1 /*
2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp"
27 #include "gc_implementation/parNew/parGCAllocBuffer.hpp"
28 #include "gc_implementation/parNew/parNewGeneration.hpp"
29 #include "gc_implementation/parNew/parOopClosures.inline.hpp"
30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp"
31 #include "gc_implementation/shared/ageTable.hpp"
32 #include "gc_implementation/shared/spaceDecorator.hpp"
33 #include "memory/defNewGeneration.inline.hpp"
34 #include "memory/genCollectedHeap.hpp"
35 #include "memory/genOopClosures.inline.hpp"
36 #include "memory/generation.hpp"
37 #include "memory/generation.inline.hpp"
38 #include "memory/referencePolicy.hpp"
39 #include "memory/resourceArea.hpp"
40 #include "memory/sharedHeap.hpp"
41 #include "memory/space.hpp"
42 #include "oops/objArrayOop.hpp"
43 #include "oops/oop.inline.hpp"
44 #include "oops/oop.pcgc.inline.hpp"
45 #include "runtime/handles.hpp"
46 #include "runtime/handles.inline.hpp"
47 #include "runtime/java.hpp"
48 #include "runtime/thread.hpp"
49 #include "utilities/copy.hpp"
50 #include "utilities/globalDefinitions.hpp"
51 #include "utilities/workgroup.hpp"
52
53 #ifdef _MSC_VER
54 #pragma warning( push )
55 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
56 #endif
57 ParScanThreadState::ParScanThreadState(Space* to_space_,
58 ParNewGeneration* gen_,
59 Generation* old_gen_,
60 int thread_num_,
61 ObjToScanQueueSet* work_queue_set_,
62 Stack<oop>* overflow_stacks_,
63 size_t desired_plab_sz_,
64 ParallelTaskTerminator& term_) :
65 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_),
66 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false),
67 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
68 _ageTable(false), // false ==> not the global age table, no perf data.
69 _to_space_alloc_buffer(desired_plab_sz_),
70 _to_space_closure(gen_, this), _old_gen_closure(gen_, this),
71 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this),
72 _older_gen_closure(gen_, this),
73 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
74 &_to_space_root_closure, gen_, &_old_gen_root_closure,
75 work_queue_set_, &term_),
76 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this),
77 _keep_alive_closure(&_scan_weak_ref_closure),
78 _promotion_failure_size(0),
79 _strong_roots_time(0.0), _term_time(0.0)
80 {
81 #if TASKQUEUE_STATS
82 _term_attempts = 0;
83 _overflow_refills = 0;
84 _overflow_refill_objs = 0;
85 #endif // TASKQUEUE_STATS
86
87 _survivor_chunk_array =
88 (ChunkArray*) old_gen()->get_data_recorder(thread_num());
89 _hash_seed = 17; // Might want to take time-based random value.
90 _start = os::elapsedTime();
91 _old_gen_closure.set_generation(old_gen_);
92 _old_gen_root_closure.set_generation(old_gen_);
93 }
94 #ifdef _MSC_VER
95 #pragma warning( pop )
96 #endif
97
98 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
99 size_t plab_word_size) {
100 ChunkArray* sca = survivor_chunk_array();
101 if (sca != NULL) {
102 // A non-null SCA implies that we want the PLAB data recorded.
103 sca->record_sample(plab_start, plab_word_size);
104 }
105 }
106
107 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
108 return new_obj->is_objArray() &&
109 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
110 new_obj != old_obj;
111 }
112
113 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
114 assert(old->is_objArray(), "must be obj array");
115 assert(old->is_forwarded(), "must be forwarded");
116 assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
117 assert(!old_gen()->is_in(old), "must be in young generation.");
118
119 objArrayOop obj = objArrayOop(old->forwardee());
120 // Process ParGCArrayScanChunk elements now
121 // and push the remainder back onto queue
122 int start = arrayOop(old)->length();
123 int end = obj->length();
124 int remainder = end - start;
125 assert(start <= end, "just checking");
126 if (remainder > 2 * ParGCArrayScanChunk) {
127 // Test above combines last partial chunk with a full chunk
128 end = start + ParGCArrayScanChunk;
129 arrayOop(old)->set_length(end);
130 // Push remainder.
131 bool ok = work_queue()->push(old);
132 assert(ok, "just popped, push must be okay");
133 } else {
134 // Restore length so that it can be used if there
135 // is a promotion failure and forwarding pointers
136 // must be removed.
137 arrayOop(old)->set_length(end);
138 }
139
140 // process our set of indices (include header in first chunk)
141 // should make sure end is even (aligned to HeapWord in case of compressed oops)
142 if ((HeapWord *)obj < young_old_boundary()) {
143 // object is in to_space
144 obj->oop_iterate_range(&_to_space_closure, start, end);
145 } else {
146 // object is in old generation
147 obj->oop_iterate_range(&_old_gen_closure, start, end);
148 }
149 }
150
151
152 void ParScanThreadState::trim_queues(int max_size) {
153 ObjToScanQueue* queue = work_queue();
154 do {
155 while (queue->size() > (juint)max_size) {
156 oop obj_to_scan;
157 if (queue->pop_local(obj_to_scan)) {
158 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
159 if (obj_to_scan->is_objArray() &&
160 obj_to_scan->is_forwarded() &&
161 obj_to_scan->forwardee() != obj_to_scan) {
162 scan_partial_array_and_push_remainder(obj_to_scan);
163 } else {
164 // object is in to_space
165 obj_to_scan->oop_iterate(&_to_space_closure);
166 }
167 } else {
168 // object is in old generation
169 obj_to_scan->oop_iterate(&_old_gen_closure);
170 }
171 }
172 }
173 // For the case of compressed oops, we have a private, non-shared
174 // overflow stack, so we eagerly drain it so as to more evenly
175 // distribute load early. Note: this may be good to do in
176 // general rather than delay for the final stealing phase.
177 // If applicable, we'll transfer a set of objects over to our
178 // work queue, allowing them to be stolen and draining our
179 // private overflow stack.
180 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
181 }
182
183 bool ParScanThreadState::take_from_overflow_stack() {
184 assert(ParGCUseLocalOverflow, "Else should not call");
185 assert(young_gen()->overflow_list() == NULL, "Error");
186 ObjToScanQueue* queue = work_queue();
187 Stack<oop>* const of_stack = overflow_stack();
188 const size_t num_overflow_elems = of_stack->size();
189 const size_t space_available = queue->max_elems() - queue->size();
190 const size_t num_take_elems = MIN3(space_available / 4,
191 ParGCDesiredObjsFromOverflowList,
192 num_overflow_elems);
193 // Transfer the most recent num_take_elems from the overflow
194 // stack to our work queue.
195 for (size_t i = 0; i != num_take_elems; i++) {
196 oop cur = of_stack->pop();
197 oop obj_to_push = cur->forwardee();
198 assert(Universe::heap()->is_in_reserved(cur), "Should be in heap");
199 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
200 assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap");
201 if (should_be_partially_scanned(obj_to_push, cur)) {
202 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
203 obj_to_push = cur;
204 }
205 bool ok = queue->push(obj_to_push);
206 assert(ok, "Should have succeeded");
207 }
208 assert(young_gen()->overflow_list() == NULL, "Error");
209 return num_take_elems > 0; // was something transferred?
210 }
211
212 void ParScanThreadState::push_on_overflow_stack(oop p) {
213 assert(ParGCUseLocalOverflow, "Else should not call");
214 overflow_stack()->push(p);
215 assert(young_gen()->overflow_list() == NULL, "Error");
216 }
217
218 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
219
220 // Otherwise, if the object is small enough, try to reallocate the
221 // buffer.
222 HeapWord* obj = NULL;
223 if (!_to_space_full) {
224 ParGCAllocBuffer* const plab = to_space_alloc_buffer();
225 Space* const sp = to_space();
226 if (word_sz * 100 <
227 ParallelGCBufferWastePct * plab->word_sz()) {
228 // Is small enough; abandon this buffer and start a new one.
229 plab->retire(false, false);
230 size_t buf_size = plab->word_sz();
231 HeapWord* buf_space = sp->par_allocate(buf_size);
232 if (buf_space == NULL) {
233 const size_t min_bytes =
234 ParGCAllocBuffer::min_size() << LogHeapWordSize;
235 size_t free_bytes = sp->free();
236 while(buf_space == NULL && free_bytes >= min_bytes) {
237 buf_size = free_bytes >> LogHeapWordSize;
238 assert(buf_size == (size_t)align_object_size(buf_size),
239 "Invariant");
240 buf_space = sp->par_allocate(buf_size);
241 free_bytes = sp->free();
242 }
243 }
244 if (buf_space != NULL) {
245 plab->set_word_size(buf_size);
246 plab->set_buf(buf_space);
247 record_survivor_plab(buf_space, buf_size);
248 obj = plab->allocate(word_sz);
249 // Note that we cannot compare buf_size < word_sz below
250 // because of AlignmentReserve (see ParGCAllocBuffer::allocate()).
251 assert(obj != NULL || plab->words_remaining() < word_sz,
252 "Else should have been able to allocate");
253 // It's conceivable that we may be able to use the
254 // buffer we just grabbed for subsequent small requests
255 // even if not for this one.
256 } else {
257 // We're used up.
258 _to_space_full = true;
259 }
260
261 } else {
262 // Too large; allocate the object individually.
263 obj = sp->par_allocate(word_sz);
264 }
265 }
266 return obj;
267 }
268
269
270 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj,
271 size_t word_sz) {
272 // Is the alloc in the current alloc buffer?
273 if (to_space_alloc_buffer()->contains(obj)) {
274 assert(to_space_alloc_buffer()->contains(obj + word_sz - 1),
275 "Should contain whole object.");
276 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
277 } else {
278 CollectedHeap::fill_with_object(obj, word_sz);
279 }
280 }
281
282 void ParScanThreadState::print_and_clear_promotion_failure_size() {
283 if (_promotion_failure_size != 0) {
284 if (PrintPromotionFailure) {
285 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ",
286 _thread_num, _promotion_failure_size);
287 }
288 _promotion_failure_size = 0;
289 }
290 }
291
292 class ParScanThreadStateSet: private ResourceArray {
293 public:
294 // Initializes states for the specified number of threads;
295 ParScanThreadStateSet(int num_threads,
296 Space& to_space,
297 ParNewGeneration& gen,
298 Generation& old_gen,
299 ObjToScanQueueSet& queue_set,
300 Stack<oop>* overflow_stacks_,
301 size_t desired_plab_sz,
302 ParallelTaskTerminator& term);
303
304 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
305
306 inline ParScanThreadState& thread_state(int i);
307
308 void reset(int active_workers, bool promotion_failed);
309 void flush();
310
311 #if TASKQUEUE_STATS
312 static void
313 print_termination_stats_hdr(outputStream* const st = gclog_or_tty);
314 void print_termination_stats(outputStream* const st = gclog_or_tty);
315 static void
316 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty);
317 void print_taskqueue_stats(outputStream* const st = gclog_or_tty);
318 void reset_stats();
319 #endif // TASKQUEUE_STATS
320
321 private:
322 ParallelTaskTerminator& _term;
323 ParNewGeneration& _gen;
324 Generation& _next_gen;
325 public:
326 bool is_valid(int id) const { return id < length(); }
327 ParallelTaskTerminator* terminator() { return &_term; }
328 };
329
330
331 ParScanThreadStateSet::ParScanThreadStateSet(
332 int num_threads, Space& to_space, ParNewGeneration& gen,
333 Generation& old_gen, ObjToScanQueueSet& queue_set,
334 Stack<oop>* overflow_stacks,
335 size_t desired_plab_sz, ParallelTaskTerminator& term)
336 : ResourceArray(sizeof(ParScanThreadState), num_threads),
337 _gen(gen), _next_gen(old_gen), _term(term)
338 {
339 assert(num_threads > 0, "sanity check!");
340 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
341 "overflow_stack allocation mismatch");
342 // Initialize states.
343 for (int i = 0; i < num_threads; ++i) {
344 new ((ParScanThreadState*)_data + i)
345 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set,
346 overflow_stacks, desired_plab_sz, term);
347 }
348 }
349
350 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i)
351 {
352 assert(i >= 0 && i < length(), "sanity check!");
353 return ((ParScanThreadState*)_data)[i];
354 }
355
356
357 void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed)
358 {
359 _term.reset_for_reuse(active_threads);
360 if (promotion_failed) {
361 for (int i = 0; i < length(); ++i) {
362 thread_state(i).print_and_clear_promotion_failure_size();
363 }
364 }
365 }
366
367 #if TASKQUEUE_STATS
368 void
369 ParScanThreadState::reset_stats()
370 {
371 taskqueue_stats().reset();
372 _term_attempts = 0;
373 _overflow_refills = 0;
374 _overflow_refill_objs = 0;
375 }
376
377 void ParScanThreadStateSet::reset_stats()
378 {
379 for (int i = 0; i < length(); ++i) {
380 thread_state(i).reset_stats();
381 }
382 }
383
384 void
385 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st)
386 {
387 st->print_raw_cr("GC Termination Stats");
388 st->print_raw_cr(" elapsed --strong roots-- "
389 "-------termination-------");
390 st->print_raw_cr("thr ms ms % "
391 " ms % attempts");
392 st->print_raw_cr("--- --------- --------- ------ "
393 "--------- ------ --------");
394 }
395
396 void ParScanThreadStateSet::print_termination_stats(outputStream* const st)
397 {
398 print_termination_stats_hdr(st);
399
400 for (int i = 0; i < length(); ++i) {
401 const ParScanThreadState & pss = thread_state(i);
402 const double elapsed_ms = pss.elapsed_time() * 1000.0;
403 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
404 const double term_ms = pss.term_time() * 1000.0;
405 st->print_cr("%3d %9.2f %9.2f %6.2f "
406 "%9.2f %6.2f " SIZE_FORMAT_W(8),
407 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
408 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
409 }
410 }
411
412 // Print stats related to work queue activity.
413 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st)
414 {
415 st->print_raw_cr("GC Task Stats");
416 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
417 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
418 }
419
420 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st)
421 {
422 print_taskqueue_stats_hdr(st);
423
424 TaskQueueStats totals;
425 for (int i = 0; i < length(); ++i) {
426 const ParScanThreadState & pss = thread_state(i);
427 const TaskQueueStats & stats = pss.taskqueue_stats();
428 st->print("%3d ", i); stats.print(st); st->cr();
429 totals += stats;
430
431 if (pss.overflow_refills() > 0) {
432 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
433 SIZE_FORMAT_W(10) " overflow objects",
434 pss.overflow_refills(), pss.overflow_refill_objs());
435 }
436 }
437 st->print("tot "); totals.print(st); st->cr();
438
439 DEBUG_ONLY(totals.verify());
440 }
441 #endif // TASKQUEUE_STATS
442
443 void ParScanThreadStateSet::flush()
444 {
445 // Work in this loop should be kept as lightweight as
446 // possible since this might otherwise become a bottleneck
447 // to scaling. Should we add heavy-weight work into this
448 // loop, consider parallelizing the loop into the worker threads.
449 for (int i = 0; i < length(); ++i) {
450 ParScanThreadState& par_scan_state = thread_state(i);
451
452 // Flush stats related to To-space PLAB activity and
453 // retire the last buffer.
454 par_scan_state.to_space_alloc_buffer()->
455 flush_stats_and_retire(_gen.plab_stats(),
456 false /* !retain */);
457
458 // Every thread has its own age table. We need to merge
459 // them all into one.
460 ageTable *local_table = par_scan_state.age_table();
461 _gen.age_table()->merge(local_table);
462
463 // Inform old gen that we're done.
464 _next_gen.par_promote_alloc_done(i);
465 _next_gen.par_oop_since_save_marks_iterate_done(i);
466 }
467
468 if (UseConcMarkSweepGC && ParallelGCThreads > 0) {
469 // We need to call this even when ResizeOldPLAB is disabled
470 // so as to avoid breaking some asserts. While we may be able
471 // to avoid this by reorganizing the code a bit, I am loathe
472 // to do that unless we find cases where ergo leads to bad
473 // performance.
474 CFLS_LAB::compute_desired_plab_size();
475 }
476 }
477
478 ParScanClosure::ParScanClosure(ParNewGeneration* g,
479 ParScanThreadState* par_scan_state) :
480 OopsInGenClosure(g), _par_scan_state(par_scan_state), _g(g)
481 {
482 assert(_g->level() == 0, "Optimized for youngest generation");
483 _boundary = _g->reserved().end();
484 }
485
486 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
487 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
488
489 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
490 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
491
492 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
493 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
494
495 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
496 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
497
498 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
499 ParScanThreadState* par_scan_state)
500 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
501 {}
502
503 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
504 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
505
506 #ifdef WIN32
507 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
508 #endif
509
510 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
511 ParScanThreadState* par_scan_state_,
512 ParScanWithoutBarrierClosure* to_space_closure_,
513 ParScanWithBarrierClosure* old_gen_closure_,
514 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
515 ParNewGeneration* par_gen_,
516 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
517 ObjToScanQueueSet* task_queues_,
518 ParallelTaskTerminator* terminator_) :
519
520 _par_scan_state(par_scan_state_),
521 _to_space_closure(to_space_closure_),
522 _old_gen_closure(old_gen_closure_),
523 _to_space_root_closure(to_space_root_closure_),
524 _old_gen_root_closure(old_gen_root_closure_),
525 _par_gen(par_gen_),
526 _task_queues(task_queues_),
527 _terminator(terminator_)
528 {}
529
530 void ParEvacuateFollowersClosure::do_void() {
531 ObjToScanQueue* work_q = par_scan_state()->work_queue();
532
533 while (true) {
534
535 // Scan to-space and old-gen objs until we run out of both.
536 oop obj_to_scan;
537 par_scan_state()->trim_queues(0);
538
539 // We have no local work, attempt to steal from other threads.
540
541 // attempt to steal work from promoted.
542 if (task_queues()->steal(par_scan_state()->thread_num(),
543 par_scan_state()->hash_seed(),
544 obj_to_scan)) {
545 bool res = work_q->push(obj_to_scan);
546 assert(res, "Empty queue should have room for a push.");
547
548 // if successful, goto Start.
549 continue;
550
551 // try global overflow list.
552 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
553 continue;
554 }
555
556 // Otherwise, offer termination.
557 par_scan_state()->start_term_time();
558 if (terminator()->offer_termination()) break;
559 par_scan_state()->end_term_time();
560 }
561 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
562 "Broken overflow list?");
563 // Finish the last termination pause.
564 par_scan_state()->end_term_time();
565 }
566
567 ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen,
568 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) :
569 AbstractGangTask("ParNewGeneration collection"),
570 _gen(gen), _next_gen(next_gen),
571 _young_old_boundary(young_old_boundary),
572 _state_set(state_set)
573 {}
574
575 // Reset the terminator for the given number of
576 // active threads.
577 void ParNewGenTask::set_for_termination(int active_workers) {
578 _state_set->reset(active_workers, _gen->promotion_failed());
579 // Should the heap be passed in? There's only 1 for now so
580 // grab it instead.
581 GenCollectedHeap* gch = GenCollectedHeap::heap();
582 gch->set_n_termination(active_workers);
583 }
584
585 // The "i" passed to this method is the part of the work for
586 // this thread. It is not the worker ID. The "i" is derived
587 // from _started_workers which is incremented in internal_note_start()
588 // called in GangWorker loop() and which is called under the
589 // which is called under the protection of the gang monitor and is
590 // called after a task is started. So "i" is based on
591 // first-come-first-served.
592
593 void ParNewGenTask::work(uint worker_id) {
594 GenCollectedHeap* gch = GenCollectedHeap::heap();
595 // Since this is being done in a separate thread, need new resource
596 // and handle marks.
597 ResourceMark rm;
598 HandleMark hm;
599 // We would need multiple old-gen queues otherwise.
600 assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen.");
601
602 Generation* old_gen = gch->next_gen(_gen);
603
604 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
605 assert(_state_set->is_valid(worker_id), "Should not have been called");
606
607 par_scan_state.set_young_old_boundary(_young_old_boundary);
608
609 par_scan_state.start_strong_roots();
610 gch->gen_process_strong_roots(_gen->level(),
611 true, // Process younger gens, if any,
612 // as strong roots.
613 false, // no scope; this is parallel code
614 false, // not collecting perm generation.
615 SharedHeap::SO_AllClasses,
616 &par_scan_state.to_space_root_closure(),
617 true, // walk *all* scavengable nmethods
618 &par_scan_state.older_gen_closure());
619 par_scan_state.end_strong_roots();
620
621 // "evacuate followers".
622 par_scan_state.evacuate_followers_closure().do_void();
623 }
624
625 #ifdef _MSC_VER
626 #pragma warning( push )
627 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
628 #endif
629 ParNewGeneration::
630 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level)
631 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"),
632 _overflow_list(NULL),
633 _is_alive_closure(this),
634 _plab_stats(YoungPLABSize, PLABWeight)
635 {
636 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
637 NOT_PRODUCT(_num_par_pushes = 0;)
638 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
639 guarantee(_task_queues != NULL, "task_queues allocation failure.");
640
641 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
642 ObjToScanQueue *q = new ObjToScanQueue();
643 guarantee(q != NULL, "work_queue Allocation failure.");
644 _task_queues->register_queue(i1, q);
645 }
646
647 for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
648 _task_queues->queue(i2)->initialize();
649
650 _overflow_stacks = NULL;
651 if (ParGCUseLocalOverflow) {
652 _overflow_stacks = NEW_C_HEAP_ARRAY(Stack<oop>, ParallelGCThreads);
653 for (size_t i = 0; i < ParallelGCThreads; ++i) {
654 new (_overflow_stacks + i) Stack<oop>();
655 }
656 }
657
658 if (UsePerfData) {
659 EXCEPTION_MARK;
660 ResourceMark rm;
661
662 const char* cname =
663 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
664 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
665 ParallelGCThreads, CHECK);
666 }
667 }
668 #ifdef _MSC_VER
669 #pragma warning( pop )
670 #endif
671
672 // ParNewGeneration::
673 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
674 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
675
676 template <class T>
677 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
678 #ifdef ASSERT
679 {
680 assert(!oopDesc::is_null(*p), "expected non-null ref");
681 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
682 // We never expect to see a null reference being processed
683 // as a weak reference.
684 assert(obj->is_oop(), "expected an oop while scanning weak refs");
685 }
686 #endif // ASSERT
687
688 _par_cl->do_oop_nv(p);
689
690 if (Universe::heap()->is_in_reserved(p)) {
691 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
692 _rs->write_ref_field_gc_par(p, obj);
693 }
694 }
695
696 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
697 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
698
699 // ParNewGeneration::
700 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
701 DefNewGeneration::KeepAliveClosure(cl) {}
702
703 template <class T>
704 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
705 #ifdef ASSERT
706 {
707 assert(!oopDesc::is_null(*p), "expected non-null ref");
708 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
709 // We never expect to see a null reference being processed
710 // as a weak reference.
711 assert(obj->is_oop(), "expected an oop while scanning weak refs");
712 }
713 #endif // ASSERT
714
715 _cl->do_oop_nv(p);
716
717 if (Universe::heap()->is_in_reserved(p)) {
718 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
719 _rs->write_ref_field_gc_par(p, obj);
720 }
721 }
722
723 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
724 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
725
726 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
727 T heap_oop = oopDesc::load_heap_oop(p);
728 if (!oopDesc::is_null(heap_oop)) {
729 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
730 if ((HeapWord*)obj < _boundary) {
731 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
732 oop new_obj = obj->is_forwarded()
733 ? obj->forwardee()
734 : _g->DefNewGeneration::copy_to_survivor_space(obj);
735 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
736 }
737 if (_gc_barrier) {
738 // If p points to a younger generation, mark the card.
739 if ((HeapWord*)obj < _gen_boundary) {
740 _rs->write_ref_field_gc_par(p, obj);
741 }
742 }
743 }
744 }
745
746 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
747 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
748
749 class ParNewRefProcTaskProxy: public AbstractGangTask {
750 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
751 public:
752 ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen,
753 Generation& next_gen,
754 HeapWord* young_old_boundary,
755 ParScanThreadStateSet& state_set);
756
757 private:
758 virtual void work(uint worker_id);
759 virtual void set_for_termination(int active_workers) {
760 _state_set.terminator()->reset_for_reuse(active_workers);
761 }
762 private:
763 ParNewGeneration& _gen;
764 ProcessTask& _task;
765 Generation& _next_gen;
766 HeapWord* _young_old_boundary;
767 ParScanThreadStateSet& _state_set;
768 };
769
770 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(
771 ProcessTask& task, ParNewGeneration& gen,
772 Generation& next_gen,
773 HeapWord* young_old_boundary,
774 ParScanThreadStateSet& state_set)
775 : AbstractGangTask("ParNewGeneration parallel reference processing"),
776 _gen(gen),
777 _task(task),
778 _next_gen(next_gen),
779 _young_old_boundary(young_old_boundary),
780 _state_set(state_set)
781 {
782 }
783
784 void ParNewRefProcTaskProxy::work(uint worker_id)
785 {
786 ResourceMark rm;
787 HandleMark hm;
788 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
789 par_scan_state.set_young_old_boundary(_young_old_boundary);
790 _task.work(worker_id, par_scan_state.is_alive_closure(),
791 par_scan_state.keep_alive_closure(),
792 par_scan_state.evacuate_followers_closure());
793 }
794
795 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
796 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
797 EnqueueTask& _task;
798
799 public:
800 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
801 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
802 _task(task)
803 { }
804
805 virtual void work(uint worker_id)
806 {
807 _task.work(worker_id);
808 }
809 };
810
811
812 void ParNewRefProcTaskExecutor::execute(ProcessTask& task)
813 {
814 GenCollectedHeap* gch = GenCollectedHeap::heap();
815 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
816 "not a generational heap");
817 FlexibleWorkGang* workers = gch->workers();
818 assert(workers != NULL, "Need parallel worker threads.");
819 _state_set.reset(workers->active_workers(), _generation.promotion_failed());
820 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(),
821 _generation.reserved().end(), _state_set);
822 workers->run_task(&rp_task);
823 _state_set.reset(0 /* bad value in debug if not reset */,
824 _generation.promotion_failed());
825 }
826
827 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task)
828 {
829 GenCollectedHeap* gch = GenCollectedHeap::heap();
830 FlexibleWorkGang* workers = gch->workers();
831 assert(workers != NULL, "Need parallel worker threads.");
832 ParNewRefEnqueueTaskProxy enq_task(task);
833 workers->run_task(&enq_task);
834 }
835
836 void ParNewRefProcTaskExecutor::set_single_threaded_mode()
837 {
838 _state_set.flush();
839 GenCollectedHeap* gch = GenCollectedHeap::heap();
840 gch->set_par_threads(0); // 0 ==> non-parallel.
841 gch->save_marks();
842 }
843
844 ScanClosureWithParBarrier::
845 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
846 ScanClosure(g, gc_barrier) {}
847
848 EvacuateFollowersClosureGeneral::
849 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
850 OopsInGenClosure* cur,
851 OopsInGenClosure* older) :
852 _gch(gch), _level(level),
853 _scan_cur_or_nonheap(cur), _scan_older(older)
854 {}
855
856 void EvacuateFollowersClosureGeneral::do_void() {
857 do {
858 // Beware: this call will lead to closure applications via virtual
859 // calls.
860 _gch->oop_since_save_marks_iterate(_level,
861 _scan_cur_or_nonheap,
862 _scan_older);
863 } while (!_gch->no_allocs_since_save_marks(_level));
864 }
865
866
867 bool ParNewGeneration::_avoid_promotion_undo = false;
868
869 void ParNewGeneration::adjust_desired_tenuring_threshold() {
870 // Set the desired survivor size to half the real survivor space
871 _tenuring_threshold =
872 age_table()->compute_tenuring_threshold(to()->capacity()/HeapWordSize);
873 }
874
875 // A Generation that does parallel young-gen collection.
876
877 void ParNewGeneration::collect(bool full,
878 bool clear_all_soft_refs,
879 size_t size,
880 bool is_tlab) {
881 assert(full || size > 0, "otherwise we don't want to collect");
882 GenCollectedHeap* gch = GenCollectedHeap::heap();
883 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
884 "not a CMS generational heap");
885 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
886 FlexibleWorkGang* workers = gch->workers();
887 assert(workers != NULL, "Need workgang for parallel work");
888 int active_workers =
889 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
890 workers->active_workers(),
891 Threads::number_of_non_daemon_threads());
892 workers->set_active_workers(active_workers);
893 _next_gen = gch->next_gen(this);
894 assert(_next_gen != NULL,
895 "This must be the youngest gen, and not the only gen");
896 assert(gch->n_gens() == 2,
897 "Par collection currently only works with single older gen.");
898 // Do we have to avoid promotion_undo?
899 if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) {
900 set_avoid_promotion_undo(true);
901 }
902
903 // If the next generation is too full to accomodate 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 init_assuming_no_promotion_failure();
913
914 if (UseAdaptiveSizePolicy) {
915 set_survivor_overflow(false);
916 size_policy->minor_collection_begin();
917 }
918
919 TraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, gclog_or_tty);
920 // Capture heap used before collection (for printing).
921 size_t gch_prev_used = gch->used();
922
923 SpecializationStats::clear();
924
925 age_table()->clear();
926 to()->clear(SpaceDecorator::Mangle);
927
928 gch->save_marks();
929 assert(workers != NULL, "Need parallel worker threads.");
930 int n_workers = active_workers;
931
932 // Set the correct parallelism (number of queues) in the reference processor
933 ref_processor()->set_active_mt_degree(n_workers);
934
935 // Always set the terminator for the active number of workers
936 // because only those workers go through the termination protocol.
937 ParallelTaskTerminator _term(n_workers, task_queues());
938 ParScanThreadStateSet thread_state_set(workers->active_workers(),
939 *to(), *this, *_next_gen, *task_queues(),
940 _overflow_stacks, desired_plab_sz(), _term);
941
942 ParNewGenTask tsk(this, _next_gen, reserved().end(), &thread_state_set);
943 gch->set_par_threads(n_workers);
944 gch->rem_set()->prepare_for_younger_refs_iterate(true);
945 // It turns out that even when we're using 1 thread, doing the work in a
946 // separate thread causes wide variance in run times. We can't help this
947 // in the multi-threaded case, but we special-case n=1 here to get
948 // repeatable measurements of the 1-thread overhead of the parallel code.
949 if (n_workers > 1) {
950 GenCollectedHeap::StrongRootsScope srs(gch);
951 workers->run_task(&tsk);
952 } else {
953 GenCollectedHeap::StrongRootsScope srs(gch);
954 tsk.work(0);
955 }
956 thread_state_set.reset(0 /* Bad value in debug if not reset */,
957 promotion_failed());
958
959 // Process (weak) reference objects found during scavenge.
960 ReferenceProcessor* rp = ref_processor();
961 IsAliveClosure is_alive(this);
962 ScanWeakRefClosure scan_weak_ref(this);
963 KeepAliveClosure keep_alive(&scan_weak_ref);
964 ScanClosure scan_without_gc_barrier(this, false);
965 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
966 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
967 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level,
968 &scan_without_gc_barrier, &scan_with_gc_barrier);
969 rp->setup_policy(clear_all_soft_refs);
970 // Can the mt_degree be set later (at run_task() time would be best)?
971 rp->set_active_mt_degree(active_workers);
972 if (rp->processing_is_mt()) {
973 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
974 rp->process_discovered_references(&is_alive, &keep_alive,
975 &evacuate_followers, &task_executor);
976 } else {
977 thread_state_set.flush();
978 gch->set_par_threads(0); // 0 ==> non-parallel.
979 gch->save_marks();
980 rp->process_discovered_references(&is_alive, &keep_alive,
981 &evacuate_followers, NULL);
982 }
983 if (!promotion_failed()) {
984 // Swap the survivor spaces.
985 eden()->clear(SpaceDecorator::Mangle);
986 from()->clear(SpaceDecorator::Mangle);
987 if (ZapUnusedHeapArea) {
988 // This is now done here because of the piece-meal mangling which
989 // can check for valid mangling at intermediate points in the
990 // collection(s). When a minor collection fails to collect
991 // sufficient space resizing of the young generation can occur
992 // an redistribute the spaces in the young generation. Mangle
993 // here so that unzapped regions don't get distributed to
994 // other spaces.
995 to()->mangle_unused_area();
996 }
997 swap_spaces();
998
999 // A successful scavenge should restart the GC time limit count which is
1000 // for full GC's.
1001 size_policy->reset_gc_overhead_limit_count();
1002
1003 assert(to()->is_empty(), "to space should be empty now");
1004 } else {
1005 assert(_promo_failure_scan_stack.is_empty(), "post condition");
1006 _promo_failure_scan_stack.clear(true); // Clear cached segments.
1007
1008 remove_forwarding_pointers();
1009 if (PrintGCDetails) {
1010 gclog_or_tty->print(" (promotion failed)");
1011 }
1012 // All the spaces are in play for mark-sweep.
1013 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
1014 from()->set_next_compaction_space(to());
1015 gch->set_incremental_collection_failed();
1016 // Inform the next generation that a promotion failure occurred.
1017 _next_gen->promotion_failure_occurred();
1018
1019 // Reset the PromotionFailureALot counters.
1020 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();)
1021 }
1022 // set new iteration safe limit for the survivor spaces
1023 from()->set_concurrent_iteration_safe_limit(from()->top());
1024 to()->set_concurrent_iteration_safe_limit(to()->top());
1025
1026 adjust_desired_tenuring_threshold();
1027 if (ResizePLAB) {
1028 plab_stats()->adjust_desired_plab_sz();
1029 }
1030
1031 if (PrintGC && !PrintGCDetails) {
1032 gch->print_heap_change(gch_prev_used);
1033 }
1034
1035 if (PrintGCDetails && ParallelGCVerbose) {
1036 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
1037 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
1038 }
1039
1040 if (UseAdaptiveSizePolicy) {
1041 size_policy->minor_collection_end(gch->gc_cause());
1042 size_policy->avg_survived()->sample(from()->used());
1043 }
1044
1045 // We need to use a monotonically non-deccreasing time in ms
1046 // or we will see time-warp warnings and os::javaTimeMillis()
1047 // does not guarantee monotonicity.
1048 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1049 update_time_of_last_gc(now);
1050
1051 SpecializationStats::print();
1052
1053 rp->set_enqueuing_is_done(true);
1054 if (rp->processing_is_mt()) {
1055 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1056 rp->enqueue_discovered_references(&task_executor);
1057 } else {
1058 rp->enqueue_discovered_references(NULL);
1059 }
1060 rp->verify_no_references_recorded();
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 = oop(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 #ifdef ASSERT
1099 bool ParNewGeneration::is_legal_forward_ptr(oop p) {
1100 return
1101 (_avoid_promotion_undo && p == ClaimedForwardPtr)
1102 || Universe::heap()->is_in_reserved(p);
1103 }
1104 #endif
1105
1106 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
1107 if (m->must_be_preserved_for_promotion_failure(obj)) {
1108 // We should really have separate per-worker stacks, rather
1109 // than use locking of a common pair of stacks.
1110 MutexLocker ml(ParGCRareEvent_lock);
1111 preserve_mark(obj, m);
1112 }
1113 }
1114
1115 // Multiple GC threads may try to promote an object. If the object
1116 // is successfully promoted, a forwarding pointer will be installed in
1117 // the object in the young generation. This method claims the right
1118 // to install the forwarding pointer before it copies the object,
1119 // thus avoiding the need to undo the copy as in
1120 // copy_to_survivor_space_avoiding_with_undo.
1121
1122 oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo(
1123 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1124 // In the sequential version, this assert also says that the object is
1125 // not forwarded. That might not be the case here. It is the case that
1126 // the caller observed it to be not forwarded at some time in the past.
1127 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1128
1129 // The sequential code read "old->age()" below. That doesn't work here,
1130 // since the age is in the mark word, and that might be overwritten with
1131 // a forwarding pointer by a parallel thread. So we must save the mark
1132 // word in a local and then analyze it.
1133 oopDesc dummyOld;
1134 dummyOld.set_mark(m);
1135 assert(!dummyOld.is_forwarded(),
1136 "should not be called with forwarding pointer mark word.");
1137
1138 oop new_obj = NULL;
1139 oop forward_ptr;
1140
1141 // Try allocating obj in to-space (unless too old)
1142 if (dummyOld.age() < tenuring_threshold()) {
1143 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1144 if (new_obj == NULL) {
1145 set_survivor_overflow(true);
1146 }
1147 }
1148
1149 if (new_obj == NULL) {
1150 // Either to-space is full or we decided to promote
1151 // try allocating obj tenured
1152
1153 // Attempt to install a null forwarding pointer (atomically),
1154 // to claim the right to install the real forwarding pointer.
1155 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1156 if (forward_ptr != NULL) {
1157 // someone else beat us to it.
1158 return real_forwardee(old);
1159 }
1160
1161 new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
1162 old, m, sz);
1163
1164 if (new_obj == NULL) {
1165 // promotion failed, forward to self
1166 _promotion_failed = true;
1167 new_obj = old;
1168
1169 preserve_mark_if_necessary(old, m);
1170 // Log the size of the maiden promotion failure
1171 par_scan_state->log_promotion_failure(sz);
1172 }
1173
1174 old->forward_to(new_obj);
1175 forward_ptr = NULL;
1176 } else {
1177 // Is in to-space; do copying ourselves.
1178 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1179 forward_ptr = old->forward_to_atomic(new_obj);
1180 // Restore the mark word copied above.
1181 new_obj->set_mark(m);
1182 // Increment age if obj still in new generation
1183 new_obj->incr_age();
1184 par_scan_state->age_table()->add(new_obj, sz);
1185 }
1186 assert(new_obj != NULL, "just checking");
1187
1188 if (forward_ptr == NULL) {
1189 oop obj_to_push = new_obj;
1190 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1191 // Length field used as index of next element to be scanned.
1192 // Real length can be obtained from real_forwardee()
1193 arrayOop(old)->set_length(0);
1194 obj_to_push = old;
1195 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1196 "push forwarded object");
1197 }
1198 // Push it on one of the queues of to-be-scanned objects.
1199 bool simulate_overflow = false;
1200 NOT_PRODUCT(
1201 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1202 // simulate a stack overflow
1203 simulate_overflow = true;
1204 }
1205 )
1206 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1207 // Add stats for overflow pushes.
1208 if (Verbose && PrintGCDetails) {
1209 gclog_or_tty->print("queue overflow!\n");
1210 }
1211 push_on_overflow_list(old, par_scan_state);
1212 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1213 }
1214
1215 return new_obj;
1216 }
1217
1218 // Oops. Someone beat us to it. Undo the allocation. Where did we
1219 // allocate it?
1220 if (is_in_reserved(new_obj)) {
1221 // Must be in to_space.
1222 assert(to()->is_in_reserved(new_obj), "Checking");
1223 if (forward_ptr == ClaimedForwardPtr) {
1224 // Wait to get the real forwarding pointer value.
1225 forward_ptr = real_forwardee(old);
1226 }
1227 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1228 }
1229
1230 return forward_ptr;
1231 }
1232
1233
1234 // Multiple GC threads may try to promote the same object. If two
1235 // or more GC threads copy the object, only one wins the race to install
1236 // the forwarding pointer. The other threads have to undo their copy.
1237
1238 oop ParNewGeneration::copy_to_survivor_space_with_undo(
1239 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1240
1241 // In the sequential version, this assert also says that the object is
1242 // not forwarded. That might not be the case here. It is the case that
1243 // the caller observed it to be not forwarded at some time in the past.
1244 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1245
1246 // The sequential code read "old->age()" below. That doesn't work here,
1247 // since the age is in the mark word, and that might be overwritten with
1248 // a forwarding pointer by a parallel thread. So we must save the mark
1249 // word here, install it in a local oopDesc, and then analyze it.
1250 oopDesc dummyOld;
1251 dummyOld.set_mark(m);
1252 assert(!dummyOld.is_forwarded(),
1253 "should not be called with forwarding pointer mark word.");
1254
1255 bool failed_to_promote = false;
1256 oop new_obj = NULL;
1257 oop forward_ptr;
1258
1259 // Try allocating obj in to-space (unless too old)
1260 if (dummyOld.age() < tenuring_threshold()) {
1261 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1262 if (new_obj == NULL) {
1263 set_survivor_overflow(true);
1264 }
1265 }
1266
1267 if (new_obj == NULL) {
1268 // Either to-space is full or we decided to promote
1269 // try allocating obj tenured
1270 new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
1271 old, m, sz);
1272
1273 if (new_obj == NULL) {
1274 // promotion failed, forward to self
1275 forward_ptr = old->forward_to_atomic(old);
1276 new_obj = old;
1277
1278 if (forward_ptr != NULL) {
1279 return forward_ptr; // someone else succeeded
1280 }
1281
1282 _promotion_failed = true;
1283 failed_to_promote = true;
1284
1285 preserve_mark_if_necessary(old, m);
1286 // Log the size of the maiden promotion failure
1287 par_scan_state->log_promotion_failure(sz);
1288 }
1289 } else {
1290 // Is in to-space; do copying ourselves.
1291 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1292 // Restore the mark word copied above.
1293 new_obj->set_mark(m);
1294 // Increment age if new_obj still in new generation
1295 new_obj->incr_age();
1296 par_scan_state->age_table()->add(new_obj, sz);
1297 }
1298 assert(new_obj != NULL, "just checking");
1299
1300 // Now attempt to install the forwarding pointer (atomically).
1301 // We have to copy the mark word before overwriting with forwarding
1302 // ptr, so we can restore it below in the copy.
1303 if (!failed_to_promote) {
1304 forward_ptr = old->forward_to_atomic(new_obj);
1305 }
1306
1307 if (forward_ptr == NULL) {
1308 oop obj_to_push = new_obj;
1309 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1310 // Length field used as index of next element to be scanned.
1311 // Real length can be obtained from real_forwardee()
1312 arrayOop(old)->set_length(0);
1313 obj_to_push = old;
1314 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1315 "push forwarded object");
1316 }
1317 // Push it on one of the queues of to-be-scanned objects.
1318 bool simulate_overflow = false;
1319 NOT_PRODUCT(
1320 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1321 // simulate a stack overflow
1322 simulate_overflow = true;
1323 }
1324 )
1325 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1326 // Add stats for overflow pushes.
1327 push_on_overflow_list(old, par_scan_state);
1328 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1329 }
1330
1331 return new_obj;
1332 }
1333
1334 // Oops. Someone beat us to it. Undo the allocation. Where did we
1335 // allocate it?
1336 if (is_in_reserved(new_obj)) {
1337 // Must be in to_space.
1338 assert(to()->is_in_reserved(new_obj), "Checking");
1339 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1340 } else {
1341 assert(!_avoid_promotion_undo, "Should not be here if avoiding.");
1342 _next_gen->par_promote_alloc_undo(par_scan_state->thread_num(),
1343 (HeapWord*)new_obj, sz);
1344 }
1345
1346 return forward_ptr;
1347 }
1348
1349 #ifndef PRODUCT
1350 // It's OK to call this multi-threaded; the worst thing
1351 // that can happen is that we'll get a bunch of closely
1352 // spaced simulated oveflows, but that's OK, in fact
1353 // probably good as it would exercise the overflow code
1354 // under contention.
1355 bool ParNewGeneration::should_simulate_overflow() {
1356 if (_overflow_counter-- <= 0) { // just being defensive
1357 _overflow_counter = ParGCWorkQueueOverflowInterval;
1358 return true;
1359 } else {
1360 return false;
1361 }
1362 }
1363 #endif
1364
1365 // In case we are using compressed oops, we need to be careful.
1366 // If the object being pushed is an object array, then its length
1367 // field keeps track of the "grey boundary" at which the next
1368 // incremental scan will be done (see ParGCArrayScanChunk).
1369 // When using compressed oops, this length field is kept in the
1370 // lower 32 bits of the erstwhile klass word and cannot be used
1371 // for the overflow chaining pointer (OCP below). As such the OCP
1372 // would itself need to be compressed into the top 32-bits in this
1373 // case. Unfortunately, see below, in the event that we have a
1374 // promotion failure, the node to be pushed on the list can be
1375 // outside of the Java heap, so the heap-based pointer compression
1376 // would not work (we would have potential aliasing between C-heap
1377 // and Java-heap pointers). For this reason, when using compressed
1378 // oops, we simply use a worker-thread-local, non-shared overflow
1379 // list in the form of a growable array, with a slightly different
1380 // overflow stack draining strategy. If/when we start using fat
1381 // stacks here, we can go back to using (fat) pointer chains
1382 // (although some performance comparisons would be useful since
1383 // single global lists have their own performance disadvantages
1384 // as we were made painfully aware not long ago, see 6786503).
1385 #define BUSY (oop(0x1aff1aff))
1386 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1387 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1388 if (ParGCUseLocalOverflow) {
1389 // In the case of compressed oops, we use a private, not-shared
1390 // overflow stack.
1391 par_scan_state->push_on_overflow_stack(from_space_obj);
1392 } else {
1393 assert(!UseCompressedOops, "Error");
1394 // if the object has been forwarded to itself, then we cannot
1395 // use the klass pointer for the linked list. Instead we have
1396 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1397 // XXX This is horribly inefficient when a promotion failure occurs
1398 // and should be fixed. XXX FIX ME !!!
1399 #ifndef PRODUCT
1400 Atomic::inc_ptr(&_num_par_pushes);
1401 assert(_num_par_pushes > 0, "Tautology");
1402 #endif
1403 if (from_space_obj->forwardee() == from_space_obj) {
1404 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1);
1405 listhead->forward_to(from_space_obj);
1406 from_space_obj = listhead;
1407 }
1408 oop observed_overflow_list = _overflow_list;
1409 oop cur_overflow_list;
1410 do {
1411 cur_overflow_list = observed_overflow_list;
1412 if (cur_overflow_list != BUSY) {
1413 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1414 } else {
1415 from_space_obj->set_klass_to_list_ptr(NULL);
1416 }
1417 observed_overflow_list =
1418 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1419 } while (cur_overflow_list != observed_overflow_list);
1420 }
1421 }
1422
1423 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1424 bool res;
1425
1426 if (ParGCUseLocalOverflow) {
1427 res = par_scan_state->take_from_overflow_stack();
1428 } else {
1429 assert(!UseCompressedOops, "Error");
1430 res = take_from_overflow_list_work(par_scan_state);
1431 }
1432 return res;
1433 }
1434
1435
1436 // *NOTE*: The overflow list manipulation code here and
1437 // in CMSCollector:: are very similar in shape,
1438 // except that in the CMS case we thread the objects
1439 // directly into the list via their mark word, and do
1440 // not need to deal with special cases below related
1441 // to chunking of object arrays and promotion failure
1442 // handling.
1443 // CR 6797058 has been filed to attempt consolidation of
1444 // the common code.
1445 // Because of the common code, if you make any changes in
1446 // the code below, please check the CMS version to see if
1447 // similar changes might be needed.
1448 // See CMSCollector::par_take_from_overflow_list() for
1449 // more extensive documentation comments.
1450 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1451 ObjToScanQueue* work_q = par_scan_state->work_queue();
1452 // How many to take?
1453 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1454 (size_t)ParGCDesiredObjsFromOverflowList);
1455
1456 assert(!UseCompressedOops, "Error");
1457 assert(par_scan_state->overflow_stack() == NULL, "Error");
1458 if (_overflow_list == NULL) return false;
1459
1460 // Otherwise, there was something there; try claiming the list.
1461 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
1462 // Trim off a prefix of at most objsFromOverflow items
1463 Thread* tid = Thread::current();
1464 size_t spin_count = (size_t)ParallelGCThreads;
1465 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1466 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1467 // someone grabbed it before we did ...
1468 // ... we spin for a short while...
1469 os::sleep(tid, sleep_time_millis, false);
1470 if (_overflow_list == NULL) {
1471 // nothing left to take
1472 return false;
1473 } else if (_overflow_list != BUSY) {
1474 // try and grab the prefix
1475 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
1476 }
1477 }
1478 if (prefix == NULL || prefix == BUSY) {
1479 // Nothing to take or waited long enough
1480 if (prefix == NULL) {
1481 // Write back the NULL in case we overwrote it with BUSY above
1482 // and it is still the same value.
1483 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1484 }
1485 return false;
1486 }
1487 assert(prefix != NULL && prefix != BUSY, "Error");
1488 size_t i = 1;
1489 oop cur = prefix;
1490 while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1491 i++; cur = oop(cur->klass());
1492 }
1493
1494 // Reattach remaining (suffix) to overflow list
1495 if (cur->klass_or_null() == NULL) {
1496 // Write back the NULL in lieu of the BUSY we wrote
1497 // above and it is still the same value.
1498 if (_overflow_list == BUSY) {
1499 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1500 }
1501 } else {
1502 assert(cur->klass_or_null() != BUSY, "Error");
1503 oop suffix = oop(cur->klass()); // suffix will be put back on global list
1504 cur->set_klass_to_list_ptr(NULL); // break off suffix
1505 // It's possible that the list is still in the empty(busy) state
1506 // we left it in a short while ago; in that case we may be
1507 // able to place back the suffix.
1508 oop observed_overflow_list = _overflow_list;
1509 oop cur_overflow_list = observed_overflow_list;
1510 bool attached = false;
1511 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1512 observed_overflow_list =
1513 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1514 if (cur_overflow_list == observed_overflow_list) {
1515 attached = true;
1516 break;
1517 } else cur_overflow_list = observed_overflow_list;
1518 }
1519 if (!attached) {
1520 // Too bad, someone else got in in between; we'll need to do a splice.
1521 // Find the last item of suffix list
1522 oop last = suffix;
1523 while (last->klass_or_null() != NULL) {
1524 last = oop(last->klass());
1525 }
1526 // Atomically prepend suffix to current overflow list
1527 observed_overflow_list = _overflow_list;
1528 do {
1529 cur_overflow_list = observed_overflow_list;
1530 if (cur_overflow_list != BUSY) {
1531 // Do the splice ...
1532 last->set_klass_to_list_ptr(cur_overflow_list);
1533 } else { // cur_overflow_list == BUSY
1534 last->set_klass_to_list_ptr(NULL);
1535 }
1536 observed_overflow_list =
1537 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1538 } while (cur_overflow_list != observed_overflow_list);
1539 }
1540 }
1541
1542 // Push objects on prefix list onto this thread's work queue
1543 assert(prefix != NULL && prefix != BUSY, "program logic");
1544 cur = prefix;
1545 ssize_t n = 0;
1546 while (cur != NULL) {
1547 oop obj_to_push = cur->forwardee();
1548 oop next = oop(cur->klass_or_null());
1549 cur->set_klass(obj_to_push->klass());
1550 // This may be an array object that is self-forwarded. In that case, the list pointer
1551 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1552 if (!is_in_reserved(cur)) {
1553 // This can become a scaling bottleneck when there is work queue overflow coincident
1554 // with promotion failure.
1555 oopDesc* f = cur;
1556 FREE_C_HEAP_ARRAY(oopDesc, f);
1557 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1558 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1559 obj_to_push = cur;
1560 }
1561 bool ok = work_q->push(obj_to_push);
1562 assert(ok, "Should have succeeded");
1563 cur = next;
1564 n++;
1565 }
1566 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1567 #ifndef PRODUCT
1568 assert(_num_par_pushes >= n, "Too many pops?");
1569 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1570 #endif
1571 return true;
1572 }
1573 #undef BUSY
1574
1575 void ParNewGeneration::ref_processor_init()
1576 {
1577 if (_ref_processor == NULL) {
1578 // Allocate and initialize a reference processor
1579 _ref_processor =
1580 new ReferenceProcessor(_reserved, // span
1581 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1582 (int) ParallelGCThreads, // mt processing degree
1583 refs_discovery_is_mt(), // mt discovery
1584 (int) ParallelGCThreads, // mt discovery degree
1585 refs_discovery_is_atomic(), // atomic_discovery
1586 NULL, // is_alive_non_header
1587 false); // write barrier for next field updates
1588 }
1589 }
1590
1591 const char* ParNewGeneration::name() const {
1592 return "par new generation";
1593 }
1594
1595 bool ParNewGeneration::in_use() {
1596 return UseParNewGC && ParallelGCThreads > 0;
1597 }