1 /*
2 * Copyright (c) 2001, 2014, 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/concurrentMarkSweepGeneration.hpp"
27 #include "gc_implementation/parNew/parNewGeneration.hpp"
28 #include "gc_implementation/parNew/parOopClosures.inline.hpp"
29 #include "gc_implementation/shared/adaptiveSizePolicy.hpp"
30 #include "gc_implementation/shared/ageTable.hpp"
31 #include "gc_implementation/shared/copyFailedInfo.hpp"
32 #include "gc_implementation/shared/gcHeapSummary.hpp"
33 #include "gc_implementation/shared/gcTimer.hpp"
34 #include "gc_implementation/shared/gcTrace.hpp"
35 #include "gc_implementation/shared/gcTraceTime.hpp"
36 #include "gc_implementation/shared/parGCAllocBuffer.inline.hpp"
37 #include "gc_implementation/shared/spaceDecorator.hpp"
38 #include "memory/defNewGeneration.inline.hpp"
39 #include "memory/genCollectedHeap.hpp"
40 #include "memory/genOopClosures.inline.hpp"
41 #include "memory/generation.hpp"
42 #include "memory/generation.inline.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 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
60
61 #ifdef _MSC_VER
62 #pragma warning( push )
63 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
64 #endif
65 ParScanThreadState::ParScanThreadState(Space* to_space_,
66 ParNewGeneration* young_gen_,
67 Generation* old_gen_,
68 int thread_num_,
69 ObjToScanQueueSet* work_queue_set_,
70 Stack<oop, mtGC>* overflow_stacks_,
71 size_t desired_plab_sz_,
72 ParallelTaskTerminator& term_)
73 : _to_space(to_space_),
74 _old_gen(old_gen_),
75 _young_gen(young_gen_),
76 _thread_num(thread_num_),
77 _work_queue(work_queue_set_->queue(thread_num_)),
78 _to_space_full(false),
79 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
80 _ageTable(false), // false ==> not the global age table, no perf data.
81 _to_space_alloc_buffer(desired_plab_sz_),
82 _to_space_closure(young_gen_, this),
83 _old_gen_closure(young_gen_, this),
84 _to_space_root_closure(young_gen_, this),
85 _old_gen_root_closure(young_gen_, this),
86 _older_gen_closure(young_gen_, this),
87 _evacuate_followers(this,
88 &_to_space_closure,
89 &_old_gen_closure,
90 &_to_space_root_closure,
91 young_gen_,
92 &_old_gen_root_closure,
93 work_queue_set_,
94 &term_),
95 _is_alive_closure(young_gen_),
96 _scan_weak_ref_closure(young_gen_, this),
97 _keep_alive_closure(&_scan_weak_ref_closure),
98 _strong_roots_time(0.0),
99 _term_time(0.0) {
100 #if TASKQUEUE_STATS
101 _term_attempts = 0;
102 _overflow_refills = 0;
103 _overflow_refill_objs = 0;
104 #endif // TASKQUEUE_STATS
105
106 _survivor_chunk_array = (ChunkArray*) old_gen()->get_data_recorder(thread_num());
107 _hash_seed = 17; // Might want to take time-based random value.
108 _start = os::elapsedTime();
109 _old_gen_closure.set_generation(old_gen_);
110 _old_gen_root_closure.set_generation(old_gen_);
111 }
112 #ifdef _MSC_VER
113 #pragma warning( pop )
114 #endif
115
116 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
117 size_t plab_word_size) {
118 ChunkArray* sca = survivor_chunk_array();
119 if (sca != NULL) {
120 // A non-null SCA implies that we want the PLAB data recorded.
121 sca->record_sample(plab_start, plab_word_size);
122 }
123 }
124
125 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
126 return new_obj->is_objArray() &&
127 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
128 new_obj != old_obj;
129 }
130
131 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
132 assert(old->is_objArray(), "must be obj array");
133 assert(old->is_forwarded(), "must be forwarded");
134 assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
135 assert(!old_gen()->is_in(old), "must be in young generation.");
136
137 objArrayOop obj = objArrayOop(old->forwardee());
138 // Process ParGCArrayScanChunk elements now
139 // and push the remainder back onto queue
140 int start = arrayOop(old)->length();
141 int end = obj->length();
142 int remainder = end - start;
143 assert(start <= end, "just checking");
144 if (remainder > 2 * ParGCArrayScanChunk) {
145 // Test above combines last partial chunk with a full chunk
146 end = start + ParGCArrayScanChunk;
147 arrayOop(old)->set_length(end);
148 // Push remainder.
149 bool ok = work_queue()->push(old);
150 assert(ok, "just popped, push must be okay");
151 } else {
152 // Restore length so that it can be used if there
153 // is a promotion failure and forwarding pointers
154 // must be removed.
155 arrayOop(old)->set_length(end);
156 }
157
158 // process our set of indices (include header in first chunk)
159 // should make sure end is even (aligned to HeapWord in case of compressed oops)
160 if ((HeapWord *)obj < young_old_boundary()) {
161 // object is in to_space
162 obj->oop_iterate_range(&_to_space_closure, start, end);
163 } else {
164 // object is in old generation
165 obj->oop_iterate_range(&_old_gen_closure, start, end);
166 }
167 }
168
169 void ParScanThreadState::trim_queues(int max_size) {
170 ObjToScanQueue* queue = work_queue();
171 do {
172 while (queue->size() > (juint)max_size) {
173 oop obj_to_scan;
174 if (queue->pop_local(obj_to_scan)) {
175 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
176 if (obj_to_scan->is_objArray() &&
177 obj_to_scan->is_forwarded() &&
178 obj_to_scan->forwardee() != obj_to_scan) {
179 scan_partial_array_and_push_remainder(obj_to_scan);
180 } else {
181 // object is in to_space
182 obj_to_scan->oop_iterate(&_to_space_closure);
183 }
184 } else {
185 // object is in old generation
186 obj_to_scan->oop_iterate(&_old_gen_closure);
187 }
188 }
189 }
190 // For the case of compressed oops, we have a private, non-shared
191 // overflow stack, so we eagerly drain it so as to more evenly
192 // distribute load early. Note: this may be good to do in
193 // general rather than delay for the final stealing phase.
194 // If applicable, we'll transfer a set of objects over to our
195 // work queue, allowing them to be stolen and draining our
196 // private overflow stack.
197 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
198 }
199
200 bool ParScanThreadState::take_from_overflow_stack() {
201 assert(ParGCUseLocalOverflow, "Else should not call");
202 assert(young_gen()->overflow_list() == NULL, "Error");
203 ObjToScanQueue* queue = work_queue();
204 Stack<oop, mtGC>* const of_stack = overflow_stack();
205 const size_t num_overflow_elems = of_stack->size();
206 const size_t space_available = queue->max_elems() - queue->size();
207 const size_t num_take_elems = MIN3(space_available / 4,
208 ParGCDesiredObjsFromOverflowList,
209 num_overflow_elems);
210 // Transfer the most recent num_take_elems from the overflow
211 // stack to our work queue.
212 for (size_t i = 0; i != num_take_elems; i++) {
213 oop cur = of_stack->pop();
214 oop obj_to_push = cur->forwardee();
215 assert(Universe::heap()->is_in_reserved(cur), "Should be in heap");
216 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
217 assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap");
218 if (should_be_partially_scanned(obj_to_push, cur)) {
219 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
220 obj_to_push = cur;
221 }
222 bool ok = queue->push(obj_to_push);
223 assert(ok, "Should have succeeded");
224 }
225 assert(young_gen()->overflow_list() == NULL, "Error");
226 return num_take_elems > 0; // was something transferred?
227 }
228
229 void ParScanThreadState::push_on_overflow_stack(oop p) {
230 assert(ParGCUseLocalOverflow, "Else should not call");
231 overflow_stack()->push(p);
232 assert(young_gen()->overflow_list() == NULL, "Error");
233 }
234
235 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
236 // If the object is small enough, try to reallocate the buffer.
237 HeapWord* obj = NULL;
238 if (!_to_space_full) {
239 ParGCAllocBuffer* const plab = to_space_alloc_buffer();
240 Space* const sp = to_space();
241 if (word_sz * 100 < ParallelGCBufferWastePct * plab->word_sz()) {
242 // Is small enough; abandon this buffer and start a new one.
243 plab->retire(false, false);
244 size_t buf_size = plab->word_sz();
245 HeapWord* buf_space = sp->par_allocate(buf_size);
246 if (buf_space == NULL) {
247 const size_t min_bytes =
248 ParGCAllocBuffer::min_size() << LogHeapWordSize;
249 size_t free_bytes = sp->free();
250 while(buf_space == NULL && free_bytes >= min_bytes) {
251 buf_size = free_bytes >> LogHeapWordSize;
252 assert(buf_size == (size_t)align_object_size(buf_size),
253 "Invariant");
254 buf_space = sp->par_allocate(buf_size);
255 free_bytes = sp->free();
256 }
257 }
258 if (buf_space != NULL) {
259 plab->set_word_size(buf_size);
260 plab->set_buf(buf_space);
261 record_survivor_plab(buf_space, buf_size);
262 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
263 // Note that we cannot compare buf_size < word_sz below
264 // because of AlignmentReserve (see ParGCAllocBuffer::allocate()).
265 assert(obj != NULL || plab->words_remaining() < word_sz,
266 "Else should have been able to allocate");
267 // It's conceivable that we may be able to use the
268 // buffer we just grabbed for subsequent small requests
269 // even if not for this one.
270 } else {
271 // We're used up.
272 _to_space_full = true;
273 }
274
275 } else {
276 // Too large; allocate the object individually.
277 obj = sp->par_allocate(word_sz);
278 }
279 }
280 return obj;
281 }
282
283 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) {
284 // Is the alloc in the current alloc buffer?
285 if (to_space_alloc_buffer()->contains(obj)) {
286 assert(to_space_alloc_buffer()->contains(obj + word_sz - 1),
287 "Should contain whole object.");
288 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
289 } else {
290 CollectedHeap::fill_with_object(obj, word_sz);
291 }
292 }
293
294 void ParScanThreadState::print_promotion_failure_size() {
295 if (_promotion_failed_info.has_failed() && PrintPromotionFailure) {
296 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ",
297 _thread_num, _promotion_failed_info.first_size());
298 }
299 }
300
301 class ParScanThreadStateSet: private ResourceArray {
302 public:
303 // Initializes states for the specified number of threads;
304 ParScanThreadStateSet(int num_threads,
305 Space& to_space,
306 ParNewGeneration& gen,
307 Generation& old_gen,
308 ObjToScanQueueSet& queue_set,
309 Stack<oop, mtGC>* overflow_stacks,
310 size_t desired_plab_sz,
311 ParallelTaskTerminator& term);
312
313 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
314
315 inline ParScanThreadState& thread_state(int i);
316
317 void trace_promotion_failed(YoungGCTracer& gc_tracer);
318 void reset(int active_workers, bool promotion_failed);
319 void flush();
320
321 #if TASKQUEUE_STATS
322 static void
323 print_termination_stats_hdr(outputStream* const st = gclog_or_tty);
324 void print_termination_stats(outputStream* const st = gclog_or_tty);
325 static void
326 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty);
327 void print_taskqueue_stats(outputStream* const st = gclog_or_tty);
328 void reset_stats();
329 #endif // TASKQUEUE_STATS
330
331 private:
332 ParallelTaskTerminator& _term;
333 ParNewGeneration& _gen;
334 Generation& _old_gen;
335 public:
336 bool is_valid(int id) const { return id < length(); }
337 ParallelTaskTerminator* terminator() { return &_term; }
338 };
339
340 ParScanThreadStateSet::ParScanThreadStateSet(int num_threads,
341 Space& to_space,
342 ParNewGeneration& gen,
343 Generation& old_gen,
344 ObjToScanQueueSet& queue_set,
345 Stack<oop, mtGC>* overflow_stacks,
346 size_t desired_plab_sz,
347 ParallelTaskTerminator& term)
348 : ResourceArray(sizeof(ParScanThreadState), num_threads),
349 _gen(gen),
350 _old_gen(old_gen),
351 _term(term) {
352 assert(num_threads > 0, "sanity check!");
353 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
354 "overflow_stack allocation mismatch");
355 // Initialize states.
356 for (int i = 0; i < num_threads; ++i) {
357 new ((ParScanThreadState*)_data + i)
358 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set,
359 overflow_stacks, desired_plab_sz, term);
360 }
361 }
362
363 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) {
364 assert(i >= 0 && i < length(), "sanity check!");
365 return ((ParScanThreadState*)_data)[i];
366 }
367
368 void ParScanThreadStateSet::trace_promotion_failed(YoungGCTracer& gc_tracer) {
369 for (int i = 0; i < length(); ++i) {
370 if (thread_state(i).promotion_failed()) {
371 gc_tracer.report_promotion_failed(thread_state(i).promotion_failed_info());
372 thread_state(i).promotion_failed_info().reset();
373 }
374 }
375 }
376
377 void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed) {
378 _term.reset_for_reuse(active_threads);
379 if (promotion_failed) {
380 for (int i = 0; i < length(); ++i) {
381 thread_state(i).print_promotion_failure_size();
382 }
383 }
384 }
385
386 #if TASKQUEUE_STATS
387 void
388 ParScanThreadState::reset_stats() {
389 taskqueue_stats().reset();
390 _term_attempts = 0;
391 _overflow_refills = 0;
392 _overflow_refill_objs = 0;
393 }
394
395 void ParScanThreadStateSet::reset_stats() {
396 for (int i = 0; i < length(); ++i) {
397 thread_state(i).reset_stats();
398 }
399 }
400
401 void ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) {
402 st->print_raw_cr("GC Termination Stats");
403 st->print_raw_cr(" elapsed --strong roots-- "
404 "-------termination-------");
405 st->print_raw_cr("thr ms ms % "
406 " ms % attempts");
407 st->print_raw_cr("--- --------- --------- ------ "
408 "--------- ------ --------");
409 }
410
411 void ParScanThreadStateSet::print_termination_stats(outputStream* const st) {
412 print_termination_stats_hdr(st);
413
414 for (int i = 0; i < length(); ++i) {
415 const ParScanThreadState & pss = thread_state(i);
416 const double elapsed_ms = pss.elapsed_time() * 1000.0;
417 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
418 const double term_ms = pss.term_time() * 1000.0;
419 st->print_cr("%3d %9.2f %9.2f %6.2f "
420 "%9.2f %6.2f " SIZE_FORMAT_W(8),
421 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
422 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
423 }
424 }
425
426 // Print stats related to work queue activity.
427 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) {
428 st->print_raw_cr("GC Task Stats");
429 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
430 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
431 }
432
433 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st) {
434 print_taskqueue_stats_hdr(st);
435
436 TaskQueueStats totals;
437 for (int i = 0; i < length(); ++i) {
438 const ParScanThreadState & pss = thread_state(i);
439 const TaskQueueStats & stats = pss.taskqueue_stats();
440 st->print("%3d ", i); stats.print(st); st->cr();
441 totals += stats;
442
443 if (pss.overflow_refills() > 0) {
444 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
445 SIZE_FORMAT_W(10) " overflow objects",
446 pss.overflow_refills(), pss.overflow_refill_objs());
447 }
448 }
449 st->print("tot "); totals.print(st); st->cr();
450
451 DEBUG_ONLY(totals.verify());
452 }
453 #endif // TASKQUEUE_STATS
454
455 void ParScanThreadStateSet::flush() {
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()->
466 flush_stats_and_retire(_gen.plab_stats(),
467 true /* end_of_gc */,
468 false /* retain */);
469
470 // Every thread has its own age table. We need to merge
471 // them all into one.
472 ageTable *local_table = par_scan_state.age_table();
473 _gen.age_table()->merge(local_table);
474
475 // Inform old gen that we're done.
476 _old_gen.par_promote_alloc_done(i);
477 _old_gen.par_oop_since_save_marks_iterate_done(i);
478 }
479
480 if (UseConcMarkSweepGC && ParallelGCThreads > 0) {
481 // We need to call this even when ResizeOldPLAB is disabled
482 // so as to avoid breaking some asserts. While we may be able
483 // to avoid this by reorganizing the code a bit, I am loathe
484 // to do that unless we find cases where ergo leads to bad
485 // performance.
486 CFLS_LAB::compute_desired_plab_size();
487 }
488 }
489
490 ParScanClosure::ParScanClosure(ParNewGeneration* g,
491 ParScanThreadState* par_scan_state)
492 : OopsInKlassOrGenClosure(g),
493 _par_scan_state(par_scan_state),
494 _g(g) {
495 _boundary = _g->reserved().end();
496 }
497
498 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
499 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
500
501 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
502 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
503
504 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
505 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
506
507 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
508 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
509
510 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
511 ParScanThreadState* par_scan_state)
512 : ScanWeakRefClosure(g),
513 _par_scan_state(par_scan_state) {
514 }
515
516 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
517 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
518
519 #ifdef WIN32
520 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
521 #endif
522
523 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(ParScanThreadState* par_scan_state,
524 ParScanWithoutBarrierClosure* to_space_closure,
525 ParScanWithBarrierClosure* old_gen_closure,
526 ParRootScanWithoutBarrierClosure* to_space_root_closure,
527 ParNewGeneration* par_gen,
528 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure,
529 ObjToScanQueueSet* task_queues,
530 ParallelTaskTerminator* terminator)
531 : _par_scan_state(par_scan_state),
532 _to_space_closure(to_space_closure),
533 _old_gen_closure(old_gen_closure),
534 _to_space_root_closure(to_space_root_closure),
535 _old_gen_root_closure(old_gen_root_closure),
536 _par_gen(par_gen),
537 _task_queues(task_queues),
538 _terminator(terminator) {
539 }
540
541 void ParEvacuateFollowersClosure::do_void() {
542 ObjToScanQueue* work_q = par_scan_state()->work_queue();
543
544 while (true) {
545 // Scan to-space and old-gen objs until we run out of both.
546 oop obj_to_scan;
547 par_scan_state()->trim_queues(0);
548
549 // We have no local work, attempt to steal from other threads.
550
551 // attempt to steal work from promoted.
552 if (task_queues()->steal(par_scan_state()->thread_num(),
553 par_scan_state()->hash_seed(),
554 obj_to_scan)) {
555 bool res = work_q->push(obj_to_scan);
556 assert(res, "Empty queue should have room for a push.");
557
558 // if successful, goto Start.
559 continue;
560
561 // try global overflow list.
562 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
563 continue;
564 }
565
566 // Otherwise, offer termination.
567 par_scan_state()->start_term_time();
568 if (terminator()->offer_termination()) break;
569 par_scan_state()->end_term_time();
570 }
571 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
572 "Broken overflow list?");
573 // Finish the last termination pause.
574 par_scan_state()->end_term_time();
575 }
576
577 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen,
578 Generation* old_gen,
579 HeapWord* young_old_boundary,
580 ParScanThreadStateSet* state_set)
581 : AbstractGangTask("ParNewGeneration collection"),
582 _young_gen(young_gen), _old_gen(old_gen),
583 _young_old_boundary(young_old_boundary),
584 _state_set(state_set) {
585 }
586
587 // Reset the terminator for the given number of
588 // active threads.
589 void ParNewGenTask::set_for_termination(int active_workers) {
590 _state_set->reset(active_workers, _young_gen->promotion_failed());
591 // Should the heap be passed in? There's only 1 for now so
592 // grab it instead.
593 GenCollectedHeap* gch = GenCollectedHeap::heap();
594 gch->set_n_termination(active_workers);
595 }
596
597 void ParNewGenTask::work(uint worker_id) {
598 GenCollectedHeap* gch = GenCollectedHeap::heap();
599 // Since this is being done in a separate thread, need new resource
600 // and handle marks.
601 ResourceMark rm;
602 HandleMark hm;
603
604 Generation* old_gen = gch->old_gen();
605
606 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
607 assert(_state_set->is_valid(worker_id), "Should not have been called");
608
609 par_scan_state.set_young_old_boundary(_young_old_boundary);
610
611 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(),
612 gch->rem_set()->klass_rem_set());
613 CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure,
614 &par_scan_state.to_space_root_closure(),
615 false);
616
617 par_scan_state.start_strong_roots();
618 gch->gen_process_roots(Generation::Young,
619 true, // Process younger gens, if any,
620 // as strong roots.
621 false, // no scope; this is parallel code
622 SharedHeap::SO_ScavengeCodeCache,
623 GenCollectedHeap::StrongAndWeakRoots,
624 &par_scan_state.to_space_root_closure(),
625 &par_scan_state.older_gen_closure(),
626 &cld_scan_closure);
627
628 par_scan_state.end_strong_roots();
629
630 // "evacuate followers".
631 par_scan_state.evacuate_followers_closure().do_void();
632 }
633
634 #ifdef _MSC_VER
635 #pragma warning( push )
636 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
637 #endif
638 ParNewGeneration::
639 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size)
640 : DefNewGeneration(rs, initial_byte_size, "PCopy"),
641 _overflow_list(NULL),
642 _is_alive_closure(this),
643 _plab_stats(YoungPLABSize, PLABWeight) {
644 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
645 NOT_PRODUCT(_num_par_pushes = 0;)
646 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
647 guarantee(_task_queues != NULL, "task_queues allocation failure.");
648
649 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
650 ObjToScanQueue *q = new ObjToScanQueue();
651 guarantee(q != NULL, "work_queue Allocation failure.");
652 _task_queues->register_queue(i1, q);
653 }
654
655 for (uint i2 = 0; i2 < ParallelGCThreads; i2++) {
656 _task_queues->queue(i2)->initialize();
657 }
658
659 _overflow_stacks = NULL;
660 if (ParGCUseLocalOverflow) {
661 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal
662 // with ','
663 typedef Stack<oop, mtGC> GCOopStack;
664
665 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
666 for (size_t i = 0; i < ParallelGCThreads; ++i) {
667 new (_overflow_stacks + i) Stack<oop, mtGC>();
668 }
669 }
670
671 if (UsePerfData) {
672 EXCEPTION_MARK;
673 ResourceMark rm;
674
675 const char* cname =
676 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
677 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
678 ParallelGCThreads, CHECK);
679 }
680 }
681 #ifdef _MSC_VER
682 #pragma warning( pop )
683 #endif
684
685 // ParNewGeneration::
686 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl)
687 : DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {
688 }
689
690 template <class T>
691 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
692 #ifdef ASSERT
693 {
694 assert(!oopDesc::is_null(*p), "expected non-null ref");
695 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
696 // We never expect to see a null reference being processed
697 // as a weak reference.
698 assert(obj->is_oop(), "expected an oop while scanning weak refs");
699 }
700 #endif // ASSERT
701
702 _par_cl->do_oop_nv(p);
703
704 if (Universe::heap()->is_in_reserved(p)) {
705 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
706 _rs->write_ref_field_gc_par(p, obj);
707 }
708 }
709
710 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
711 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
712
713 // ParNewGeneration::
714 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl)
715 : DefNewGeneration::KeepAliveClosure(cl) {
716 }
717
718 template <class T>
719 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
720 #ifdef ASSERT
721 {
722 assert(!oopDesc::is_null(*p), "expected non-null ref");
723 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
724 // We never expect to see a null reference being processed
725 // as a weak reference.
726 assert(obj->is_oop(), "expected an oop while scanning weak refs");
727 }
728 #endif // ASSERT
729
730 _cl->do_oop_nv(p);
731
732 if (Universe::heap()->is_in_reserved(p)) {
733 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
734 _rs->write_ref_field_gc_par(p, obj);
735 }
736 }
737
738 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
739 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
740
741 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
742 T heap_oop = oopDesc::load_heap_oop(p);
743 if (!oopDesc::is_null(heap_oop)) {
744 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
745 if ((HeapWord*)obj < _boundary) {
746 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
747 oop new_obj = obj->is_forwarded()
748 ? obj->forwardee()
749 : _g->DefNewGeneration::copy_to_survivor_space(obj);
750 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
751 }
752 if (_gc_barrier) {
753 // If p points to a younger generation, mark the card.
754 if ((HeapWord*)obj < _gen_boundary) {
755 _rs->write_ref_field_gc_par(p, obj);
756 }
757 }
758 }
759 }
760
761 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
762 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
763
764 class ParNewRefProcTaskProxy: public AbstractGangTask {
765 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
766 public:
767 ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen,
768 Generation& old_gen,
769 HeapWord* young_old_boundary,
770 ParScanThreadStateSet& state_set);
771
772 private:
773 virtual void work(uint worker_id);
774 virtual void set_for_termination(int active_workers) {
775 _state_set.terminator()->reset_for_reuse(active_workers);
776 }
777 private:
778 ParNewGeneration& _young_gen;
779 ProcessTask& _task;
780 Generation& _old_gen;
781 HeapWord* _young_old_boundary;
782 ParScanThreadStateSet& _state_set;
783 };
784
785 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
786 ParNewGeneration& young_gen,
787 Generation& old_gen,
788 HeapWord* young_old_boundary,
789 ParScanThreadStateSet& state_set)
790 : AbstractGangTask("ParNewGeneration parallel reference processing"),
791 _young_gen(young_gen),
792 _task(task),
793 _old_gen(old_gen),
794 _young_old_boundary(young_old_boundary),
795 _state_set(state_set) {
796 }
797
798 void ParNewRefProcTaskProxy::work(uint worker_id) {
799 ResourceMark rm;
800 HandleMark hm;
801 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
802 par_scan_state.set_young_old_boundary(_young_old_boundary);
803 _task.work(worker_id, par_scan_state.is_alive_closure(),
804 par_scan_state.keep_alive_closure(),
805 par_scan_state.evacuate_followers_closure());
806 }
807
808 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
809 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
810 EnqueueTask& _task;
811
812 public:
813 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
814 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
815 _task(task) {
816 }
817
818 virtual void work(uint worker_id) {
819 _task.work(worker_id);
820 }
821 };
822
823 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) {
824 GenCollectedHeap* gch = GenCollectedHeap::heap();
825 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
826 "not a generational heap");
827 FlexibleWorkGang* workers = gch->workers();
828 assert(workers != NULL, "Need parallel worker threads.");
829 _state_set.reset(workers->active_workers(), _generation.promotion_failed());
830 ParNewRefProcTaskProxy rp_task(task, _generation, *(gch->old_gen()),
831 _generation.reserved().end(), _state_set);
832 workers->run_task(&rp_task);
833 _state_set.reset(0 /* bad value in debug if not reset */,
834 _generation.promotion_failed());
835 }
836
837 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) {
838 GenCollectedHeap* gch = GenCollectedHeap::heap();
839 FlexibleWorkGang* workers = gch->workers();
840 assert(workers != NULL, "Need parallel worker threads.");
841 ParNewRefEnqueueTaskProxy enq_task(task);
842 workers->run_task(&enq_task);
843 }
844
845 void ParNewRefProcTaskExecutor::set_single_threaded_mode() {
846 _state_set.flush();
847 GenCollectedHeap* gch = GenCollectedHeap::heap();
848 gch->set_par_threads(0); // 0 ==> non-parallel.
849 gch->save_marks();
850 }
851
852 ScanClosureWithParBarrier::ScanClosureWithParBarrier(ParNewGeneration* g,
853 bool gc_barrier)
854 : ScanClosure(g, gc_barrier) {
855 }
856
857 EvacuateFollowersClosureGeneral::
858 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch,
859 OopsInGenClosure* cur,
860 OopsInGenClosure* older)
861 : _gch(gch),
862 _scan_cur_or_nonheap(cur),
863 _scan_older(older) {
864 }
865
866 void EvacuateFollowersClosureGeneral::do_void() {
867 do {
868 // Beware: this call will lead to closure applications via virtual
869 // calls.
870 _gch->oop_since_save_marks_iterate(Generation::Young,
871 _scan_cur_or_nonheap,
872 _scan_older);
873 } while (!_gch->no_allocs_since_save_marks(true /* include_young */));
874 }
875
876
877 // A Generation that does parallel young-gen collection.
878
879 bool ParNewGeneration::_avoid_promotion_undo = false;
880
881 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch,
882 ParScanThreadStateSet& thread_state_set,
883 ParNewTracer& gc_tracer) {
884 assert(_promo_failure_scan_stack.is_empty(), "post condition");
885 _promo_failure_scan_stack.clear(true); // Clear cached segments.
886
887 remove_forwarding_pointers();
888 if (PrintGCDetails) {
889 gclog_or_tty->print(" (promotion failed)");
890 }
891 // All the spaces are in play for mark-sweep.
892 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
893 from()->set_next_compaction_space(to());
894 gch->set_incremental_collection_failed();
895 // Inform the next generation that a promotion failure occurred.
896 _old_gen->promotion_failure_occurred();
897
898 // Trace promotion failure in the parallel GC threads
899 thread_state_set.trace_promotion_failed(gc_tracer);
900 // Single threaded code may have reported promotion failure to the global state
901 if (_promotion_failed_info.has_failed()) {
902 gc_tracer.report_promotion_failed(_promotion_failed_info);
903 }
904 // Reset the PromotionFailureALot counters.
905 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();)
906 }
907
908 void ParNewGeneration::collect(bool full,
909 bool clear_all_soft_refs,
910 size_t size,
911 bool is_tlab) {
912 assert(full || size > 0, "otherwise we don't want to collect");
913
914 GenCollectedHeap* gch = GenCollectedHeap::heap();
915
916 _gc_timer->register_gc_start();
917
918 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
919 "not a CMS generational heap");
920 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
921 FlexibleWorkGang* workers = gch->workers();
922 assert(workers != NULL, "Need workgang for parallel work");
923 int active_workers =
924 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
925 workers->active_workers(),
926 Threads::number_of_non_daemon_threads());
927 workers->set_active_workers(active_workers);
928 _old_gen = gch->old_gen();
929 // Do we have to avoid promotion_undo?
930 if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) {
931 set_avoid_promotion_undo(true);
932 }
933
934 // If the next generation is too full to accommodate worst-case promotion
935 // from this generation, pass on collection; let the next generation
936 // do it.
937 if (!collection_attempt_is_safe()) {
938 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
939 return;
940 }
941 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
942
943 ParNewTracer gc_tracer;
944 gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
945 gch->trace_heap_before_gc(&gc_tracer);
946
947 init_assuming_no_promotion_failure();
948
949 if (UseAdaptiveSizePolicy) {
950 set_survivor_overflow(false);
951 size_policy->minor_collection_begin();
952 }
953
954 GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL, gc_tracer.gc_id());
955 // Capture heap used before collection (for printing).
956 size_t gch_prev_used = gch->used();
957
958 SpecializationStats::clear();
959
960 age_table()->clear();
961 to()->clear(SpaceDecorator::Mangle);
962
963 gch->save_marks();
964 assert(workers != NULL, "Need parallel worker threads.");
965 int n_workers = active_workers;
966
967 // Set the correct parallelism (number of queues) in the reference processor
968 ref_processor()->set_active_mt_degree(n_workers);
969
970 // Always set the terminator for the active number of workers
971 // because only those workers go through the termination protocol.
972 ParallelTaskTerminator _term(n_workers, task_queues());
973 ParScanThreadStateSet thread_state_set(workers->active_workers(),
974 *to(), *this, *_old_gen, *task_queues(),
975 _overflow_stacks, desired_plab_sz(), _term);
976
977 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set);
978 gch->set_par_threads(n_workers);
979 gch->rem_set()->prepare_for_younger_refs_iterate(true);
980 // It turns out that even when we're using 1 thread, doing the work in a
981 // separate thread causes wide variance in run times. We can't help this
982 // in the multi-threaded case, but we special-case n=1 here to get
983 // repeatable measurements of the 1-thread overhead of the parallel code.
984 if (n_workers > 1) {
985 GenCollectedHeap::StrongRootsScope srs(gch);
986 workers->run_task(&tsk);
987 } else {
988 GenCollectedHeap::StrongRootsScope srs(gch);
989 tsk.work(0);
990 }
991 thread_state_set.reset(0 /* Bad value in debug if not reset */,
992 promotion_failed());
993
994 // Process (weak) reference objects found during scavenge.
995 ReferenceProcessor* rp = ref_processor();
996 IsAliveClosure is_alive(this);
997 ScanWeakRefClosure scan_weak_ref(this);
998 KeepAliveClosure keep_alive(&scan_weak_ref);
999 ScanClosure scan_without_gc_barrier(this, false);
1000 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
1001 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
1002 EvacuateFollowersClosureGeneral evacuate_followers(gch,
1003 &scan_without_gc_barrier, &scan_with_gc_barrier);
1004 rp->setup_policy(clear_all_soft_refs);
1005 // Can the mt_degree be set later (at run_task() time would be best)?
1006 rp->set_active_mt_degree(active_workers);
1007 ReferenceProcessorStats stats;
1008 if (rp->processing_is_mt()) {
1009 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1010 stats = rp->process_discovered_references(&is_alive, &keep_alive,
1011 &evacuate_followers, &task_executor,
1012 _gc_timer, gc_tracer.gc_id());
1013 } else {
1014 thread_state_set.flush();
1015 gch->set_par_threads(0); // 0 ==> non-parallel.
1016 gch->save_marks();
1017 stats = rp->process_discovered_references(&is_alive, &keep_alive,
1018 &evacuate_followers, NULL,
1019 _gc_timer, gc_tracer.gc_id());
1020 }
1021 gc_tracer.report_gc_reference_stats(stats);
1022 if (!promotion_failed()) {
1023 // Swap the survivor spaces.
1024 eden()->clear(SpaceDecorator::Mangle);
1025 from()->clear(SpaceDecorator::Mangle);
1026 if (ZapUnusedHeapArea) {
1027 // This is now done here because of the piece-meal mangling which
1028 // can check for valid mangling at intermediate points in the
1029 // collection(s). When a minor collection fails to collect
1030 // sufficient space resizing of the young generation can occur
1031 // an redistribute the spaces in the young generation. Mangle
1032 // here so that unzapped regions don't get distributed to
1033 // other spaces.
1034 to()->mangle_unused_area();
1035 }
1036 swap_spaces();
1037
1038 // A successful scavenge should restart the GC time limit count which is
1039 // for full GC's.
1040 size_policy->reset_gc_overhead_limit_count();
1041
1042 assert(to()->is_empty(), "to space should be empty now");
1043
1044 adjust_desired_tenuring_threshold();
1045 } else {
1046 handle_promotion_failed(gch, thread_state_set, gc_tracer);
1047 }
1048 // set new iteration safe limit for the survivor spaces
1049 from()->set_concurrent_iteration_safe_limit(from()->top());
1050 to()->set_concurrent_iteration_safe_limit(to()->top());
1051
1052 if (ResizePLAB) {
1053 plab_stats()->adjust_desired_plab_sz(n_workers);
1054 }
1055
1056 if (PrintGC && !PrintGCDetails) {
1057 gch->print_heap_change(gch_prev_used);
1058 }
1059
1060 TASKQUEUE_STATS_ONLY(if (PrintTerminationStats) thread_state_set.print_termination_stats());
1061 TASKQUEUE_STATS_ONLY(if (PrintTaskqueue) thread_state_set.print_taskqueue_stats());
1062
1063 if (UseAdaptiveSizePolicy) {
1064 size_policy->minor_collection_end(gch->gc_cause());
1065 size_policy->avg_survived()->sample(from()->used());
1066 }
1067
1068 // We need to use a monotonically non-decreasing time in ms
1069 // or we will see time-warp warnings and os::javaTimeMillis()
1070 // does not guarantee monotonicity.
1071 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1072 update_time_of_last_gc(now);
1073
1074 SpecializationStats::print();
1075
1076 rp->set_enqueuing_is_done(true);
1077 if (rp->processing_is_mt()) {
1078 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1079 rp->enqueue_discovered_references(&task_executor);
1080 } else {
1081 rp->enqueue_discovered_references(NULL);
1082 }
1083 rp->verify_no_references_recorded();
1084
1085 gch->trace_heap_after_gc(&gc_tracer);
1086 gc_tracer.report_tenuring_threshold(tenuring_threshold());
1087
1088 _gc_timer->register_gc_end();
1089
1090 gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1091 }
1092
1093 static int sum;
1094 void ParNewGeneration::waste_some_time() {
1095 for (int i = 0; i < 100; i++) {
1096 sum += i;
1097 }
1098 }
1099
1100 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1101
1102 // Because of concurrency, there are times where an object for which
1103 // "is_forwarded()" is true contains an "interim" forwarding pointer
1104 // value. Such a value will soon be overwritten with a real value.
1105 // This method requires "obj" to have a forwarding pointer, and waits, if
1106 // necessary for a real one to be inserted, and returns it.
1107
1108 oop ParNewGeneration::real_forwardee(oop obj) {
1109 oop forward_ptr = obj->forwardee();
1110 if (forward_ptr != ClaimedForwardPtr) {
1111 return forward_ptr;
1112 } else {
1113 return real_forwardee_slow(obj);
1114 }
1115 }
1116
1117 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1118 // Spin-read if it is claimed but not yet written by another thread.
1119 oop forward_ptr = obj->forwardee();
1120 while (forward_ptr == ClaimedForwardPtr) {
1121 waste_some_time();
1122 assert(obj->is_forwarded(), "precondition");
1123 forward_ptr = obj->forwardee();
1124 }
1125 return forward_ptr;
1126 }
1127
1128 #ifdef ASSERT
1129 bool ParNewGeneration::is_legal_forward_ptr(oop p) {
1130 return
1131 (_avoid_promotion_undo && p == ClaimedForwardPtr)
1132 || Universe::heap()->is_in_reserved(p);
1133 }
1134 #endif
1135
1136 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
1137 if (m->must_be_preserved_for_promotion_failure(obj)) {
1138 // We should really have separate per-worker stacks, rather
1139 // than use locking of a common pair of stacks.
1140 MutexLocker ml(ParGCRareEvent_lock);
1141 preserve_mark(obj, m);
1142 }
1143 }
1144
1145 // Multiple GC threads may try to promote an object. If the object
1146 // is successfully promoted, a forwarding pointer will be installed in
1147 // the object in the young generation. This method claims the right
1148 // to install the forwarding pointer before it copies the object,
1149 // thus avoiding the need to undo the copy as in
1150 // copy_to_survivor_space_avoiding_with_undo.
1151
1152 oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo(
1153 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1154 // In the sequential version, this assert also says that the object is
1155 // not forwarded. That might not be the case here. It is the case that
1156 // the caller observed it to be not forwarded at some time in the past.
1157 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1158
1159 // The sequential code read "old->age()" below. That doesn't work here,
1160 // since the age is in the mark word, and that might be overwritten with
1161 // a forwarding pointer by a parallel thread. So we must save the mark
1162 // word in a local and then analyze it.
1163 oopDesc dummyOld;
1164 dummyOld.set_mark(m);
1165 assert(!dummyOld.is_forwarded(),
1166 "should not be called with forwarding pointer mark word.");
1167
1168 oop new_obj = NULL;
1169 oop forward_ptr;
1170
1171 // Try allocating obj in to-space (unless too old)
1172 if (dummyOld.age() < tenuring_threshold()) {
1173 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1174 if (new_obj == NULL) {
1175 set_survivor_overflow(true);
1176 }
1177 }
1178
1179 if (new_obj == NULL) {
1180 // Either to-space is full or we decided to promote
1181 // try allocating obj tenured
1182
1183 // Attempt to install a null forwarding pointer (atomically),
1184 // to claim the right to install the real forwarding pointer.
1185 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1186 if (forward_ptr != NULL) {
1187 // someone else beat us to it.
1188 return real_forwardee(old);
1189 }
1190
1191 new_obj = _old_gen->par_promote(par_scan_state->thread_num(), old, m, sz);
1192
1193 if (new_obj == NULL) {
1194 // promotion failed, forward to self
1195 _promotion_failed = true;
1196 new_obj = old;
1197
1198 preserve_mark_if_necessary(old, m);
1199 par_scan_state->register_promotion_failure(sz);
1200 }
1201
1202 old->forward_to(new_obj);
1203 forward_ptr = NULL;
1204 } else {
1205 // Is in to-space; do copying ourselves.
1206 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1207 forward_ptr = old->forward_to_atomic(new_obj);
1208 // Restore the mark word copied above.
1209 new_obj->set_mark(m);
1210 // Increment age if obj still in new generation
1211 new_obj->incr_age();
1212 par_scan_state->age_table()->add(new_obj, sz);
1213 }
1214 assert(new_obj != NULL, "just checking");
1215
1216 #ifndef PRODUCT
1217 // This code must come after the CAS test, or it will print incorrect
1218 // information.
1219 if (TraceScavenge) {
1220 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1221 is_in_reserved(new_obj) ? "copying" : "tenuring",
1222 new_obj->klass()->internal_name(),
1223 (void *)old,
1224 (void *)new_obj,
1225 new_obj->size());
1226 }
1227 #endif
1228
1229 if (forward_ptr == NULL) {
1230 oop obj_to_push = new_obj;
1231 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1232 // Length field used as index of next element to be scanned.
1233 // Real length can be obtained from real_forwardee()
1234 arrayOop(old)->set_length(0);
1235 obj_to_push = old;
1236 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1237 "push forwarded object");
1238 }
1239 // Push it on one of the queues of to-be-scanned objects.
1240 bool simulate_overflow = false;
1241 NOT_PRODUCT(
1242 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1243 // simulate a stack overflow
1244 simulate_overflow = true;
1245 }
1246 )
1247 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1248 // Add stats for overflow pushes.
1249 if (Verbose && PrintGCDetails) {
1250 gclog_or_tty->print("queue overflow!\n");
1251 }
1252 push_on_overflow_list(old, par_scan_state);
1253 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1254 }
1255
1256 return new_obj;
1257 }
1258
1259 // Oops. Someone beat us to it. Undo the allocation. Where did we
1260 // allocate it?
1261 if (is_in_reserved(new_obj)) {
1262 // Must be in to_space.
1263 assert(to()->is_in_reserved(new_obj), "Checking");
1264 if (forward_ptr == ClaimedForwardPtr) {
1265 // Wait to get the real forwarding pointer value.
1266 forward_ptr = real_forwardee(old);
1267 }
1268 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1269 }
1270
1271 return forward_ptr;
1272 }
1273
1274
1275 // Multiple GC threads may try to promote the same object. If two
1276 // or more GC threads copy the object, only one wins the race to install
1277 // the forwarding pointer. The other threads have to undo their copy.
1278
1279 oop ParNewGeneration::copy_to_survivor_space_with_undo(
1280 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1281
1282 // In the sequential version, this assert also says that the object is
1283 // not forwarded. That might not be the case here. It is the case that
1284 // the caller observed it to be not forwarded at some time in the past.
1285 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1286
1287 // The sequential code read "old->age()" below. That doesn't work here,
1288 // since the age is in the mark word, and that might be overwritten with
1289 // a forwarding pointer by a parallel thread. So we must save the mark
1290 // word here, install it in a local oopDesc, and then analyze it.
1291 oopDesc dummyOld;
1292 dummyOld.set_mark(m);
1293 assert(!dummyOld.is_forwarded(),
1294 "should not be called with forwarding pointer mark word.");
1295
1296 bool failed_to_promote = false;
1297 oop new_obj = NULL;
1298 oop forward_ptr;
1299
1300 // Try allocating obj in to-space (unless too old)
1301 if (dummyOld.age() < tenuring_threshold()) {
1302 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1303 if (new_obj == NULL) {
1304 set_survivor_overflow(true);
1305 }
1306 }
1307
1308 if (new_obj == NULL) {
1309 // Either to-space is full or we decided to promote
1310 // try allocating obj tenured
1311 new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1312 old, m, sz);
1313
1314 if (new_obj == NULL) {
1315 // promotion failed, forward to self
1316 forward_ptr = old->forward_to_atomic(old);
1317 new_obj = old;
1318
1319 if (forward_ptr != NULL) {
1320 return forward_ptr; // someone else succeeded
1321 }
1322
1323 _promotion_failed = true;
1324 failed_to_promote = true;
1325
1326 preserve_mark_if_necessary(old, m);
1327 par_scan_state->register_promotion_failure(sz);
1328 }
1329 } else {
1330 // Is in to-space; do copying ourselves.
1331 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1332 // Restore the mark word copied above.
1333 new_obj->set_mark(m);
1334 // Increment age if new_obj still in new generation
1335 new_obj->incr_age();
1336 par_scan_state->age_table()->add(new_obj, sz);
1337 }
1338 assert(new_obj != NULL, "just checking");
1339
1340 #ifndef PRODUCT
1341 // This code must come after the CAS test, or it will print incorrect
1342 // information.
1343 if (TraceScavenge) {
1344 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1345 is_in_reserved(new_obj) ? "copying" : "tenuring",
1346 new_obj->klass()->internal_name(),
1347 (void*)old,
1348 (void*)new_obj,
1349 new_obj->size());
1350 }
1351 #endif
1352
1353 // Now attempt to install the forwarding pointer (atomically).
1354 // We have to copy the mark word before overwriting with forwarding
1355 // ptr, so we can restore it below in the copy.
1356 if (!failed_to_promote) {
1357 forward_ptr = old->forward_to_atomic(new_obj);
1358 }
1359
1360 if (forward_ptr == NULL) {
1361 oop obj_to_push = new_obj;
1362 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1363 // Length field used as index of next element to be scanned.
1364 // Real length can be obtained from real_forwardee()
1365 arrayOop(old)->set_length(0);
1366 obj_to_push = old;
1367 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1368 "push forwarded object");
1369 }
1370 // Push it on one of the queues of to-be-scanned objects.
1371 bool simulate_overflow = false;
1372 NOT_PRODUCT(
1373 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1374 // simulate a stack overflow
1375 simulate_overflow = true;
1376 }
1377 )
1378 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1379 // Add stats for overflow pushes.
1380 push_on_overflow_list(old, par_scan_state);
1381 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1382 }
1383
1384 return new_obj;
1385 }
1386
1387 // Oops. Someone beat us to it. Undo the allocation. Where did we
1388 // allocate it?
1389 if (is_in_reserved(new_obj)) {
1390 // Must be in to_space.
1391 assert(to()->is_in_reserved(new_obj), "Checking");
1392 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1393 } else {
1394 assert(!_avoid_promotion_undo, "Should not be here if avoiding.");
1395 _old_gen->par_promote_alloc_undo(par_scan_state->thread_num(),
1396 (HeapWord*)new_obj, sz);
1397 }
1398
1399 return forward_ptr;
1400 }
1401
1402 #ifndef PRODUCT
1403 // It's OK to call this multi-threaded; the worst thing
1404 // that can happen is that we'll get a bunch of closely
1405 // spaced simulated overflows, but that's OK, in fact
1406 // probably good as it would exercise the overflow code
1407 // under contention.
1408 bool ParNewGeneration::should_simulate_overflow() {
1409 if (_overflow_counter-- <= 0) { // just being defensive
1410 _overflow_counter = ParGCWorkQueueOverflowInterval;
1411 return true;
1412 } else {
1413 return false;
1414 }
1415 }
1416 #endif
1417
1418 // In case we are using compressed oops, we need to be careful.
1419 // If the object being pushed is an object array, then its length
1420 // field keeps track of the "grey boundary" at which the next
1421 // incremental scan will be done (see ParGCArrayScanChunk).
1422 // When using compressed oops, this length field is kept in the
1423 // lower 32 bits of the erstwhile klass word and cannot be used
1424 // for the overflow chaining pointer (OCP below). As such the OCP
1425 // would itself need to be compressed into the top 32-bits in this
1426 // case. Unfortunately, see below, in the event that we have a
1427 // promotion failure, the node to be pushed on the list can be
1428 // outside of the Java heap, so the heap-based pointer compression
1429 // would not work (we would have potential aliasing between C-heap
1430 // and Java-heap pointers). For this reason, when using compressed
1431 // oops, we simply use a worker-thread-local, non-shared overflow
1432 // list in the form of a growable array, with a slightly different
1433 // overflow stack draining strategy. If/when we start using fat
1434 // stacks here, we can go back to using (fat) pointer chains
1435 // (although some performance comparisons would be useful since
1436 // single global lists have their own performance disadvantages
1437 // as we were made painfully aware not long ago, see 6786503).
1438 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1439 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1440 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1441 if (ParGCUseLocalOverflow) {
1442 // In the case of compressed oops, we use a private, not-shared
1443 // overflow stack.
1444 par_scan_state->push_on_overflow_stack(from_space_obj);
1445 } else {
1446 assert(!UseCompressedOops, "Error");
1447 // if the object has been forwarded to itself, then we cannot
1448 // use the klass pointer for the linked list. Instead we have
1449 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1450 // XXX This is horribly inefficient when a promotion failure occurs
1451 // and should be fixed. XXX FIX ME !!!
1452 #ifndef PRODUCT
1453 Atomic::inc_ptr(&_num_par_pushes);
1454 assert(_num_par_pushes > 0, "Tautology");
1455 #endif
1456 if (from_space_obj->forwardee() == from_space_obj) {
1457 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1458 listhead->forward_to(from_space_obj);
1459 from_space_obj = listhead;
1460 }
1461 oop observed_overflow_list = _overflow_list;
1462 oop cur_overflow_list;
1463 do {
1464 cur_overflow_list = observed_overflow_list;
1465 if (cur_overflow_list != BUSY) {
1466 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1467 } else {
1468 from_space_obj->set_klass_to_list_ptr(NULL);
1469 }
1470 observed_overflow_list =
1471 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1472 } while (cur_overflow_list != observed_overflow_list);
1473 }
1474 }
1475
1476 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1477 bool res;
1478
1479 if (ParGCUseLocalOverflow) {
1480 res = par_scan_state->take_from_overflow_stack();
1481 } else {
1482 assert(!UseCompressedOops, "Error");
1483 res = take_from_overflow_list_work(par_scan_state);
1484 }
1485 return res;
1486 }
1487
1488
1489 // *NOTE*: The overflow list manipulation code here and
1490 // in CMSCollector:: are very similar in shape,
1491 // except that in the CMS case we thread the objects
1492 // directly into the list via their mark word, and do
1493 // not need to deal with special cases below related
1494 // to chunking of object arrays and promotion failure
1495 // handling.
1496 // CR 6797058 has been filed to attempt consolidation of
1497 // the common code.
1498 // Because of the common code, if you make any changes in
1499 // the code below, please check the CMS version to see if
1500 // similar changes might be needed.
1501 // See CMSCollector::par_take_from_overflow_list() for
1502 // more extensive documentation comments.
1503 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1504 ObjToScanQueue* work_q = par_scan_state->work_queue();
1505 // How many to take?
1506 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1507 (size_t)ParGCDesiredObjsFromOverflowList);
1508
1509 assert(!UseCompressedOops, "Error");
1510 assert(par_scan_state->overflow_stack() == NULL, "Error");
1511 if (_overflow_list == NULL) {
1512 return false;
1513 }
1514
1515 // Otherwise, there was something there; try claiming the list.
1516 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1517 // Trim off a prefix of at most objsFromOverflow items
1518 Thread* tid = Thread::current();
1519 size_t spin_count = (size_t)ParallelGCThreads;
1520 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1521 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1522 // someone grabbed it before we did ...
1523 // ... we spin for a short while...
1524 os::sleep(tid, sleep_time_millis, false);
1525 if (_overflow_list == NULL) {
1526 // nothing left to take
1527 return false;
1528 } else if (_overflow_list != BUSY) {
1529 // try and grab the prefix
1530 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1531 }
1532 }
1533 if (prefix == NULL || prefix == BUSY) {
1534 // Nothing to take or waited long enough
1535 if (prefix == NULL) {
1536 // Write back the NULL in case we overwrote it with BUSY above
1537 // and it is still the same value.
1538 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1539 }
1540 return false;
1541 }
1542 assert(prefix != NULL && prefix != BUSY, "Error");
1543 size_t i = 1;
1544 oop cur = prefix;
1545 while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1546 i++; cur = cur->list_ptr_from_klass();
1547 }
1548
1549 // Reattach remaining (suffix) to overflow list
1550 if (cur->klass_or_null() == NULL) {
1551 // Write back the NULL in lieu of the BUSY we wrote
1552 // above and it is still the same value.
1553 if (_overflow_list == BUSY) {
1554 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1555 }
1556 } else {
1557 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error");
1558 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list
1559 cur->set_klass_to_list_ptr(NULL); // break off suffix
1560 // It's possible that the list is still in the empty(busy) state
1561 // we left it in a short while ago; in that case we may be
1562 // able to place back the suffix.
1563 oop observed_overflow_list = _overflow_list;
1564 oop cur_overflow_list = observed_overflow_list;
1565 bool attached = false;
1566 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1567 observed_overflow_list =
1568 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1569 if (cur_overflow_list == observed_overflow_list) {
1570 attached = true;
1571 break;
1572 } else cur_overflow_list = observed_overflow_list;
1573 }
1574 if (!attached) {
1575 // Too bad, someone else got in in between; we'll need to do a splice.
1576 // Find the last item of suffix list
1577 oop last = suffix;
1578 while (last->klass_or_null() != NULL) {
1579 last = last->list_ptr_from_klass();
1580 }
1581 // Atomically prepend suffix to current overflow list
1582 observed_overflow_list = _overflow_list;
1583 do {
1584 cur_overflow_list = observed_overflow_list;
1585 if (cur_overflow_list != BUSY) {
1586 // Do the splice ...
1587 last->set_klass_to_list_ptr(cur_overflow_list);
1588 } else { // cur_overflow_list == BUSY
1589 last->set_klass_to_list_ptr(NULL);
1590 }
1591 observed_overflow_list =
1592 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1593 } while (cur_overflow_list != observed_overflow_list);
1594 }
1595 }
1596
1597 // Push objects on prefix list onto this thread's work queue
1598 assert(prefix != NULL && prefix != BUSY, "program logic");
1599 cur = prefix;
1600 ssize_t n = 0;
1601 while (cur != NULL) {
1602 oop obj_to_push = cur->forwardee();
1603 oop next = cur->list_ptr_from_klass();
1604 cur->set_klass(obj_to_push->klass());
1605 // This may be an array object that is self-forwarded. In that case, the list pointer
1606 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1607 if (!is_in_reserved(cur)) {
1608 // This can become a scaling bottleneck when there is work queue overflow coincident
1609 // with promotion failure.
1610 oopDesc* f = cur;
1611 FREE_C_HEAP_ARRAY(oopDesc, f, mtGC);
1612 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1613 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1614 obj_to_push = cur;
1615 }
1616 bool ok = work_q->push(obj_to_push);
1617 assert(ok, "Should have succeeded");
1618 cur = next;
1619 n++;
1620 }
1621 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1622 #ifndef PRODUCT
1623 assert(_num_par_pushes >= n, "Too many pops?");
1624 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1625 #endif
1626 return true;
1627 }
1628 #undef BUSY
1629
1630 void ParNewGeneration::ref_processor_init() {
1631 if (_ref_processor == NULL) {
1632 // Allocate and initialize a reference processor
1633 _ref_processor =
1634 new ReferenceProcessor(_reserved, // span
1635 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1636 (int) ParallelGCThreads, // mt processing degree
1637 refs_discovery_is_mt(), // mt discovery
1638 (int) ParallelGCThreads, // mt discovery degree
1639 refs_discovery_is_atomic(), // atomic_discovery
1640 NULL); // is_alive_non_header
1641 }
1642 }
1643
1644 const char* ParNewGeneration::name() const {
1645 return "par new generation";
1646 }