1 /*
2 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/cms/compactibleFreeListSpace.hpp"
27 #include "gc/cms/concurrentMarkSweepGeneration.hpp"
28 #include "gc/cms/parNewGeneration.inline.hpp"
29 #include "gc/cms/parOopClosures.inline.hpp"
30 #include "gc/serial/defNewGeneration.inline.hpp"
31 #include "gc/shared/adaptiveSizePolicy.hpp"
32 #include "gc/shared/ageTable.inline.hpp"
33 #include "gc/shared/copyFailedInfo.hpp"
34 #include "gc/shared/gcHeapSummary.hpp"
35 #include "gc/shared/gcTimer.hpp"
36 #include "gc/shared/gcTrace.hpp"
37 #include "gc/shared/gcTraceTime.inline.hpp"
38 #include "gc/shared/genCollectedHeap.hpp"
39 #include "gc/shared/genOopClosures.inline.hpp"
40 #include "gc/shared/generation.hpp"
41 #include "gc/shared/plab.inline.hpp"
42 #include "gc/shared/preservedMarks.inline.hpp"
43 #include "gc/shared/referencePolicy.hpp"
44 #include "gc/shared/space.hpp"
45 #include "gc/shared/spaceDecorator.hpp"
46 #include "gc/shared/strongRootsScope.hpp"
47 #include "gc/shared/taskqueue.inline.hpp"
48 #include "gc/shared/workgroup.hpp"
49 #include "logging/log.hpp"
50 #include "memory/resourceArea.hpp"
51 #include "oops/objArrayOop.hpp"
52 #include "oops/oop.inline.hpp"
53 #include "runtime/atomic.inline.hpp"
54 #include "runtime/handles.hpp"
55 #include "runtime/handles.inline.hpp"
56 #include "runtime/java.hpp"
57 #include "runtime/thread.inline.hpp"
58 #include "utilities/copy.hpp"
59 #include "utilities/globalDefinitions.hpp"
60 #include "utilities/stack.inline.hpp"
61
62 ParScanThreadState::ParScanThreadState(Space* to_space_,
63 ParNewGeneration* young_gen_,
64 Generation* old_gen_,
65 int thread_num_,
66 ObjToScanQueueSet* work_queue_set_,
67 Stack<oop, mtGC>* overflow_stacks_,
68 PreservedMarks* preserved_marks_,
69 size_t desired_plab_sz_,
70 ParallelTaskTerminator& term_) :
71 _to_space(to_space_),
72 _old_gen(old_gen_),
73 _young_gen(young_gen_),
74 _thread_num(thread_num_),
75 _work_queue(work_queue_set_->queue(thread_num_)),
76 _to_space_full(false),
77 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
78 _preserved_marks(preserved_marks_),
79 _ageTable(false), // false ==> not the global age table, no perf data.
80 _to_space_alloc_buffer(desired_plab_sz_),
81 _to_space_closure(young_gen_, this),
82 _old_gen_closure(young_gen_, this),
83 _to_space_root_closure(young_gen_, this),
84 _old_gen_root_closure(young_gen_, this),
85 _older_gen_closure(young_gen_, this),
86 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
87 &_to_space_root_closure, young_gen_, &_old_gen_root_closure,
88 work_queue_set_, &term_),
89 _is_alive_closure(young_gen_),
90 _scan_weak_ref_closure(young_gen_, this),
91 _keep_alive_closure(&_scan_weak_ref_closure),
92 _strong_roots_time(0.0),
93 _term_time(0.0)
94 {
95 #if TASKQUEUE_STATS
96 _term_attempts = 0;
97 _overflow_refills = 0;
98 _overflow_refill_objs = 0;
99 #endif // TASKQUEUE_STATS
100
101 _survivor_chunk_array = (ChunkArray*) old_gen()->get_data_recorder(thread_num());
102 _hash_seed = 17; // Might want to take time-based random value.
103 _start = os::elapsedTime();
104 _old_gen_closure.set_generation(old_gen_);
105 _old_gen_root_closure.set_generation(old_gen_);
106 }
107
108 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
109 size_t plab_word_size) {
110 ChunkArray* sca = survivor_chunk_array();
111 if (sca != NULL) {
112 // A non-null SCA implies that we want the PLAB data recorded.
113 sca->record_sample(plab_start, plab_word_size);
114 }
115 }
116
117 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
118 return new_obj->is_objArray() &&
119 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
120 new_obj != old_obj;
121 }
122
123 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
124 assert(old->is_objArray(), "must be obj array");
125 assert(old->is_forwarded(), "must be forwarded");
126 assert(GenCollectedHeap::heap()->is_in_reserved(old), "must be in heap.");
127 assert(!old_gen()->is_in(old), "must be in young generation.");
128
129 objArrayOop obj = objArrayOop(old->forwardee());
130 // Process ParGCArrayScanChunk elements now
131 // and push the remainder back onto queue
132 int start = arrayOop(old)->length();
133 int end = obj->length();
134 int remainder = end - start;
135 assert(start <= end, "just checking");
136 if (remainder > 2 * ParGCArrayScanChunk) {
137 // Test above combines last partial chunk with a full chunk
138 end = start + ParGCArrayScanChunk;
139 arrayOop(old)->set_length(end);
140 // Push remainder.
141 bool ok = work_queue()->push(old);
142 assert(ok, "just popped, push must be okay");
143 } else {
144 // Restore length so that it can be used if there
145 // is a promotion failure and forwarding pointers
146 // must be removed.
147 arrayOop(old)->set_length(end);
148 }
149
150 // process our set of indices (include header in first chunk)
151 // should make sure end is even (aligned to HeapWord in case of compressed oops)
152 if ((HeapWord *)obj < young_old_boundary()) {
153 // object is in to_space
154 obj->oop_iterate_range(&_to_space_closure, start, end);
155 } else {
156 // object is in old generation
157 obj->oop_iterate_range(&_old_gen_closure, start, end);
158 }
159 }
160
161 void ParScanThreadState::trim_queues(int max_size) {
162 ObjToScanQueue* queue = work_queue();
163 do {
164 while (queue->size() > (juint)max_size) {
165 oop obj_to_scan;
166 if (queue->pop_local(obj_to_scan)) {
167 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
168 if (obj_to_scan->is_objArray() &&
169 obj_to_scan->is_forwarded() &&
170 obj_to_scan->forwardee() != obj_to_scan) {
171 scan_partial_array_and_push_remainder(obj_to_scan);
172 } else {
173 // object is in to_space
174 obj_to_scan->oop_iterate(&_to_space_closure);
175 }
176 } else {
177 // object is in old generation
178 obj_to_scan->oop_iterate(&_old_gen_closure);
179 }
180 }
181 }
182 // For the case of compressed oops, we have a private, non-shared
183 // overflow stack, so we eagerly drain it so as to more evenly
184 // distribute load early. Note: this may be good to do in
185 // general rather than delay for the final stealing phase.
186 // If applicable, we'll transfer a set of objects over to our
187 // work queue, allowing them to be stolen and draining our
188 // private overflow stack.
189 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
190 }
191
192 bool ParScanThreadState::take_from_overflow_stack() {
193 assert(ParGCUseLocalOverflow, "Else should not call");
194 assert(young_gen()->overflow_list() == NULL, "Error");
195 ObjToScanQueue* queue = work_queue();
196 Stack<oop, mtGC>* const of_stack = overflow_stack();
197 const size_t num_overflow_elems = of_stack->size();
198 const size_t space_available = queue->max_elems() - queue->size();
199 const size_t num_take_elems = MIN3(space_available / 4,
200 ParGCDesiredObjsFromOverflowList,
201 num_overflow_elems);
202 // Transfer the most recent num_take_elems from the overflow
203 // stack to our work queue.
204 for (size_t i = 0; i != num_take_elems; i++) {
205 oop cur = of_stack->pop();
206 oop obj_to_push = cur->forwardee();
207 assert(GenCollectedHeap::heap()->is_in_reserved(cur), "Should be in heap");
208 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
209 assert(GenCollectedHeap::heap()->is_in_reserved(obj_to_push), "Should be in heap");
210 if (should_be_partially_scanned(obj_to_push, cur)) {
211 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
212 obj_to_push = cur;
213 }
214 bool ok = queue->push(obj_to_push);
215 assert(ok, "Should have succeeded");
216 }
217 assert(young_gen()->overflow_list() == NULL, "Error");
218 return num_take_elems > 0; // was something transferred?
219 }
220
221 void ParScanThreadState::push_on_overflow_stack(oop p) {
222 assert(ParGCUseLocalOverflow, "Else should not call");
223 overflow_stack()->push(p);
224 assert(young_gen()->overflow_list() == NULL, "Error");
225 }
226
227 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
228 // If the object is small enough, try to reallocate the buffer.
229 HeapWord* obj = NULL;
230 if (!_to_space_full) {
231 PLAB* const plab = to_space_alloc_buffer();
232 Space* const sp = to_space();
233 if (word_sz * 100 < ParallelGCBufferWastePct * plab->word_sz()) {
234 // Is small enough; abandon this buffer and start a new one.
235 plab->retire();
236 size_t buf_size = plab->word_sz();
237 HeapWord* buf_space = sp->par_allocate(buf_size);
238 if (buf_space == NULL) {
239 const size_t min_bytes =
240 PLAB::min_size() << LogHeapWordSize;
241 size_t free_bytes = sp->free();
242 while(buf_space == NULL && free_bytes >= min_bytes) {
243 buf_size = free_bytes >> LogHeapWordSize;
244 assert(buf_size == (size_t)align_object_size(buf_size), "Invariant");
245 buf_space = sp->par_allocate(buf_size);
246 free_bytes = sp->free();
247 }
248 }
249 if (buf_space != NULL) {
250 plab->set_buf(buf_space, buf_size);
251 record_survivor_plab(buf_space, buf_size);
252 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
253 // Note that we cannot compare buf_size < word_sz below
254 // because of AlignmentReserve (see PLAB::allocate()).
255 assert(obj != NULL || plab->words_remaining() < word_sz,
256 "Else should have been able to allocate");
257 // It's conceivable that we may be able to use the
258 // buffer we just grabbed for subsequent small requests
259 // even if not for this one.
260 } else {
261 // We're used up.
262 _to_space_full = true;
263 }
264 } else {
265 // Too large; allocate the object individually.
266 obj = sp->par_allocate(word_sz);
267 }
268 }
269 return obj;
270 }
271
272 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) {
273 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
274 }
275
276 void ParScanThreadState::print_promotion_failure_size() {
277 if (_promotion_failed_info.has_failed()) {
278 log_trace(gc, promotion)(" (%d: promotion failure size = " SIZE_FORMAT ") ",
279 _thread_num, _promotion_failed_info.first_size());
280 }
281 }
282
283 class ParScanThreadStateSet: private GenericGrowableArray {
284 public:
285 // Initializes states for the specified number of threads;
286 ParScanThreadStateSet(int num_threads,
287 Space& to_space,
288 ParNewGeneration& young_gen,
289 Generation& old_gen,
290 ObjToScanQueueSet& queue_set,
291 Stack<oop, mtGC>* overflow_stacks_,
292 PreservedMarksSet& preserved_marks_set,
293 size_t desired_plab_sz,
294 ParallelTaskTerminator& term);
295
296 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
297
298 inline ParScanThreadState& thread_state(int i);
299
300 void trace_promotion_failed(const YoungGCTracer* gc_tracer);
301 void reset(uint active_workers, bool promotion_failed);
302 void flush();
303
304 #if TASKQUEUE_STATS
305 static void
306 print_termination_stats_hdr(outputStream* const st);
307 void print_termination_stats();
308 static void
309 print_taskqueue_stats_hdr(outputStream* const st);
310 void print_taskqueue_stats();
311 void reset_stats();
312 #endif // TASKQUEUE_STATS
313
314 private:
315 ParallelTaskTerminator& _term;
316 ParNewGeneration& _young_gen;
317 Generation& _old_gen;
318 ParScanThreadState* _data;
319 int length() const { return _len; };
320 public:
321 bool is_valid(int id) const { return id < length(); }
322 ParallelTaskTerminator* terminator() { return &_term; }
323 };
324
325 ParScanThreadStateSet::ParScanThreadStateSet(int num_threads,
326 Space& to_space,
327 ParNewGeneration& young_gen,
328 Generation& old_gen,
329 ObjToScanQueueSet& queue_set,
330 Stack<oop, mtGC>* overflow_stacks,
331 PreservedMarksSet& preserved_marks_set,
332 size_t desired_plab_sz,
333 ParallelTaskTerminator& term)
334 : GenericGrowableArray(num_threads, num_threads, false, mtInternal),
335 _young_gen(young_gen),
336 _old_gen(old_gen),
337 _term(term)
338 {
339 assert(num_threads > 0, "sanity check!");
340 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
341 "overflow_stack allocation mismatch");
342 // Initialize states.
343 _data = (ParScanThreadState*) raw_allocate(sizeof(ParScanThreadState));
344 for (int i = 0; i < num_threads; ++i) {
345 ::new((void*) &_data[i])
346 ParScanThreadState(&to_space, &young_gen, &old_gen, i, &queue_set,
347 overflow_stacks, preserved_marks_set.get(i),
348 desired_plab_sz, term);
349 }
350 }
351
352 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) {
353 assert(i >= 0 && i < length(), "sanity check!");
354 return _data[i];
355 }
356
357 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) {
358 for (int i = 0; i < length(); ++i) {
359 if (thread_state(i).promotion_failed()) {
360 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info());
361 thread_state(i).promotion_failed_info().reset();
362 }
363 }
364 }
365
366 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed) {
367 _term.reset_for_reuse(active_threads);
368 if (promotion_failed) {
369 for (int i = 0; i < length(); ++i) {
370 thread_state(i).print_promotion_failure_size();
371 }
372 }
373 }
374
375 #if TASKQUEUE_STATS
376 void ParScanThreadState::reset_stats() {
377 taskqueue_stats().reset();
378 _term_attempts = 0;
379 _overflow_refills = 0;
380 _overflow_refill_objs = 0;
381 }
382
383 void ParScanThreadStateSet::reset_stats() {
384 for (int i = 0; i < length(); ++i) {
385 thread_state(i).reset_stats();
386 }
387 }
388
389 void ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) {
390 st->print_raw_cr("GC Termination Stats");
391 st->print_raw_cr(" elapsed --strong roots-- -------termination-------");
392 st->print_raw_cr("thr ms ms % ms % attempts");
393 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------");
394 }
395
396 void ParScanThreadStateSet::print_termination_stats() {
397 LogHandle(gc, task, stats) log;
398 if (!log.is_debug()) {
399 return;
400 }
401
402 ResourceMark rm;
403 outputStream* st = log.debug_stream();
404
405 print_termination_stats_hdr(st);
406
407 for (int i = 0; i < length(); ++i) {
408 const ParScanThreadState & pss = thread_state(i);
409 const double elapsed_ms = pss.elapsed_time() * 1000.0;
410 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
411 const double term_ms = pss.term_time() * 1000.0;
412 st->print_cr("%3d %9.2f %9.2f %6.2f %9.2f %6.2f " SIZE_FORMAT_W(8),
413 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
414 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
415 }
416 }
417
418 // Print stats related to work queue activity.
419 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) {
420 st->print_raw_cr("GC Task Stats");
421 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
422 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
423 }
424
425 void ParScanThreadStateSet::print_taskqueue_stats() {
426 if (!log_develop_is_enabled(Trace, gc, task, stats)) {
427 return;
428 }
429 LogHandle(gc, task, stats) log;
430 ResourceMark rm;
431 outputStream* st = log.trace_stream();
432 print_taskqueue_stats_hdr(st);
433
434 TaskQueueStats totals;
435 for (int i = 0; i < length(); ++i) {
436 const ParScanThreadState & pss = thread_state(i);
437 const TaskQueueStats & stats = pss.taskqueue_stats();
438 st->print("%3d ", i); stats.print(st); st->cr();
439 totals += stats;
440
441 if (pss.overflow_refills() > 0) {
442 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
443 SIZE_FORMAT_W(10) " overflow objects",
444 pss.overflow_refills(), pss.overflow_refill_objs());
445 }
446 }
447 st->print("tot "); totals.print(st); st->cr();
448
449 DEBUG_ONLY(totals.verify());
450 }
451 #endif // TASKQUEUE_STATS
452
453 void ParScanThreadStateSet::flush() {
454 // Work in this loop should be kept as lightweight as
455 // possible since this might otherwise become a bottleneck
456 // to scaling. Should we add heavy-weight work into this
457 // loop, consider parallelizing the loop into the worker threads.
458 for (int i = 0; i < length(); ++i) {
459 ParScanThreadState& par_scan_state = thread_state(i);
460
461 // Flush stats related to To-space PLAB activity and
462 // retire the last buffer.
463 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_young_gen.plab_stats());
464
465 // Every thread has its own age table. We need to merge
466 // them all into one.
467 AgeTable *local_table = par_scan_state.age_table();
468 _young_gen.age_table()->merge(local_table);
469
470 // Inform old gen that we're done.
471 _old_gen.par_promote_alloc_done(i);
472 _old_gen.par_oop_since_save_marks_iterate_done(i);
473 }
474
475 if (UseConcMarkSweepGC) {
476 // We need to call this even when ResizeOldPLAB is disabled
477 // so as to avoid breaking some asserts. While we may be able
478 // to avoid this by reorganizing the code a bit, I am loathe
479 // to do that unless we find cases where ergo leads to bad
480 // performance.
481 CompactibleFreeListSpaceLAB::compute_desired_plab_size();
482 }
483 }
484
485 ParScanClosure::ParScanClosure(ParNewGeneration* g,
486 ParScanThreadState* par_scan_state) :
487 OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) {
488 _boundary = _g->reserved().end();
489 }
490
491 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
492 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
493
494 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
495 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
496
497 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
498 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
499
500 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
501 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
502
503 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
504 ParScanThreadState* par_scan_state)
505 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
506 {}
507
508 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
509 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
510
511 #ifdef WIN32
512 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
513 #endif
514
515 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
516 ParScanThreadState* par_scan_state_,
517 ParScanWithoutBarrierClosure* to_space_closure_,
518 ParScanWithBarrierClosure* old_gen_closure_,
519 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
520 ParNewGeneration* par_gen_,
521 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
522 ObjToScanQueueSet* task_queues_,
523 ParallelTaskTerminator* terminator_) :
524
525 _par_scan_state(par_scan_state_),
526 _to_space_closure(to_space_closure_),
527 _old_gen_closure(old_gen_closure_),
528 _to_space_root_closure(to_space_root_closure_),
529 _old_gen_root_closure(old_gen_root_closure_),
530 _par_gen(par_gen_),
531 _task_queues(task_queues_),
532 _terminator(terminator_)
533 {}
534
535 void ParEvacuateFollowersClosure::do_void() {
536 ObjToScanQueue* work_q = par_scan_state()->work_queue();
537
538 while (true) {
539 // Scan to-space and old-gen objs until we run out of both.
540 oop obj_to_scan;
541 par_scan_state()->trim_queues(0);
542
543 // We have no local work, attempt to steal from other threads.
544
545 // Attempt to steal work from promoted.
546 if (task_queues()->steal(par_scan_state()->thread_num(),
547 par_scan_state()->hash_seed(),
548 obj_to_scan)) {
549 bool res = work_q->push(obj_to_scan);
550 assert(res, "Empty queue should have room for a push.");
551
552 // If successful, goto Start.
553 continue;
554
555 // Try global overflow list.
556 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
557 continue;
558 }
559
560 // Otherwise, offer termination.
561 par_scan_state()->start_term_time();
562 if (terminator()->offer_termination()) break;
563 par_scan_state()->end_term_time();
564 }
565 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
566 "Broken overflow list?");
567 // Finish the last termination pause.
568 par_scan_state()->end_term_time();
569 }
570
571 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen,
572 Generation* old_gen,
573 HeapWord* young_old_boundary,
574 ParScanThreadStateSet* state_set,
575 StrongRootsScope* strong_roots_scope) :
576 AbstractGangTask("ParNewGeneration collection"),
577 _young_gen(young_gen), _old_gen(old_gen),
578 _young_old_boundary(young_old_boundary),
579 _state_set(state_set),
580 _strong_roots_scope(strong_roots_scope)
581 {}
582
583 void ParNewGenTask::work(uint worker_id) {
584 GenCollectedHeap* gch = GenCollectedHeap::heap();
585 // Since this is being done in a separate thread, need new resource
586 // and handle marks.
587 ResourceMark rm;
588 HandleMark hm;
589
590 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
591 assert(_state_set->is_valid(worker_id), "Should not have been called");
592
593 par_scan_state.set_young_old_boundary(_young_old_boundary);
594
595 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(),
596 gch->rem_set()->klass_rem_set());
597 CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure,
598 &par_scan_state.to_space_root_closure(),
599 false);
600
601 par_scan_state.start_strong_roots();
602 gch->gen_process_roots(_strong_roots_scope,
603 GenCollectedHeap::YoungGen,
604 true, // Process younger gens, if any, as strong roots.
605 GenCollectedHeap::SO_ScavengeCodeCache,
606 GenCollectedHeap::StrongAndWeakRoots,
607 &par_scan_state.to_space_root_closure(),
608 &par_scan_state.older_gen_closure(),
609 &cld_scan_closure);
610
611 par_scan_state.end_strong_roots();
612
613 // "evacuate followers".
614 par_scan_state.evacuate_followers_closure().do_void();
615 }
616
617 ParNewGeneration::ParNewGeneration(ReservedSpace rs, size_t initial_byte_size)
618 : DefNewGeneration(rs, initial_byte_size, "PCopy"),
619 _overflow_list(NULL),
620 _is_alive_closure(this),
621 _plab_stats("Young", YoungPLABSize, PLABWeight)
622 {
623 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
624 NOT_PRODUCT(_num_par_pushes = 0;)
625 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
626 guarantee(_task_queues != NULL, "task_queues allocation failure.");
627
628 for (uint i = 0; i < ParallelGCThreads; i++) {
629 ObjToScanQueue *q = new ObjToScanQueue();
630 guarantee(q != NULL, "work_queue Allocation failure.");
631 _task_queues->register_queue(i, q);
632 }
633
634 for (uint i = 0; i < ParallelGCThreads; i++) {
635 _task_queues->queue(i)->initialize();
636 }
637
638 _overflow_stacks = NULL;
639 if (ParGCUseLocalOverflow) {
640 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal with ','
641 typedef Stack<oop, mtGC> GCOopStack;
642
643 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
644 for (size_t i = 0; i < ParallelGCThreads; ++i) {
645 new (_overflow_stacks + i) Stack<oop, mtGC>();
646 }
647 }
648
649 if (UsePerfData) {
650 EXCEPTION_MARK;
651 ResourceMark rm;
652
653 const char* cname =
654 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
655 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
656 ParallelGCThreads, CHECK);
657 }
658 }
659
660 // ParNewGeneration::
661 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
662 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
663
664 template <class T>
665 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
666 #ifdef ASSERT
667 {
668 assert(!oopDesc::is_null(*p), "expected non-null ref");
669 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
670 // We never expect to see a null reference being processed
671 // as a weak reference.
672 assert(obj->is_oop(), "expected an oop while scanning weak refs");
673 }
674 #endif // ASSERT
675
676 _par_cl->do_oop_nv(p);
677
678 if (GenCollectedHeap::heap()->is_in_reserved(p)) {
679 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
680 _rs->write_ref_field_gc_par(p, obj);
681 }
682 }
683
684 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
685 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
686
687 // ParNewGeneration::
688 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
689 DefNewGeneration::KeepAliveClosure(cl) {}
690
691 template <class T>
692 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
693 #ifdef ASSERT
694 {
695 assert(!oopDesc::is_null(*p), "expected non-null ref");
696 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
697 // We never expect to see a null reference being processed
698 // as a weak reference.
699 assert(obj->is_oop(), "expected an oop while scanning weak refs");
700 }
701 #endif // ASSERT
702
703 _cl->do_oop_nv(p);
704
705 if (GenCollectedHeap::heap()->is_in_reserved(p)) {
706 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
707 _rs->write_ref_field_gc_par(p, obj);
708 }
709 }
710
711 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
712 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
713
714 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
715 T heap_oop = oopDesc::load_heap_oop(p);
716 if (!oopDesc::is_null(heap_oop)) {
717 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
718 if ((HeapWord*)obj < _boundary) {
719 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
720 oop new_obj = obj->is_forwarded()
721 ? obj->forwardee()
722 : _g->DefNewGeneration::copy_to_survivor_space(obj);
723 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
724 }
725 if (_gc_barrier) {
726 // If p points to a younger generation, mark the card.
727 if ((HeapWord*)obj < _gen_boundary) {
728 _rs->write_ref_field_gc_par(p, obj);
729 }
730 }
731 }
732 }
733
734 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
735 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
736
737 class ParNewRefProcTaskProxy: public AbstractGangTask {
738 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
739 public:
740 ParNewRefProcTaskProxy(ProcessTask& task,
741 ParNewGeneration& young_gen,
742 Generation& old_gen,
743 HeapWord* young_old_boundary,
744 ParScanThreadStateSet& state_set);
745
746 private:
747 virtual void work(uint worker_id);
748 private:
749 ParNewGeneration& _young_gen;
750 ProcessTask& _task;
751 Generation& _old_gen;
752 HeapWord* _young_old_boundary;
753 ParScanThreadStateSet& _state_set;
754 };
755
756 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
757 ParNewGeneration& young_gen,
758 Generation& old_gen,
759 HeapWord* young_old_boundary,
760 ParScanThreadStateSet& state_set)
761 : AbstractGangTask("ParNewGeneration parallel reference processing"),
762 _young_gen(young_gen),
763 _task(task),
764 _old_gen(old_gen),
765 _young_old_boundary(young_old_boundary),
766 _state_set(state_set)
767 { }
768
769 void ParNewRefProcTaskProxy::work(uint worker_id) {
770 ResourceMark rm;
771 HandleMark hm;
772 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
773 par_scan_state.set_young_old_boundary(_young_old_boundary);
774 _task.work(worker_id, par_scan_state.is_alive_closure(),
775 par_scan_state.keep_alive_closure(),
776 par_scan_state.evacuate_followers_closure());
777 }
778
779 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
780 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
781 EnqueueTask& _task;
782
783 public:
784 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
785 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
786 _task(task)
787 { }
788
789 virtual void work(uint worker_id) {
790 _task.work(worker_id);
791 }
792 };
793
794 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) {
795 GenCollectedHeap* gch = GenCollectedHeap::heap();
796 WorkGang* workers = gch->workers();
797 assert(workers != NULL, "Need parallel worker threads.");
798 _state_set.reset(workers->active_workers(), _young_gen.promotion_failed());
799 ParNewRefProcTaskProxy rp_task(task, _young_gen, _old_gen,
800 _young_gen.reserved().end(), _state_set);
801 workers->run_task(&rp_task);
802 _state_set.reset(0 /* bad value in debug if not reset */,
803 _young_gen.promotion_failed());
804 }
805
806 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task) {
807 GenCollectedHeap* gch = GenCollectedHeap::heap();
808 WorkGang* workers = gch->workers();
809 assert(workers != NULL, "Need parallel worker threads.");
810 ParNewRefEnqueueTaskProxy enq_task(task);
811 workers->run_task(&enq_task);
812 }
813
814 void ParNewRefProcTaskExecutor::set_single_threaded_mode() {
815 _state_set.flush();
816 GenCollectedHeap* gch = GenCollectedHeap::heap();
817 gch->save_marks();
818 }
819
820 ScanClosureWithParBarrier::
821 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
822 ScanClosure(g, gc_barrier)
823 { }
824
825 EvacuateFollowersClosureGeneral::
826 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch,
827 OopsInGenClosure* cur,
828 OopsInGenClosure* older) :
829 _gch(gch),
830 _scan_cur_or_nonheap(cur), _scan_older(older)
831 { }
832
833 void EvacuateFollowersClosureGeneral::do_void() {
834 do {
835 // Beware: this call will lead to closure applications via virtual
836 // calls.
837 _gch->oop_since_save_marks_iterate(GenCollectedHeap::YoungGen,
838 _scan_cur_or_nonheap,
839 _scan_older);
840 } while (!_gch->no_allocs_since_save_marks());
841 }
842
843 // A Generation that does parallel young-gen collection.
844
845 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set) {
846 assert(_promo_failure_scan_stack.is_empty(), "post condition");
847 _promo_failure_scan_stack.clear(true); // Clear cached segments.
848
849 remove_forwarding_pointers();
850 log_info(gc, promotion)("Promotion failed");
851 // All the spaces are in play for mark-sweep.
852 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
853 from()->set_next_compaction_space(to());
854 gch->set_incremental_collection_failed();
855 // Inform the next generation that a promotion failure occurred.
856 _old_gen->promotion_failure_occurred();
857
858 // Trace promotion failure in the parallel GC threads
859 thread_state_set.trace_promotion_failed(gc_tracer());
860 // Single threaded code may have reported promotion failure to the global state
861 if (_promotion_failed_info.has_failed()) {
862 _gc_tracer.report_promotion_failed(_promotion_failed_info);
863 }
864 // Reset the PromotionFailureALot counters.
865 NOT_PRODUCT(gch->reset_promotion_should_fail();)
866 }
867
868 void ParNewGeneration::collect(bool full,
869 bool clear_all_soft_refs,
870 size_t size,
871 bool is_tlab) {
872 assert(full || size > 0, "otherwise we don't want to collect");
873
874 GenCollectedHeap* gch = GenCollectedHeap::heap();
875
876 _gc_timer->register_gc_start();
877
878 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
879 WorkGang* workers = gch->workers();
880 assert(workers != NULL, "Need workgang for parallel work");
881 uint active_workers =
882 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
883 workers->active_workers(),
884 Threads::number_of_non_daemon_threads());
885 workers->set_active_workers(active_workers);
886 _old_gen = gch->old_gen();
887
888 // If the next generation is too full to accommodate worst-case promotion
889 // from this generation, pass on collection; let the next generation
890 // do it.
891 if (!collection_attempt_is_safe()) {
892 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
893 return;
894 }
895 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
896
897 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
898 gch->trace_heap_before_gc(gc_tracer());
899
900 init_assuming_no_promotion_failure();
901
902 if (UseAdaptiveSizePolicy) {
903 set_survivor_overflow(false);
904 size_policy->minor_collection_begin();
905 }
906
907 GCTraceTime(Trace, gc) t1("ParNew", NULL, gch->gc_cause());
908
909 age_table()->clear();
910 to()->clear(SpaceDecorator::Mangle);
911
912 gch->save_marks();
913
914 // Set the correct parallelism (number of queues) in the reference processor
915 ref_processor()->set_active_mt_degree(active_workers);
916
917 // Need to initialize the preserved marks before the ThreadStateSet c'tor.
918 _preserved_marks_set.init(active_workers);
919
920 // Always set the terminator for the active number of workers
921 // because only those workers go through the termination protocol.
922 ParallelTaskTerminator _term(active_workers, task_queues());
923 ParScanThreadStateSet thread_state_set(active_workers,
924 *to(), *this, *_old_gen, *task_queues(),
925 _overflow_stacks, _preserved_marks_set,
926 desired_plab_sz(), _term);
927
928 thread_state_set.reset(active_workers, promotion_failed());
929
930 {
931 StrongRootsScope srs(active_workers);
932
933 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs);
934 gch->rem_set()->prepare_for_younger_refs_iterate(true);
935 // It turns out that even when we're using 1 thread, doing the work in a
936 // separate thread causes wide variance in run times. We can't help this
937 // in the multi-threaded case, but we special-case n=1 here to get
938 // repeatable measurements of the 1-thread overhead of the parallel code.
939 if (active_workers > 1) {
940 workers->run_task(&tsk);
941 } else {
942 tsk.work(0);
943 }
944 }
945
946 thread_state_set.reset(0 /* Bad value in debug if not reset */,
947 promotion_failed());
948
949 // Trace and reset failed promotion info.
950 if (promotion_failed()) {
951 thread_state_set.trace_promotion_failed(gc_tracer());
952 }
953
954 // Process (weak) reference objects found during scavenge.
955 ReferenceProcessor* rp = ref_processor();
956 IsAliveClosure is_alive(this);
957 ScanWeakRefClosure scan_weak_ref(this);
958 KeepAliveClosure keep_alive(&scan_weak_ref);
959 ScanClosure scan_without_gc_barrier(this, false);
960 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
961 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
962 EvacuateFollowersClosureGeneral evacuate_followers(gch,
963 &scan_without_gc_barrier, &scan_with_gc_barrier);
964 rp->setup_policy(clear_all_soft_refs);
965 // Can the mt_degree be set later (at run_task() time would be best)?
966 rp->set_active_mt_degree(active_workers);
967 ReferenceProcessorStats stats;
968 if (rp->processing_is_mt()) {
969 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set);
970 stats = rp->process_discovered_references(&is_alive, &keep_alive,
971 &evacuate_followers, &task_executor,
972 _gc_timer);
973 } else {
974 thread_state_set.flush();
975 gch->save_marks();
976 stats = rp->process_discovered_references(&is_alive, &keep_alive,
977 &evacuate_followers, NULL,
978 _gc_timer);
979 }
980 _gc_tracer.report_gc_reference_stats(stats);
981 _gc_tracer.report_tenuring_threshold(tenuring_threshold());
982
983 if (!promotion_failed()) {
984 // Swap the survivor spaces.
985 eden()->clear(SpaceDecorator::Mangle);
986 from()->clear(SpaceDecorator::Mangle);
987 if (ZapUnusedHeapArea) {
988 // This is now done here because of the piece-meal mangling which
989 // can check for valid mangling at intermediate points in the
990 // collection(s). When a young collection fails to collect
991 // sufficient space resizing of the young generation can occur
992 // and redistribute the spaces in the young generation. Mangle
993 // here so that unzapped regions don't get distributed to
994 // other spaces.
995 to()->mangle_unused_area();
996 }
997 swap_spaces();
998
999 // A successful scavenge should restart the GC time limit count which is
1000 // for full GC's.
1001 size_policy->reset_gc_overhead_limit_count();
1002
1003 assert(to()->is_empty(), "to space should be empty now");
1004
1005 adjust_desired_tenuring_threshold();
1006 } else {
1007 handle_promotion_failed(gch, thread_state_set);
1008 }
1009 _preserved_marks_set.reclaim();
1010 // set new iteration safe limit for the survivor spaces
1011 from()->set_concurrent_iteration_safe_limit(from()->top());
1012 to()->set_concurrent_iteration_safe_limit(to()->top());
1013
1014 plab_stats()->adjust_desired_plab_sz();
1015
1016 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
1017 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
1018
1019 if (UseAdaptiveSizePolicy) {
1020 size_policy->minor_collection_end(gch->gc_cause());
1021 size_policy->avg_survived()->sample(from()->used());
1022 }
1023
1024 // We need to use a monotonically non-decreasing time in ms
1025 // or we will see time-warp warnings and os::javaTimeMillis()
1026 // does not guarantee monotonicity.
1027 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1028 update_time_of_last_gc(now);
1029
1030 rp->set_enqueuing_is_done(true);
1031 if (rp->processing_is_mt()) {
1032 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set);
1033 rp->enqueue_discovered_references(&task_executor);
1034 } else {
1035 rp->enqueue_discovered_references(NULL);
1036 }
1037 rp->verify_no_references_recorded();
1038
1039 gch->trace_heap_after_gc(gc_tracer());
1040
1041 _gc_timer->register_gc_end();
1042
1043 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1044 }
1045
1046 size_t ParNewGeneration::desired_plab_sz() {
1047 return _plab_stats.desired_plab_sz(GenCollectedHeap::heap()->workers()->active_workers());
1048 }
1049
1050 static int sum;
1051 void ParNewGeneration::waste_some_time() {
1052 for (int i = 0; i < 100; i++) {
1053 sum += i;
1054 }
1055 }
1056
1057 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1058
1059 // Because of concurrency, there are times where an object for which
1060 // "is_forwarded()" is true contains an "interim" forwarding pointer
1061 // value. Such a value will soon be overwritten with a real value.
1062 // This method requires "obj" to have a forwarding pointer, and waits, if
1063 // necessary for a real one to be inserted, and returns it.
1064
1065 oop ParNewGeneration::real_forwardee(oop obj) {
1066 oop forward_ptr = obj->forwardee();
1067 if (forward_ptr != ClaimedForwardPtr) {
1068 return forward_ptr;
1069 } else {
1070 return real_forwardee_slow(obj);
1071 }
1072 }
1073
1074 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1075 // Spin-read if it is claimed but not yet written by another thread.
1076 oop forward_ptr = obj->forwardee();
1077 while (forward_ptr == ClaimedForwardPtr) {
1078 waste_some_time();
1079 assert(obj->is_forwarded(), "precondition");
1080 forward_ptr = obj->forwardee();
1081 }
1082 return forward_ptr;
1083 }
1084
1085 // Multiple GC threads may try to promote an object. If the object
1086 // is successfully promoted, a forwarding pointer will be installed in
1087 // the object in the young generation. This method claims the right
1088 // to install the forwarding pointer before it copies the object,
1089 // thus avoiding the need to undo the copy as in
1090 // copy_to_survivor_space_avoiding_with_undo.
1091
1092 oop ParNewGeneration::copy_to_survivor_space(ParScanThreadState* par_scan_state,
1093 oop old,
1094 size_t sz,
1095 markOop m) {
1096 // In the sequential version, this assert also says that the object is
1097 // not forwarded. That might not be the case here. It is the case that
1098 // the caller observed it to be not forwarded at some time in the past.
1099 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1100
1101 // The sequential code read "old->age()" below. That doesn't work here,
1102 // since the age is in the mark word, and that might be overwritten with
1103 // a forwarding pointer by a parallel thread. So we must save the mark
1104 // word in a local and then analyze it.
1105 oopDesc dummyOld;
1106 dummyOld.set_mark(m);
1107 assert(!dummyOld.is_forwarded(),
1108 "should not be called with forwarding pointer mark word.");
1109
1110 oop new_obj = NULL;
1111 oop forward_ptr;
1112
1113 // Try allocating obj in to-space (unless too old)
1114 if (dummyOld.age() < tenuring_threshold()) {
1115 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1116 if (new_obj == NULL) {
1117 set_survivor_overflow(true);
1118 }
1119 }
1120
1121 if (new_obj == NULL) {
1122 // Either to-space is full or we decided to promote try allocating obj tenured
1123
1124 // Attempt to install a null forwarding pointer (atomically),
1125 // to claim the right to install the real forwarding pointer.
1126 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1127 if (forward_ptr != NULL) {
1128 // someone else beat us to it.
1129 return real_forwardee(old);
1130 }
1131
1132 if (!_promotion_failed) {
1133 new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1134 old, m, sz);
1135 }
1136
1137 if (new_obj == NULL) {
1138 // promotion failed, forward to self
1139 _promotion_failed = true;
1140 new_obj = old;
1141
1142 par_scan_state->preserved_marks()->push_if_necessary(old, m);
1143 par_scan_state->register_promotion_failure(sz);
1144 }
1145
1146 old->forward_to(new_obj);
1147 forward_ptr = NULL;
1148 } else {
1149 // Is in to-space; do copying ourselves.
1150 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1151 assert(GenCollectedHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1152 forward_ptr = old->forward_to_atomic(new_obj);
1153 // Restore the mark word copied above.
1154 new_obj->set_mark(m);
1155 // Increment age if obj still in new generation
1156 new_obj->incr_age();
1157 par_scan_state->age_table()->add(new_obj, sz);
1158 }
1159 assert(new_obj != NULL, "just checking");
1160
1161 // This code must come after the CAS test, or it will print incorrect
1162 // information.
1163 log_develop_trace(gc, scavenge)("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1164 is_in_reserved(new_obj) ? "copying" : "tenuring",
1165 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size());
1166
1167 if (forward_ptr == NULL) {
1168 oop obj_to_push = new_obj;
1169 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1170 // Length field used as index of next element to be scanned.
1171 // Real length can be obtained from real_forwardee()
1172 arrayOop(old)->set_length(0);
1173 obj_to_push = old;
1174 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1175 "push forwarded object");
1176 }
1177 // Push it on one of the queues of to-be-scanned objects.
1178 bool simulate_overflow = false;
1179 NOT_PRODUCT(
1180 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1181 // simulate a stack overflow
1182 simulate_overflow = true;
1183 }
1184 )
1185 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1186 // Add stats for overflow pushes.
1187 log_develop_trace(gc)("Queue Overflow");
1188 push_on_overflow_list(old, par_scan_state);
1189 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1190 }
1191
1192 return new_obj;
1193 }
1194
1195 // Oops. Someone beat us to it. Undo the allocation. Where did we
1196 // allocate it?
1197 if (is_in_reserved(new_obj)) {
1198 // Must be in to_space.
1199 assert(to()->is_in_reserved(new_obj), "Checking");
1200 if (forward_ptr == ClaimedForwardPtr) {
1201 // Wait to get the real forwarding pointer value.
1202 forward_ptr = real_forwardee(old);
1203 }
1204 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1205 }
1206
1207 return forward_ptr;
1208 }
1209
1210 #ifndef PRODUCT
1211 // It's OK to call this multi-threaded; the worst thing
1212 // that can happen is that we'll get a bunch of closely
1213 // spaced simulated overflows, but that's OK, in fact
1214 // probably good as it would exercise the overflow code
1215 // under contention.
1216 bool ParNewGeneration::should_simulate_overflow() {
1217 if (_overflow_counter-- <= 0) { // just being defensive
1218 _overflow_counter = ParGCWorkQueueOverflowInterval;
1219 return true;
1220 } else {
1221 return false;
1222 }
1223 }
1224 #endif
1225
1226 // In case we are using compressed oops, we need to be careful.
1227 // If the object being pushed is an object array, then its length
1228 // field keeps track of the "grey boundary" at which the next
1229 // incremental scan will be done (see ParGCArrayScanChunk).
1230 // When using compressed oops, this length field is kept in the
1231 // lower 32 bits of the erstwhile klass word and cannot be used
1232 // for the overflow chaining pointer (OCP below). As such the OCP
1233 // would itself need to be compressed into the top 32-bits in this
1234 // case. Unfortunately, see below, in the event that we have a
1235 // promotion failure, the node to be pushed on the list can be
1236 // outside of the Java heap, so the heap-based pointer compression
1237 // would not work (we would have potential aliasing between C-heap
1238 // and Java-heap pointers). For this reason, when using compressed
1239 // oops, we simply use a worker-thread-local, non-shared overflow
1240 // list in the form of a growable array, with a slightly different
1241 // overflow stack draining strategy. If/when we start using fat
1242 // stacks here, we can go back to using (fat) pointer chains
1243 // (although some performance comparisons would be useful since
1244 // single global lists have their own performance disadvantages
1245 // as we were made painfully aware not long ago, see 6786503).
1246 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1247 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1248 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1249 if (ParGCUseLocalOverflow) {
1250 // In the case of compressed oops, we use a private, not-shared
1251 // overflow stack.
1252 par_scan_state->push_on_overflow_stack(from_space_obj);
1253 } else {
1254 assert(!UseCompressedOops, "Error");
1255 // if the object has been forwarded to itself, then we cannot
1256 // use the klass pointer for the linked list. Instead we have
1257 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1258 // XXX This is horribly inefficient when a promotion failure occurs
1259 // and should be fixed. XXX FIX ME !!!
1260 #ifndef PRODUCT
1261 Atomic::inc_ptr(&_num_par_pushes);
1262 assert(_num_par_pushes > 0, "Tautology");
1263 #endif
1264 if (from_space_obj->forwardee() == from_space_obj) {
1265 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1266 listhead->forward_to(from_space_obj);
1267 from_space_obj = listhead;
1268 }
1269 oop observed_overflow_list = _overflow_list;
1270 oop cur_overflow_list;
1271 do {
1272 cur_overflow_list = observed_overflow_list;
1273 if (cur_overflow_list != BUSY) {
1274 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1275 } else {
1276 from_space_obj->set_klass_to_list_ptr(NULL);
1277 }
1278 observed_overflow_list =
1279 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1280 } while (cur_overflow_list != observed_overflow_list);
1281 }
1282 }
1283
1284 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1285 bool res;
1286
1287 if (ParGCUseLocalOverflow) {
1288 res = par_scan_state->take_from_overflow_stack();
1289 } else {
1290 assert(!UseCompressedOops, "Error");
1291 res = take_from_overflow_list_work(par_scan_state);
1292 }
1293 return res;
1294 }
1295
1296
1297 // *NOTE*: The overflow list manipulation code here and
1298 // in CMSCollector:: are very similar in shape,
1299 // except that in the CMS case we thread the objects
1300 // directly into the list via their mark word, and do
1301 // not need to deal with special cases below related
1302 // to chunking of object arrays and promotion failure
1303 // handling.
1304 // CR 6797058 has been filed to attempt consolidation of
1305 // the common code.
1306 // Because of the common code, if you make any changes in
1307 // the code below, please check the CMS version to see if
1308 // similar changes might be needed.
1309 // See CMSCollector::par_take_from_overflow_list() for
1310 // more extensive documentation comments.
1311 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1312 ObjToScanQueue* work_q = par_scan_state->work_queue();
1313 // How many to take?
1314 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1315 (size_t)ParGCDesiredObjsFromOverflowList);
1316
1317 assert(!UseCompressedOops, "Error");
1318 assert(par_scan_state->overflow_stack() == NULL, "Error");
1319 if (_overflow_list == NULL) return false;
1320
1321 // Otherwise, there was something there; try claiming the list.
1322 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1323 // Trim off a prefix of at most objsFromOverflow items
1324 Thread* tid = Thread::current();
1325 size_t spin_count = ParallelGCThreads;
1326 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1327 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1328 // someone grabbed it before we did ...
1329 // ... we spin for a short while...
1330 os::sleep(tid, sleep_time_millis, false);
1331 if (_overflow_list == NULL) {
1332 // nothing left to take
1333 return false;
1334 } else if (_overflow_list != BUSY) {
1335 // try and grab the prefix
1336 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1337 }
1338 }
1339 if (prefix == NULL || prefix == BUSY) {
1340 // Nothing to take or waited long enough
1341 if (prefix == NULL) {
1342 // Write back the NULL in case we overwrote it with BUSY above
1343 // and it is still the same value.
1344 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1345 }
1346 return false;
1347 }
1348 assert(prefix != NULL && prefix != BUSY, "Error");
1349 size_t i = 1;
1350 oop cur = prefix;
1351 while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1352 i++; cur = cur->list_ptr_from_klass();
1353 }
1354
1355 // Reattach remaining (suffix) to overflow list
1356 if (cur->klass_or_null() == NULL) {
1357 // Write back the NULL in lieu of the BUSY we wrote
1358 // above and it is still the same value.
1359 if (_overflow_list == BUSY) {
1360 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1361 }
1362 } else {
1363 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error");
1364 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list
1365 cur->set_klass_to_list_ptr(NULL); // break off suffix
1366 // It's possible that the list is still in the empty(busy) state
1367 // we left it in a short while ago; in that case we may be
1368 // able to place back the suffix.
1369 oop observed_overflow_list = _overflow_list;
1370 oop cur_overflow_list = observed_overflow_list;
1371 bool attached = false;
1372 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1373 observed_overflow_list =
1374 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1375 if (cur_overflow_list == observed_overflow_list) {
1376 attached = true;
1377 break;
1378 } else cur_overflow_list = observed_overflow_list;
1379 }
1380 if (!attached) {
1381 // Too bad, someone else got in in between; we'll need to do a splice.
1382 // Find the last item of suffix list
1383 oop last = suffix;
1384 while (last->klass_or_null() != NULL) {
1385 last = last->list_ptr_from_klass();
1386 }
1387 // Atomically prepend suffix to current overflow list
1388 observed_overflow_list = _overflow_list;
1389 do {
1390 cur_overflow_list = observed_overflow_list;
1391 if (cur_overflow_list != BUSY) {
1392 // Do the splice ...
1393 last->set_klass_to_list_ptr(cur_overflow_list);
1394 } else { // cur_overflow_list == BUSY
1395 last->set_klass_to_list_ptr(NULL);
1396 }
1397 observed_overflow_list =
1398 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1399 } while (cur_overflow_list != observed_overflow_list);
1400 }
1401 }
1402
1403 // Push objects on prefix list onto this thread's work queue
1404 assert(prefix != NULL && prefix != BUSY, "program logic");
1405 cur = prefix;
1406 ssize_t n = 0;
1407 while (cur != NULL) {
1408 oop obj_to_push = cur->forwardee();
1409 oop next = cur->list_ptr_from_klass();
1410 cur->set_klass(obj_to_push->klass());
1411 // This may be an array object that is self-forwarded. In that case, the list pointer
1412 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1413 if (!is_in_reserved(cur)) {
1414 // This can become a scaling bottleneck when there is work queue overflow coincident
1415 // with promotion failure.
1416 oopDesc* f = cur;
1417 FREE_C_HEAP_ARRAY(oopDesc, f);
1418 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1419 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1420 obj_to_push = cur;
1421 }
1422 bool ok = work_q->push(obj_to_push);
1423 assert(ok, "Should have succeeded");
1424 cur = next;
1425 n++;
1426 }
1427 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1428 #ifndef PRODUCT
1429 assert(_num_par_pushes >= n, "Too many pops?");
1430 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1431 #endif
1432 return true;
1433 }
1434 #undef BUSY
1435
1436 void ParNewGeneration::ref_processor_init() {
1437 if (_ref_processor == NULL) {
1438 // Allocate and initialize a reference processor
1439 _ref_processor =
1440 new ReferenceProcessor(_reserved, // span
1441 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1442 ParallelGCThreads, // mt processing degree
1443 refs_discovery_is_mt(), // mt discovery
1444 ParallelGCThreads, // mt discovery degree
1445 refs_discovery_is_atomic(), // atomic_discovery
1446 NULL); // is_alive_non_header
1447 }
1448 }
1449
1450 const char* ParNewGeneration::name() const {
1451 return "par new generation";
1452 }