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_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
27 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp"
28 #include "gc_implementation/parNew/parNewGeneration.hpp"
29 #include "gc_implementation/parNew/parOopClosures.inline.hpp"
30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp"
31 #include "gc_implementation/shared/ageTable.hpp"
32 #include "gc_implementation/shared/copyFailedInfo.hpp"
33 #include "gc_implementation/shared/gcHeapSummary.hpp"
34 #include "gc_implementation/shared/gcTimer.hpp"
35 #include "gc_implementation/shared/gcTrace.hpp"
36 #include "gc_implementation/shared/gcTraceTime.hpp"
37 #include "gc_implementation/shared/parGCAllocBuffer.inline.hpp"
38 #include "gc_implementation/shared/spaceDecorator.hpp"
39 #include "memory/defNewGeneration.inline.hpp"
40 #include "memory/genCollectedHeap.hpp"
41 #include "memory/genOopClosures.inline.hpp"
42 #include "memory/generation.hpp"
43 #include "memory/referencePolicy.hpp"
44 #include "memory/resourceArea.hpp"
45 #include "memory/sharedHeap.hpp"
46 #include "memory/space.hpp"
47 #include "oops/objArrayOop.hpp"
48 #include "oops/oop.inline.hpp"
49 #include "oops/oop.pcgc.inline.hpp"
50 #include "runtime/atomic.inline.hpp"
51 #include "runtime/handles.hpp"
52 #include "runtime/handles.inline.hpp"
53 #include "runtime/java.hpp"
54 #include "runtime/thread.inline.hpp"
55 #include "utilities/copy.hpp"
56 #include "utilities/globalDefinitions.hpp"
57 #include "utilities/workgroup.hpp"
58
59 #ifdef _MSC_VER
60 #pragma warning( push )
61 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
62 #endif
63 ParScanThreadState::ParScanThreadState(Space* to_space_,
64 ParNewGeneration* gen_,
65 Generation* old_gen_,
66 int thread_num_,
67 ObjToScanQueueSet* work_queue_set_,
68 Stack<oop, mtGC>* overflow_stacks_,
69 size_t desired_plab_sz_,
70 ParallelTaskTerminator& term_) :
71 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_),
72 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false),
73 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
74 _ageTable(false), // false ==> not the global age table, no perf data.
75 _to_space_alloc_buffer(desired_plab_sz_),
76 _to_space_closure(gen_, this), _old_gen_closure(gen_, this),
77 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this),
78 _older_gen_closure(gen_, this),
79 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
80 &_to_space_root_closure, gen_, &_old_gen_root_closure,
81 work_queue_set_, &term_),
82 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this),
83 _keep_alive_closure(&_scan_weak_ref_closure),
84 _strong_roots_time(0.0), _term_time(0.0)
85 {
86 #if TASKQUEUE_STATS
87 _term_attempts = 0;
88 _overflow_refills = 0;
89 _overflow_refill_objs = 0;
90 #endif // TASKQUEUE_STATS
91
92 _survivor_chunk_array =
93 (ChunkArray*) old_gen()->get_data_recorder(thread_num());
94 _hash_seed = 17; // Might want to take time-based random value.
95 _start = os::elapsedTime();
96 _old_gen_closure.set_generation(old_gen_);
97 _old_gen_root_closure.set_generation(old_gen_);
98 }
99 #ifdef _MSC_VER
100 #pragma warning( pop )
101 #endif
102
103 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
104 size_t plab_word_size) {
105 ChunkArray* sca = survivor_chunk_array();
106 if (sca != NULL) {
107 // A non-null SCA implies that we want the PLAB data recorded.
108 sca->record_sample(plab_start, plab_word_size);
109 }
110 }
111
112 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
113 return new_obj->is_objArray() &&
114 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
115 new_obj != old_obj;
116 }
117
118 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
119 assert(old->is_objArray(), "must be obj array");
120 assert(old->is_forwarded(), "must be forwarded");
121 assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
122 assert(!old_gen()->is_in(old), "must be in young generation.");
123
124 objArrayOop obj = objArrayOop(old->forwardee());
125 // Process ParGCArrayScanChunk elements now
126 // and push the remainder back onto queue
127 int start = arrayOop(old)->length();
128 int end = obj->length();
129 int remainder = end - start;
130 assert(start <= end, "just checking");
131 if (remainder > 2 * ParGCArrayScanChunk) {
132 // Test above combines last partial chunk with a full chunk
133 end = start + ParGCArrayScanChunk;
134 arrayOop(old)->set_length(end);
135 // Push remainder.
136 bool ok = work_queue()->push(old);
137 assert(ok, "just popped, push must be okay");
138 } else {
139 // Restore length so that it can be used if there
140 // is a promotion failure and forwarding pointers
141 // must be removed.
142 arrayOop(old)->set_length(end);
143 }
144
145 // process our set of indices (include header in first chunk)
146 // should make sure end is even (aligned to HeapWord in case of compressed oops)
147 if ((HeapWord *)obj < young_old_boundary()) {
148 // object is in to_space
149 obj->oop_iterate_range(&_to_space_closure, start, end);
150 } else {
151 // object is in old generation
152 obj->oop_iterate_range(&_old_gen_closure, start, end);
153 }
154 }
155
156
157 void ParScanThreadState::trim_queues(int max_size) {
158 ObjToScanQueue* queue = work_queue();
159 do {
160 while (queue->size() > (juint)max_size) {
161 oop obj_to_scan;
162 if (queue->pop_local(obj_to_scan)) {
163 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
164 if (obj_to_scan->is_objArray() &&
165 obj_to_scan->is_forwarded() &&
166 obj_to_scan->forwardee() != obj_to_scan) {
167 scan_partial_array_and_push_remainder(obj_to_scan);
168 } else {
169 // object is in to_space
170 obj_to_scan->oop_iterate(&_to_space_closure);
171 }
172 } else {
173 // object is in old generation
174 obj_to_scan->oop_iterate(&_old_gen_closure);
175 }
176 }
177 }
178 // For the case of compressed oops, we have a private, non-shared
179 // overflow stack, so we eagerly drain it so as to more evenly
180 // distribute load early. Note: this may be good to do in
181 // general rather than delay for the final stealing phase.
182 // If applicable, we'll transfer a set of objects over to our
183 // work queue, allowing them to be stolen and draining our
184 // private overflow stack.
185 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
186 }
187
188 bool ParScanThreadState::take_from_overflow_stack() {
189 assert(ParGCUseLocalOverflow, "Else should not call");
190 assert(young_gen()->overflow_list() == NULL, "Error");
191 ObjToScanQueue* queue = work_queue();
192 Stack<oop, mtGC>* const of_stack = overflow_stack();
193 const size_t num_overflow_elems = of_stack->size();
194 const size_t space_available = queue->max_elems() - queue->size();
195 const size_t num_take_elems = MIN3(space_available / 4,
196 ParGCDesiredObjsFromOverflowList,
197 num_overflow_elems);
198 // Transfer the most recent num_take_elems from the overflow
199 // stack to our work queue.
200 for (size_t i = 0; i != num_take_elems; i++) {
201 oop cur = of_stack->pop();
202 oop obj_to_push = cur->forwardee();
203 assert(Universe::heap()->is_in_reserved(cur), "Should be in heap");
204 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
205 assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap");
206 if (should_be_partially_scanned(obj_to_push, cur)) {
207 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
208 obj_to_push = cur;
209 }
210 bool ok = queue->push(obj_to_push);
211 assert(ok, "Should have succeeded");
212 }
213 assert(young_gen()->overflow_list() == NULL, "Error");
214 return num_take_elems > 0; // was something transferred?
215 }
216
217 void ParScanThreadState::push_on_overflow_stack(oop p) {
218 assert(ParGCUseLocalOverflow, "Else should not call");
219 overflow_stack()->push(p);
220 assert(young_gen()->overflow_list() == NULL, "Error");
221 }
222
223 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
224
225 // Otherwise, if the object is small enough, try to reallocate the
226 // buffer.
227 HeapWord* obj = NULL;
228 if (!_to_space_full) {
229 ParGCAllocBuffer* const plab = to_space_alloc_buffer();
230 Space* const sp = to_space();
231 if (word_sz * 100 <
232 ParallelGCBufferWastePct * plab->word_sz()) {
233 // Is small enough; abandon this buffer and start a new one.
234 plab->retire();
235 size_t buf_size = plab->word_sz();
236 HeapWord* buf_space = sp->par_allocate(buf_size);
237 if (buf_space == NULL) {
238 const size_t min_bytes =
239 ParGCAllocBuffer::min_size() << LogHeapWordSize;
240 size_t free_bytes = sp->free();
241 while(buf_space == NULL && free_bytes >= min_bytes) {
242 buf_size = free_bytes >> LogHeapWordSize;
243 assert(buf_size == (size_t)align_object_size(buf_size),
244 "Invariant");
245 buf_space = sp->par_allocate(buf_size);
246 free_bytes = sp->free();
247 }
248 }
249 if (buf_space != NULL) {
250 plab->set_word_size(buf_size);
251 plab->set_buf(buf_space);
252 record_survivor_plab(buf_space, buf_size);
253 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
254 // Note that we cannot compare buf_size < word_sz below
255 // because of AlignmentReserve (see ParGCAllocBuffer::allocate()).
256 assert(obj != NULL || plab->words_remaining() < word_sz,
257 "Else should have been able to allocate");
258 // It's conceivable that we may be able to use the
259 // buffer we just grabbed for subsequent small requests
260 // even if not for this one.
261 } else {
262 // We're used up.
263 _to_space_full = true;
264 }
265
266 } else {
267 // Too large; allocate the object individually.
268 obj = sp->par_allocate(word_sz);
269 }
270 }
271 return obj;
272 }
273
274
275 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj,
276 size_t word_sz) {
277 // Is the alloc in the current alloc buffer?
278 if (to_space_alloc_buffer()->contains(obj)) {
279 assert(to_space_alloc_buffer()->contains(obj + word_sz - 1),
280 "Should contain whole object.");
281 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
282 } else {
283 CollectedHeap::fill_with_object(obj, word_sz);
284 }
285 }
286
287 void ParScanThreadState::print_promotion_failure_size() {
288 if (_promotion_failed_info.has_failed() && PrintPromotionFailure) {
289 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ",
290 _thread_num, _promotion_failed_info.first_size());
291 }
292 }
293
294 class ParScanThreadStateSet: private ResourceArray {
295 public:
296 // Initializes states for the specified number of threads;
297 ParScanThreadStateSet(int num_threads,
298 Space& to_space,
299 ParNewGeneration& gen,
300 Generation& old_gen,
301 ObjToScanQueueSet& queue_set,
302 Stack<oop, mtGC>* overflow_stacks_,
303 size_t desired_plab_sz,
304 ParallelTaskTerminator& term);
305
306 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
307
308 inline ParScanThreadState& thread_state(int i);
309
310 void trace_promotion_failed(const YoungGCTracer* gc_tracer);
311 void reset(int active_workers, bool promotion_failed);
312 void flush();
313
314 #if TASKQUEUE_STATS
315 static void
316 print_termination_stats_hdr(outputStream* const st = gclog_or_tty);
317 void print_termination_stats(outputStream* const st = gclog_or_tty);
318 static void
319 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty);
320 void print_taskqueue_stats(outputStream* const st = gclog_or_tty);
321 void reset_stats();
322 #endif // TASKQUEUE_STATS
323
324 private:
325 ParallelTaskTerminator& _term;
326 ParNewGeneration& _gen;
327 Generation& _old_gen;
328 public:
329 bool is_valid(int id) const { return id < length(); }
330 ParallelTaskTerminator* terminator() { return &_term; }
331 };
332
333
334 ParScanThreadStateSet::ParScanThreadStateSet(
335 int num_threads, Space& to_space, ParNewGeneration& gen,
336 Generation& old_gen, ObjToScanQueueSet& queue_set,
337 Stack<oop, mtGC>* overflow_stacks,
338 size_t desired_plab_sz, ParallelTaskTerminator& term)
339 : ResourceArray(sizeof(ParScanThreadState), num_threads),
340 _gen(gen), _old_gen(old_gen), _term(term)
341 {
342 assert(num_threads > 0, "sanity check!");
343 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
344 "overflow_stack allocation mismatch");
345 // Initialize states.
346 for (int i = 0; i < num_threads; ++i) {
347 new ((ParScanThreadState*)_data + i)
348 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set,
349 overflow_stacks, desired_plab_sz, term);
350 }
351 }
352
353 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i)
354 {
355 assert(i >= 0 && i < length(), "sanity check!");
356 return ((ParScanThreadState*)_data)[i];
357 }
358
359 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) {
360 for (int i = 0; i < length(); ++i) {
361 if (thread_state(i).promotion_failed()) {
362 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info());
363 thread_state(i).promotion_failed_info().reset();
364 }
365 }
366 }
367
368 void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed)
369 {
370 _term.reset_for_reuse(active_threads);
371 if (promotion_failed) {
372 for (int i = 0; i < length(); ++i) {
373 thread_state(i).print_promotion_failure_size();
374 }
375 }
376 }
377
378 #if TASKQUEUE_STATS
379 void
380 ParScanThreadState::reset_stats()
381 {
382 taskqueue_stats().reset();
383 _term_attempts = 0;
384 _overflow_refills = 0;
385 _overflow_refill_objs = 0;
386 }
387
388 void ParScanThreadStateSet::reset_stats()
389 {
390 for (int i = 0; i < length(); ++i) {
391 thread_state(i).reset_stats();
392 }
393 }
394
395 void
396 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st)
397 {
398 st->print_raw_cr("GC Termination Stats");
399 st->print_raw_cr(" elapsed --strong roots-- "
400 "-------termination-------");
401 st->print_raw_cr("thr ms ms % "
402 " ms % attempts");
403 st->print_raw_cr("--- --------- --------- ------ "
404 "--------- ------ --------");
405 }
406
407 void ParScanThreadStateSet::print_termination_stats(outputStream* const st)
408 {
409 print_termination_stats_hdr(st);
410
411 for (int i = 0; i < length(); ++i) {
412 const ParScanThreadState & pss = thread_state(i);
413 const double elapsed_ms = pss.elapsed_time() * 1000.0;
414 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
415 const double term_ms = pss.term_time() * 1000.0;
416 st->print_cr("%3d %9.2f %9.2f %6.2f "
417 "%9.2f %6.2f " SIZE_FORMAT_W(8),
418 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
419 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
420 }
421 }
422
423 // Print stats related to work queue activity.
424 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st)
425 {
426 st->print_raw_cr("GC Task Stats");
427 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
428 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
429 }
430
431 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st)
432 {
433 print_taskqueue_stats_hdr(st);
434
435 TaskQueueStats totals;
436 for (int i = 0; i < length(); ++i) {
437 const ParScanThreadState & pss = thread_state(i);
438 const TaskQueueStats & stats = pss.taskqueue_stats();
439 st->print("%3d ", i); stats.print(st); st->cr();
440 totals += stats;
441
442 if (pss.overflow_refills() > 0) {
443 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
444 SIZE_FORMAT_W(10) " overflow objects",
445 pss.overflow_refills(), pss.overflow_refill_objs());
446 }
447 }
448 st->print("tot "); totals.print(st); st->cr();
449
450 DEBUG_ONLY(totals.verify());
451 }
452 #endif // TASKQUEUE_STATS
453
454 void ParScanThreadStateSet::flush()
455 {
456 // Work in this loop should be kept as lightweight as
457 // possible since this might otherwise become a bottleneck
458 // to scaling. Should we add heavy-weight work into this
459 // loop, consider parallelizing the loop into the worker threads.
460 for (int i = 0; i < length(); ++i) {
461 ParScanThreadState& par_scan_state = thread_state(i);
462
463 // Flush stats related to To-space PLAB activity and
464 // retire the last buffer.
465 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_gen.plab_stats());
466
467 // Every thread has its own age table. We need to merge
468 // them all into one.
469 ageTable *local_table = par_scan_state.age_table();
470 _gen.age_table()->merge(local_table);
471
472 // Inform old gen that we're done.
473 _old_gen.par_promote_alloc_done(i);
474 _old_gen.par_oop_since_save_marks_iterate_done(i);
475 }
476
477 if (UseConcMarkSweepGC) {
478 // We need to call this even when ResizeOldPLAB is disabled
479 // so as to avoid breaking some asserts. While we may be able
480 // to avoid this by reorganizing the code a bit, I am loathe
481 // to do that unless we find cases where ergo leads to bad
482 // performance.
483 CFLS_LAB::compute_desired_plab_size();
484 }
485 }
486
487 ParScanClosure::ParScanClosure(ParNewGeneration* g,
488 ParScanThreadState* par_scan_state) :
489 OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g)
490 {
491 assert(_g->level() == 0, "Optimized for youngest generation");
492 _boundary = _g->reserved().end();
493 }
494
495 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
496 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
497
498 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
499 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
500
501 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
502 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
503
504 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
505 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
506
507 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
508 ParScanThreadState* par_scan_state)
509 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
510 {}
511
512 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
513 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
514
515 #ifdef WIN32
516 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
517 #endif
518
519 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
520 ParScanThreadState* par_scan_state_,
521 ParScanWithoutBarrierClosure* to_space_closure_,
522 ParScanWithBarrierClosure* old_gen_closure_,
523 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
524 ParNewGeneration* par_gen_,
525 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
526 ObjToScanQueueSet* task_queues_,
527 ParallelTaskTerminator* terminator_) :
528
529 _par_scan_state(par_scan_state_),
530 _to_space_closure(to_space_closure_),
531 _old_gen_closure(old_gen_closure_),
532 _to_space_root_closure(to_space_root_closure_),
533 _old_gen_root_closure(old_gen_root_closure_),
534 _par_gen(par_gen_),
535 _task_queues(task_queues_),
536 _terminator(terminator_)
537 {}
538
539 void ParEvacuateFollowersClosure::do_void() {
540 ObjToScanQueue* work_q = par_scan_state()->work_queue();
541
542 while (true) {
543
544 // Scan to-space and old-gen objs until we run out of both.
545 oop obj_to_scan;
546 par_scan_state()->trim_queues(0);
547
548 // We have no local work, attempt to steal from other threads.
549
550 // attempt to steal work from promoted.
551 if (task_queues()->steal(par_scan_state()->thread_num(),
552 par_scan_state()->hash_seed(),
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* gen, Generation* old_gen,
577 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) :
578 AbstractGangTask("ParNewGeneration collection"),
579 _gen(gen), _old_gen(old_gen),
580 _young_old_boundary(young_old_boundary),
581 _state_set(state_set)
582 {}
583
584 // Reset the terminator for the given number of
585 // active threads.
586 void ParNewGenTask::set_for_termination(int active_workers) {
587 _state_set->reset(active_workers, _gen->promotion_failed());
588 // Should the heap be passed in? There's only 1 for now so
589 // grab it instead.
590 GenCollectedHeap* gch = GenCollectedHeap::heap();
591 gch->set_n_termination(active_workers);
592 }
593
594 void ParNewGenTask::work(uint worker_id) {
595 GenCollectedHeap* gch = GenCollectedHeap::heap();
596 // Since this is being done in a separate thread, need new resource
597 // and handle marks.
598 ResourceMark rm;
599 HandleMark hm;
600 // We would need multiple old-gen queues otherwise.
601 assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen.");
602
603 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
604 assert(_state_set->is_valid(worker_id), "Should not have been called");
605
606 par_scan_state.set_young_old_boundary(_young_old_boundary);
607
608 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(),
609 gch->rem_set()->klass_rem_set());
610 CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure,
611 &par_scan_state.to_space_root_closure(),
612 false);
613
614 par_scan_state.start_strong_roots();
615 gch->gen_process_roots(_gen->level(),
616 true, // Process younger gens, if any,
617 // as strong roots.
618 false, // no scope; this is parallel code
619 GenCollectedHeap::SO_ScavengeCodeCache,
620 GenCollectedHeap::StrongAndWeakRoots,
621 &par_scan_state.to_space_root_closure(),
622 &par_scan_state.older_gen_closure(),
623 &cld_scan_closure);
624
625 par_scan_state.end_strong_roots();
626
627 // "evacuate followers".
628 par_scan_state.evacuate_followers_closure().do_void();
629 }
630
631 #ifdef _MSC_VER
632 #pragma warning( push )
633 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
634 #endif
635 ParNewGeneration::
636 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level)
637 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"),
638 _overflow_list(NULL),
639 _is_alive_closure(this),
640 _plab_stats(YoungPLABSize, PLABWeight)
641 {
642 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
643 NOT_PRODUCT(_num_par_pushes = 0;)
644 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
645 guarantee(_task_queues != NULL, "task_queues allocation failure.");
646
647 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
648 ObjToScanQueue *q = new ObjToScanQueue();
649 guarantee(q != NULL, "work_queue Allocation failure.");
650 _task_queues->register_queue(i1, q);
651 }
652
653 for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
654 _task_queues->queue(i2)->initialize();
655
656 _overflow_stacks = NULL;
657 if (ParGCUseLocalOverflow) {
658
659 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal
660 // with ','
661 typedef Stack<oop, mtGC> GCOopStack;
662
663 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
664 for (size_t i = 0; i < ParallelGCThreads; ++i) {
665 new (_overflow_stacks + i) Stack<oop, mtGC>();
666 }
667 }
668
669 if (UsePerfData) {
670 EXCEPTION_MARK;
671 ResourceMark rm;
672
673 const char* cname =
674 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
675 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
676 ParallelGCThreads, CHECK);
677 }
678 }
679 #ifdef _MSC_VER
680 #pragma warning( pop )
681 #endif
682
683 // ParNewGeneration::
684 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
685 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
686
687 template <class T>
688 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
689 #ifdef ASSERT
690 {
691 assert(!oopDesc::is_null(*p), "expected non-null ref");
692 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
693 // We never expect to see a null reference being processed
694 // as a weak reference.
695 assert(obj->is_oop(), "expected an oop while scanning weak refs");
696 }
697 #endif // ASSERT
698
699 _par_cl->do_oop_nv(p);
700
701 if (Universe::heap()->is_in_reserved(p)) {
702 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
703 _rs->write_ref_field_gc_par(p, obj);
704 }
705 }
706
707 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
708 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
709
710 // ParNewGeneration::
711 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
712 DefNewGeneration::KeepAliveClosure(cl) {}
713
714 template <class T>
715 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
716 #ifdef ASSERT
717 {
718 assert(!oopDesc::is_null(*p), "expected non-null ref");
719 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
720 // We never expect to see a null reference being processed
721 // as a weak reference.
722 assert(obj->is_oop(), "expected an oop while scanning weak refs");
723 }
724 #endif // ASSERT
725
726 _cl->do_oop_nv(p);
727
728 if (Universe::heap()->is_in_reserved(p)) {
729 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
730 _rs->write_ref_field_gc_par(p, obj);
731 }
732 }
733
734 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
735 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
736
737 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
738 T heap_oop = oopDesc::load_heap_oop(p);
739 if (!oopDesc::is_null(heap_oop)) {
740 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
741 if ((HeapWord*)obj < _boundary) {
742 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
743 oop new_obj = obj->is_forwarded()
744 ? obj->forwardee()
745 : _g->DefNewGeneration::copy_to_survivor_space(obj);
746 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
747 }
748 if (_gc_barrier) {
749 // If p points to a younger generation, mark the card.
750 if ((HeapWord*)obj < _gen_boundary) {
751 _rs->write_ref_field_gc_par(p, obj);
752 }
753 }
754 }
755 }
756
757 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
758 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
759
760 class ParNewRefProcTaskProxy: public AbstractGangTask {
761 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
762 public:
763 ParNewRefProcTaskProxy(ProcessTask& task,
764 ParNewGeneration& gen,
765 Generation& old_gen,
766 HeapWord* young_old_boundary,
767 ParScanThreadStateSet& state_set);
768
769 private:
770 virtual void work(uint worker_id);
771 virtual void set_for_termination(int active_workers) {
772 _state_set.terminator()->reset_for_reuse(active_workers);
773 }
774 private:
775 ParNewGeneration& _gen;
776 ProcessTask& _task;
777 Generation& _old_gen;
778 HeapWord* _young_old_boundary;
779 ParScanThreadStateSet& _state_set;
780 };
781
782 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
783 ParNewGeneration& gen,
784 Generation& old_gen,
785 HeapWord* young_old_boundary,
786 ParScanThreadStateSet& state_set)
787 : AbstractGangTask("ParNewGeneration parallel reference processing"),
788 _gen(gen),
789 _task(task),
790 _old_gen(old_gen),
791 _young_old_boundary(young_old_boundary),
792 _state_set(state_set)
793 {
794 }
795
796 void ParNewRefProcTaskProxy::work(uint worker_id)
797 {
798 ResourceMark rm;
799 HandleMark hm;
800 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
801 par_scan_state.set_young_old_boundary(_young_old_boundary);
802 _task.work(worker_id, par_scan_state.is_alive_closure(),
803 par_scan_state.keep_alive_closure(),
804 par_scan_state.evacuate_followers_closure());
805 }
806
807 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
808 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
809 EnqueueTask& _task;
810
811 public:
812 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
813 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
814 _task(task)
815 { }
816
817 virtual void work(uint worker_id)
818 {
819 _task.work(worker_id);
820 }
821 };
822
823
824 void ParNewRefProcTaskExecutor::execute(ProcessTask& task)
825 {
826 GenCollectedHeap* gch = GenCollectedHeap::heap();
827 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
828 "not a generational heap");
829 FlexibleWorkGang* workers = gch->workers();
830 assert(workers != NULL, "Need parallel worker threads.");
831 _state_set.reset(workers->active_workers(), _generation.promotion_failed());
832 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(),
833 _generation.reserved().end(), _state_set);
834 workers->run_task(&rp_task);
835 _state_set.reset(0 /* bad value in debug if not reset */,
836 _generation.promotion_failed());
837 }
838
839 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task)
840 {
841 GenCollectedHeap* gch = GenCollectedHeap::heap();
842 FlexibleWorkGang* workers = gch->workers();
843 assert(workers != NULL, "Need parallel worker threads.");
844 ParNewRefEnqueueTaskProxy enq_task(task);
845 workers->run_task(&enq_task);
846 }
847
848 void ParNewRefProcTaskExecutor::set_single_threaded_mode()
849 {
850 _state_set.flush();
851 GenCollectedHeap* gch = GenCollectedHeap::heap();
852 gch->set_par_threads(0); // 0 ==> non-parallel.
853 gch->save_marks();
854 }
855
856 ScanClosureWithParBarrier::
857 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
858 ScanClosure(g, gc_barrier) {}
859
860 EvacuateFollowersClosureGeneral::
861 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
862 OopsInGenClosure* cur,
863 OopsInGenClosure* older) :
864 _gch(gch), _level(level),
865 _scan_cur_or_nonheap(cur), _scan_older(older)
866 {}
867
868 void EvacuateFollowersClosureGeneral::do_void() {
869 do {
870 // Beware: this call will lead to closure applications via virtual
871 // calls.
872 _gch->oop_since_save_marks_iterate(_level,
873 _scan_cur_or_nonheap,
874 _scan_older);
875 } while (!_gch->no_allocs_since_save_marks(_level));
876 }
877
878
879 // A Generation that does parallel young-gen collection.
880
881 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set) {
882 assert(_promo_failure_scan_stack.is_empty(), "post condition");
883 _promo_failure_scan_stack.clear(true); // Clear cached segments.
884
885 remove_forwarding_pointers();
886 if (PrintGCDetails) {
887 gclog_or_tty->print(" (promotion failed)");
888 }
889 // All the spaces are in play for mark-sweep.
890 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
891 from()->set_next_compaction_space(to());
892 gch->set_incremental_collection_failed();
893 // Inform the next generation that a promotion failure occurred.
894 _old_gen->promotion_failure_occurred();
895
896 // Trace promotion failure in the parallel GC threads
897 thread_state_set.trace_promotion_failed(gc_tracer());
898 // Single threaded code may have reported promotion failure to the global state
899 if (_promotion_failed_info.has_failed()) {
900 _gc_tracer.report_promotion_failed(_promotion_failed_info);
901 }
902 // Reset the PromotionFailureALot counters.
903 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();)
904 }
905
906 void ParNewGeneration::collect(bool full,
907 bool clear_all_soft_refs,
908 size_t size,
909 bool is_tlab) {
910 assert(full || size > 0, "otherwise we don't want to collect");
911
912 GenCollectedHeap* gch = GenCollectedHeap::heap();
913
914 _gc_timer->register_gc_start();
915
916 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
917 "not a CMS generational heap");
918 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
919 FlexibleWorkGang* workers = gch->workers();
920 assert(workers != NULL, "Need workgang for parallel work");
921 int active_workers =
922 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
923 workers->active_workers(),
924 Threads::number_of_non_daemon_threads());
925 workers->set_active_workers(active_workers);
926 assert(gch->n_gens() == 2,
927 "Par collection currently only works with single older gen.");
928 _old_gen = gch->old_gen();
929
930 // If the next generation is too full to accommodate worst-case promotion
931 // from this generation, pass on collection; let the next generation
932 // do it.
933 if (!collection_attempt_is_safe()) {
934 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
935 return;
936 }
937 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
938
939 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
940 gch->trace_heap_before_gc(gc_tracer());
941
942 init_assuming_no_promotion_failure();
943
944 if (UseAdaptiveSizePolicy) {
945 set_survivor_overflow(false);
946 size_policy->minor_collection_begin();
947 }
948
949 GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL, _gc_tracer.gc_id());
950 // Capture heap used before collection (for printing).
951 size_t gch_prev_used = gch->used();
952
953 age_table()->clear();
954 to()->clear(SpaceDecorator::Mangle);
955
956 gch->save_marks();
957 assert(workers != NULL, "Need parallel worker threads.");
958 int n_workers = active_workers;
959
960 // Set the correct parallelism (number of queues) in the reference processor
961 ref_processor()->set_active_mt_degree(n_workers);
962
963 // Always set the terminator for the active number of workers
964 // because only those workers go through the termination protocol.
965 ParallelTaskTerminator _term(n_workers, task_queues());
966 ParScanThreadStateSet thread_state_set(workers->active_workers(),
967 *to(), *this, *_old_gen, *task_queues(),
968 _overflow_stacks, desired_plab_sz(), _term);
969
970 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set);
971 gch->set_par_threads(n_workers);
972 gch->rem_set()->prepare_for_younger_refs_iterate(true);
973 // It turns out that even when we're using 1 thread, doing the work in a
974 // separate thread causes wide variance in run times. We can't help this
975 // in the multi-threaded case, but we special-case n=1 here to get
976 // repeatable measurements of the 1-thread overhead of the parallel code.
977 if (n_workers > 1) {
978 GenCollectedHeap::StrongRootsScope srs(gch);
979 workers->run_task(&tsk);
980 } else {
981 GenCollectedHeap::StrongRootsScope srs(gch);
982 tsk.work(0);
983 }
984 thread_state_set.reset(0 /* Bad value in debug if not reset */,
985 promotion_failed());
986
987 // Trace and reset failed promotion info.
988 if (promotion_failed()) {
989 thread_state_set.trace_promotion_failed(gc_tracer());
990 }
991
992 // Process (weak) reference objects found during scavenge.
993 ReferenceProcessor* rp = ref_processor();
994 IsAliveClosure is_alive(this);
995 ScanWeakRefClosure scan_weak_ref(this);
996 KeepAliveClosure keep_alive(&scan_weak_ref);
997 ScanClosure scan_without_gc_barrier(this, false);
998 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
999 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
1000 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level,
1001 &scan_without_gc_barrier, &scan_with_gc_barrier);
1002 rp->setup_policy(clear_all_soft_refs);
1003 // Can the mt_degree be set later (at run_task() time would be best)?
1004 rp->set_active_mt_degree(active_workers);
1005 ReferenceProcessorStats stats;
1006 if (rp->processing_is_mt()) {
1007 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1008 stats = rp->process_discovered_references(&is_alive, &keep_alive,
1009 &evacuate_followers, &task_executor,
1010 _gc_timer, _gc_tracer.gc_id());
1011 } else {
1012 thread_state_set.flush();
1013 gch->set_par_threads(0); // 0 ==> non-parallel.
1014 gch->save_marks();
1015 stats = rp->process_discovered_references(&is_alive, &keep_alive,
1016 &evacuate_followers, NULL,
1017 _gc_timer, _gc_tracer.gc_id());
1018 }
1019 _gc_tracer.report_gc_reference_stats(stats);
1020 if (!promotion_failed()) {
1021 // Swap the survivor spaces.
1022 eden()->clear(SpaceDecorator::Mangle);
1023 from()->clear(SpaceDecorator::Mangle);
1024 if (ZapUnusedHeapArea) {
1025 // This is now done here because of the piece-meal mangling which
1026 // can check for valid mangling at intermediate points in the
1027 // collection(s). When a minor collection fails to collect
1028 // sufficient space resizing of the young generation can occur
1029 // an redistribute the spaces in the young generation. Mangle
1030 // here so that unzapped regions don't get distributed to
1031 // other spaces.
1032 to()->mangle_unused_area();
1033 }
1034 swap_spaces();
1035
1036 // A successful scavenge should restart the GC time limit count which is
1037 // for full GC's.
1038 size_policy->reset_gc_overhead_limit_count();
1039
1040 assert(to()->is_empty(), "to space should be empty now");
1041
1042 adjust_desired_tenuring_threshold();
1043 } else {
1044 handle_promotion_failed(gch, thread_state_set);
1045 }
1046 // set new iteration safe limit for the survivor spaces
1047 from()->set_concurrent_iteration_safe_limit(from()->top());
1048 to()->set_concurrent_iteration_safe_limit(to()->top());
1049
1050 if (ResizePLAB) {
1051 plab_stats()->adjust_desired_plab_sz(n_workers);
1052 }
1053
1054 if (PrintGC && !PrintGCDetails) {
1055 gch->print_heap_change(gch_prev_used);
1056 }
1057
1058 TASKQUEUE_STATS_ONLY(if (PrintTerminationStats) thread_state_set.print_termination_stats());
1059 TASKQUEUE_STATS_ONLY(if (PrintTaskqueue) thread_state_set.print_taskqueue_stats());
1060
1061 if (UseAdaptiveSizePolicy) {
1062 size_policy->minor_collection_end(gch->gc_cause());
1063 size_policy->avg_survived()->sample(from()->used());
1064 }
1065
1066 // We need to use a monotonically non-decreasing time in ms
1067 // or we will see time-warp warnings and os::javaTimeMillis()
1068 // does not guarantee monotonicity.
1069 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1070 update_time_of_last_gc(now);
1071
1072 rp->set_enqueuing_is_done(true);
1073 if (rp->processing_is_mt()) {
1074 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1075 rp->enqueue_discovered_references(&task_executor);
1076 } else {
1077 rp->enqueue_discovered_references(NULL);
1078 }
1079 rp->verify_no_references_recorded();
1080
1081 gch->trace_heap_after_gc(gc_tracer());
1082 _gc_tracer.report_tenuring_threshold(tenuring_threshold());
1083
1084 _gc_timer->register_gc_end();
1085
1086 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1087 }
1088
1089 static int sum;
1090 void ParNewGeneration::waste_some_time() {
1091 for (int i = 0; i < 100; i++) {
1092 sum += i;
1093 }
1094 }
1095
1096 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1097
1098 // Because of concurrency, there are times where an object for which
1099 // "is_forwarded()" is true contains an "interim" forwarding pointer
1100 // value. Such a value will soon be overwritten with a real value.
1101 // This method requires "obj" to have a forwarding pointer, and waits, if
1102 // necessary for a real one to be inserted, and returns it.
1103
1104 oop ParNewGeneration::real_forwardee(oop obj) {
1105 oop forward_ptr = obj->forwardee();
1106 if (forward_ptr != ClaimedForwardPtr) {
1107 return forward_ptr;
1108 } else {
1109 return real_forwardee_slow(obj);
1110 }
1111 }
1112
1113 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1114 // Spin-read if it is claimed but not yet written by another thread.
1115 oop forward_ptr = obj->forwardee();
1116 while (forward_ptr == ClaimedForwardPtr) {
1117 waste_some_time();
1118 assert(obj->is_forwarded(), "precondition");
1119 forward_ptr = obj->forwardee();
1120 }
1121 return forward_ptr;
1122 }
1123
1124 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
1125 if (m->must_be_preserved_for_promotion_failure(obj)) {
1126 // We should really have separate per-worker stacks, rather
1127 // than use locking of a common pair of stacks.
1128 MutexLocker ml(ParGCRareEvent_lock);
1129 preserve_mark(obj, m);
1130 }
1131 }
1132
1133 // Multiple GC threads may try to promote an object. If the object
1134 // is successfully promoted, a forwarding pointer will be installed in
1135 // the object in the young generation. This method claims the right
1136 // to install the forwarding pointer before it copies the object,
1137 // thus avoiding the need to undo the copy as in
1138 // copy_to_survivor_space_avoiding_with_undo.
1139
1140 oop ParNewGeneration::copy_to_survivor_space(
1141 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1142 // In the sequential version, this assert also says that the object is
1143 // not forwarded. That might not be the case here. It is the case that
1144 // the caller observed it to be not forwarded at some time in the past.
1145 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1146
1147 // The sequential code read "old->age()" below. That doesn't work here,
1148 // since the age is in the mark word, and that might be overwritten with
1149 // a forwarding pointer by a parallel thread. So we must save the mark
1150 // word in a local and then analyze it.
1151 oopDesc dummyOld;
1152 dummyOld.set_mark(m);
1153 assert(!dummyOld.is_forwarded(),
1154 "should not be called with forwarding pointer mark word.");
1155
1156 oop new_obj = NULL;
1157 oop forward_ptr;
1158
1159 // Try allocating obj in to-space (unless too old)
1160 if (dummyOld.age() < tenuring_threshold()) {
1161 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1162 if (new_obj == NULL) {
1163 set_survivor_overflow(true);
1164 }
1165 }
1166
1167 if (new_obj == NULL) {
1168 // Either to-space is full or we decided to promote
1169 // try allocating obj tenured
1170
1171 // Attempt to install a null forwarding pointer (atomically),
1172 // to claim the right to install the real forwarding pointer.
1173 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1174 if (forward_ptr != NULL) {
1175 // someone else beat us to it.
1176 return real_forwardee(old);
1177 }
1178
1179 if (!_promotion_failed) {
1180 new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1181 old, m, sz);
1182 }
1183
1184 if (new_obj == NULL) {
1185 // promotion failed, forward to self
1186 _promotion_failed = true;
1187 new_obj = old;
1188
1189 preserve_mark_if_necessary(old, m);
1190 par_scan_state->register_promotion_failure(sz);
1191 }
1192
1193 old->forward_to(new_obj);
1194 forward_ptr = NULL;
1195 } else {
1196 // Is in to-space; do copying ourselves.
1197 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1198 assert(Universe::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1199 forward_ptr = old->forward_to_atomic(new_obj);
1200 // Restore the mark word copied above.
1201 new_obj->set_mark(m);
1202 // Increment age if obj still in new generation
1203 new_obj->incr_age();
1204 par_scan_state->age_table()->add(new_obj, sz);
1205 }
1206 assert(new_obj != NULL, "just checking");
1207
1208 #ifndef PRODUCT
1209 // This code must come after the CAS test, or it will print incorrect
1210 // information.
1211 if (TraceScavenge) {
1212 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1213 is_in_reserved(new_obj) ? "copying" : "tenuring",
1214 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size());
1215 }
1216 #endif
1217
1218 if (forward_ptr == NULL) {
1219 oop obj_to_push = new_obj;
1220 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1221 // Length field used as index of next element to be scanned.
1222 // Real length can be obtained from real_forwardee()
1223 arrayOop(old)->set_length(0);
1224 obj_to_push = old;
1225 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1226 "push forwarded object");
1227 }
1228 // Push it on one of the queues of to-be-scanned objects.
1229 bool simulate_overflow = false;
1230 NOT_PRODUCT(
1231 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1232 // simulate a stack overflow
1233 simulate_overflow = true;
1234 }
1235 )
1236 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1237 // Add stats for overflow pushes.
1238 if (Verbose && PrintGCDetails) {
1239 gclog_or_tty->print("queue overflow!\n");
1240 }
1241 push_on_overflow_list(old, par_scan_state);
1242 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1243 }
1244
1245 return new_obj;
1246 }
1247
1248 // Oops. Someone beat us to it. Undo the allocation. Where did we
1249 // allocate it?
1250 if (is_in_reserved(new_obj)) {
1251 // Must be in to_space.
1252 assert(to()->is_in_reserved(new_obj), "Checking");
1253 if (forward_ptr == ClaimedForwardPtr) {
1254 // Wait to get the real forwarding pointer value.
1255 forward_ptr = real_forwardee(old);
1256 }
1257 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1258 }
1259
1260 return forward_ptr;
1261 }
1262
1263 #ifndef PRODUCT
1264 // It's OK to call this multi-threaded; the worst thing
1265 // that can happen is that we'll get a bunch of closely
1266 // spaced simulated overflows, but that's OK, in fact
1267 // probably good as it would exercise the overflow code
1268 // under contention.
1269 bool ParNewGeneration::should_simulate_overflow() {
1270 if (_overflow_counter-- <= 0) { // just being defensive
1271 _overflow_counter = ParGCWorkQueueOverflowInterval;
1272 return true;
1273 } else {
1274 return false;
1275 }
1276 }
1277 #endif
1278
1279 // In case we are using compressed oops, we need to be careful.
1280 // If the object being pushed is an object array, then its length
1281 // field keeps track of the "grey boundary" at which the next
1282 // incremental scan will be done (see ParGCArrayScanChunk).
1283 // When using compressed oops, this length field is kept in the
1284 // lower 32 bits of the erstwhile klass word and cannot be used
1285 // for the overflow chaining pointer (OCP below). As such the OCP
1286 // would itself need to be compressed into the top 32-bits in this
1287 // case. Unfortunately, see below, in the event that we have a
1288 // promotion failure, the node to be pushed on the list can be
1289 // outside of the Java heap, so the heap-based pointer compression
1290 // would not work (we would have potential aliasing between C-heap
1291 // and Java-heap pointers). For this reason, when using compressed
1292 // oops, we simply use a worker-thread-local, non-shared overflow
1293 // list in the form of a growable array, with a slightly different
1294 // overflow stack draining strategy. If/when we start using fat
1295 // stacks here, we can go back to using (fat) pointer chains
1296 // (although some performance comparisons would be useful since
1297 // single global lists have their own performance disadvantages
1298 // as we were made painfully aware not long ago, see 6786503).
1299 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1300 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1301 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1302 if (ParGCUseLocalOverflow) {
1303 // In the case of compressed oops, we use a private, not-shared
1304 // overflow stack.
1305 par_scan_state->push_on_overflow_stack(from_space_obj);
1306 } else {
1307 assert(!UseCompressedOops, "Error");
1308 // if the object has been forwarded to itself, then we cannot
1309 // use the klass pointer for the linked list. Instead we have
1310 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1311 // XXX This is horribly inefficient when a promotion failure occurs
1312 // and should be fixed. XXX FIX ME !!!
1313 #ifndef PRODUCT
1314 Atomic::inc_ptr(&_num_par_pushes);
1315 assert(_num_par_pushes > 0, "Tautology");
1316 #endif
1317 if (from_space_obj->forwardee() == from_space_obj) {
1318 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1319 listhead->forward_to(from_space_obj);
1320 from_space_obj = listhead;
1321 }
1322 oop observed_overflow_list = _overflow_list;
1323 oop cur_overflow_list;
1324 do {
1325 cur_overflow_list = observed_overflow_list;
1326 if (cur_overflow_list != BUSY) {
1327 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1328 } else {
1329 from_space_obj->set_klass_to_list_ptr(NULL);
1330 }
1331 observed_overflow_list =
1332 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1333 } while (cur_overflow_list != observed_overflow_list);
1334 }
1335 }
1336
1337 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1338 bool res;
1339
1340 if (ParGCUseLocalOverflow) {
1341 res = par_scan_state->take_from_overflow_stack();
1342 } else {
1343 assert(!UseCompressedOops, "Error");
1344 res = take_from_overflow_list_work(par_scan_state);
1345 }
1346 return res;
1347 }
1348
1349
1350 // *NOTE*: The overflow list manipulation code here and
1351 // in CMSCollector:: are very similar in shape,
1352 // except that in the CMS case we thread the objects
1353 // directly into the list via their mark word, and do
1354 // not need to deal with special cases below related
1355 // to chunking of object arrays and promotion failure
1356 // handling.
1357 // CR 6797058 has been filed to attempt consolidation of
1358 // the common code.
1359 // Because of the common code, if you make any changes in
1360 // the code below, please check the CMS version to see if
1361 // similar changes might be needed.
1362 // See CMSCollector::par_take_from_overflow_list() for
1363 // more extensive documentation comments.
1364 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1365 ObjToScanQueue* work_q = par_scan_state->work_queue();
1366 // How many to take?
1367 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1368 (size_t)ParGCDesiredObjsFromOverflowList);
1369
1370 assert(!UseCompressedOops, "Error");
1371 assert(par_scan_state->overflow_stack() == NULL, "Error");
1372 if (_overflow_list == NULL) return false;
1373
1374 // Otherwise, there was something there; try claiming the list.
1375 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1376 // Trim off a prefix of at most objsFromOverflow items
1377 Thread* tid = Thread::current();
1378 size_t spin_count = (size_t)ParallelGCThreads;
1379 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1380 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1381 // someone grabbed it before we did ...
1382 // ... we spin for a short while...
1383 os::sleep(tid, sleep_time_millis, false);
1384 if (_overflow_list == NULL) {
1385 // nothing left to take
1386 return false;
1387 } else if (_overflow_list != BUSY) {
1388 // try and grab the prefix
1389 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1390 }
1391 }
1392 if (prefix == NULL || prefix == BUSY) {
1393 // Nothing to take or waited long enough
1394 if (prefix == NULL) {
1395 // Write back the NULL in case we overwrote it with BUSY above
1396 // and it is still the same value.
1397 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1398 }
1399 return false;
1400 }
1401 assert(prefix != NULL && prefix != BUSY, "Error");
1402 size_t i = 1;
1403 oop cur = prefix;
1404 while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1405 i++; cur = cur->list_ptr_from_klass();
1406 }
1407
1408 // Reattach remaining (suffix) to overflow list
1409 if (cur->klass_or_null() == NULL) {
1410 // Write back the NULL in lieu of the BUSY we wrote
1411 // above and it is still the same value.
1412 if (_overflow_list == BUSY) {
1413 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1414 }
1415 } else {
1416 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error");
1417 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list
1418 cur->set_klass_to_list_ptr(NULL); // break off suffix
1419 // It's possible that the list is still in the empty(busy) state
1420 // we left it in a short while ago; in that case we may be
1421 // able to place back the suffix.
1422 oop observed_overflow_list = _overflow_list;
1423 oop cur_overflow_list = observed_overflow_list;
1424 bool attached = false;
1425 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1426 observed_overflow_list =
1427 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1428 if (cur_overflow_list == observed_overflow_list) {
1429 attached = true;
1430 break;
1431 } else cur_overflow_list = observed_overflow_list;
1432 }
1433 if (!attached) {
1434 // Too bad, someone else got in in between; we'll need to do a splice.
1435 // Find the last item of suffix list
1436 oop last = suffix;
1437 while (last->klass_or_null() != NULL) {
1438 last = last->list_ptr_from_klass();
1439 }
1440 // Atomically prepend suffix to current overflow list
1441 observed_overflow_list = _overflow_list;
1442 do {
1443 cur_overflow_list = observed_overflow_list;
1444 if (cur_overflow_list != BUSY) {
1445 // Do the splice ...
1446 last->set_klass_to_list_ptr(cur_overflow_list);
1447 } else { // cur_overflow_list == BUSY
1448 last->set_klass_to_list_ptr(NULL);
1449 }
1450 observed_overflow_list =
1451 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1452 } while (cur_overflow_list != observed_overflow_list);
1453 }
1454 }
1455
1456 // Push objects on prefix list onto this thread's work queue
1457 assert(prefix != NULL && prefix != BUSY, "program logic");
1458 cur = prefix;
1459 ssize_t n = 0;
1460 while (cur != NULL) {
1461 oop obj_to_push = cur->forwardee();
1462 oop next = cur->list_ptr_from_klass();
1463 cur->set_klass(obj_to_push->klass());
1464 // This may be an array object that is self-forwarded. In that case, the list pointer
1465 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1466 if (!is_in_reserved(cur)) {
1467 // This can become a scaling bottleneck when there is work queue overflow coincident
1468 // with promotion failure.
1469 oopDesc* f = cur;
1470 FREE_C_HEAP_ARRAY(oopDesc, f);
1471 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1472 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1473 obj_to_push = cur;
1474 }
1475 bool ok = work_q->push(obj_to_push);
1476 assert(ok, "Should have succeeded");
1477 cur = next;
1478 n++;
1479 }
1480 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1481 #ifndef PRODUCT
1482 assert(_num_par_pushes >= n, "Too many pops?");
1483 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1484 #endif
1485 return true;
1486 }
1487 #undef BUSY
1488
1489 void ParNewGeneration::ref_processor_init() {
1490 if (_ref_processor == NULL) {
1491 // Allocate and initialize a reference processor
1492 _ref_processor =
1493 new ReferenceProcessor(_reserved, // span
1494 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1495 (int) ParallelGCThreads, // mt processing degree
1496 refs_discovery_is_mt(), // mt discovery
1497 (int) ParallelGCThreads, // mt discovery degree
1498 refs_discovery_is_atomic(), // atomic_discovery
1499 NULL); // is_alive_non_header
1500 }
1501 }
1502
1503 const char* ParNewGeneration::name() const {
1504 return "par new generation";
1505 }