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