1 /*
2 * Copyright (c) 2001, 2015, 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.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.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/referencePolicy.hpp"
43 #include "gc/shared/space.hpp"
44 #include "gc/shared/spaceDecorator.hpp"
45 #include "gc/shared/strongRootsScope.hpp"
46 #include "gc/shared/taskqueue.inline.hpp"
47 #include "gc/shared/workgroup.hpp"
48 #include "memory/resourceArea.hpp"
49 #include "oops/objArrayOop.hpp"
50 #include "oops/oop.inline.hpp"
51 #include "runtime/atomic.inline.hpp"
52 #include "runtime/handles.hpp"
53 #include "runtime/handles.inline.hpp"
54 #include "runtime/java.hpp"
55 #include "runtime/thread.inline.hpp"
56 #include "utilities/copy.hpp"
57 #include "utilities/globalDefinitions.hpp"
58 #include "utilities/stack.inline.hpp"
59
60 #ifdef _MSC_VER
61 #pragma warning( push )
62 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
63 #endif
64 ParScanThreadState::ParScanThreadState(Space* to_space_,
65 ParNewGeneration* young_gen_,
66 Generation* old_gen_,
67 int thread_num_,
68 ObjToScanQueueSet* work_queue_set_,
69 Stack<oop, mtGC>* overflow_stacks_,
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 _ageTable(false), // false ==> not the global age table, no perf data.
80 _to_space_alloc_buffer(desired_plab_sz_),
81 _to_space_closure(young_gen_, this),
82 _old_gen_closure(young_gen_, this),
83 _to_space_root_closure(young_gen_, this),
84 _old_gen_root_closure(young_gen_, this),
85 _older_gen_closure(young_gen_, this),
86 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
87 &_to_space_root_closure, young_gen_, &_old_gen_root_closure,
88 work_queue_set_, &term_),
89 _is_alive_closure(young_gen_),
90 _scan_weak_ref_closure(young_gen_, this),
91 _keep_alive_closure(&_scan_weak_ref_closure),
92 _strong_roots_time(0.0),
93 _term_time(0.0)
94 {
95 #if TASKQUEUE_STATS
96 _term_attempts = 0;
97 _overflow_refills = 0;
98 _overflow_refill_objs = 0;
99 #endif // TASKQUEUE_STATS
100
101 _survivor_chunk_array = (ChunkArray*) old_gen()->get_data_recorder(thread_num());
102 _hash_seed = 17; // Might want to take time-based random value.
103 _start = os::elapsedTime();
104 _old_gen_closure.set_generation(old_gen_);
105 _old_gen_root_closure.set_generation(old_gen_);
106 }
107 #ifdef _MSC_VER
108 #pragma warning( pop )
109 #endif
110
111 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
112 size_t plab_word_size) {
113 ChunkArray* sca = survivor_chunk_array();
114 if (sca != NULL) {
115 // A non-null SCA implies that we want the PLAB data recorded.
116 sca->record_sample(plab_start, plab_word_size);
117 }
118 }
119
120 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
121 return new_obj->is_objArray() &&
122 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
123 new_obj != old_obj;
124 }
125
126 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
127 assert(old->is_objArray(), "must be obj array");
128 assert(old->is_forwarded(), "must be forwarded");
129 assert(GenCollectedHeap::heap()->is_in_reserved(old), "must be in heap.");
130 assert(!old_gen()->is_in(old), "must be in young generation.");
131
132 objArrayOop obj = objArrayOop(old->forwardee());
133 // Process ParGCArrayScanChunk elements now
134 // and push the remainder back onto queue
135 int start = arrayOop(old)->length();
136 int end = obj->length();
137 int remainder = end - start;
138 assert(start <= end, "just checking");
139 if (remainder > 2 * ParGCArrayScanChunk) {
140 // Test above combines last partial chunk with a full chunk
141 end = start + ParGCArrayScanChunk;
142 arrayOop(old)->set_length(end);
143 // Push remainder.
144 bool ok = work_queue()->push(old);
145 assert(ok, "just popped, push must be okay");
146 } else {
147 // Restore length so that it can be used if there
148 // is a promotion failure and forwarding pointers
149 // must be removed.
150 arrayOop(old)->set_length(end);
151 }
152
153 // process our set of indices (include header in first chunk)
154 // should make sure end is even (aligned to HeapWord in case of compressed oops)
155 if ((HeapWord *)obj < young_old_boundary()) {
156 // object is in to_space
157 obj->oop_iterate_range(&_to_space_closure, start, end);
158 } else {
159 // object is in old generation
160 obj->oop_iterate_range(&_old_gen_closure, start, end);
161 }
162 }
163
164 void ParScanThreadState::trim_queues(int max_size) {
165 ObjToScanQueue* queue = work_queue();
166 do {
167 while (queue->size() > (juint)max_size) {
168 oop obj_to_scan;
169 if (queue->pop_local(obj_to_scan)) {
170 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
171 if (obj_to_scan->is_objArray() &&
172 obj_to_scan->is_forwarded() &&
173 obj_to_scan->forwardee() != obj_to_scan) {
174 scan_partial_array_and_push_remainder(obj_to_scan);
175 } else {
176 // object is in to_space
177 obj_to_scan->oop_iterate(&_to_space_closure);
178 }
179 } else {
180 // object is in old generation
181 obj_to_scan->oop_iterate(&_old_gen_closure);
182 }
183 }
184 }
185 // For the case of compressed oops, we have a private, non-shared
186 // overflow stack, so we eagerly drain it so as to more evenly
187 // distribute load early. Note: this may be good to do in
188 // general rather than delay for the final stealing phase.
189 // If applicable, we'll transfer a set of objects over to our
190 // work queue, allowing them to be stolen and draining our
191 // private overflow stack.
192 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
193 }
194
195 bool ParScanThreadState::take_from_overflow_stack() {
196 assert(ParGCUseLocalOverflow, "Else should not call");
197 assert(young_gen()->overflow_list() == NULL, "Error");
198 ObjToScanQueue* queue = work_queue();
199 Stack<oop, mtGC>* const of_stack = overflow_stack();
200 const size_t num_overflow_elems = of_stack->size();
201 const size_t space_available = queue->max_elems() - queue->size();
202 const size_t num_take_elems = MIN3(space_available / 4,
203 ParGCDesiredObjsFromOverflowList,
204 num_overflow_elems);
205 // Transfer the most recent num_take_elems from the overflow
206 // stack to our work queue.
207 for (size_t i = 0; i != num_take_elems; i++) {
208 oop cur = of_stack->pop();
209 oop obj_to_push = cur->forwardee();
210 assert(GenCollectedHeap::heap()->is_in_reserved(cur), "Should be in heap");
211 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
212 assert(GenCollectedHeap::heap()->is_in_reserved(obj_to_push), "Should be in heap");
213 if (should_be_partially_scanned(obj_to_push, cur)) {
214 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
215 obj_to_push = cur;
216 }
217 bool ok = queue->push(obj_to_push);
218 assert(ok, "Should have succeeded");
219 }
220 assert(young_gen()->overflow_list() == NULL, "Error");
221 return num_take_elems > 0; // was something transferred?
222 }
223
224 void ParScanThreadState::push_on_overflow_stack(oop p) {
225 assert(ParGCUseLocalOverflow, "Else should not call");
226 overflow_stack()->push(p);
227 assert(young_gen()->overflow_list() == NULL, "Error");
228 }
229
230 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
231 // If the object is small enough, try to reallocate the buffer.
232 HeapWord* obj = NULL;
233 if (!_to_space_full) {
234 PLAB* const plab = to_space_alloc_buffer();
235 Space* const sp = to_space();
236 if (word_sz * 100 < ParallelGCBufferWastePct * plab->word_sz()) {
237 // Is small enough; abandon this buffer and start a new one.
238 plab->retire();
239 size_t buf_size = plab->word_sz();
240 HeapWord* buf_space = sp->par_allocate(buf_size);
241 if (buf_space == NULL) {
242 const size_t min_bytes =
243 PLAB::min_size() << LogHeapWordSize;
244 size_t free_bytes = sp->free();
245 while(buf_space == NULL && free_bytes >= min_bytes) {
246 buf_size = free_bytes >> LogHeapWordSize;
247 assert(buf_size == (size_t)align_object_size(buf_size), "Invariant");
248 buf_space = sp->par_allocate(buf_size);
249 free_bytes = sp->free();
250 }
251 }
252 if (buf_space != NULL) {
253 plab->set_buf(buf_space, buf_size);
254 record_survivor_plab(buf_space, buf_size);
255 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
256 // Note that we cannot compare buf_size < word_sz below
257 // because of AlignmentReserve (see PLAB::allocate()).
258 assert(obj != NULL || plab->words_remaining() < word_sz,
259 "Else should have been able to allocate");
260 // It's conceivable that we may be able to use the
261 // buffer we just grabbed for subsequent small requests
262 // even if not for this one.
263 } else {
264 // We're used up.
265 _to_space_full = true;
266 }
267 } else {
268 // Too large; allocate the object individually.
269 obj = sp->par_allocate(word_sz);
270 }
271 }
272 return obj;
273 }
274
275 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) {
276 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
277 }
278
279 void ParScanThreadState::print_promotion_failure_size() {
280 if (_promotion_failed_info.has_failed() && PrintPromotionFailure) {
281 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ",
282 _thread_num, _promotion_failed_info.first_size());
283 }
284 }
285
286 class ParScanThreadStateSet: private ResourceArray {
287 public:
288 // Initializes states for the specified number of threads;
289 ParScanThreadStateSet(int num_threads,
290 Space& to_space,
291 ParNewGeneration& young_gen,
292 Generation& old_gen,
293 ObjToScanQueueSet& queue_set,
294 Stack<oop, mtGC>* overflow_stacks_,
295 size_t desired_plab_sz,
296 ParallelTaskTerminator& term);
297
298 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
299
300 inline ParScanThreadState& thread_state(int i);
301
302 void trace_promotion_failed(const YoungGCTracer* gc_tracer);
303 void reset(uint active_workers, bool promotion_failed);
304 void flush();
305
306 #if TASKQUEUE_STATS
307 static void
308 print_termination_stats_hdr(outputStream* const st = gclog_or_tty);
309 void print_termination_stats(outputStream* const st = gclog_or_tty);
310 static void
311 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty);
312 void print_taskqueue_stats(outputStream* const st = gclog_or_tty);
313 void reset_stats();
314 #endif // TASKQUEUE_STATS
315
316 private:
317 ParallelTaskTerminator& _term;
318 ParNewGeneration& _young_gen;
319 Generation& _old_gen;
320 public:
321 bool is_valid(int id) const { return id < length(); }
322 ParallelTaskTerminator* terminator() { return &_term; }
323 };
324
325 ParScanThreadStateSet::ParScanThreadStateSet(int num_threads,
326 Space& to_space,
327 ParNewGeneration& young_gen,
328 Generation& old_gen,
329 ObjToScanQueueSet& queue_set,
330 Stack<oop, mtGC>* overflow_stacks,
331 size_t desired_plab_sz,
332 ParallelTaskTerminator& term)
333 : ResourceArray(sizeof(ParScanThreadState), num_threads),
334 _young_gen(young_gen),
335 _old_gen(old_gen),
336 _term(term)
337 {
338 assert(num_threads > 0, "sanity check!");
339 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
340 "overflow_stack allocation mismatch");
341 // Initialize states.
342 for (int i = 0; i < num_threads; ++i) {
343 new ((ParScanThreadState*)_data + i)
344 ParScanThreadState(&to_space, &young_gen, &old_gen, i, &queue_set,
345 overflow_stacks, desired_plab_sz, term);
346 }
347 }
348
349 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) {
350 assert(i >= 0 && i < length(), "sanity check!");
351 return ((ParScanThreadState*)_data)[i];
352 }
353
354 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) {
355 for (int i = 0; i < length(); ++i) {
356 if (thread_state(i).promotion_failed()) {
357 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info());
358 thread_state(i).promotion_failed_info().reset();
359 }
360 }
361 }
362
363 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed) {
364 _term.reset_for_reuse(active_threads);
365 if (promotion_failed) {
366 for (int i = 0; i < length(); ++i) {
367 thread_state(i).print_promotion_failure_size();
368 }
369 }
370 }
371
372 #if TASKQUEUE_STATS
373 void ParScanThreadState::reset_stats() {
374 taskqueue_stats().reset();
375 _term_attempts = 0;
376 _overflow_refills = 0;
377 _overflow_refill_objs = 0;
378 }
379
380 void ParScanThreadStateSet::reset_stats() {
381 for (int i = 0; i < length(); ++i) {
382 thread_state(i).reset_stats();
383 }
384 }
385
386 void ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) {
387 st->print_raw_cr("GC Termination Stats");
388 st->print_raw_cr(" elapsed --strong roots-- -------termination-------");
389 st->print_raw_cr("thr ms ms % ms % attempts");
390 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------");
391 }
392
393 void ParScanThreadStateSet::print_termination_stats(outputStream* const st) {
394 print_termination_stats_hdr(st);
395
396 for (int i = 0; i < length(); ++i) {
397 const ParScanThreadState & pss = thread_state(i);
398 const double elapsed_ms = pss.elapsed_time() * 1000.0;
399 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
400 const double term_ms = pss.term_time() * 1000.0;
401 st->print_cr("%3d %9.2f %9.2f %6.2f %9.2f %6.2f " SIZE_FORMAT_W(8),
402 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
403 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
404 }
405 }
406
407 // Print stats related to work queue activity.
408 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) {
409 st->print_raw_cr("GC Task Stats");
410 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
411 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
412 }
413
414 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st) {
415 print_taskqueue_stats_hdr(st);
416
417 TaskQueueStats totals;
418 for (int i = 0; i < length(); ++i) {
419 const ParScanThreadState & pss = thread_state(i);
420 const TaskQueueStats & stats = pss.taskqueue_stats();
421 st->print("%3d ", i); stats.print(st); st->cr();
422 totals += stats;
423
424 if (pss.overflow_refills() > 0) {
425 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
426 SIZE_FORMAT_W(10) " overflow objects",
427 pss.overflow_refills(), pss.overflow_refill_objs());
428 }
429 }
430 st->print("tot "); totals.print(st); st->cr();
431
432 DEBUG_ONLY(totals.verify());
433 }
434 #endif // TASKQUEUE_STATS
435
436 void ParScanThreadStateSet::flush() {
437 // Work in this loop should be kept as lightweight as
438 // possible since this might otherwise become a bottleneck
439 // to scaling. Should we add heavy-weight work into this
440 // loop, consider parallelizing the loop into the worker threads.
441 for (int i = 0; i < length(); ++i) {
442 ParScanThreadState& par_scan_state = thread_state(i);
443
444 // Flush stats related to To-space PLAB activity and
445 // retire the last buffer.
446 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_young_gen.plab_stats());
447
448 // Every thread has its own age table. We need to merge
449 // them all into one.
450 ageTable *local_table = par_scan_state.age_table();
451 _young_gen.age_table()->merge(local_table);
452
453 // Inform old gen that we're done.
454 _old_gen.par_promote_alloc_done(i);
455 _old_gen.par_oop_since_save_marks_iterate_done(i);
456 }
457
458 if (UseConcMarkSweepGC) {
459 // We need to call this even when ResizeOldPLAB is disabled
460 // so as to avoid breaking some asserts. While we may be able
461 // to avoid this by reorganizing the code a bit, I am loathe
462 // to do that unless we find cases where ergo leads to bad
463 // performance.
464 CFLS_LAB::compute_desired_plab_size();
465 }
466 }
467
468 ParScanClosure::ParScanClosure(ParNewGeneration* g,
469 ParScanThreadState* par_scan_state) :
470 OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) {
471 _boundary = _g->reserved().end();
472 }
473
474 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
475 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
476
477 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
478 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
479
480 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
481 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
482
483 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
484 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
485
486 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
487 ParScanThreadState* par_scan_state)
488 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
489 {}
490
491 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
492 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
493
494 #ifdef WIN32
495 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
496 #endif
497
498 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
499 ParScanThreadState* par_scan_state_,
500 ParScanWithoutBarrierClosure* to_space_closure_,
501 ParScanWithBarrierClosure* old_gen_closure_,
502 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
503 ParNewGeneration* par_gen_,
504 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
505 ObjToScanQueueSet* task_queues_,
506 ParallelTaskTerminator* terminator_) :
507
508 _par_scan_state(par_scan_state_),
509 _to_space_closure(to_space_closure_),
510 _old_gen_closure(old_gen_closure_),
511 _to_space_root_closure(to_space_root_closure_),
512 _old_gen_root_closure(old_gen_root_closure_),
513 _par_gen(par_gen_),
514 _task_queues(task_queues_),
515 _terminator(terminator_)
516 {}
517
518 void ParEvacuateFollowersClosure::do_void() {
519 ObjToScanQueue* work_q = par_scan_state()->work_queue();
520
521 while (true) {
522 // Scan to-space and old-gen objs until we run out of both.
523 oop obj_to_scan;
524 par_scan_state()->trim_queues(0);
525
526 // We have no local work, attempt to steal from other threads.
527
528 // Attempt to steal work from promoted.
529 if (task_queues()->steal(par_scan_state()->thread_num(),
530 par_scan_state()->hash_seed(),
531 obj_to_scan)) {
532 bool res = work_q->push(obj_to_scan);
533 assert(res, "Empty queue should have room for a push.");
534
535 // If successful, goto Start.
536 continue;
537
538 // Try global overflow list.
539 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
540 continue;
541 }
542
543 // Otherwise, offer termination.
544 par_scan_state()->start_term_time();
545 if (terminator()->offer_termination()) break;
546 par_scan_state()->end_term_time();
547 }
548 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
549 "Broken overflow list?");
550 // Finish the last termination pause.
551 par_scan_state()->end_term_time();
552 }
553
554 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen,
555 Generation* old_gen,
556 HeapWord* young_old_boundary,
557 ParScanThreadStateSet* state_set,
558 StrongRootsScope* strong_roots_scope) :
559 AbstractGangTask("ParNewGeneration collection"),
560 _young_gen(young_gen), _old_gen(old_gen),
561 _young_old_boundary(young_old_boundary),
562 _state_set(state_set),
563 _strong_roots_scope(strong_roots_scope)
564 {}
565
566 void ParNewGenTask::work(uint worker_id) {
567 GenCollectedHeap* gch = GenCollectedHeap::heap();
568 // Since this is being done in a separate thread, need new resource
569 // and handle marks.
570 ResourceMark rm;
571 HandleMark hm;
572
573 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
574 assert(_state_set->is_valid(worker_id), "Should not have been called");
575
576 par_scan_state.set_young_old_boundary(_young_old_boundary);
577
578 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(),
579 gch->rem_set()->klass_rem_set());
580 CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure,
581 &par_scan_state.to_space_root_closure(),
582 false);
583
584 par_scan_state.start_strong_roots();
585 gch->gen_process_roots(_strong_roots_scope,
586 GenCollectedHeap::YoungGen,
587 true, // Process younger gens, if any, as strong roots.
588 GenCollectedHeap::SO_ScavengeCodeCache,
589 GenCollectedHeap::StrongAndWeakRoots,
590 &par_scan_state.to_space_root_closure(),
591 &par_scan_state.older_gen_closure(),
592 &cld_scan_closure);
593
594 par_scan_state.end_strong_roots();
595
596 // "evacuate followers".
597 par_scan_state.evacuate_followers_closure().do_void();
598 }
599
600 #ifdef _MSC_VER
601 #pragma warning( push )
602 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
603 #endif
604 ParNewGeneration::ParNewGeneration(ReservedSpace rs, size_t initial_byte_size)
605 : DefNewGeneration(rs, initial_byte_size, "PCopy"),
606 _overflow_list(NULL),
607 _is_alive_closure(this),
608 _plab_stats(YoungPLABSize, PLABWeight)
609 {
610 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
611 NOT_PRODUCT(_num_par_pushes = 0;)
612 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
613 guarantee(_task_queues != NULL, "task_queues allocation failure.");
614
615 for (uint i = 0; i < ParallelGCThreads; i++) {
616 ObjToScanQueue *q = new ObjToScanQueue();
617 guarantee(q != NULL, "work_queue Allocation failure.");
618 _task_queues->register_queue(i, q);
619 }
620
621 for (uint i = 0; i < ParallelGCThreads; i++) {
622 _task_queues->queue(i)->initialize();
623 }
624
625 _overflow_stacks = NULL;
626 if (ParGCUseLocalOverflow) {
627 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal with ','
628 typedef Stack<oop, mtGC> GCOopStack;
629
630 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
631 for (size_t i = 0; i < ParallelGCThreads; ++i) {
632 new (_overflow_stacks + i) Stack<oop, mtGC>();
633 }
634 }
635
636 if (UsePerfData) {
637 EXCEPTION_MARK;
638 ResourceMark rm;
639
640 const char* cname =
641 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
642 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
643 ParallelGCThreads, CHECK);
644 }
645 }
646 #ifdef _MSC_VER
647 #pragma warning( pop )
648 #endif
649
650 // ParNewGeneration::
651 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
652 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
653
654 template <class T>
655 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
656 #ifdef ASSERT
657 {
658 assert(!oopDesc::is_null(*p), "expected non-null ref");
659 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
660 // We never expect to see a null reference being processed
661 // as a weak reference.
662 assert(obj->is_oop(), "expected an oop while scanning weak refs");
663 }
664 #endif // ASSERT
665
666 _par_cl->do_oop_nv(p);
667
668 if (GenCollectedHeap::heap()->is_in_reserved(p)) {
669 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
670 _rs->write_ref_field_gc_par(p, obj);
671 }
672 }
673
674 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
675 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
676
677 // ParNewGeneration::
678 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
679 DefNewGeneration::KeepAliveClosure(cl) {}
680
681 template <class T>
682 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
683 #ifdef ASSERT
684 {
685 assert(!oopDesc::is_null(*p), "expected non-null ref");
686 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
687 // We never expect to see a null reference being processed
688 // as a weak reference.
689 assert(obj->is_oop(), "expected an oop while scanning weak refs");
690 }
691 #endif // ASSERT
692
693 _cl->do_oop_nv(p);
694
695 if (GenCollectedHeap::heap()->is_in_reserved(p)) {
696 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
697 _rs->write_ref_field_gc_par(p, obj);
698 }
699 }
700
701 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
702 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
703
704 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
705 T heap_oop = oopDesc::load_heap_oop(p);
706 if (!oopDesc::is_null(heap_oop)) {
707 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
708 if ((HeapWord*)obj < _boundary) {
709 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
710 oop new_obj = obj->is_forwarded()
711 ? obj->forwardee()
712 : _g->DefNewGeneration::copy_to_survivor_space(obj);
713 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
714 }
715 if (_gc_barrier) {
716 // If p points to a younger generation, mark the card.
717 if ((HeapWord*)obj < _gen_boundary) {
718 _rs->write_ref_field_gc_par(p, obj);
719 }
720 }
721 }
722 }
723
724 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
725 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
726
727 class ParNewRefProcTaskProxy: public AbstractGangTask {
728 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
729 public:
730 ParNewRefProcTaskProxy(ProcessTask& task,
731 ParNewGeneration& young_gen,
732 Generation& old_gen,
733 HeapWord* young_old_boundary,
734 ParScanThreadStateSet& state_set);
735
736 private:
737 virtual void work(uint worker_id);
738 private:
739 ParNewGeneration& _young_gen;
740 ProcessTask& _task;
741 Generation& _old_gen;
742 HeapWord* _young_old_boundary;
743 ParScanThreadStateSet& _state_set;
744 };
745
746 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
747 ParNewGeneration& young_gen,
748 Generation& old_gen,
749 HeapWord* young_old_boundary,
750 ParScanThreadStateSet& state_set)
751 : AbstractGangTask("ParNewGeneration parallel reference processing"),
752 _young_gen(young_gen),
753 _task(task),
754 _old_gen(old_gen),
755 _young_old_boundary(young_old_boundary),
756 _state_set(state_set)
757 { }
758
759 void ParNewRefProcTaskProxy::work(uint worker_id) {
760 ResourceMark rm;
761 HandleMark hm;
762 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
763 par_scan_state.set_young_old_boundary(_young_old_boundary);
764 _task.work(worker_id, par_scan_state.is_alive_closure(),
765 par_scan_state.keep_alive_closure(),
766 par_scan_state.evacuate_followers_closure());
767 }
768
769 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
770 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
771 EnqueueTask& _task;
772
773 public:
774 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
775 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
776 _task(task)
777 { }
778
779 virtual void work(uint worker_id) {
780 _task.work(worker_id);
781 }
782 };
783
784 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) {
785 GenCollectedHeap* gch = GenCollectedHeap::heap();
786 WorkGang* workers = gch->workers();
787 assert(workers != NULL, "Need parallel worker threads.");
788 _state_set.reset(workers->active_workers(), _young_gen.promotion_failed());
789 ParNewRefProcTaskProxy rp_task(task, _young_gen, _old_gen,
790 _young_gen.reserved().end(), _state_set);
791 workers->run_task(&rp_task);
792 _state_set.reset(0 /* bad value in debug if not reset */,
793 _young_gen.promotion_failed());
794 }
795
796 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) {
797 GenCollectedHeap* gch = GenCollectedHeap::heap();
798 WorkGang* workers = gch->workers();
799 assert(workers != NULL, "Need parallel worker threads.");
800 ParNewRefEnqueueTaskProxy enq_task(task);
801 workers->run_task(&enq_task);
802 }
803
804 void ParNewRefProcTaskExecutor::set_single_threaded_mode() {
805 _state_set.flush();
806 GenCollectedHeap* gch = GenCollectedHeap::heap();
807 gch->save_marks();
808 }
809
810 ScanClosureWithParBarrier::
811 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
812 ScanClosure(g, gc_barrier)
813 { }
814
815 EvacuateFollowersClosureGeneral::
816 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch,
817 OopsInGenClosure* cur,
818 OopsInGenClosure* older) :
819 _gch(gch),
820 _scan_cur_or_nonheap(cur), _scan_older(older)
821 { }
822
823 void EvacuateFollowersClosureGeneral::do_void() {
824 do {
825 // Beware: this call will lead to closure applications via virtual
826 // calls.
827 _gch->oop_since_save_marks_iterate(GenCollectedHeap::YoungGen,
828 _scan_cur_or_nonheap,
829 _scan_older);
830 } while (!_gch->no_allocs_since_save_marks());
831 }
832
833 // A Generation that does parallel young-gen collection.
834
835 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set) {
836 assert(_promo_failure_scan_stack.is_empty(), "post condition");
837 _promo_failure_scan_stack.clear(true); // Clear cached segments.
838
839 remove_forwarding_pointers();
840 if (PrintGCDetails) {
841 gclog_or_tty->print(" (promotion failed)");
842 }
843 // All the spaces are in play for mark-sweep.
844 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
845 from()->set_next_compaction_space(to());
846 gch->set_incremental_collection_failed();
847 // Inform the next generation that a promotion failure occurred.
848 _old_gen->promotion_failure_occurred();
849
850 // Trace promotion failure in the parallel GC threads
851 thread_state_set.trace_promotion_failed(gc_tracer());
852 // Single threaded code may have reported promotion failure to the global state
853 if (_promotion_failed_info.has_failed()) {
854 _gc_tracer.report_promotion_failed(_promotion_failed_info);
855 }
856 // Reset the PromotionFailureALot counters.
857 NOT_PRODUCT(gch->reset_promotion_should_fail();)
858 }
859
860 void ParNewGeneration::collect(bool full,
861 bool clear_all_soft_refs,
862 size_t size,
863 bool is_tlab) {
864 assert(full || size > 0, "otherwise we don't want to collect");
865
866 GenCollectedHeap* gch = GenCollectedHeap::heap();
867
868 _gc_timer->register_gc_start();
869
870 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
871 WorkGang* workers = gch->workers();
872 assert(workers != NULL, "Need workgang for parallel work");
873 uint active_workers =
874 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
875 workers->active_workers(),
876 Threads::number_of_non_daemon_threads());
877 workers->set_active_workers(active_workers);
878 _old_gen = gch->old_gen();
879
880 // If the next generation is too full to accommodate worst-case promotion
881 // from this generation, pass on collection; let the next generation
882 // do it.
883 if (!collection_attempt_is_safe()) {
884 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
885 return;
886 }
887 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
888
889 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
890 gch->trace_heap_before_gc(gc_tracer());
891
892 init_assuming_no_promotion_failure();
893
894 if (UseAdaptiveSizePolicy) {
895 set_survivor_overflow(false);
896 size_policy->minor_collection_begin();
897 }
898
899 GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL, _gc_tracer.gc_id());
900 // Capture heap used before collection (for printing).
901 size_t gch_prev_used = gch->used();
902
903 age_table()->clear();
904 to()->clear(SpaceDecorator::Mangle);
905
906 gch->save_marks();
907
908 // Set the correct parallelism (number of queues) in the reference processor
909 ref_processor()->set_active_mt_degree(active_workers);
910
911 // Always set the terminator for the active number of workers
912 // because only those workers go through the termination protocol.
913 ParallelTaskTerminator _term(active_workers, task_queues());
914 ParScanThreadStateSet thread_state_set(active_workers,
915 *to(), *this, *_old_gen, *task_queues(),
916 _overflow_stacks, desired_plab_sz(), _term);
917
918 thread_state_set.reset(active_workers, promotion_failed());
919
920 {
921 StrongRootsScope srs(active_workers);
922
923 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs);
924 gch->rem_set()->prepare_for_younger_refs_iterate(true);
925 // It turns out that even when we're using 1 thread, doing the work in a
926 // separate thread causes wide variance in run times. We can't help this
927 // in the multi-threaded case, but we special-case n=1 here to get
928 // repeatable measurements of the 1-thread overhead of the parallel code.
929 if (active_workers > 1) {
930 workers->run_task(&tsk);
931 } else {
932 tsk.work(0);
933 }
934 }
935
936 thread_state_set.reset(0 /* Bad value in debug if not reset */,
937 promotion_failed());
938
939 // Trace and reset failed promotion info.
940 if (promotion_failed()) {
941 thread_state_set.trace_promotion_failed(gc_tracer());
942 }
943
944 // Process (weak) reference objects found during scavenge.
945 ReferenceProcessor* rp = ref_processor();
946 IsAliveClosure is_alive(this);
947 ScanWeakRefClosure scan_weak_ref(this);
948 KeepAliveClosure keep_alive(&scan_weak_ref);
949 ScanClosure scan_without_gc_barrier(this, false);
950 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
951 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
952 EvacuateFollowersClosureGeneral evacuate_followers(gch,
953 &scan_without_gc_barrier, &scan_with_gc_barrier);
954 rp->setup_policy(clear_all_soft_refs);
955 // Can the mt_degree be set later (at run_task() time would be best)?
956 rp->set_active_mt_degree(active_workers);
957 ReferenceProcessorStats stats;
958 if (rp->processing_is_mt()) {
959 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set);
960 stats = rp->process_discovered_references(&is_alive, &keep_alive,
961 &evacuate_followers, &task_executor,
962 _gc_timer, _gc_tracer.gc_id());
963 } else {
964 thread_state_set.flush();
965 gch->save_marks();
966 stats = rp->process_discovered_references(&is_alive, &keep_alive,
967 &evacuate_followers, NULL,
968 _gc_timer, _gc_tracer.gc_id());
969 }
970 _gc_tracer.report_gc_reference_stats(stats);
971 if (!promotion_failed()) {
972 // Swap the survivor spaces.
973 eden()->clear(SpaceDecorator::Mangle);
974 from()->clear(SpaceDecorator::Mangle);
975 if (ZapUnusedHeapArea) {
976 // This is now done here because of the piece-meal mangling which
977 // can check for valid mangling at intermediate points in the
978 // collection(s). When a young collection fails to collect
979 // sufficient space resizing of the young generation can occur
980 // and redistribute the spaces in the young generation. Mangle
981 // here so that unzapped regions don't get distributed to
982 // other spaces.
983 to()->mangle_unused_area();
984 }
985 swap_spaces();
986
987 // A successful scavenge should restart the GC time limit count which is
988 // for full GC's.
989 size_policy->reset_gc_overhead_limit_count();
990
991 assert(to()->is_empty(), "to space should be empty now");
992
993 adjust_desired_tenuring_threshold();
994 } else {
995 handle_promotion_failed(gch, thread_state_set);
996 }
997 // set new iteration safe limit for the survivor spaces
998 from()->set_concurrent_iteration_safe_limit(from()->top());
999 to()->set_concurrent_iteration_safe_limit(to()->top());
1000
1001 if (ResizePLAB) {
1002 plab_stats()->adjust_desired_plab_sz();
1003 }
1004
1005 if (PrintGC && !PrintGCDetails) {
1006 gch->print_heap_change(gch_prev_used);
1007 }
1008
1009 TASKQUEUE_STATS_ONLY(if (PrintTerminationStats) thread_state_set.print_termination_stats());
1010 TASKQUEUE_STATS_ONLY(if (PrintTaskqueue) thread_state_set.print_taskqueue_stats());
1011
1012 if (UseAdaptiveSizePolicy) {
1013 size_policy->minor_collection_end(gch->gc_cause());
1014 size_policy->avg_survived()->sample(from()->used());
1015 }
1016
1017 // We need to use a monotonically non-decreasing time in ms
1018 // or we will see time-warp warnings and os::javaTimeMillis()
1019 // does not guarantee monotonicity.
1020 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1021 update_time_of_last_gc(now);
1022
1023 rp->set_enqueuing_is_done(true);
1024 if (rp->processing_is_mt()) {
1025 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set);
1026 rp->enqueue_discovered_references(&task_executor);
1027 } else {
1028 rp->enqueue_discovered_references(NULL);
1029 }
1030 rp->verify_no_references_recorded();
1031
1032 gch->trace_heap_after_gc(gc_tracer());
1033 _gc_tracer.report_tenuring_threshold(tenuring_threshold());
1034
1035 _gc_timer->register_gc_end();
1036
1037 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1038 }
1039
1040 size_t ParNewGeneration::desired_plab_sz() {
1041 return _plab_stats.desired_plab_sz(GenCollectedHeap::heap()->workers()->active_workers());
1042 }
1043
1044 static int sum;
1045 void ParNewGeneration::waste_some_time() {
1046 for (int i = 0; i < 100; i++) {
1047 sum += i;
1048 }
1049 }
1050
1051 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1052
1053 // Because of concurrency, there are times where an object for which
1054 // "is_forwarded()" is true contains an "interim" forwarding pointer
1055 // value. Such a value will soon be overwritten with a real value.
1056 // This method requires "obj" to have a forwarding pointer, and waits, if
1057 // necessary for a real one to be inserted, and returns it.
1058
1059 oop ParNewGeneration::real_forwardee(oop obj) {
1060 oop forward_ptr = obj->forwardee();
1061 if (forward_ptr != ClaimedForwardPtr) {
1062 return forward_ptr;
1063 } else {
1064 return real_forwardee_slow(obj);
1065 }
1066 }
1067
1068 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1069 // Spin-read if it is claimed but not yet written by another thread.
1070 oop forward_ptr = obj->forwardee();
1071 while (forward_ptr == ClaimedForwardPtr) {
1072 waste_some_time();
1073 assert(obj->is_forwarded(), "precondition");
1074 forward_ptr = obj->forwardee();
1075 }
1076 return forward_ptr;
1077 }
1078
1079 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
1080 if (m->must_be_preserved_for_promotion_failure(obj)) {
1081 // We should really have separate per-worker stacks, rather
1082 // than use locking of a common pair of stacks.
1083 MutexLocker ml(ParGCRareEvent_lock);
1084 preserve_mark(obj, m);
1085 }
1086 }
1087
1088 // Multiple GC threads may try to promote an object. If the object
1089 // is successfully promoted, a forwarding pointer will be installed in
1090 // the object in the young generation. This method claims the right
1091 // to install the forwarding pointer before it copies the object,
1092 // thus avoiding the need to undo the copy as in
1093 // copy_to_survivor_space_avoiding_with_undo.
1094
1095 oop ParNewGeneration::copy_to_survivor_space(ParScanThreadState* par_scan_state,
1096 oop old,
1097 size_t sz,
1098 markOop m) {
1099 // In the sequential version, this assert also says that the object is
1100 // not forwarded. That might not be the case here. It is the case that
1101 // the caller observed it to be not forwarded at some time in the past.
1102 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1103
1104 // The sequential code read "old->age()" below. That doesn't work here,
1105 // since the age is in the mark word, and that might be overwritten with
1106 // a forwarding pointer by a parallel thread. So we must save the mark
1107 // word in a local and then analyze it.
1108 oopDesc dummyOld;
1109 dummyOld.set_mark(m);
1110 assert(!dummyOld.is_forwarded(),
1111 "should not be called with forwarding pointer mark word.");
1112
1113 oop new_obj = NULL;
1114 oop forward_ptr;
1115
1116 // Try allocating obj in to-space (unless too old)
1117 if (dummyOld.age() < tenuring_threshold()) {
1118 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1119 if (new_obj == NULL) {
1120 set_survivor_overflow(true);
1121 }
1122 }
1123
1124 if (new_obj == NULL) {
1125 // Either to-space is full or we decided to promote try allocating obj tenured
1126
1127 // Attempt to install a null forwarding pointer (atomically),
1128 // to claim the right to install the real forwarding pointer.
1129 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1130 if (forward_ptr != NULL) {
1131 // someone else beat us to it.
1132 return real_forwardee(old);
1133 }
1134
1135 if (!_promotion_failed) {
1136 new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1137 old, m, sz);
1138 }
1139
1140 if (new_obj == NULL) {
1141 // promotion failed, forward to self
1142 _promotion_failed = true;
1143 new_obj = old;
1144
1145 preserve_mark_if_necessary(old, m);
1146 par_scan_state->register_promotion_failure(sz);
1147 }
1148
1149 old->forward_to(new_obj);
1150 forward_ptr = NULL;
1151 } else {
1152 // Is in to-space; do copying ourselves.
1153 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1154 assert(GenCollectedHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1155 forward_ptr = old->forward_to_atomic(new_obj);
1156 // Restore the mark word copied above.
1157 new_obj->set_mark(m);
1158 // Increment age if obj still in new generation
1159 new_obj->incr_age();
1160 par_scan_state->age_table()->add(new_obj, sz);
1161 }
1162 assert(new_obj != NULL, "just checking");
1163
1164 #ifndef PRODUCT
1165 // This code must come after the CAS test, or it will print incorrect
1166 // information.
1167 if (TraceScavenge) {
1168 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1169 is_in_reserved(new_obj) ? "copying" : "tenuring",
1170 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size());
1171 }
1172 #endif
1173
1174 if (forward_ptr == NULL) {
1175 oop obj_to_push = new_obj;
1176 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1177 // Length field used as index of next element to be scanned.
1178 // Real length can be obtained from real_forwardee()
1179 arrayOop(old)->set_length(0);
1180 obj_to_push = old;
1181 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1182 "push forwarded object");
1183 }
1184 // Push it on one of the queues of to-be-scanned objects.
1185 bool simulate_overflow = false;
1186 NOT_PRODUCT(
1187 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1188 // simulate a stack overflow
1189 simulate_overflow = true;
1190 }
1191 )
1192 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1193 // Add stats for overflow pushes.
1194 if (Verbose && PrintGCDetails) {
1195 gclog_or_tty->print("queue overflow!\n");
1196 }
1197 push_on_overflow_list(old, par_scan_state);
1198 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1199 }
1200
1201 return new_obj;
1202 }
1203
1204 // Oops. Someone beat us to it. Undo the allocation. Where did we
1205 // allocate it?
1206 if (is_in_reserved(new_obj)) {
1207 // Must be in to_space.
1208 assert(to()->is_in_reserved(new_obj), "Checking");
1209 if (forward_ptr == ClaimedForwardPtr) {
1210 // Wait to get the real forwarding pointer value.
1211 forward_ptr = real_forwardee(old);
1212 }
1213 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1214 }
1215
1216 return forward_ptr;
1217 }
1218
1219 #ifndef PRODUCT
1220 // It's OK to call this multi-threaded; the worst thing
1221 // that can happen is that we'll get a bunch of closely
1222 // spaced simulated overflows, but that's OK, in fact
1223 // probably good as it would exercise the overflow code
1224 // under contention.
1225 bool ParNewGeneration::should_simulate_overflow() {
1226 if (_overflow_counter-- <= 0) { // just being defensive
1227 _overflow_counter = ParGCWorkQueueOverflowInterval;
1228 return true;
1229 } else {
1230 return false;
1231 }
1232 }
1233 #endif
1234
1235 // In case we are using compressed oops, we need to be careful.
1236 // If the object being pushed is an object array, then its length
1237 // field keeps track of the "grey boundary" at which the next
1238 // incremental scan will be done (see ParGCArrayScanChunk).
1239 // When using compressed oops, this length field is kept in the
1240 // lower 32 bits of the erstwhile klass word and cannot be used
1241 // for the overflow chaining pointer (OCP below). As such the OCP
1242 // would itself need to be compressed into the top 32-bits in this
1243 // case. Unfortunately, see below, in the event that we have a
1244 // promotion failure, the node to be pushed on the list can be
1245 // outside of the Java heap, so the heap-based pointer compression
1246 // would not work (we would have potential aliasing between C-heap
1247 // and Java-heap pointers). For this reason, when using compressed
1248 // oops, we simply use a worker-thread-local, non-shared overflow
1249 // list in the form of a growable array, with a slightly different
1250 // overflow stack draining strategy. If/when we start using fat
1251 // stacks here, we can go back to using (fat) pointer chains
1252 // (although some performance comparisons would be useful since
1253 // single global lists have their own performance disadvantages
1254 // as we were made painfully aware not long ago, see 6786503).
1255 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1256 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1257 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1258 if (ParGCUseLocalOverflow) {
1259 // In the case of compressed oops, we use a private, not-shared
1260 // overflow stack.
1261 par_scan_state->push_on_overflow_stack(from_space_obj);
1262 } else {
1263 assert(!UseCompressedOops, "Error");
1264 // if the object has been forwarded to itself, then we cannot
1265 // use the klass pointer for the linked list. Instead we have
1266 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1267 // XXX This is horribly inefficient when a promotion failure occurs
1268 // and should be fixed. XXX FIX ME !!!
1269 #ifndef PRODUCT
1270 Atomic::inc_ptr(&_num_par_pushes);
1271 assert(_num_par_pushes > 0, "Tautology");
1272 #endif
1273 if (from_space_obj->forwardee() == from_space_obj) {
1274 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1275 listhead->forward_to(from_space_obj);
1276 from_space_obj = listhead;
1277 }
1278 oop observed_overflow_list = _overflow_list;
1279 oop cur_overflow_list;
1280 do {
1281 cur_overflow_list = observed_overflow_list;
1282 if (cur_overflow_list != BUSY) {
1283 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1284 } else {
1285 from_space_obj->set_klass_to_list_ptr(NULL);
1286 }
1287 observed_overflow_list =
1288 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1289 } while (cur_overflow_list != observed_overflow_list);
1290 }
1291 }
1292
1293 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1294 bool res;
1295
1296 if (ParGCUseLocalOverflow) {
1297 res = par_scan_state->take_from_overflow_stack();
1298 } else {
1299 assert(!UseCompressedOops, "Error");
1300 res = take_from_overflow_list_work(par_scan_state);
1301 }
1302 return res;
1303 }
1304
1305
1306 // *NOTE*: The overflow list manipulation code here and
1307 // in CMSCollector:: are very similar in shape,
1308 // except that in the CMS case we thread the objects
1309 // directly into the list via their mark word, and do
1310 // not need to deal with special cases below related
1311 // to chunking of object arrays and promotion failure
1312 // handling.
1313 // CR 6797058 has been filed to attempt consolidation of
1314 // the common code.
1315 // Because of the common code, if you make any changes in
1316 // the code below, please check the CMS version to see if
1317 // similar changes might be needed.
1318 // See CMSCollector::par_take_from_overflow_list() for
1319 // more extensive documentation comments.
1320 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1321 ObjToScanQueue* work_q = par_scan_state->work_queue();
1322 // How many to take?
1323 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1324 (size_t)ParGCDesiredObjsFromOverflowList);
1325
1326 assert(!UseCompressedOops, "Error");
1327 assert(par_scan_state->overflow_stack() == NULL, "Error");
1328 if (_overflow_list == NULL) return false;
1329
1330 // Otherwise, there was something there; try claiming the list.
1331 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1332 // Trim off a prefix of at most objsFromOverflow items
1333 Thread* tid = Thread::current();
1334 size_t spin_count = ParallelGCThreads;
1335 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1336 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1337 // someone grabbed it before we did ...
1338 // ... we spin for a short while...
1339 os::sleep(tid, sleep_time_millis, false);
1340 if (_overflow_list == NULL) {
1341 // nothing left to take
1342 return false;
1343 } else if (_overflow_list != BUSY) {
1344 // try and grab the prefix
1345 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1346 }
1347 }
1348 if (prefix == NULL || prefix == BUSY) {
1349 // Nothing to take or waited long enough
1350 if (prefix == NULL) {
1351 // Write back the NULL in case we overwrote it with BUSY above
1352 // and it is still the same value.
1353 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1354 }
1355 return false;
1356 }
1357 assert(prefix != NULL && prefix != BUSY, "Error");
1358 size_t i = 1;
1359 oop cur = prefix;
1360 while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1361 i++; cur = cur->list_ptr_from_klass();
1362 }
1363
1364 // Reattach remaining (suffix) to overflow list
1365 if (cur->klass_or_null() == NULL) {
1366 // Write back the NULL in lieu of the BUSY we wrote
1367 // above and it is still the same value.
1368 if (_overflow_list == BUSY) {
1369 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1370 }
1371 } else {
1372 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error");
1373 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list
1374 cur->set_klass_to_list_ptr(NULL); // break off suffix
1375 // It's possible that the list is still in the empty(busy) state
1376 // we left it in a short while ago; in that case we may be
1377 // able to place back the suffix.
1378 oop observed_overflow_list = _overflow_list;
1379 oop cur_overflow_list = observed_overflow_list;
1380 bool attached = false;
1381 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1382 observed_overflow_list =
1383 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1384 if (cur_overflow_list == observed_overflow_list) {
1385 attached = true;
1386 break;
1387 } else cur_overflow_list = observed_overflow_list;
1388 }
1389 if (!attached) {
1390 // Too bad, someone else got in in between; we'll need to do a splice.
1391 // Find the last item of suffix list
1392 oop last = suffix;
1393 while (last->klass_or_null() != NULL) {
1394 last = last->list_ptr_from_klass();
1395 }
1396 // Atomically prepend suffix to current overflow list
1397 observed_overflow_list = _overflow_list;
1398 do {
1399 cur_overflow_list = observed_overflow_list;
1400 if (cur_overflow_list != BUSY) {
1401 // Do the splice ...
1402 last->set_klass_to_list_ptr(cur_overflow_list);
1403 } else { // cur_overflow_list == BUSY
1404 last->set_klass_to_list_ptr(NULL);
1405 }
1406 observed_overflow_list =
1407 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1408 } while (cur_overflow_list != observed_overflow_list);
1409 }
1410 }
1411
1412 // Push objects on prefix list onto this thread's work queue
1413 assert(prefix != NULL && prefix != BUSY, "program logic");
1414 cur = prefix;
1415 ssize_t n = 0;
1416 while (cur != NULL) {
1417 oop obj_to_push = cur->forwardee();
1418 oop next = cur->list_ptr_from_klass();
1419 cur->set_klass(obj_to_push->klass());
1420 // This may be an array object that is self-forwarded. In that case, the list pointer
1421 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1422 if (!is_in_reserved(cur)) {
1423 // This can become a scaling bottleneck when there is work queue overflow coincident
1424 // with promotion failure.
1425 oopDesc* f = cur;
1426 FREE_C_HEAP_ARRAY(oopDesc, f);
1427 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1428 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1429 obj_to_push = cur;
1430 }
1431 bool ok = work_q->push(obj_to_push);
1432 assert(ok, "Should have succeeded");
1433 cur = next;
1434 n++;
1435 }
1436 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1437 #ifndef PRODUCT
1438 assert(_num_par_pushes >= n, "Too many pops?");
1439 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1440 #endif
1441 return true;
1442 }
1443 #undef BUSY
1444
1445 void ParNewGeneration::ref_processor_init() {
1446 if (_ref_processor == NULL) {
1447 // Allocate and initialize a reference processor
1448 _ref_processor =
1449 new ReferenceProcessor(_reserved, // span
1450 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1451 ParallelGCThreads, // mt processing degree
1452 refs_discovery_is_mt(), // mt discovery
1453 ParallelGCThreads, // mt discovery degree
1454 refs_discovery_is_atomic(), // atomic_discovery
1455 NULL); // is_alive_non_header
1456 }
1457 }
1458
1459 const char* ParNewGeneration::name() const {
1460 return "par new generation";
1461 }