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 {
625 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
626 NOT_PRODUCT(_num_par_pushes = 0;)
627 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
628 guarantee(_task_queues != NULL, "task_queues allocation failure.");
629
630 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
631 ObjToScanQueue *q = new ObjToScanQueue();
632 guarantee(q != NULL, "work_queue Allocation failure.");
633 _task_queues->register_queue(i1, q);
634 }
635
636 for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
637 _task_queues->queue(i2)->initialize();
638
639 _overflow_stacks = NULL;
640 if (ParGCUseLocalOverflow) {
641
642 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal
643 // with ','
644 typedef Stack<oop, mtGC> GCOopStack;
645
646 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
647 for (size_t i = 0; i < ParallelGCThreads; ++i) {
648 new (_overflow_stacks + i) Stack<oop, mtGC>();
649 }
650 }
651
652 if (UsePerfData) {
653 EXCEPTION_MARK;
654 ResourceMark rm;
655
656 const char* cname =
657 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
658 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
659 ParallelGCThreads, CHECK);
660 }
661 }
662 #ifdef _MSC_VER
663 #pragma warning( pop )
664 #endif
665
666 // ParNewGeneration::
667 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
668 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
669
670 template <class T>
671 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
672 #ifdef ASSERT
673 {
674 assert(!oopDesc::is_null(*p), "expected non-null ref");
675 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
676 // We never expect to see a null reference being processed
677 // as a weak reference.
678 assert(obj->is_oop(), "expected an oop while scanning weak refs");
679 }
680 #endif // ASSERT
681
682 _par_cl->do_oop_nv(p);
683
684 if (GenCollectedHeap::heap()->is_in_reserved(p)) {
685 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
686 _rs->write_ref_field_gc_par(p, obj);
687 }
688 }
689
690 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
691 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
692
693 // ParNewGeneration::
694 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
695 DefNewGeneration::KeepAliveClosure(cl) {}
696
697 template <class T>
698 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
699 #ifdef ASSERT
700 {
701 assert(!oopDesc::is_null(*p), "expected non-null ref");
702 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
703 // We never expect to see a null reference being processed
704 // as a weak reference.
705 assert(obj->is_oop(), "expected an oop while scanning weak refs");
706 }
707 #endif // ASSERT
708
709 _cl->do_oop_nv(p);
710
711 if (GenCollectedHeap::heap()->is_in_reserved(p)) {
712 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
713 _rs->write_ref_field_gc_par(p, obj);
714 }
715 }
716
717 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
718 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
719
720 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
721 T heap_oop = oopDesc::load_heap_oop(p);
722 if (!oopDesc::is_null(heap_oop)) {
723 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
724 if ((HeapWord*)obj < _boundary) {
725 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
726 oop new_obj = obj->is_forwarded()
727 ? obj->forwardee()
728 : _g->DefNewGeneration::copy_to_survivor_space(obj);
729 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
730 }
731 if (_gc_barrier) {
732 // If p points to a younger generation, mark the card.
733 if ((HeapWord*)obj < _gen_boundary) {
734 _rs->write_ref_field_gc_par(p, obj);
735 }
736 }
737 }
738 }
739
740 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
741 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
742
743 class ParNewRefProcTaskProxy: public AbstractGangTask {
744 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
745 public:
746 ParNewRefProcTaskProxy(ProcessTask& task,
747 ParNewGeneration& gen,
748 Generation& old_gen,
749 HeapWord* young_old_boundary,
750 ParScanThreadStateSet& state_set);
751
752 private:
753 virtual void work(uint worker_id);
754 private:
755 ParNewGeneration& _gen;
756 ProcessTask& _task;
757 Generation& _old_gen;
758 HeapWord* _young_old_boundary;
759 ParScanThreadStateSet& _state_set;
760 };
761
762 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
763 ParNewGeneration& gen,
764 Generation& old_gen,
765 HeapWord* young_old_boundary,
766 ParScanThreadStateSet& state_set)
767 : AbstractGangTask("ParNewGeneration parallel reference processing"),
768 _gen(gen),
769 _task(task),
770 _old_gen(old_gen),
771 _young_old_boundary(young_old_boundary),
772 _state_set(state_set)
773 {
774 }
775
776 void ParNewRefProcTaskProxy::work(uint worker_id)
777 {
778 ResourceMark rm;
779 HandleMark hm;
780 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
781 par_scan_state.set_young_old_boundary(_young_old_boundary);
782 _task.work(worker_id, par_scan_state.is_alive_closure(),
783 par_scan_state.keep_alive_closure(),
784 par_scan_state.evacuate_followers_closure());
785 }
786
787 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
788 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
789 EnqueueTask& _task;
790
791 public:
792 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
793 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
794 _task(task)
795 { }
796
797 virtual void work(uint worker_id)
798 {
799 _task.work(worker_id);
800 }
801 };
802
803
804 void ParNewRefProcTaskExecutor::execute(ProcessTask& task)
805 {
806 GenCollectedHeap* gch = GenCollectedHeap::heap();
807 FlexibleWorkGang* workers = gch->workers();
808 assert(workers != NULL, "Need parallel worker threads.");
809 _state_set.reset(workers->active_workers(), _generation.promotion_failed());
810 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(),
811 _generation.reserved().end(), _state_set);
812 workers->run_task(&rp_task);
813 _state_set.reset(0 /* bad value in debug if not reset */,
814 _generation.promotion_failed());
815 }
816
817 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task)
818 {
819 GenCollectedHeap* gch = GenCollectedHeap::heap();
820 FlexibleWorkGang* workers = gch->workers();
821 assert(workers != NULL, "Need parallel worker threads.");
822 ParNewRefEnqueueTaskProxy enq_task(task);
823 workers->run_task(&enq_task);
824 }
825
826 void ParNewRefProcTaskExecutor::set_single_threaded_mode()
827 {
828 _state_set.flush();
829 GenCollectedHeap* gch = GenCollectedHeap::heap();
830 gch->save_marks();
831 }
832
833 ScanClosureWithParBarrier::
834 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
835 ScanClosure(g, gc_barrier) {}
836
837 EvacuateFollowersClosureGeneral::
838 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
839 OopsInGenClosure* cur,
840 OopsInGenClosure* older) :
841 _gch(gch), _level(level),
842 _scan_cur_or_nonheap(cur), _scan_older(older)
843 {}
844
845 void EvacuateFollowersClosureGeneral::do_void() {
846 do {
847 // Beware: this call will lead to closure applications via virtual
848 // calls.
849 _gch->oop_since_save_marks_iterate(_level,
850 _scan_cur_or_nonheap,
851 _scan_older);
852 } while (!_gch->no_allocs_since_save_marks(_level));
853 }
854
855
856 // A Generation that does parallel young-gen collection.
857
858 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set) {
859 assert(_promo_failure_scan_stack.is_empty(), "post condition");
860 _promo_failure_scan_stack.clear(true); // Clear cached segments.
861
862 remove_forwarding_pointers();
863 if (PrintGCDetails) {
864 gclog_or_tty->print(" (promotion failed)");
865 }
866 // All the spaces are in play for mark-sweep.
867 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
868 from()->set_next_compaction_space(to());
869 gch->set_incremental_collection_failed();
870 // Inform the next generation that a promotion failure occurred.
871 _old_gen->promotion_failure_occurred();
872
873 // Trace promotion failure in the parallel GC threads
874 thread_state_set.trace_promotion_failed(gc_tracer());
875 // Single threaded code may have reported promotion failure to the global state
876 if (_promotion_failed_info.has_failed()) {
877 _gc_tracer.report_promotion_failed(_promotion_failed_info);
878 }
879 // Reset the PromotionFailureALot counters.
880 NOT_PRODUCT(gch->reset_promotion_should_fail();)
881 }
882
883 void ParNewGeneration::collect(bool full,
884 bool clear_all_soft_refs,
885 size_t size,
886 bool is_tlab) {
887 assert(full || size > 0, "otherwise we don't want to collect");
888
889 GenCollectedHeap* gch = GenCollectedHeap::heap();
890
891 _gc_timer->register_gc_start();
892
893 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
894 FlexibleWorkGang* workers = gch->workers();
895 assert(workers != NULL, "Need workgang for parallel work");
896 uint active_workers =
897 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
898 workers->active_workers(),
899 Threads::number_of_non_daemon_threads());
900 workers->set_active_workers(active_workers);
901 _old_gen = gch->old_gen();
902
903 // If the next generation is too full to accommodate worst-case promotion
904 // from this generation, pass on collection; let the next generation
905 // do it.
906 if (!collection_attempt_is_safe()) {
907 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
908 return;
909 }
910 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
911
912 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
913 gch->trace_heap_before_gc(gc_tracer());
914
915 init_assuming_no_promotion_failure();
916
917 if (UseAdaptiveSizePolicy) {
918 set_survivor_overflow(false);
919 size_policy->minor_collection_begin();
920 }
921
922 GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL, _gc_tracer.gc_id());
923 // Capture heap used before collection (for printing).
924 size_t gch_prev_used = gch->used();
925
926 age_table()->clear();
927 to()->clear(SpaceDecorator::Mangle);
928
929 gch->save_marks();
930
931 // Set the correct parallelism (number of queues) in the reference processor
932 ref_processor()->set_active_mt_degree(active_workers);
933
934 // Always set the terminator for the active number of workers
935 // because only those workers go through the termination protocol.
936 ParallelTaskTerminator _term(active_workers, task_queues());
937 ParScanThreadStateSet thread_state_set(active_workers,
938 *to(), *this, *_old_gen, *task_queues(),
939 _overflow_stacks, desired_plab_sz(), _term);
940
941 thread_state_set.reset(active_workers, promotion_failed());
942
943 {
944 StrongRootsScope srs(active_workers);
945
946 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs);
947 gch->rem_set()->prepare_for_younger_refs_iterate(true);
948 // It turns out that even when we're using 1 thread, doing the work in a
949 // separate thread causes wide variance in run times. We can't help this
950 // in the multi-threaded case, but we special-case n=1 here to get
951 // repeatable measurements of the 1-thread overhead of the parallel code.
952 if (active_workers > 1) {
953 workers->run_task(&tsk);
954 } else {
955 tsk.work(0);
956 }
957 }
958
959 thread_state_set.reset(0 /* Bad value in debug if not reset */,
960 promotion_failed());
961
962 // Trace and reset failed promotion info.
963 if (promotion_failed()) {
964 thread_state_set.trace_promotion_failed(gc_tracer());
965 }
966
967 // Process (weak) reference objects found during scavenge.
968 ReferenceProcessor* rp = ref_processor();
969 IsAliveClosure is_alive(this);
970 ScanWeakRefClosure scan_weak_ref(this);
971 KeepAliveClosure keep_alive(&scan_weak_ref);
972 ScanClosure scan_without_gc_barrier(this, false);
973 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
974 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
975 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level,
976 &scan_without_gc_barrier, &scan_with_gc_barrier);
977 rp->setup_policy(clear_all_soft_refs);
978 // Can the mt_degree be set later (at run_task() time would be best)?
979 rp->set_active_mt_degree(active_workers);
980 ReferenceProcessorStats stats;
981 if (rp->processing_is_mt()) {
982 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
983 stats = rp->process_discovered_references(&is_alive, &keep_alive,
984 &evacuate_followers, &task_executor,
985 _gc_timer, _gc_tracer.gc_id());
986 } else {
987 thread_state_set.flush();
988 gch->save_marks();
989 stats = rp->process_discovered_references(&is_alive, &keep_alive,
990 &evacuate_followers, NULL,
991 _gc_timer, _gc_tracer.gc_id());
992 }
993 _gc_tracer.report_gc_reference_stats(stats);
994 if (!promotion_failed()) {
995 // Swap the survivor spaces.
996 eden()->clear(SpaceDecorator::Mangle);
997 from()->clear(SpaceDecorator::Mangle);
998 if (ZapUnusedHeapArea) {
999 // This is now done here because of the piece-meal mangling which
1000 // can check for valid mangling at intermediate points in the
1001 // collection(s). When a minor collection fails to collect
1002 // sufficient space resizing of the young generation can occur
1003 // an redistribute the spaces in the young generation. Mangle
1004 // here so that unzapped regions don't get distributed to
1005 // other spaces.
1006 to()->mangle_unused_area();
1007 }
1008 swap_spaces();
1009
1010 // A successful scavenge should restart the GC time limit count which is
1011 // for full GC's.
1012 size_policy->reset_gc_overhead_limit_count();
1013
1014 assert(to()->is_empty(), "to space should be empty now");
1015
1016 adjust_desired_tenuring_threshold();
1017 } else {
1018 handle_promotion_failed(gch, thread_state_set);
1019 }
1020 // set new iteration safe limit for the survivor spaces
1021 from()->set_concurrent_iteration_safe_limit(from()->top());
1022 to()->set_concurrent_iteration_safe_limit(to()->top());
1023
1024 if (ResizePLAB) {
1025 plab_stats()->adjust_desired_plab_sz(active_workers);
1026 }
1027
1028 if (PrintGC && !PrintGCDetails) {
1029 gch->print_heap_change(gch_prev_used);
1030 }
1031
1032 TASKQUEUE_STATS_ONLY(if (PrintTerminationStats) thread_state_set.print_termination_stats());
1033 TASKQUEUE_STATS_ONLY(if (PrintTaskqueue) thread_state_set.print_taskqueue_stats());
1034
1035 if (UseAdaptiveSizePolicy) {
1036 size_policy->minor_collection_end(gch->gc_cause());
1037 size_policy->avg_survived()->sample(from()->used());
1038 }
1039
1040 // We need to use a monotonically non-decreasing time in ms
1041 // or we will see time-warp warnings and os::javaTimeMillis()
1042 // does not guarantee monotonicity.
1043 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1044 update_time_of_last_gc(now);
1045
1046 rp->set_enqueuing_is_done(true);
1047 if (rp->processing_is_mt()) {
1048 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1049 rp->enqueue_discovered_references(&task_executor);
1050 } else {
1051 rp->enqueue_discovered_references(NULL);
1052 }
1053 rp->verify_no_references_recorded();
1054
1055 gch->trace_heap_after_gc(gc_tracer());
1056 _gc_tracer.report_tenuring_threshold(tenuring_threshold());
1057
1058 _gc_timer->register_gc_end();
1059
1060 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1061 }
1062
1063 static int sum;
1064 void ParNewGeneration::waste_some_time() {
1065 for (int i = 0; i < 100; i++) {
1066 sum += i;
1067 }
1068 }
1069
1070 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1071
1072 // Because of concurrency, there are times where an object for which
1073 // "is_forwarded()" is true contains an "interim" forwarding pointer
1074 // value. Such a value will soon be overwritten with a real value.
1075 // This method requires "obj" to have a forwarding pointer, and waits, if
1076 // necessary for a real one to be inserted, and returns it.
1077
1078 oop ParNewGeneration::real_forwardee(oop obj) {
1079 oop forward_ptr = obj->forwardee();
1080 if (forward_ptr != ClaimedForwardPtr) {
1081 return forward_ptr;
1082 } else {
1083 return real_forwardee_slow(obj);
1084 }
1085 }
1086
1087 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1088 // Spin-read if it is claimed but not yet written by another thread.
1089 oop forward_ptr = obj->forwardee();
1090 while (forward_ptr == ClaimedForwardPtr) {
1091 waste_some_time();
1092 assert(obj->is_forwarded(), "precondition");
1093 forward_ptr = obj->forwardee();
1094 }
1095 return forward_ptr;
1096 }
1097
1098 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
1099 if (m->must_be_preserved_for_promotion_failure(obj)) {
1100 // We should really have separate per-worker stacks, rather
1101 // than use locking of a common pair of stacks.
1102 MutexLocker ml(ParGCRareEvent_lock);
1103 preserve_mark(obj, m);
1104 }
1105 }
1106
1107 // Multiple GC threads may try to promote an object. If the object
1108 // is successfully promoted, a forwarding pointer will be installed in
1109 // the object in the young generation. This method claims the right
1110 // to install the forwarding pointer before it copies the object,
1111 // thus avoiding the need to undo the copy as in
1112 // copy_to_survivor_space_avoiding_with_undo.
1113
1114 oop ParNewGeneration::copy_to_survivor_space(
1115 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1116 // In the sequential version, this assert also says that the object is
1117 // not forwarded. That might not be the case here. It is the case that
1118 // the caller observed it to be not forwarded at some time in the past.
1119 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1120
1121 // The sequential code read "old->age()" below. That doesn't work here,
1122 // since the age is in the mark word, and that might be overwritten with
1123 // a forwarding pointer by a parallel thread. So we must save the mark
1124 // word in a local and then analyze it.
1125 oopDesc dummyOld;
1126 dummyOld.set_mark(m);
1127 assert(!dummyOld.is_forwarded(),
1128 "should not be called with forwarding pointer mark word.");
1129
1130 oop new_obj = NULL;
1131 oop forward_ptr;
1132
1133 // Try allocating obj in to-space (unless too old)
1134 if (dummyOld.age() < tenuring_threshold()) {
1135 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1136 if (new_obj == NULL) {
1137 set_survivor_overflow(true);
1138 }
1139 }
1140
1141 if (new_obj == NULL) {
1142 // Either to-space is full or we decided to promote
1143 // try allocating obj tenured
1144
1145 // Attempt to install a null forwarding pointer (atomically),
1146 // to claim the right to install the real forwarding pointer.
1147 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1148 if (forward_ptr != NULL) {
1149 // someone else beat us to it.
1150 return real_forwardee(old);
1151 }
1152
1153 if (!_promotion_failed) {
1154 new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1155 old, m, sz);
1156 }
1157
1158 if (new_obj == NULL) {
1159 // promotion failed, forward to self
1160 _promotion_failed = true;
1161 new_obj = old;
1162
1163 preserve_mark_if_necessary(old, m);
1164 par_scan_state->register_promotion_failure(sz);
1165 }
1166
1167 old->forward_to(new_obj);
1168 forward_ptr = NULL;
1169 } else {
1170 // Is in to-space; do copying ourselves.
1171 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1172 assert(GenCollectedHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1173 forward_ptr = old->forward_to_atomic(new_obj);
1174 // Restore the mark word copied above.
1175 new_obj->set_mark(m);
1176 // Increment age if obj still in new generation
1177 new_obj->incr_age();
1178 par_scan_state->age_table()->add(new_obj, sz);
1179 }
1180 assert(new_obj != NULL, "just checking");
1181
1182 #ifndef PRODUCT
1183 // This code must come after the CAS test, or it will print incorrect
1184 // information.
1185 if (TraceScavenge) {
1186 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1187 is_in_reserved(new_obj) ? "copying" : "tenuring",
1188 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size());
1189 }
1190 #endif
1191
1192 if (forward_ptr == NULL) {
1193 oop obj_to_push = new_obj;
1194 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1195 // Length field used as index of next element to be scanned.
1196 // Real length can be obtained from real_forwardee()
1197 arrayOop(old)->set_length(0);
1198 obj_to_push = old;
1199 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1200 "push forwarded object");
1201 }
1202 // Push it on one of the queues of to-be-scanned objects.
1203 bool simulate_overflow = false;
1204 NOT_PRODUCT(
1205 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1206 // simulate a stack overflow
1207 simulate_overflow = true;
1208 }
1209 )
1210 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1211 // Add stats for overflow pushes.
1212 if (Verbose && PrintGCDetails) {
1213 gclog_or_tty->print("queue overflow!\n");
1214 }
1215 push_on_overflow_list(old, par_scan_state);
1216 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1217 }
1218
1219 return new_obj;
1220 }
1221
1222 // Oops. Someone beat us to it. Undo the allocation. Where did we
1223 // allocate it?
1224 if (is_in_reserved(new_obj)) {
1225 // Must be in to_space.
1226 assert(to()->is_in_reserved(new_obj), "Checking");
1227 if (forward_ptr == ClaimedForwardPtr) {
1228 // Wait to get the real forwarding pointer value.
1229 forward_ptr = real_forwardee(old);
1230 }
1231 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1232 }
1233
1234 return forward_ptr;
1235 }
1236
1237 #ifndef PRODUCT
1238 // It's OK to call this multi-threaded; the worst thing
1239 // that can happen is that we'll get a bunch of closely
1240 // spaced simulated overflows, but that's OK, in fact
1241 // probably good as it would exercise the overflow code
1242 // under contention.
1243 bool ParNewGeneration::should_simulate_overflow() {
1244 if (_overflow_counter-- <= 0) { // just being defensive
1245 _overflow_counter = ParGCWorkQueueOverflowInterval;
1246 return true;
1247 } else {
1248 return false;
1249 }
1250 }
1251 #endif
1252
1253 // In case we are using compressed oops, we need to be careful.
1254 // If the object being pushed is an object array, then its length
1255 // field keeps track of the "grey boundary" at which the next
1256 // incremental scan will be done (see ParGCArrayScanChunk).
1257 // When using compressed oops, this length field is kept in the
1258 // lower 32 bits of the erstwhile klass word and cannot be used
1259 // for the overflow chaining pointer (OCP below). As such the OCP
1260 // would itself need to be compressed into the top 32-bits in this
1261 // case. Unfortunately, see below, in the event that we have a
1262 // promotion failure, the node to be pushed on the list can be
1263 // outside of the Java heap, so the heap-based pointer compression
1264 // would not work (we would have potential aliasing between C-heap
1265 // and Java-heap pointers). For this reason, when using compressed
1266 // oops, we simply use a worker-thread-local, non-shared overflow
1267 // list in the form of a growable array, with a slightly different
1268 // overflow stack draining strategy. If/when we start using fat
1269 // stacks here, we can go back to using (fat) pointer chains
1270 // (although some performance comparisons would be useful since
1271 // single global lists have their own performance disadvantages
1272 // as we were made painfully aware not long ago, see 6786503).
1273 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1274 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1275 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1276 if (ParGCUseLocalOverflow) {
1277 // In the case of compressed oops, we use a private, not-shared
1278 // overflow stack.
1279 par_scan_state->push_on_overflow_stack(from_space_obj);
1280 } else {
1281 assert(!UseCompressedOops, "Error");
1282 // if the object has been forwarded to itself, then we cannot
1283 // use the klass pointer for the linked list. Instead we have
1284 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1285 // XXX This is horribly inefficient when a promotion failure occurs
1286 // and should be fixed. XXX FIX ME !!!
1287 #ifndef PRODUCT
1288 Atomic::inc_ptr(&_num_par_pushes);
1289 assert(_num_par_pushes > 0, "Tautology");
1290 #endif
1291 if (from_space_obj->forwardee() == from_space_obj) {
1292 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1293 listhead->forward_to(from_space_obj);
1294 from_space_obj = listhead;
1295 }
1296 oop observed_overflow_list = _overflow_list;
1297 oop cur_overflow_list;
1298 do {
1299 cur_overflow_list = observed_overflow_list;
1300 if (cur_overflow_list != BUSY) {
1301 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1302 } else {
1303 from_space_obj->set_klass_to_list_ptr(NULL);
1304 }
1305 observed_overflow_list =
1306 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1307 } while (cur_overflow_list != observed_overflow_list);
1308 }
1309 }
1310
1311 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1312 bool res;
1313
1314 if (ParGCUseLocalOverflow) {
1315 res = par_scan_state->take_from_overflow_stack();
1316 } else {
1317 assert(!UseCompressedOops, "Error");
1318 res = take_from_overflow_list_work(par_scan_state);
1319 }
1320 return res;
1321 }
1322
1323
1324 // *NOTE*: The overflow list manipulation code here and
1325 // in CMSCollector:: are very similar in shape,
1326 // except that in the CMS case we thread the objects
1327 // directly into the list via their mark word, and do
1328 // not need to deal with special cases below related
1329 // to chunking of object arrays and promotion failure
1330 // handling.
1331 // CR 6797058 has been filed to attempt consolidation of
1332 // the common code.
1333 // Because of the common code, if you make any changes in
1334 // the code below, please check the CMS version to see if
1335 // similar changes might be needed.
1336 // See CMSCollector::par_take_from_overflow_list() for
1337 // more extensive documentation comments.
1338 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1339 ObjToScanQueue* work_q = par_scan_state->work_queue();
1340 // How many to take?
1341 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1342 (size_t)ParGCDesiredObjsFromOverflowList);
1343
1344 assert(!UseCompressedOops, "Error");
1345 assert(par_scan_state->overflow_stack() == NULL, "Error");
1346 if (_overflow_list == NULL) return false;
1347
1348 // Otherwise, there was something there; try claiming the list.
1349 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1350 // Trim off a prefix of at most objsFromOverflow items
1351 Thread* tid = Thread::current();
1352 size_t spin_count = (size_t)ParallelGCThreads;
1353 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1354 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1355 // someone grabbed it before we did ...
1356 // ... we spin for a short while...
1357 os::sleep(tid, sleep_time_millis, false);
1358 if (_overflow_list == NULL) {
1359 // nothing left to take
1360 return false;
1361 } else if (_overflow_list != BUSY) {
1362 // try and grab the prefix
1363 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1364 }
1365 }
1366 if (prefix == NULL || prefix == BUSY) {
1367 // Nothing to take or waited long enough
1368 if (prefix == NULL) {
1369 // Write back the NULL in case we overwrote it with BUSY above
1370 // and it is still the same value.
1371 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1372 }
1373 return false;
1374 }
1375 assert(prefix != NULL && prefix != BUSY, "Error");
1376 size_t i = 1;
1377 oop cur = prefix;
1378 while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1379 i++; cur = cur->list_ptr_from_klass();
1380 }
1381
1382 // Reattach remaining (suffix) to overflow list
1383 if (cur->klass_or_null() == NULL) {
1384 // Write back the NULL in lieu of the BUSY we wrote
1385 // above and it is still the same value.
1386 if (_overflow_list == BUSY) {
1387 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1388 }
1389 } else {
1390 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error");
1391 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list
1392 cur->set_klass_to_list_ptr(NULL); // break off suffix
1393 // It's possible that the list is still in the empty(busy) state
1394 // we left it in a short while ago; in that case we may be
1395 // able to place back the suffix.
1396 oop observed_overflow_list = _overflow_list;
1397 oop cur_overflow_list = observed_overflow_list;
1398 bool attached = false;
1399 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1400 observed_overflow_list =
1401 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1402 if (cur_overflow_list == observed_overflow_list) {
1403 attached = true;
1404 break;
1405 } else cur_overflow_list = observed_overflow_list;
1406 }
1407 if (!attached) {
1408 // Too bad, someone else got in in between; we'll need to do a splice.
1409 // Find the last item of suffix list
1410 oop last = suffix;
1411 while (last->klass_or_null() != NULL) {
1412 last = last->list_ptr_from_klass();
1413 }
1414 // Atomically prepend suffix to current overflow list
1415 observed_overflow_list = _overflow_list;
1416 do {
1417 cur_overflow_list = observed_overflow_list;
1418 if (cur_overflow_list != BUSY) {
1419 // Do the splice ...
1420 last->set_klass_to_list_ptr(cur_overflow_list);
1421 } else { // cur_overflow_list == BUSY
1422 last->set_klass_to_list_ptr(NULL);
1423 }
1424 observed_overflow_list =
1425 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1426 } while (cur_overflow_list != observed_overflow_list);
1427 }
1428 }
1429
1430 // Push objects on prefix list onto this thread's work queue
1431 assert(prefix != NULL && prefix != BUSY, "program logic");
1432 cur = prefix;
1433 ssize_t n = 0;
1434 while (cur != NULL) {
1435 oop obj_to_push = cur->forwardee();
1436 oop next = cur->list_ptr_from_klass();
1437 cur->set_klass(obj_to_push->klass());
1438 // This may be an array object that is self-forwarded. In that case, the list pointer
1439 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1440 if (!is_in_reserved(cur)) {
1441 // This can become a scaling bottleneck when there is work queue overflow coincident
1442 // with promotion failure.
1443 oopDesc* f = cur;
1444 FREE_C_HEAP_ARRAY(oopDesc, f);
1445 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1446 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1447 obj_to_push = cur;
1448 }
1449 bool ok = work_q->push(obj_to_push);
1450 assert(ok, "Should have succeeded");
1451 cur = next;
1452 n++;
1453 }
1454 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1455 #ifndef PRODUCT
1456 assert(_num_par_pushes >= n, "Too many pops?");
1457 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1458 #endif
1459 return true;
1460 }
1461 #undef BUSY
1462
1463 void ParNewGeneration::ref_processor_init() {
1464 if (_ref_processor == NULL) {
1465 // Allocate and initialize a reference processor
1466 _ref_processor =
1467 new ReferenceProcessor(_reserved, // span
1468 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1469 (uint) ParallelGCThreads, // mt processing degree
1470 refs_discovery_is_mt(), // mt discovery
1471 (uint) ParallelGCThreads, // mt discovery degree
1472 refs_discovery_is_atomic(), // atomic_discovery
1473 NULL); // is_alive_non_header
1474 }
1475 }
1476
1477 const char* ParNewGeneration::name() const {
1478 return "par new generation";
1479 }