cld_scan_closure(&par_scan_state.to_space_root_closure(),
602 heap->rem_set()->cld_rem_set()->accumulate_modified_oops());
603
604 par_scan_state.start_strong_roots();
605 heap->young_process_roots(_strong_roots_scope,
606 &par_scan_state.to_space_root_closure(),
607 &par_scan_state.older_gen_closure(),
608 &cld_scan_closure,
609 &_par_state_string);
610
611 par_scan_state.end_strong_roots();
612
613 // "evacuate followers".
614 par_scan_state.evacuate_followers_closure().do_void();
615
616 // This will collapse this worker's promoted object list that's
617 // created during the main ParNew parallel phase of ParNew. This has
618 // to be called after all workers have finished promoting objects
619 // and scanning promoted objects. It should be safe calling it from
620 // here, given that we can only reach here after all thread have
621 // offered termination, i.e., after there is no more work to be
622 // done. It will also disable promotion tracking for the rest of
623 // this GC as it's not necessary to be on during reference processing.
624 _old_gen->par_oop_since_save_marks_iterate_done((int) worker_id);
625 }
626
627 ParNewGeneration::ParNewGeneration(ReservedSpace rs, size_t initial_byte_size)
628 : DefNewGeneration(rs, initial_byte_size, "PCopy"),
629 _plab_stats("Young", YoungPLABSize, PLABWeight),
630 _overflow_list(NULL),
631 _is_alive_closure(this)
632 {
633 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
634 NOT_PRODUCT(_num_par_pushes = 0;)
635 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
636 guarantee(_task_queues != NULL, "task_queues allocation failure.");
637
638 for (uint i = 0; i < ParallelGCThreads; i++) {
639 ObjToScanQueue *q = new ObjToScanQueue();
640 guarantee(q != NULL, "work_queue Allocation failure.");
641 _task_queues->register_queue(i, q);
642 }
643
644 for (uint i = 0; i < ParallelGCThreads; i++) {
645 _task_queues->queue(i)->initialize();
646 }
647
648 _overflow_stacks = NULL;
649 if (ParGCUseLocalOverflow) {
650 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal with ','
651 typedef Stack<oop, mtGC> GCOopStack;
652
653 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
654 for (size_t i = 0; i < ParallelGCThreads; ++i) {
655 new (_overflow_stacks + i) Stack<oop, mtGC>();
656 }
657 }
658
659 if (UsePerfData) {
660 EXCEPTION_MARK;
661 ResourceMark rm;
662
663 const char* cname =
664 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
665 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
666 ParallelGCThreads, CHECK);
667 }
668 }
669
670 // ParNewGeneration::
671 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
672 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
673
674 template <class T>
675 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
676 #ifdef ASSERT
677 {
678 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
679 // We never expect to see a null reference being processed
680 // as a weak reference.
681 assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs");
682 }
683 #endif // ASSERT
684
685 Devirtualizer::do_oop_no_verify(_par_cl, p);
686
687 if (CMSHeap::heap()->is_in_reserved(p)) {
688 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);;
689 _rs->write_ref_field_gc_par(p, obj);
690 }
691 }
692
693 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
694 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
695
696 // ParNewGeneration::
697 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
698 DefNewGeneration::KeepAliveClosure(cl) {}
699
700 template <class T>
701 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
702 #ifdef ASSERT
703 {
704 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
705 // We never expect to see a null reference being processed
706 // as a weak reference.
707 assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs");
708 }
709 #endif // ASSERT
710
711 Devirtualizer::do_oop_no_verify(_cl, p);
712
713 if (CMSHeap::heap()->is_in_reserved(p)) {
714 oop obj = RawAccess<IS_NOT_NULL>::oop_load(p);
715 _rs->write_ref_field_gc_par(p, obj);
716 }
717 }
718
719 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
720 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
721
722 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
723 T heap_oop = RawAccess<>::oop_load(p);
724 if (!CompressedOops::is_null(heap_oop)) {
725 oop obj = CompressedOops::decode_not_null(heap_oop);
726 if ((HeapWord*)obj < _boundary) {
727 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
728 oop new_obj = obj->is_forwarded()
729 ? obj->forwardee()
730 : _g->DefNewGeneration::copy_to_survivor_space(obj);
731 RawAccess<IS_NOT_NULL>::oop_store(p, new_obj);
732 }
733 if (_gc_barrier) {
734 // If p points to a younger generation, mark the card.
735 if ((HeapWord*)obj < _gen_boundary) {
736 _rs->write_ref_field_gc_par(p, obj);
737 }
738 }
739 }
740 }
741
742 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
743 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
744
745 class ParNewRefProcTaskProxy: public AbstractGangTask {
746 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
747 public:
748 ParNewRefProcTaskProxy(ProcessTask& task,
749 ParNewGeneration& young_gen,
750 Generation& old_gen,
751 HeapWord* young_old_boundary,
752 ParScanThreadStateSet& state_set);
753
754 private:
755 virtual void work(uint worker_id);
756 private:
757 ParNewGeneration& _young_gen;
758 ProcessTask& _task;
759 Generation& _old_gen;
760 HeapWord* _young_old_boundary;
761 ParScanThreadStateSet& _state_set;
762 };
763
764 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
765 ParNewGeneration& young_gen,
766 Generation& old_gen,
767 HeapWord* young_old_boundary,
768 ParScanThreadStateSet& state_set)
769 : AbstractGangTask("ParNewGeneration parallel reference processing"),
770 _young_gen(young_gen),
771 _task(task),
772 _old_gen(old_gen),
773 _young_old_boundary(young_old_boundary),
774 _state_set(state_set)
775 { }
776
777 void ParNewRefProcTaskProxy::work(uint worker_id) {
778 ResourceMark rm;
779 HandleMark hm;
780 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
781 par_scan_state.set_young_old_boundary(_young_old_boundary);
782 _task.work(worker_id, par_scan_state.is_alive_closure(),
783 par_scan_state.keep_alive_closure(),
784 par_scan_state.evacuate_followers_closure());
785 }
786
787 void ParNewRefProcTaskExecutor::execute(ProcessTask& task, uint ergo_workers) {
788 CMSHeap* gch = CMSHeap::heap();
789 WorkGang* workers = gch->workers();
790 assert(workers != NULL, "Need parallel worker threads.");
791 assert(workers->active_workers() == ergo_workers,
792 "Ergonomically chosen workers (%u) must be equal to active workers (%u)",
793 ergo_workers, workers->active_workers());
794 _state_set.reset(workers->active_workers(), _young_gen.promotion_failed());
795 ParNewRefProcTaskProxy rp_task(task, _young_gen, _old_gen,
796 _young_gen.reserved().end(), _state_set);
797 workers->run_task(&rp_task, workers->active_workers());
798 _state_set.reset(0 /* bad value in debug if not reset */,
799 _young_gen.promotion_failed());
800 }
801
802 void ParNewRefProcTaskExecutor::set_single_threaded_mode() {
803 _state_set.flush();
804 CMSHeap* heap = CMSHeap::heap();
805 heap->save_marks();
806 }
807
808 ScanClosureWithParBarrier::
809 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
810 OopsInClassLoaderDataOrGenClosure(g), _g(g), _boundary(g->reserved().end()), _gc_barrier(gc_barrier)
811 { }
812
813 template <typename OopClosureType1, typename OopClosureType2>
814 EvacuateFollowersClosureGeneral<OopClosureType1, OopClosureType2>::
815 EvacuateFollowersClosureGeneral(CMSHeap* heap,
816 OopClosureType1* cur,
817 OopClosureType2* older) :
818 _heap(heap),
819 _scan_cur_or_nonheap(cur), _scan_older(older)
820 { }
821
822 template <typename OopClosureType1, typename OopClosureType2>
823 void EvacuateFollowersClosureGeneral<OopClosureType1, OopClosureType2>::do_void() {
824 do {
825 _heap->oop_since_save_marks_iterate(_scan_cur_or_nonheap,
826 _scan_older);
827 } while (!_heap->no_allocs_since_save_marks());
828 }
829
830 // A Generation that does parallel young-gen collection.
831
832 void ParNewGeneration::handle_promotion_failed(CMSHeap* gch, ParScanThreadStateSet& thread_state_set) {
833 assert(_promo_failure_scan_stack.is_empty(), "post condition");
834 _promo_failure_scan_stack.clear(true); // Clear cached segments.
835
836 remove_forwarding_pointers();
837 log_info(gc, promotion)("Promotion failed");
838 // All the spaces are in play for mark-sweep.
839 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
840 from()->set_next_compaction_space(to());
841 gch->set_incremental_collection_failed();
842 // Inform the next generation that a promotion failure occurred.
843 _old_gen->promotion_failure_occurred();
844
845 // Trace promotion failure in the parallel GC threads
846 thread_state_set.trace_promotion_failed(gc_tracer());
847 // Single threaded code may have reported promotion failure to the global state
848 if (_promotion_failed_info.has_failed()) {
849 _gc_tracer.report_promotion_failed(_promotion_failed_info);
850 }
851 // Reset the PromotionFailureALot counters.
852 NOT_PRODUCT(gch->reset_promotion_should_fail();)
853 }
854
855 void ParNewGeneration::collect(bool full,
856 bool clear_all_soft_refs,
857 size_t size,
858 bool is_tlab) {
859 assert(full || size > 0, "otherwise we don't want to collect");
860
861 CMSHeap* gch = CMSHeap::heap();
862
863 _gc_timer->register_gc_start();
864
865 AdaptiveSizePolicy* size_policy = gch->size_policy();
866 WorkGang* workers = gch->workers();
867 assert(workers != NULL, "Need workgang for parallel work");
868 uint active_workers =
869 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
870 workers->active_workers(),
871 Threads::number_of_non_daemon_threads());
872 active_workers = workers->update_active_workers(active_workers);
873 log_info(gc,task)("Using %u workers of %u for evacuation", active_workers, workers->total_workers());
874
875 _old_gen = gch->old_gen();
876
877 // If the next generation is too full to accommodate worst-case promotion
878 // from this generation, pass on collection; let the next generation
879 // do it.
880 if (!collection_attempt_is_safe()) {
881 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
882 return;
883 }
884 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
885
886 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
887 gch->trace_heap_before_gc(gc_tracer());
888
889 init_assuming_no_promotion_failure();
890
891 if (UseAdaptiveSizePolicy) {
892 set_survivor_overflow(false);
893 size_policy->minor_collection_begin();
894 }
895
896 GCTraceTime(Trace, gc, phases) t1("ParNew", NULL, gch->gc_cause());
897
898 age_table()->clear();
899 to()->clear(SpaceDecorator::Mangle);
900
901 gch->save_marks();
902
903 // Set the correct parallelism (number of queues) in the reference processor
904 ref_processor()->set_active_mt_degree(active_workers);
905
906 // Need to initialize the preserved marks before the ThreadStateSet c'tor.
907 _preserved_marks_set.init(active_workers);
908
909 // Always set the terminator for the active number of workers
910 // because only those workers go through the termination protocol.
911 ParallelTaskTerminator _term(active_workers, task_queues());
912 ParScanThreadStateSet thread_state_set(active_workers,
913 *to(), *this, *_old_gen, *task_queues(),
914 _overflow_stacks, _preserved_marks_set,
915 desired_plab_sz(), _term);
916
917 thread_state_set.reset(active_workers, promotion_failed());
918
919 {
920 StrongRootsScope srs(active_workers);
921
922 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs);
923 gch->rem_set()->prepare_for_younger_refs_iterate(true);
924 // It turns out that even when we're using 1 thread, doing the work in a
925 // separate thread causes wide variance in run times. We can't help this
926 // in the multi-threaded case, but we special-case n=1 here to get
927 // repeatable measurements of the 1-thread overhead of the parallel code.
928 // Might multiple workers ever be used? If yes, initialization
929 // has been done such that the single threaded path should not be used.
930 if (workers->total_workers() > 1) {
931 workers->run_task(&tsk);
932 } else {
933 tsk.work(0);
934 }
935 }
936
937 thread_state_set.reset(0 /* Bad value in debug if not reset */,
938 promotion_failed());
939
940 // Trace and reset failed promotion info.
941 if (promotion_failed()) {
942 thread_state_set.trace_promotion_failed(gc_tracer());
943 }
944
945 // Process (weak) reference objects found during scavenge.
946 ReferenceProcessor* rp = ref_processor();
947 IsAliveClosure is_alive(this);
948 ScanWeakRefClosure scan_weak_ref(this);
949 KeepAliveClosure keep_alive(&scan_weak_ref);
950 ScanClosure scan_without_gc_barrier(this, false);
951 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
952 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
953 EvacuateFollowersClosureGeneral<ScanClosure, ScanClosureWithParBarrier> evacuate_followers(
954 gch, &scan_without_gc_barrier, &scan_with_gc_barrier);
955 rp->setup_policy(clear_all_soft_refs);
956 // Can the mt_degree be set later (at run_task() time would be best)?
957 rp->set_active_mt_degree(active_workers);
958 ReferenceProcessorStats stats;
959 ReferenceProcessorPhaseTimes pt(_gc_timer, rp->max_num_queues());
960 if (rp->processing_is_mt()) {
961 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set);
962 stats = rp->process_discovered_references(&is_alive, &keep_alive,
963 &evacuate_followers, &task_executor,
964 &pt);
965 } else {
966 thread_state_set.flush();
967 gch->save_marks();
968 stats = rp->process_discovered_references(&is_alive, &keep_alive,
969 &evacuate_followers, NULL,
970 &pt);
971 }
972 _gc_tracer.report_gc_reference_stats(stats);
973 _gc_tracer.report_tenuring_threshold(tenuring_threshold());
974 pt.print_all_references();
975
976 assert(gch->no_allocs_since_save_marks(), "evacuation should be done at this point");
977
978 WeakProcessor::weak_oops_do(&is_alive, &keep_alive);
979
980 // Verify that the usage of keep_alive only forwarded
981 // the oops and did not find anything new to copy.
982 assert(gch->no_allocs_since_save_marks(), "unexpectedly copied objects");
983
984 if (!promotion_failed()) {
985 // Swap the survivor spaces.
986 eden()->clear(SpaceDecorator::Mangle);
987 from()->clear(SpaceDecorator::Mangle);
988 if (ZapUnusedHeapArea) {
989 // This is now done here because of the piece-meal mangling which
990 // can check for valid mangling at intermediate points in the
991 // collection(s). When a young collection fails to collect
992 // sufficient space resizing of the young generation can occur
993 // and redistribute the spaces in the young generation. Mangle
994 // here so that unzapped regions don't get distributed to
995 // other spaces.
996 to()->mangle_unused_area();
997 }
998 swap_spaces();
999
1000 // A successful scavenge should restart the GC time limit count which is
1001 // for full GC's.
1002 size_policy->reset_gc_overhead_limit_count();
1003
1004 assert(to()->is_empty(), "to space should be empty now");
1005
1006 adjust_desired_tenuring_threshold();
1007 } else {
1008 handle_promotion_failed(gch, thread_state_set);
1009 }
1010 _preserved_marks_set.reclaim();
1011 // set new iteration safe limit for the survivor spaces
1012 from()->set_concurrent_iteration_safe_limit(from()->top());
1013 to()->set_concurrent_iteration_safe_limit(to()->top());
1014
1015 plab_stats()->adjust_desired_plab_sz();
1016
1017 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
1018 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
1019
1020 if (UseAdaptiveSizePolicy) {
1021 size_policy->minor_collection_end(gch->gc_cause());
1022 size_policy->avg_survived()->sample(from()->used());
1023 }
1024
1025 // We need to use a monotonically non-decreasing time in ms
1026 // or we will see time-warp warnings and os::javaTimeMillis()
1027 // does not guarantee monotonicity.
1028 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1029 update_time_of_last_gc(now);
1030
1031 rp->set_enqueuing_is_done(true);
1032 rp->verify_no_references_recorded();
1033
1034 gch->trace_heap_after_gc(gc_tracer());
1035
1036 _gc_timer->register_gc_end();
1037
1038 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1039 }
1040
1041 size_t ParNewGeneration::desired_plab_sz() {
1042 return _plab_stats.desired_plab_sz(CMSHeap::heap()->workers()->active_workers());
1043 }
1044
1045 static int sum;
1046 void ParNewGeneration::waste_some_time() {
1047 for (int i = 0; i < 100; i++) {
1048 sum += i;
1049 }
1050 }
1051
1052 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1053
1054 // Because of concurrency, there are times where an object for which
1055 // "is_forwarded()" is true contains an "interim" forwarding pointer
1056 // value. Such a value will soon be overwritten with a real value.
1057 // This method requires "obj" to have a forwarding pointer, and waits, if
1058 // necessary for a real one to be inserted, and returns it.
1059
1060 oop ParNewGeneration::real_forwardee(oop obj) {
1061 oop forward_ptr = obj->forwardee();
1062 if (forward_ptr != ClaimedForwardPtr) {
1063 return forward_ptr;
1064 } else {
1065 return real_forwardee_slow(obj);
1066 }
1067 }
1068
1069 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1070 // Spin-read if it is claimed but not yet written by another thread.
1071 oop forward_ptr = obj->forwardee();
1072 while (forward_ptr == ClaimedForwardPtr) {
1073 waste_some_time();
1074 assert(obj->is_forwarded(), "precondition");
1075 forward_ptr = obj->forwardee();
1076 }
1077 return forward_ptr;
1078 }
1079
1080 // Multiple GC threads may try to promote an object. If the object
1081 // is successfully promoted, a forwarding pointer will be installed in
1082 // the object in the young generation. This method claims the right
1083 // to install the forwarding pointer before it copies the object,
1084 // thus avoiding the need to undo the copy as in
1085 // copy_to_survivor_space_avoiding_with_undo.
1086
1087 oop ParNewGeneration::copy_to_survivor_space(ParScanThreadState* par_scan_state,
1088 oop old,
1089 size_t sz,
1090 markOop m) {
1091 // In the sequential version, this assert also says that the object is
1092 // not forwarded. That might not be the case here. It is the case that
1093 // the caller observed it to be not forwarded at some time in the past.
1094 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1095
1096 // The sequential code read "old->age()" below. That doesn't work here,
1097 // since the age is in the mark word, and that might be overwritten with
1098 // a forwarding pointer by a parallel thread. So we must save the mark
1099 // word in a local and then analyze it.
1100 oopDesc dummyOld;
1101 dummyOld.set_mark_raw(m);
1102 assert(!dummyOld.is_forwarded(),
1103 "should not be called with forwarding pointer mark word.");
1104
1105 oop new_obj = NULL;
1106 oop forward_ptr;
1107
1108 // Try allocating obj in to-space (unless too old)
1109 if (dummyOld.age() < tenuring_threshold()) {
1110 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1111 if (new_obj == NULL) {
1112 set_survivor_overflow(true);
1113 }
1114 }
1115
1116 if (new_obj == NULL) {
1117 // Either to-space is full or we decided to promote try allocating obj tenured
1118
1119 // Attempt to install a null forwarding pointer (atomically),
1120 // to claim the right to install the real forwarding pointer.
1121 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr, m);
1122 if (forward_ptr != NULL) {
1123 // someone else beat us to it.
1124 return real_forwardee(old);
1125 }
1126
1127 if (!_promotion_failed) {
1128 new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1129 old, m, sz);
1130 }
1131
1132 if (new_obj == NULL) {
1133 // promotion failed, forward to self
1134 _promotion_failed = true;
1135 new_obj = old;
1136
1137 par_scan_state->preserved_marks()->push_if_necessary(old, m);
1138 par_scan_state->register_promotion_failure(sz);
1139 }
1140
1141 old->forward_to(new_obj);
1142 forward_ptr = NULL;
1143 } else {
1144 // Is in to-space; do copying ourselves.
1145 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1146 assert(CMSHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1147 forward_ptr = old->forward_to_atomic(new_obj, m);
1148 // Restore the mark word copied above.
1149 new_obj->set_mark_raw(m);
1150 // Increment age if obj still in new generation
1151 new_obj->incr_age();
1152 par_scan_state->age_table()->add(new_obj, sz);
1153 }
1154 assert(new_obj != NULL, "just checking");
1155
1156 // This code must come after the CAS test, or it will print incorrect
1157 // information.
1158 log_develop_trace(gc, scavenge)("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1159 is_in_reserved(new_obj) ? "copying" : "tenuring",
1160 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size());
1161
1162 if (forward_ptr == NULL) {
1163 oop obj_to_push = new_obj;
1164 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1165 // Length field used as index of next element to be scanned.
1166 // Real length can be obtained from real_forwardee()
1167 arrayOop(old)->set_length(0);
1168 obj_to_push = old;
1169 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1170 "push forwarded object");
1171 }
1172 // Push it on one of the queues of to-be-scanned objects.
1173 bool simulate_overflow = false;
1174 NOT_PRODUCT(
1175 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1176 // simulate a stack overflow
1177 simulate_overflow = true;
1178 }
1179 )
1180 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1181 // Add stats for overflow pushes.
1182 log_develop_trace(gc)("Queue Overflow");
1183 push_on_overflow_list(old, par_scan_state);
1184 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1185 }
1186
1187 return new_obj;
1188 }
1189
1190 // Oops. Someone beat us to it. Undo the allocation. Where did we
1191 // allocate it?
1192 if (is_in_reserved(new_obj)) {
1193 // Must be in to_space.
1194 assert(to()->is_in_reserved(new_obj), "Checking");
1195 if (forward_ptr == ClaimedForwardPtr) {
1196 // Wait to get the real forwarding pointer value.
1197 forward_ptr = real_forwardee(old);
1198 }
1199 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1200 }
1201
1202 return forward_ptr;
1203 }
1204
1205 #ifndef PRODUCT
1206 // It's OK to call this multi-threaded; the worst thing
1207 // that can happen is that we'll get a bunch of closely
1208 // spaced simulated overflows, but that's OK, in fact
1209 // probably good as it would exercise the overflow code
1210 // under contention.
1211 bool ParNewGeneration::should_simulate_overflow() {
1212 if (_overflow_counter-- <= 0) { // just being defensive
1213 _overflow_counter = ParGCWorkQueueOverflowInterval;
1214 return true;
1215 } else {
1216 return false;
1217 }
1218 }
1219 #endif
1220
1221 // In case we are using compressed oops, we need to be careful.
1222 // If the object being pushed is an object array, then its length
1223 // field keeps track of the "grey boundary" at which the next
1224 // incremental scan will be done (see ParGCArrayScanChunk).
1225 // When using compressed oops, this length field is kept in the
1226 // lower 32 bits of the erstwhile klass word and cannot be used
1227 // for the overflow chaining pointer (OCP below). As such the OCP
1228 // would itself need to be compressed into the top 32-bits in this
1229 // case. Unfortunately, see below, in the event that we have a
1230 // promotion failure, the node to be pushed on the list can be
1231 // outside of the Java heap, so the heap-based pointer compression
1232 // would not work (we would have potential aliasing between C-heap
1233 // and Java-heap pointers). For this reason, when using compressed
1234 // oops, we simply use a worker-thread-local, non-shared overflow
1235 // list in the form of a growable array, with a slightly different
1236 // overflow stack draining strategy. If/when we start using fat
1237 // stacks here, we can go back to using (fat) pointer chains
1238 // (although some performance comparisons would be useful since
1239 // single global lists have their own performance disadvantages
1240 // as we were made painfully aware not long ago, see 6786503).
1241 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1242 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1243 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1244 if (ParGCUseLocalOverflow) {
1245 // In the case of compressed oops, we use a private, not-shared
1246 // overflow stack.
1247 par_scan_state->push_on_overflow_stack(from_space_obj);
1248 } else {
1249 assert(!UseCompressedOops, "Error");
1250 // if the object has been forwarded to itself, then we cannot
1251 // use the klass pointer for the linked list. Instead we have
1252 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1253 // XXX This is horribly inefficient when a promotion failure occurs
1254 // and should be fixed. XXX FIX ME !!!
1255 #ifndef PRODUCT
1256 Atomic::inc(&_num_par_pushes);
1257 assert(_num_par_pushes > 0, "Tautology");
1258 #endif
1259 if (from_space_obj->forwardee() == from_space_obj) {
1260 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1261 listhead->forward_to(from_space_obj);
1262 from_space_obj = listhead;
1263 }
1264 oop observed_overflow_list = _overflow_list;
1265 oop cur_overflow_list;
1266 do {
1267 cur_overflow_list = observed_overflow_list;
1268 if (cur_overflow_list != BUSY) {
1269 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1270 } else {
1271 from_space_obj->set_klass_to_list_ptr(NULL);
1272 }
1273 observed_overflow_list =
1274 Atomic::cmpxchg((oopDesc*)from_space_obj, &_overflow_list, (oopDesc*)cur_overflow_list);
1275 } while (cur_overflow_list != observed_overflow_list);
1276 }
1277 }
1278
1279 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1280 bool res;
1281
1282 if (ParGCUseLocalOverflow) {
1283 res = par_scan_state->take_from_overflow_stack();
1284 } else {
1285 assert(!UseCompressedOops, "Error");
1286 res = take_from_overflow_list_work(par_scan_state);
1287 }
1288 return res;
1289 }
1290
1291
1292 // *NOTE*: The overflow list manipulation code here and
1293 // in CMSCollector:: are very similar in shape,
1294 // except that in the CMS case we thread the objects
1295 // directly into the list via their mark word, and do
1296 // not need to deal with special cases below related
1297 // to chunking of object arrays and promotion failure
1298 // handling.
1299 // CR 6797058 has been filed to attempt consolidation of
1300 // the common code.
1301 // Because of the common code, if you make any changes in
1302 // the code below, please check the CMS version to see if
1303 // similar changes might be needed.
1304 // See CMSCollector::par_take_from_overflow_list() for
1305 // more extensive documentation comments.
1306 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1307 ObjToScanQueue* work_q = par_scan_state->work_queue();
1308 // How many to take?
1309 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1310 (size_t)ParGCDesiredObjsFromOverflowList);
1311
1312 assert(!UseCompressedOops, "Error");
1313 assert(par_scan_state->overflow_stack() == NULL, "Error");
1314 if (_overflow_list == NULL) return false;
1315
1316 // Otherwise, there was something there; try claiming the list.
1317 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
1318 // Trim off a prefix of at most objsFromOverflow items
1319 Thread* tid = Thread::current();
1320 size_t spin_count = ParallelGCThreads;
1321 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1322 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1323 // someone grabbed it before we did ...
1324 // ... we spin for a short while...
1325 os::sleep(tid, sleep_time_millis, false);
1326 if (_overflow_list == NULL) {
1327 // nothing left to take
1328 return false;
1329 } else if (_overflow_list != BUSY) {
1330 // try and grab the prefix
1331 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
1332 }
1333 }
1334 if (prefix == NULL || prefix == BUSY) {
1335 // Nothing to take or waited long enough
1336 if (prefix == NULL) {
1337 // Write back the NULL in case we overwrote it with BUSY above
1338 // and it is still the same value.
1339 (void) Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
1340 }
1341 return false;
1342 }
1343 assert(prefix != NULL && prefix != BUSY, "Error");
1344 oop cur = prefix;
1345 for (size_t i = 1; i < objsFromOverflow; ++i) {
1346 oop next = cur->list_ptr_from_klass();
1347 if (next == NULL) break;
1348 cur = next;
1349 }
1350 assert(cur != NULL, "Loop postcondition");
1351
1352 // Reattach remaining (suffix) to overflow list
1353 oop suffix = cur->list_ptr_from_klass();
1354 if (suffix == NULL) {
1355 // Write back the NULL in lieu of the BUSY we wrote
1356 // above and it is still the same value.
1357 if (_overflow_list == BUSY) {
1358 (void) Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
1359 }
1360 } else {
1361 assert(suffix != BUSY, "Error");
1362 // suffix will be put back on global list
1363 cur->set_klass_to_list_ptr(NULL); // break off suffix
1364 // It's possible that the list is still in the empty(busy) state
1365 // we left it in a short while ago; in that case we may be
1366 // able to place back the suffix.
1367 oop observed_overflow_list = _overflow_list;
1368 oop cur_overflow_list = observed_overflow_list;
1369 bool attached = false;
1370 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1371 observed_overflow_list =
1372 Atomic::cmpxchg((oopDesc*)suffix, &_overflow_list, (oopDesc*)cur_overflow_list);
1373 if (cur_overflow_list == observed_overflow_list) {
1374 attached = true;
1375 break;
1376 } else cur_overflow_list = observed_overflow_list;
1377 }
1378 if (!attached) {
1379 // Too bad, someone else got in in between; we'll need to do a splice.
1380 // Find the last item of suffix list
1381 oop last = suffix;
1382 while (true) {
1383 oop next = last->list_ptr_from_klass();
1384 if (next == NULL) break;
1385 last = next;
1386 }
1387 // Atomically prepend suffix to current overflow list
1388 observed_overflow_list = _overflow_list;
1389 do {
1390 cur_overflow_list = observed_overflow_list;
1391 if (cur_overflow_list != BUSY) {
1392 // Do the splice ...
1393 last->set_klass_to_list_ptr(cur_overflow_list);
1394 } else { // cur_overflow_list == BUSY
1395 last->set_klass_to_list_ptr(NULL);
1396 }
1397 observed_overflow_list =
1398 Atomic::cmpxchg((oopDesc*)suffix, &_overflow_list, (oopDesc*)cur_overflow_list);
1399 } while (cur_overflow_list != observed_overflow_list);
1400 }
1401 }
1402
1403 // Push objects on prefix list onto this thread's work queue
1404 assert(prefix != NULL && prefix != BUSY, "program logic");
1405 cur = prefix;
1406 ssize_t n = 0;
1407 while (cur != NULL) {
1408 oop obj_to_push = cur->forwardee();
1409 oop next = cur->list_ptr_from_klass();
1410 cur->set_klass(obj_to_push->klass());
1411 // This may be an array object that is self-forwarded. In that case, the list pointer
1412 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1413 if (!is_in_reserved(cur)) {
1414 // This can become a scaling bottleneck when there is work queue overflow coincident
1415 // with promotion failure.
1416 oopDesc* f = cur;
1417 FREE_C_HEAP_ARRAY(oopDesc, f);
1418 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1419 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1420 obj_to_push = cur;
1421 }
1422 bool ok = work_q->push(obj_to_push);
1423 assert(ok, "Should have succeeded");
1424 cur = next;
1425 n++;
1426 }
1427 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1428 #ifndef PRODUCT
1429 assert(_num_par_pushes >= n, "Too many pops?");
1430 Atomic::sub(n, &_num_par_pushes);
1431 #endif
1432 return true;
1433 }
1434 #undef BUSY
1435
1436 void ParNewGeneration::ref_processor_init() {
1437 if (_ref_processor == NULL) {
1438 // Allocate and initialize a reference processor
1439 _span_based_discoverer.set_span(_reserved);
1440 _ref_processor =
1441 new ReferenceProcessor(&_span_based_discoverer, // span
1442 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1443 ParallelGCThreads, // mt processing degree
1444 refs_discovery_is_mt(), // mt discovery
1445 ParallelGCThreads, // mt discovery degree
1446 refs_discovery_is_atomic(), // atomic_discovery
1447 NULL, // is_alive_non_header
1448 false); // disable adjusting number of processing threads
1449 }
1450 }
1451
1452 const char* ParNewGeneration::name() const {
1453 return "par new generation";
1454 }
1455
1456 void ParNewGeneration::restore_preserved_marks() {
1457 SharedRestorePreservedMarksTaskExecutor task_executor(CMSHeap::heap()->workers());
1458 _preserved_marks_set.restore(&task_executor);
1459 }