1 /*
2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc/cms/compactibleFreeListSpace.hpp"
27 #include "gc/cms/concurrentMarkSweepGeneration.hpp"
28 #include "gc/cms/parNewGeneration.hpp"
29 #include "gc/cms/parOopClosures.inline.hpp"
30 #include "gc/serial/defNewGeneration.inline.hpp"
31 #include "gc/shared/adaptiveSizePolicy.hpp"
32 #include "gc/shared/ageTable.hpp"
33 #include "gc/shared/copyFailedInfo.hpp"
34 #include "gc/shared/gcHeapSummary.hpp"
35 #include "gc/shared/gcTimer.hpp"
36 #include "gc/shared/gcTrace.hpp"
37 #include "gc/shared/gcTraceTime.hpp"
38 #include "gc/shared/genCollectedHeap.hpp"
39 #include "gc/shared/genOopClosures.inline.hpp"
40 #include "gc/shared/generation.hpp"
41 #include "gc/shared/plab.inline.hpp"
42 #include "gc/shared/referencePolicy.hpp"
43 #include "gc/shared/space.hpp"
44 #include "gc/shared/spaceDecorator.hpp"
45 #include "gc/shared/strongRootsScope.hpp"
46 #include "gc/shared/taskqueue.inline.hpp"
47 #include "gc/shared/workgroup.hpp"
48 #include "memory/resourceArea.hpp"
49 #include "oops/objArrayOop.hpp"
50 #include "oops/oop.inline.hpp"
51 #include "runtime/atomic.inline.hpp"
52 #include "runtime/handles.hpp"
53 #include "runtime/handles.inline.hpp"
54 #include "runtime/java.hpp"
55 #include "runtime/thread.inline.hpp"
56 #include "utilities/copy.hpp"
57 #include "utilities/globalDefinitions.hpp"
58 #include "utilities/stack.inline.hpp"
59
60 #ifdef _MSC_VER
61 #pragma warning( push )
62 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
63 #endif
64 ParScanThreadState::ParScanThreadState(Space* to_space_,
65 ParNewGeneration* gen_,
66 Generation* old_gen_,
67 int thread_num_,
68 ObjToScanQueueSet* work_queue_set_,
69 Stack<oop, mtGC>* overflow_stacks_,
70 size_t desired_plab_sz_,
71 ParallelTaskTerminator& term_) :
72 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_),
73 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false),
74 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
75 _ageTable(false), // false ==> not the global age table, no perf data.
76 _to_space_alloc_buffer(desired_plab_sz_),
77 _to_space_closure(gen_, this), _old_gen_closure(gen_, this),
78 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this),
79 _older_gen_closure(gen_, this),
80 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
81 &_to_space_root_closure, gen_, &_old_gen_root_closure,
82 work_queue_set_, &term_),
83 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this),
84 _keep_alive_closure(&_scan_weak_ref_closure),
85 _strong_roots_time(0.0), _term_time(0.0)
86 {
87 #if TASKQUEUE_STATS
88 _term_attempts = 0;
89 _overflow_refills = 0;
90 _overflow_refill_objs = 0;
91 #endif // TASKQUEUE_STATS
92
93 _survivor_chunk_array =
94 (ChunkArray*) old_gen()->get_data_recorder(thread_num());
95 _hash_seed = 17; // Might want to take time-based random value.
96 _start = os::elapsedTime();
97 _old_gen_closure.set_generation(old_gen_);
98 _old_gen_root_closure.set_generation(old_gen_);
99 }
100 #ifdef _MSC_VER
101 #pragma warning( pop )
102 #endif
103
104 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
105 size_t plab_word_size) {
106 ChunkArray* sca = survivor_chunk_array();
107 if (sca != NULL) {
108 // A non-null SCA implies that we want the PLAB data recorded.
109 sca->record_sample(plab_start, plab_word_size);
110 }
111 }
112
113 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
114 return new_obj->is_objArray() &&
115 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
116 new_obj != old_obj;
117 }
118
119 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
120 assert(old->is_objArray(), "must be obj array");
121 assert(old->is_forwarded(), "must be forwarded");
122 assert(GenCollectedHeap::heap()->is_in_reserved(old), "must be in heap.");
123 assert(!old_gen()->is_in(old), "must be in young generation.");
124
125 objArrayOop obj = objArrayOop(old->forwardee());
126 // Process ParGCArrayScanChunk elements now
127 // and push the remainder back onto queue
128 int start = arrayOop(old)->length();
129 int end = obj->length();
130 int remainder = end - start;
131 assert(start <= end, "just checking");
132 if (remainder > 2 * ParGCArrayScanChunk) {
133 // Test above combines last partial chunk with a full chunk
134 end = start + ParGCArrayScanChunk;
135 arrayOop(old)->set_length(end);
136 // Push remainder.
137 bool ok = work_queue()->push(old);
138 assert(ok, "just popped, push must be okay");
139 } else {
140 // Restore length so that it can be used if there
141 // is a promotion failure and forwarding pointers
142 // must be removed.
143 arrayOop(old)->set_length(end);
144 }
145
146 // process our set of indices (include header in first chunk)
147 // should make sure end is even (aligned to HeapWord in case of compressed oops)
148 if ((HeapWord *)obj < young_old_boundary()) {
149 // object is in to_space
150 obj->oop_iterate_range(&_to_space_closure, start, end);
151 } else {
152 // object is in old generation
153 obj->oop_iterate_range(&_old_gen_closure, start, end);
154 }
155 }
156
157
158 void ParScanThreadState::trim_queues(int max_size) {
159 ObjToScanQueue* queue = work_queue();
160 do {
161 while (queue->size() > (juint)max_size) {
162 oop obj_to_scan;
163 if (queue->pop_local(obj_to_scan)) {
164 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
165 if (obj_to_scan->is_objArray() &&
166 obj_to_scan->is_forwarded() &&
167 obj_to_scan->forwardee() != obj_to_scan) {
168 scan_partial_array_and_push_remainder(obj_to_scan);
169 } else {
170 // object is in to_space
171 obj_to_scan->oop_iterate(&_to_space_closure);
172 }
173 } else {
174 // object is in old generation
175 obj_to_scan->oop_iterate(&_old_gen_closure);
176 }
177 }
178 }
179 // For the case of compressed oops, we have a private, non-shared
180 // overflow stack, so we eagerly drain it so as to more evenly
181 // distribute load early. Note: this may be good to do in
182 // general rather than delay for the final stealing phase.
183 // If applicable, we'll transfer a set of objects over to our
184 // work queue, allowing them to be stolen and draining our
185 // private overflow stack.
186 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
187 }
188
189 bool ParScanThreadState::take_from_overflow_stack() {
190 assert(ParGCUseLocalOverflow, "Else should not call");
191 assert(young_gen()->overflow_list() == NULL, "Error");
192 ObjToScanQueue* queue = work_queue();
193 Stack<oop, mtGC>* const of_stack = overflow_stack();
194 const size_t num_overflow_elems = of_stack->size();
195 const size_t space_available = queue->max_elems() - queue->size();
196 const size_t num_take_elems = MIN3(space_available / 4,
197 ParGCDesiredObjsFromOverflowList,
198 num_overflow_elems);
199 // Transfer the most recent num_take_elems from the overflow
200 // stack to our work queue.
201 for (size_t i = 0; i != num_take_elems; i++) {
202 oop cur = of_stack->pop();
203 oop obj_to_push = cur->forwardee();
204 assert(GenCollectedHeap::heap()->is_in_reserved(cur), "Should be in heap");
205 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
206 assert(GenCollectedHeap::heap()->is_in_reserved(obj_to_push), "Should be in heap");
207 if (should_be_partially_scanned(obj_to_push, cur)) {
208 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
209 obj_to_push = cur;
210 }
211 bool ok = queue->push(obj_to_push);
212 assert(ok, "Should have succeeded");
213 }
214 assert(young_gen()->overflow_list() == NULL, "Error");
215 return num_take_elems > 0; // was something transferred?
216 }
217
218 void ParScanThreadState::push_on_overflow_stack(oop p) {
219 assert(ParGCUseLocalOverflow, "Else should not call");
220 overflow_stack()->push(p);
221 assert(young_gen()->overflow_list() == NULL, "Error");
222 }
223
224 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
225
226 // Otherwise, if the object is small enough, try to reallocate the
227 // buffer.
228 HeapWord* obj = NULL;
229 if (!_to_space_full) {
230 PLAB* const plab = to_space_alloc_buffer();
231 Space* const sp = to_space();
232 if (word_sz * 100 <
233 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),
245 "Invariant");
246 buf_space = sp->par_allocate(buf_size);
247 free_bytes = sp->free();
248 }
249 }
250 if (buf_space != NULL) {
251 plab->set_word_size(buf_size);
252 plab->set_buf(buf_space);
253 record_survivor_plab(buf_space, buf_size);
254 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
255 // Note that we cannot compare buf_size < word_sz below
256 // because of AlignmentReserve (see PLAB::allocate()).
257 assert(obj != NULL || plab->words_remaining() < word_sz,
258 "Else should have been able to allocate");
259 // It's conceivable that we may be able to use the
260 // buffer we just grabbed for subsequent small requests
261 // even if not for this one.
262 } else {
263 // We're used up.
264 _to_space_full = true;
265 }
266
267 } else {
268 // Too large; allocate the object individually.
269 obj = sp->par_allocate(word_sz);
270 }
271 }
272 return obj;
273 }
274
275
276 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) {
277 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
278 }
279
280 void ParScanThreadState::print_promotion_failure_size() {
281 if (_promotion_failed_info.has_failed() && PrintPromotionFailure) {
282 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ",
283 _thread_num, _promotion_failed_info.first_size());
284 }
285 }
286
287 class ParScanThreadStateSet: private ResourceArray {
288 public:
289 // Initializes states for the specified number of threads;
290 ParScanThreadStateSet(int num_threads,
291 Space& to_space,
292 ParNewGeneration& gen,
293 Generation& old_gen,
294 ObjToScanQueueSet& queue_set,
295 Stack<oop, mtGC>* overflow_stacks_,
296 size_t desired_plab_sz,
297 ParallelTaskTerminator& term);
298
299 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
300
301 inline ParScanThreadState& thread_state(int i);
302
303 void trace_promotion_failed(const YoungGCTracer* gc_tracer);
304 void reset(uint active_workers, bool promotion_failed);
305 void flush();
306
307 #if TASKQUEUE_STATS
308 static void
309 print_termination_stats_hdr(outputStream* const st = gclog_or_tty);
310 void print_termination_stats(outputStream* const st = gclog_or_tty);
311 static void
312 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty);
313 void print_taskqueue_stats(outputStream* const st = gclog_or_tty);
314 void reset_stats();
315 #endif // TASKQUEUE_STATS
316
317 private:
318 ParallelTaskTerminator& _term;
319 ParNewGeneration& _gen;
320 Generation& _old_gen;
321 public:
322 bool is_valid(int id) const { return id < length(); }
323 ParallelTaskTerminator* terminator() { return &_term; }
324 };
325
326
327 ParScanThreadStateSet::ParScanThreadStateSet(
328 int num_threads, Space& to_space, ParNewGeneration& gen,
329 Generation& old_gen, ObjToScanQueueSet& queue_set,
330 Stack<oop, mtGC>* overflow_stacks,
331 size_t desired_plab_sz, ParallelTaskTerminator& term)
332 : ResourceArray(sizeof(ParScanThreadState), num_threads),
333 _gen(gen), _old_gen(old_gen), _term(term)
334 {
335 assert(num_threads > 0, "sanity check!");
336 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
337 "overflow_stack allocation mismatch");
338 // Initialize states.
339 for (int i = 0; i < num_threads; ++i) {
340 new ((ParScanThreadState*)_data + i)
341 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set,
342 overflow_stacks, desired_plab_sz, term);
343 }
344 }
345
346 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i)
347 {
348 assert(i >= 0 && i < length(), "sanity check!");
349 return ((ParScanThreadState*)_data)[i];
350 }
351
352 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) {
353 for (int i = 0; i < length(); ++i) {
354 if (thread_state(i).promotion_failed()) {
355 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info());
356 thread_state(i).promotion_failed_info().reset();
357 }
358 }
359 }
360
361 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed)
362 {
363 _term.reset_for_reuse(active_threads);
364 if (promotion_failed) {
365 for (int i = 0; i < length(); ++i) {
366 thread_state(i).print_promotion_failure_size();
367 }
368 }
369 }
370
371 #if TASKQUEUE_STATS
372 void
373 ParScanThreadState::reset_stats()
374 {
375 taskqueue_stats().reset();
376 _term_attempts = 0;
377 _overflow_refills = 0;
378 _overflow_refill_objs = 0;
379 }
380
381 void ParScanThreadStateSet::reset_stats()
382 {
383 for (int i = 0; i < length(); ++i) {
384 thread_state(i).reset_stats();
385 }
386 }
387
388 void
389 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st)
390 {
391 st->print_raw_cr("GC Termination Stats");
392 st->print_raw_cr(" elapsed --strong roots-- "
393 "-------termination-------");
394 st->print_raw_cr("thr ms ms % "
395 " ms % attempts");
396 st->print_raw_cr("--- --------- --------- ------ "
397 "--------- ------ --------");
398 }
399
400 void ParScanThreadStateSet::print_termination_stats(outputStream* const st)
401 {
402 print_termination_stats_hdr(st);
403
404 for (int i = 0; i < length(); ++i) {
405 const ParScanThreadState & pss = thread_state(i);
406 const double elapsed_ms = pss.elapsed_time() * 1000.0;
407 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
408 const double term_ms = pss.term_time() * 1000.0;
409 st->print_cr("%3d %9.2f %9.2f %6.2f "
410 "%9.2f %6.2f " SIZE_FORMAT_W(8),
411 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
412 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
413 }
414 }
415
416 // Print stats related to work queue activity.
417 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st)
418 {
419 st->print_raw_cr("GC Task Stats");
420 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
421 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
422 }
423
424 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st)
425 {
426 print_taskqueue_stats_hdr(st);
427
428 TaskQueueStats totals;
429 for (int i = 0; i < length(); ++i) {
430 const ParScanThreadState & pss = thread_state(i);
431 const TaskQueueStats & stats = pss.taskqueue_stats();
432 st->print("%3d ", i); stats.print(st); st->cr();
433 totals += stats;
434
435 if (pss.overflow_refills() > 0) {
436 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
437 SIZE_FORMAT_W(10) " overflow objects",
438 pss.overflow_refills(), pss.overflow_refill_objs());
439 }
440 }
441 st->print("tot "); totals.print(st); st->cr();
442
443 DEBUG_ONLY(totals.verify());
444 }
445 #endif // TASKQUEUE_STATS
446
447 void ParScanThreadStateSet::flush()
448 {
449 // Work in this loop should be kept as lightweight as
450 // possible since this might otherwise become a bottleneck
451 // to scaling. Should we add heavy-weight work into this
452 // loop, consider parallelizing the loop into the worker threads.
453 for (int i = 0; i < length(); ++i) {
454 ParScanThreadState& par_scan_state = thread_state(i);
455
456 // Flush stats related to To-space PLAB activity and
457 // retire the last buffer.
458 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_gen.plab_stats());
459
460 // Every thread has its own age table. We need to merge
461 // them all into one.
462 ageTable *local_table = par_scan_state.age_table();
463 _gen.age_table()->merge(local_table);
464
465 // Inform old gen that we're done.
466 _old_gen.par_promote_alloc_done(i);
467 _old_gen.par_oop_since_save_marks_iterate_done(i);
468 }
469
470 if (UseConcMarkSweepGC) {
471 // We need to call this even when ResizeOldPLAB is disabled
472 // so as to avoid breaking some asserts. While we may be able
473 // to avoid this by reorganizing the code a bit, I am loathe
474 // to do that unless we find cases where ergo leads to bad
475 // performance.
476 CFLS_LAB::compute_desired_plab_size();
477 }
478 }
479
480 ParScanClosure::ParScanClosure(ParNewGeneration* g,
481 ParScanThreadState* par_scan_state) :
482 OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g)
483 {
484 assert(_g->level() == 0, "Optimized for youngest generation");
485 _boundary = _g->reserved().end();
486 }
487
488 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
489 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
490
491 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
492 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
493
494 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
495 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
496
497 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
498 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
499
500 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
501 ParScanThreadState* par_scan_state)
502 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
503 {}
504
505 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
506 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
507
508 #ifdef WIN32
509 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
510 #endif
511
512 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
513 ParScanThreadState* par_scan_state_,
514 ParScanWithoutBarrierClosure* to_space_closure_,
515 ParScanWithBarrierClosure* old_gen_closure_,
516 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
517 ParNewGeneration* par_gen_,
518 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
519 ObjToScanQueueSet* task_queues_,
520 ParallelTaskTerminator* terminator_) :
521
522 _par_scan_state(par_scan_state_),
523 _to_space_closure(to_space_closure_),
524 _old_gen_closure(old_gen_closure_),
525 _to_space_root_closure(to_space_root_closure_),
526 _old_gen_root_closure(old_gen_root_closure_),
527 _par_gen(par_gen_),
528 _task_queues(task_queues_),
529 _terminator(terminator_)
530 {}
531
532 void ParEvacuateFollowersClosure::do_void() {
533 ObjToScanQueue* work_q = par_scan_state()->work_queue();
534
535 while (true) {
536
537 // Scan to-space and old-gen objs until we run out of both.
538 oop obj_to_scan;
539 par_scan_state()->trim_queues(0);
540
541 // We have no local work, attempt to steal from other threads.
542
543 // attempt to steal work from promoted.
544 if (task_queues()->steal(par_scan_state()->thread_num(),
545 par_scan_state()->hash_seed(),
546 obj_to_scan)) {
547 bool res = work_q->push(obj_to_scan);
548 assert(res, "Empty queue should have room for a push.");
549
550 // if successful, goto Start.
551 continue;
552
553 // try global overflow list.
554 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
555 continue;
556 }
557
558 // Otherwise, offer termination.
559 par_scan_state()->start_term_time();
560 if (terminator()->offer_termination()) break;
561 par_scan_state()->end_term_time();
562 }
563 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
564 "Broken overflow list?");
565 // Finish the last termination pause.
566 par_scan_state()->end_term_time();
567 }
568
569 ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* old_gen,
570 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set,
571 StrongRootsScope* strong_roots_scope) :
572 AbstractGangTask("ParNewGeneration collection"),
573 _gen(gen), _old_gen(old_gen),
574 _young_old_boundary(young_old_boundary),
575 _state_set(state_set),
576 _strong_roots_scope(strong_roots_scope)
577 {}
578
579 void ParNewGenTask::work(uint worker_id) {
580 GenCollectedHeap* gch = GenCollectedHeap::heap();
581 // Since this is being done in a separate thread, need new resource
582 // and handle marks.
583 ResourceMark rm;
584 HandleMark hm;
585
586 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
587 assert(_state_set->is_valid(worker_id), "Should not have been called");
588
589 par_scan_state.set_young_old_boundary(_young_old_boundary);
590
591 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(),
592 gch->rem_set()->klass_rem_set());
593 CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure,
594 &par_scan_state.to_space_root_closure(),
595 false);
596
597 par_scan_state.start_strong_roots();
598 gch->gen_process_roots(_strong_roots_scope,
599 _gen->level(),
600 true, // Process younger gens, if any,
601 // as strong roots.
602 GenCollectedHeap::SO_ScavengeCodeCache,
603 GenCollectedHeap::StrongAndWeakRoots,
604 &par_scan_state.to_space_root_closure(),
605 &par_scan_state.older_gen_closure(),
606 &cld_scan_closure);
607
608 par_scan_state.end_strong_roots();
609
610 // "evacuate followers".
611 par_scan_state.evacuate_followers_closure().do_void();
612 }
613
614 #ifdef _MSC_VER
615 #pragma warning( push )
616 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
617 #endif
618 ParNewGeneration::
619 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level)
620 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"),
621 _overflow_list(NULL),
622 _is_alive_closure(this),
623 _plab_stats(YoungPLABSize, PLABWeight),
624 _should_adjust_cards_per_stride(false)
625 {
626 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
627 NOT_PRODUCT(_num_par_pushes = 0;)
628 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
629 guarantee(_task_queues != NULL, "task_queues allocation failure.");
630
631 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
632 ObjToScanQueue *q = new ObjToScanQueue();
633 guarantee(q != NULL, "work_queue Allocation failure.");
634 _task_queues->register_queue(i1, q);
635 }
636
637 for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
638 _task_queues->queue(i2)->initialize();
639
640 _overflow_stacks = NULL;
641 if (ParGCUseLocalOverflow) {
642
643 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal
644 // with ','
645 typedef Stack<oop, mtGC> GCOopStack;
646
647 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
648 for (size_t i = 0; i < ParallelGCThreads; ++i) {
649 new (_overflow_stacks + i) Stack<oop, mtGC>();
650 }
651 }
652
653 if (UsePerfData) {
654 EXCEPTION_MARK;
655 ResourceMark rm;
656
657 const char* cname =
658 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
659 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
660 ParallelGCThreads, CHECK);
661 }
662
663 // we'll automatically calculate ParGCCardsPerStrideChunk if it has
664 // not been explicitly set on the command line
665 _should_adjust_cards_per_stride = FLAG_IS_DEFAULT(ParGCCardsPerStrideChunk);
666 }
667 #ifdef _MSC_VER
668 #pragma warning( pop )
669 #endif
670
671 // ParNewGeneration::
672 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
673 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
674
675 template <class T>
676 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
677 #ifdef ASSERT
678 {
679 assert(!oopDesc::is_null(*p), "expected non-null ref");
680 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
681 // We never expect to see a null reference being processed
682 // as a weak reference.
683 assert(obj->is_oop(), "expected an oop while scanning weak refs");
684 }
685 #endif // ASSERT
686
687 _par_cl->do_oop_nv(p);
688
689 if (GenCollectedHeap::heap()->is_in_reserved(p)) {
690 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
691 _rs->write_ref_field_gc_par(p, obj);
692 }
693 }
694
695 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
696 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
697
698 // ParNewGeneration::
699 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
700 DefNewGeneration::KeepAliveClosure(cl) {}
701
702 template <class T>
703 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
704 #ifdef ASSERT
705 {
706 assert(!oopDesc::is_null(*p), "expected non-null ref");
707 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
708 // We never expect to see a null reference being processed
709 // as a weak reference.
710 assert(obj->is_oop(), "expected an oop while scanning weak refs");
711 }
712 #endif // ASSERT
713
714 _cl->do_oop_nv(p);
715
716 if (GenCollectedHeap::heap()->is_in_reserved(p)) {
717 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
718 _rs->write_ref_field_gc_par(p, obj);
719 }
720 }
721
722 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
723 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
724
725 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
726 T heap_oop = oopDesc::load_heap_oop(p);
727 if (!oopDesc::is_null(heap_oop)) {
728 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
729 if ((HeapWord*)obj < _boundary) {
730 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
731 oop new_obj = obj->is_forwarded()
732 ? obj->forwardee()
733 : _g->DefNewGeneration::copy_to_survivor_space(obj);
734 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
735 }
736 if (_gc_barrier) {
737 // If p points to a younger generation, mark the card.
738 if ((HeapWord*)obj < _gen_boundary) {
739 _rs->write_ref_field_gc_par(p, obj);
740 }
741 }
742 }
743 }
744
745 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
746 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
747
748 class ParNewRefProcTaskProxy: public AbstractGangTask {
749 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
750 public:
751 ParNewRefProcTaskProxy(ProcessTask& task,
752 ParNewGeneration& gen,
753 Generation& old_gen,
754 HeapWord* young_old_boundary,
755 ParScanThreadStateSet& state_set);
756
757 private:
758 virtual void work(uint worker_id);
759 private:
760 ParNewGeneration& _gen;
761 ProcessTask& _task;
762 Generation& _old_gen;
763 HeapWord* _young_old_boundary;
764 ParScanThreadStateSet& _state_set;
765 };
766
767 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
768 ParNewGeneration& gen,
769 Generation& old_gen,
770 HeapWord* young_old_boundary,
771 ParScanThreadStateSet& state_set)
772 : AbstractGangTask("ParNewGeneration parallel reference processing"),
773 _gen(gen),
774 _task(task),
775 _old_gen(old_gen),
776 _young_old_boundary(young_old_boundary),
777 _state_set(state_set)
778 {
779 }
780
781 void ParNewRefProcTaskProxy::work(uint worker_id)
782 {
783 ResourceMark rm;
784 HandleMark hm;
785 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
786 par_scan_state.set_young_old_boundary(_young_old_boundary);
787 _task.work(worker_id, par_scan_state.is_alive_closure(),
788 par_scan_state.keep_alive_closure(),
789 par_scan_state.evacuate_followers_closure());
790 }
791
792 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
793 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
794 EnqueueTask& _task;
795
796 public:
797 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
798 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
799 _task(task)
800 { }
801
802 virtual void work(uint worker_id)
803 {
804 _task.work(worker_id);
805 }
806 };
807
808
809 void ParNewRefProcTaskExecutor::execute(ProcessTask& task)
810 {
811 GenCollectedHeap* gch = GenCollectedHeap::heap();
812 FlexibleWorkGang* workers = gch->workers();
813 assert(workers != NULL, "Need parallel worker threads.");
814 _state_set.reset(workers->active_workers(), _generation.promotion_failed());
815 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(),
816 _generation.reserved().end(), _state_set);
817 workers->run_task(&rp_task);
818 _state_set.reset(0 /* bad value in debug if not reset */,
819 _generation.promotion_failed());
820 }
821
822 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task)
823 {
824 GenCollectedHeap* gch = GenCollectedHeap::heap();
825 FlexibleWorkGang* workers = gch->workers();
826 assert(workers != NULL, "Need parallel worker threads.");
827 ParNewRefEnqueueTaskProxy enq_task(task);
828 workers->run_task(&enq_task);
829 }
830
831 void ParNewRefProcTaskExecutor::set_single_threaded_mode()
832 {
833 _state_set.flush();
834 GenCollectedHeap* gch = GenCollectedHeap::heap();
835 gch->save_marks();
836 }
837
838 ScanClosureWithParBarrier::
839 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
840 ScanClosure(g, gc_barrier) {}
841
842 EvacuateFollowersClosureGeneral::
843 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
844 OopsInGenClosure* cur,
845 OopsInGenClosure* older) :
846 _gch(gch), _level(level),
847 _scan_cur_or_nonheap(cur), _scan_older(older)
848 {}
849
850 void EvacuateFollowersClosureGeneral::do_void() {
851 do {
852 // Beware: this call will lead to closure applications via virtual
853 // calls.
854 _gch->oop_since_save_marks_iterate(_level,
855 _scan_cur_or_nonheap,
856 _scan_older);
857 } while (!_gch->no_allocs_since_save_marks(_level));
858 }
859
860
861 // A Generation that does parallel young-gen collection.
862
863 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set) {
864 assert(_promo_failure_scan_stack.is_empty(), "post condition");
865 _promo_failure_scan_stack.clear(true); // Clear cached segments.
866
867 remove_forwarding_pointers();
868 if (PrintGCDetails) {
869 gclog_or_tty->print(" (promotion failed)");
870 }
871 // All the spaces are in play for mark-sweep.
872 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
873 from()->set_next_compaction_space(to());
874 gch->set_incremental_collection_failed();
875 // Inform the next generation that a promotion failure occurred.
876 _old_gen->promotion_failure_occurred();
877
878 // Trace promotion failure in the parallel GC threads
879 thread_state_set.trace_promotion_failed(gc_tracer());
880 // Single threaded code may have reported promotion failure to the global state
881 if (_promotion_failed_info.has_failed()) {
882 _gc_tracer.report_promotion_failed(_promotion_failed_info);
883 }
884 // Reset the PromotionFailureALot counters.
885 NOT_PRODUCT(gch->reset_promotion_should_fail();)
886 }
887
888 void ParNewGeneration::adjust_cards_per_stride(const Generation& old_gen) {
889 if (!_should_adjust_cards_per_stride) return;
890
891 const size_t MinOldGenCapacity = G;
892 const size_t MaxOldGenCapacity = 16 * G;
893 assert(MinOldGenCapacity <= MaxOldGenCapacity, "sanity");
894 const size_t OldGenCapacityDiff = MaxOldGenCapacity - MinOldGenCapacity;
895
896 const intx MinStrideChunkSize = 256;
897 const intx MaxStrideChunkSize = 8 * K;
898 assert(MinStrideChunkSize <= MaxStrideChunkSize, "sanity");
899 const intx StrideChunkSizeDiff = MaxStrideChunkSize - MinStrideChunkSize;
900
901 const size_t capacity = old_gen.capacity();
902 // adj_capacity = capacity bounded by {Min,Max}OldGenCapacity
903 size_t adj_capacity = capacity;
904 adj_capacity = MAX2(adj_capacity, MinOldGenCapacity);
905 adj_capacity = MIN2(adj_capacity, MaxOldGenCapacity);
906
907 const size_t capacity_offset = adj_capacity - MinOldGenCapacity;
908 assert(capacity_offset && capacity_offset <= OldGenCapacityDiff,
909 "sanity");
910 const double r = (double) capacity_offset / (double) OldGenCapacityDiff;
911 const intx stride_chunk_offset = (intx) (r * (double) StrideChunkSizeDiff);
912 const int stride_chunk_size_log =
913 log2_intptr((intptr_t) (MinStrideChunkSize + stride_chunk_offset));
914 const intx stride_chunk_size = (intx) 1 << stride_chunk_size_log;
915
916 assert(MinStrideChunkSize <= stride_chunk_size &&
917 stride_chunk_size <= MaxStrideChunkSize &&
918 is_power_of_2((intptr_t) stride_chunk_size),
919 err_msg("stride_chunk_size: "INTX_FORMAT, stride_chunk_size));
920
921 FLAG_SET_ERGO(intx, ParGCCardsPerStrideChunk, stride_chunk_size);
922 }
923
924 void ParNewGeneration::collect(bool full,
925 bool clear_all_soft_refs,
926 size_t size,
927 bool is_tlab) {
928 assert(full || size > 0, "otherwise we don't want to collect");
929
930 GenCollectedHeap* gch = GenCollectedHeap::heap();
931
932 _gc_timer->register_gc_start();
933
934 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
935 FlexibleWorkGang* workers = gch->workers();
936 assert(workers != NULL, "Need workgang for parallel work");
937 uint active_workers =
938 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
939 workers->active_workers(),
940 Threads::number_of_non_daemon_threads());
941 workers->set_active_workers(active_workers);
942 _old_gen = gch->old_gen();
943
944 // If the next generation is too full to accommodate worst-case promotion
945 // from this generation, pass on collection; let the next generation
946 // do it.
947 if (!collection_attempt_is_safe()) {
948 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
949 return;
950 }
951 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
952
953 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
954 gch->trace_heap_before_gc(gc_tracer());
955
956 init_assuming_no_promotion_failure();
957
958 if (UseAdaptiveSizePolicy) {
959 set_survivor_overflow(false);
960 size_policy->minor_collection_begin();
961 }
962
963 GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL, _gc_tracer.gc_id());
964 // Capture heap used before collection (for printing).
965 size_t gch_prev_used = gch->used();
966
967 age_table()->clear();
968 to()->clear(SpaceDecorator::Mangle);
969 adjust_cards_per_stride(*_old_gen);
970
971 gch->save_marks();
972
973 // Set the correct parallelism (number of queues) in the reference processor
974 ref_processor()->set_active_mt_degree(active_workers);
975
976 // Always set the terminator for the active number of workers
977 // because only those workers go through the termination protocol.
978 ParallelTaskTerminator _term(active_workers, task_queues());
979 ParScanThreadStateSet thread_state_set(active_workers,
980 *to(), *this, *_old_gen, *task_queues(),
981 _overflow_stacks, desired_plab_sz(), _term);
982
983 thread_state_set.reset(active_workers, promotion_failed());
984
985 {
986 StrongRootsScope srs(active_workers);
987
988 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs);
989 gch->rem_set()->prepare_for_younger_refs_iterate(true);
990 // It turns out that even when we're using 1 thread, doing the work in a
991 // separate thread causes wide variance in run times. We can't help this
992 // in the multi-threaded case, but we special-case n=1 here to get
993 // repeatable measurements of the 1-thread overhead of the parallel code.
994 if (active_workers > 1) {
995 workers->run_task(&tsk);
996 } else {
997 tsk.work(0);
998 }
999 }
1000
1001 thread_state_set.reset(0 /* Bad value in debug if not reset */,
1002 promotion_failed());
1003
1004 // Trace and reset failed promotion info.
1005 if (promotion_failed()) {
1006 thread_state_set.trace_promotion_failed(gc_tracer());
1007 }
1008
1009 // Process (weak) reference objects found during scavenge.
1010 ReferenceProcessor* rp = ref_processor();
1011 IsAliveClosure is_alive(this);
1012 ScanWeakRefClosure scan_weak_ref(this);
1013 KeepAliveClosure keep_alive(&scan_weak_ref);
1014 ScanClosure scan_without_gc_barrier(this, false);
1015 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
1016 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
1017 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level,
1018 &scan_without_gc_barrier, &scan_with_gc_barrier);
1019 rp->setup_policy(clear_all_soft_refs);
1020 // Can the mt_degree be set later (at run_task() time would be best)?
1021 rp->set_active_mt_degree(active_workers);
1022 ReferenceProcessorStats stats;
1023 if (rp->processing_is_mt()) {
1024 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1025 stats = rp->process_discovered_references(&is_alive, &keep_alive,
1026 &evacuate_followers, &task_executor,
1027 _gc_timer, _gc_tracer.gc_id());
1028 } else {
1029 thread_state_set.flush();
1030 gch->save_marks();
1031 stats = rp->process_discovered_references(&is_alive, &keep_alive,
1032 &evacuate_followers, NULL,
1033 _gc_timer, _gc_tracer.gc_id());
1034 }
1035 _gc_tracer.report_gc_reference_stats(stats);
1036 if (!promotion_failed()) {
1037 // Swap the survivor spaces.
1038 eden()->clear(SpaceDecorator::Mangle);
1039 from()->clear(SpaceDecorator::Mangle);
1040 if (ZapUnusedHeapArea) {
1041 // This is now done here because of the piece-meal mangling which
1042 // can check for valid mangling at intermediate points in the
1043 // collection(s). When a minor collection fails to collect
1044 // sufficient space resizing of the young generation can occur
1045 // an redistribute the spaces in the young generation. Mangle
1046 // here so that unzapped regions don't get distributed to
1047 // other spaces.
1048 to()->mangle_unused_area();
1049 }
1050 swap_spaces();
1051
1052 // A successful scavenge should restart the GC time limit count which is
1053 // for full GC's.
1054 size_policy->reset_gc_overhead_limit_count();
1055
1056 assert(to()->is_empty(), "to space should be empty now");
1057
1058 adjust_desired_tenuring_threshold();
1059 } else {
1060 handle_promotion_failed(gch, thread_state_set);
1061 }
1062 // set new iteration safe limit for the survivor spaces
1063 from()->set_concurrent_iteration_safe_limit(from()->top());
1064 to()->set_concurrent_iteration_safe_limit(to()->top());
1065
1066 if (ResizePLAB) {
1067 plab_stats()->adjust_desired_plab_sz(active_workers);
1068 }
1069
1070 if (PrintGC && !PrintGCDetails) {
1071 gch->print_heap_change(gch_prev_used);
1072 }
1073
1074 TASKQUEUE_STATS_ONLY(if (PrintTerminationStats) thread_state_set.print_termination_stats());
1075 TASKQUEUE_STATS_ONLY(if (PrintTaskqueue) thread_state_set.print_taskqueue_stats());
1076
1077 if (UseAdaptiveSizePolicy) {
1078 size_policy->minor_collection_end(gch->gc_cause());
1079 size_policy->avg_survived()->sample(from()->used());
1080 }
1081
1082 // We need to use a monotonically non-decreasing time in ms
1083 // or we will see time-warp warnings and os::javaTimeMillis()
1084 // does not guarantee monotonicity.
1085 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1086 update_time_of_last_gc(now);
1087
1088 rp->set_enqueuing_is_done(true);
1089 if (rp->processing_is_mt()) {
1090 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1091 rp->enqueue_discovered_references(&task_executor);
1092 } else {
1093 rp->enqueue_discovered_references(NULL);
1094 }
1095 rp->verify_no_references_recorded();
1096
1097 gch->trace_heap_after_gc(gc_tracer());
1098 _gc_tracer.report_tenuring_threshold(tenuring_threshold());
1099
1100 _gc_timer->register_gc_end();
1101
1102 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1103 }
1104
1105 static int sum;
1106 void ParNewGeneration::waste_some_time() {
1107 for (int i = 0; i < 100; i++) {
1108 sum += i;
1109 }
1110 }
1111
1112 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1113
1114 // Because of concurrency, there are times where an object for which
1115 // "is_forwarded()" is true contains an "interim" forwarding pointer
1116 // value. Such a value will soon be overwritten with a real value.
1117 // This method requires "obj" to have a forwarding pointer, and waits, if
1118 // necessary for a real one to be inserted, and returns it.
1119
1120 oop ParNewGeneration::real_forwardee(oop obj) {
1121 oop forward_ptr = obj->forwardee();
1122 if (forward_ptr != ClaimedForwardPtr) {
1123 return forward_ptr;
1124 } else {
1125 return real_forwardee_slow(obj);
1126 }
1127 }
1128
1129 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1130 // Spin-read if it is claimed but not yet written by another thread.
1131 oop forward_ptr = obj->forwardee();
1132 while (forward_ptr == ClaimedForwardPtr) {
1133 waste_some_time();
1134 assert(obj->is_forwarded(), "precondition");
1135 forward_ptr = obj->forwardee();
1136 }
1137 return forward_ptr;
1138 }
1139
1140 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
1141 if (m->must_be_preserved_for_promotion_failure(obj)) {
1142 // We should really have separate per-worker stacks, rather
1143 // than use locking of a common pair of stacks.
1144 MutexLocker ml(ParGCRareEvent_lock);
1145 preserve_mark(obj, m);
1146 }
1147 }
1148
1149 // Multiple GC threads may try to promote an object. If the object
1150 // is successfully promoted, a forwarding pointer will be installed in
1151 // the object in the young generation. This method claims the right
1152 // to install the forwarding pointer before it copies the object,
1153 // thus avoiding the need to undo the copy as in
1154 // copy_to_survivor_space_avoiding_with_undo.
1155
1156 oop ParNewGeneration::copy_to_survivor_space(
1157 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1158 // In the sequential version, this assert also says that the object is
1159 // not forwarded. That might not be the case here. It is the case that
1160 // the caller observed it to be not forwarded at some time in the past.
1161 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1162
1163 // The sequential code read "old->age()" below. That doesn't work here,
1164 // since the age is in the mark word, and that might be overwritten with
1165 // a forwarding pointer by a parallel thread. So we must save the mark
1166 // word in a local and then analyze it.
1167 oopDesc dummyOld;
1168 dummyOld.set_mark(m);
1169 assert(!dummyOld.is_forwarded(),
1170 "should not be called with forwarding pointer mark word.");
1171
1172 oop new_obj = NULL;
1173 oop forward_ptr;
1174
1175 // Try allocating obj in to-space (unless too old)
1176 if (dummyOld.age() < tenuring_threshold()) {
1177 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1178 if (new_obj == NULL) {
1179 set_survivor_overflow(true);
1180 }
1181 }
1182
1183 if (new_obj == NULL) {
1184 // Either to-space is full or we decided to promote
1185 // try allocating obj tenured
1186
1187 // Attempt to install a null forwarding pointer (atomically),
1188 // to claim the right to install the real forwarding pointer.
1189 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1190 if (forward_ptr != NULL) {
1191 // someone else beat us to it.
1192 return real_forwardee(old);
1193 }
1194
1195 if (!_promotion_failed) {
1196 new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1197 old, m, sz);
1198 }
1199
1200 if (new_obj == NULL) {
1201 // promotion failed, forward to self
1202 _promotion_failed = true;
1203 new_obj = old;
1204
1205 preserve_mark_if_necessary(old, m);
1206 par_scan_state->register_promotion_failure(sz);
1207 }
1208
1209 old->forward_to(new_obj);
1210 forward_ptr = NULL;
1211 } else {
1212 // Is in to-space; do copying ourselves.
1213 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1214 assert(GenCollectedHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1215 forward_ptr = old->forward_to_atomic(new_obj);
1216 // Restore the mark word copied above.
1217 new_obj->set_mark(m);
1218 // Increment age if obj still in new generation
1219 new_obj->incr_age();
1220 par_scan_state->age_table()->add(new_obj, sz);
1221 }
1222 assert(new_obj != NULL, "just checking");
1223
1224 #ifndef PRODUCT
1225 // This code must come after the CAS test, or it will print incorrect
1226 // information.
1227 if (TraceScavenge) {
1228 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1229 is_in_reserved(new_obj) ? "copying" : "tenuring",
1230 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size());
1231 }
1232 #endif
1233
1234 if (forward_ptr == NULL) {
1235 oop obj_to_push = new_obj;
1236 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1237 // Length field used as index of next element to be scanned.
1238 // Real length can be obtained from real_forwardee()
1239 arrayOop(old)->set_length(0);
1240 obj_to_push = old;
1241 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1242 "push forwarded object");
1243 }
1244 // Push it on one of the queues of to-be-scanned objects.
1245 bool simulate_overflow = false;
1246 NOT_PRODUCT(
1247 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1248 // simulate a stack overflow
1249 simulate_overflow = true;
1250 }
1251 )
1252 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1253 // Add stats for overflow pushes.
1254 if (Verbose && PrintGCDetails) {
1255 gclog_or_tty->print("queue overflow!\n");
1256 }
1257 push_on_overflow_list(old, par_scan_state);
1258 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1259 }
1260
1261 return new_obj;
1262 }
1263
1264 // Oops. Someone beat us to it. Undo the allocation. Where did we
1265 // allocate it?
1266 if (is_in_reserved(new_obj)) {
1267 // Must be in to_space.
1268 assert(to()->is_in_reserved(new_obj), "Checking");
1269 if (forward_ptr == ClaimedForwardPtr) {
1270 // Wait to get the real forwarding pointer value.
1271 forward_ptr = real_forwardee(old);
1272 }
1273 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1274 }
1275
1276 return forward_ptr;
1277 }
1278
1279 #ifndef PRODUCT
1280 // It's OK to call this multi-threaded; the worst thing
1281 // that can happen is that we'll get a bunch of closely
1282 // spaced simulated overflows, but that's OK, in fact
1283 // probably good as it would exercise the overflow code
1284 // under contention.
1285 bool ParNewGeneration::should_simulate_overflow() {
1286 if (_overflow_counter-- <= 0) { // just being defensive
1287 _overflow_counter = ParGCWorkQueueOverflowInterval;
1288 return true;
1289 } else {
1290 return false;
1291 }
1292 }
1293 #endif
1294
1295 // In case we are using compressed oops, we need to be careful.
1296 // If the object being pushed is an object array, then its length
1297 // field keeps track of the "grey boundary" at which the next
1298 // incremental scan will be done (see ParGCArrayScanChunk).
1299 // When using compressed oops, this length field is kept in the
1300 // lower 32 bits of the erstwhile klass word and cannot be used
1301 // for the overflow chaining pointer (OCP below). As such the OCP
1302 // would itself need to be compressed into the top 32-bits in this
1303 // case. Unfortunately, see below, in the event that we have a
1304 // promotion failure, the node to be pushed on the list can be
1305 // outside of the Java heap, so the heap-based pointer compression
1306 // would not work (we would have potential aliasing between C-heap
1307 // and Java-heap pointers). For this reason, when using compressed
1308 // oops, we simply use a worker-thread-local, non-shared overflow
1309 // list in the form of a growable array, with a slightly different
1310 // overflow stack draining strategy. If/when we start using fat
1311 // stacks here, we can go back to using (fat) pointer chains
1312 // (although some performance comparisons would be useful since
1313 // single global lists have their own performance disadvantages
1314 // as we were made painfully aware not long ago, see 6786503).
1315 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1316 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1317 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1318 if (ParGCUseLocalOverflow) {
1319 // In the case of compressed oops, we use a private, not-shared
1320 // overflow stack.
1321 par_scan_state->push_on_overflow_stack(from_space_obj);
1322 } else {
1323 assert(!UseCompressedOops, "Error");
1324 // if the object has been forwarded to itself, then we cannot
1325 // use the klass pointer for the linked list. Instead we have
1326 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1327 // XXX This is horribly inefficient when a promotion failure occurs
1328 // and should be fixed. XXX FIX ME !!!
1329 #ifndef PRODUCT
1330 Atomic::inc_ptr(&_num_par_pushes);
1331 assert(_num_par_pushes > 0, "Tautology");
1332 #endif
1333 if (from_space_obj->forwardee() == from_space_obj) {
1334 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1335 listhead->forward_to(from_space_obj);
1336 from_space_obj = listhead;
1337 }
1338 oop observed_overflow_list = _overflow_list;
1339 oop cur_overflow_list;
1340 do {
1341 cur_overflow_list = observed_overflow_list;
1342 if (cur_overflow_list != BUSY) {
1343 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1344 } else {
1345 from_space_obj->set_klass_to_list_ptr(NULL);
1346 }
1347 observed_overflow_list =
1348 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1349 } while (cur_overflow_list != observed_overflow_list);
1350 }
1351 }
1352
1353 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1354 bool res;
1355
1356 if (ParGCUseLocalOverflow) {
1357 res = par_scan_state->take_from_overflow_stack();
1358 } else {
1359 assert(!UseCompressedOops, "Error");
1360 res = take_from_overflow_list_work(par_scan_state);
1361 }
1362 return res;
1363 }
1364
1365
1366 // *NOTE*: The overflow list manipulation code here and
1367 // in CMSCollector:: are very similar in shape,
1368 // except that in the CMS case we thread the objects
1369 // directly into the list via their mark word, and do
1370 // not need to deal with special cases below related
1371 // to chunking of object arrays and promotion failure
1372 // handling.
1373 // CR 6797058 has been filed to attempt consolidation of
1374 // the common code.
1375 // Because of the common code, if you make any changes in
1376 // the code below, please check the CMS version to see if
1377 // similar changes might be needed.
1378 // See CMSCollector::par_take_from_overflow_list() for
1379 // more extensive documentation comments.
1380 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1381 ObjToScanQueue* work_q = par_scan_state->work_queue();
1382 // How many to take?
1383 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1384 (size_t)ParGCDesiredObjsFromOverflowList);
1385
1386 assert(!UseCompressedOops, "Error");
1387 assert(par_scan_state->overflow_stack() == NULL, "Error");
1388 if (_overflow_list == NULL) return false;
1389
1390 // Otherwise, there was something there; try claiming the list.
1391 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1392 // Trim off a prefix of at most objsFromOverflow items
1393 Thread* tid = Thread::current();
1394 size_t spin_count = (size_t)ParallelGCThreads;
1395 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1396 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1397 // someone grabbed it before we did ...
1398 // ... we spin for a short while...
1399 os::sleep(tid, sleep_time_millis, false);
1400 if (_overflow_list == NULL) {
1401 // nothing left to take
1402 return false;
1403 } else if (_overflow_list != BUSY) {
1404 // try and grab the prefix
1405 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1406 }
1407 }
1408 if (prefix == NULL || prefix == BUSY) {
1409 // Nothing to take or waited long enough
1410 if (prefix == NULL) {
1411 // Write back the NULL in case we overwrote it with BUSY above
1412 // and it is still the same value.
1413 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1414 }
1415 return false;
1416 }
1417 assert(prefix != NULL && prefix != BUSY, "Error");
1418 size_t i = 1;
1419 oop cur = prefix;
1420 while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1421 i++; cur = cur->list_ptr_from_klass();
1422 }
1423
1424 // Reattach remaining (suffix) to overflow list
1425 if (cur->klass_or_null() == NULL) {
1426 // Write back the NULL in lieu of the BUSY we wrote
1427 // above and it is still the same value.
1428 if (_overflow_list == BUSY) {
1429 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1430 }
1431 } else {
1432 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error");
1433 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list
1434 cur->set_klass_to_list_ptr(NULL); // break off suffix
1435 // It's possible that the list is still in the empty(busy) state
1436 // we left it in a short while ago; in that case we may be
1437 // able to place back the suffix.
1438 oop observed_overflow_list = _overflow_list;
1439 oop cur_overflow_list = observed_overflow_list;
1440 bool attached = false;
1441 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1442 observed_overflow_list =
1443 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1444 if (cur_overflow_list == observed_overflow_list) {
1445 attached = true;
1446 break;
1447 } else cur_overflow_list = observed_overflow_list;
1448 }
1449 if (!attached) {
1450 // Too bad, someone else got in in between; we'll need to do a splice.
1451 // Find the last item of suffix list
1452 oop last = suffix;
1453 while (last->klass_or_null() != NULL) {
1454 last = last->list_ptr_from_klass();
1455 }
1456 // Atomically prepend suffix to current overflow list
1457 observed_overflow_list = _overflow_list;
1458 do {
1459 cur_overflow_list = observed_overflow_list;
1460 if (cur_overflow_list != BUSY) {
1461 // Do the splice ...
1462 last->set_klass_to_list_ptr(cur_overflow_list);
1463 } else { // cur_overflow_list == BUSY
1464 last->set_klass_to_list_ptr(NULL);
1465 }
1466 observed_overflow_list =
1467 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1468 } while (cur_overflow_list != observed_overflow_list);
1469 }
1470 }
1471
1472 // Push objects on prefix list onto this thread's work queue
1473 assert(prefix != NULL && prefix != BUSY, "program logic");
1474 cur = prefix;
1475 ssize_t n = 0;
1476 while (cur != NULL) {
1477 oop obj_to_push = cur->forwardee();
1478 oop next = cur->list_ptr_from_klass();
1479 cur->set_klass(obj_to_push->klass());
1480 // This may be an array object that is self-forwarded. In that case, the list pointer
1481 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1482 if (!is_in_reserved(cur)) {
1483 // This can become a scaling bottleneck when there is work queue overflow coincident
1484 // with promotion failure.
1485 oopDesc* f = cur;
1486 FREE_C_HEAP_ARRAY(oopDesc, f);
1487 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1488 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1489 obj_to_push = cur;
1490 }
1491 bool ok = work_q->push(obj_to_push);
1492 assert(ok, "Should have succeeded");
1493 cur = next;
1494 n++;
1495 }
1496 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1497 #ifndef PRODUCT
1498 assert(_num_par_pushes >= n, "Too many pops?");
1499 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1500 #endif
1501 return true;
1502 }
1503 #undef BUSY
1504
1505 void ParNewGeneration::ref_processor_init() {
1506 if (_ref_processor == NULL) {
1507 // Allocate and initialize a reference processor
1508 _ref_processor =
1509 new ReferenceProcessor(_reserved, // span
1510 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1511 (uint) ParallelGCThreads, // mt processing degree
1512 refs_discovery_is_mt(), // mt discovery
1513 (uint) ParallelGCThreads, // mt discovery degree
1514 refs_discovery_is_atomic(), // atomic_discovery
1515 NULL); // is_alive_non_header
1516 }
1517 }
1518
1519 const char* ParNewGeneration::name() const {
1520 return "par new generation";
1521 }