1 /*
2 * Copyright (c) 2001, 2014, 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/generation.inline.hpp"
43 #include "memory/referencePolicy.hpp"
44 #include "memory/resourceArea.hpp"
45 #include "memory/sharedHeap.hpp"
46 #include "memory/space.hpp"
47 #include "oops/objArrayOop.hpp"
48 #include "oops/oop.inline.hpp"
49 #include "oops/oop.pcgc.inline.hpp"
50 #include "runtime/atomic.inline.hpp"
51 #include "runtime/handles.hpp"
52 #include "runtime/handles.inline.hpp"
53 #include "runtime/java.hpp"
54 #include "runtime/thread.inline.hpp"
55 #include "utilities/copy.hpp"
56 #include "utilities/globalDefinitions.hpp"
57 #include "utilities/workgroup.hpp"
58
59 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
60
61 #ifdef _MSC_VER
62 #pragma warning( push )
63 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
64 #endif
65 ParScanThreadState::ParScanThreadState(Space* to_space_,
66 ParNewGeneration* gen_,
67 Generation* old_gen_,
68 int thread_num_,
69 ObjToScanQueueSet* work_queue_set_,
70 Stack<oop, mtGC>* overflow_stacks_,
71 size_t desired_plab_sz_,
72 ParNewTracer* gc_tracer,
73 ParallelTaskTerminator& term_) :
74 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_),
75 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false),
76 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
77 _ageTable(false), // false ==> not the global age table, no perf data.
78 _to_space_alloc_buffer(desired_plab_sz_),
79 _gc_tracer(gc_tracer),
80 _to_space_closure(gen_, this), _old_gen_closure(gen_, this),
81 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this),
82 _older_gen_closure(gen_, this),
83 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
84 &_to_space_root_closure, gen_, &_old_gen_root_closure,
85 work_queue_set_, &term_),
86 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this),
87 _keep_alive_closure(&_scan_weak_ref_closure),
88 _strong_roots_time(0.0), _term_time(0.0)
89 {
90 #if TASKQUEUE_STATS
91 _term_attempts = 0;
92 _overflow_refills = 0;
93 _overflow_refill_objs = 0;
94 #endif // TASKQUEUE_STATS
95
96 _survivor_chunk_array =
97 (ChunkArray*) old_gen()->get_data_recorder(thread_num());
98 _hash_seed = 17; // Might want to take time-based random value.
99 _start = os::elapsedTime();
100 _old_gen_closure.set_generation(old_gen_);
101 _old_gen_root_closure.set_generation(old_gen_);
102 }
103 #ifdef _MSC_VER
104 #pragma warning( pop )
105 #endif
106
107 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
108 size_t plab_word_size) {
109 ChunkArray* sca = survivor_chunk_array();
110 if (sca != NULL) {
111 // A non-null SCA implies that we want the PLAB data recorded.
112 sca->record_sample(plab_start, plab_word_size);
113 }
114 }
115
116 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
117 return new_obj->is_objArray() &&
118 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
119 new_obj != old_obj;
120 }
121
122 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
123 assert(old->is_objArray(), "must be obj array");
124 assert(old->is_forwarded(), "must be forwarded");
125 assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
126 assert(!old_gen()->is_in(old), "must be in young generation.");
127
128 objArrayOop obj = objArrayOop(old->forwardee());
129 // Process ParGCArrayScanChunk elements now
130 // and push the remainder back onto queue
131 int start = arrayOop(old)->length();
132 int end = obj->length();
133 int remainder = end - start;
134 assert(start <= end, "just checking");
135 if (remainder > 2 * ParGCArrayScanChunk) {
136 // Test above combines last partial chunk with a full chunk
137 end = start + ParGCArrayScanChunk;
138 arrayOop(old)->set_length(end);
139 // Push remainder.
140 bool ok = work_queue()->push(old);
141 assert(ok, "just popped, push must be okay");
142 } else {
143 // Restore length so that it can be used if there
144 // is a promotion failure and forwarding pointers
145 // must be removed.
146 arrayOop(old)->set_length(end);
147 }
148
149 // process our set of indices (include header in first chunk)
150 // should make sure end is even (aligned to HeapWord in case of compressed oops)
151 if ((HeapWord *)obj < young_old_boundary()) {
152 // object is in to_space
153 obj->oop_iterate_range(&_to_space_closure, start, end);
154 } else {
155 // object is in old generation
156 obj->oop_iterate_range(&_old_gen_closure, start, end);
157 }
158 }
159
160
161 void ParScanThreadState::trim_queues(int max_size) {
162 ObjToScanQueue* queue = work_queue();
163 do {
164 while (queue->size() > (juint)max_size) {
165 oop obj_to_scan;
166 if (queue->pop_local(obj_to_scan)) {
167 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
168 if (obj_to_scan->is_objArray() &&
169 obj_to_scan->is_forwarded() &&
170 obj_to_scan->forwardee() != obj_to_scan) {
171 scan_partial_array_and_push_remainder(obj_to_scan);
172 } else {
173 // object is in to_space
174 obj_to_scan->oop_iterate(&_to_space_closure);
175 }
176 } else {
177 // object is in old generation
178 obj_to_scan->oop_iterate(&_old_gen_closure);
179 }
180 }
181 }
182 // For the case of compressed oops, we have a private, non-shared
183 // overflow stack, so we eagerly drain it so as to more evenly
184 // distribute load early. Note: this may be good to do in
185 // general rather than delay for the final stealing phase.
186 // If applicable, we'll transfer a set of objects over to our
187 // work queue, allowing them to be stolen and draining our
188 // private overflow stack.
189 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
190 }
191
192 bool ParScanThreadState::take_from_overflow_stack() {
193 assert(ParGCUseLocalOverflow, "Else should not call");
194 assert(young_gen()->overflow_list() == NULL, "Error");
195 ObjToScanQueue* queue = work_queue();
196 Stack<oop, mtGC>* const of_stack = overflow_stack();
197 const size_t num_overflow_elems = of_stack->size();
198 const size_t space_available = queue->max_elems() - queue->size();
199 const size_t num_take_elems = MIN3(space_available / 4,
200 ParGCDesiredObjsFromOverflowList,
201 num_overflow_elems);
202 // Transfer the most recent num_take_elems from the overflow
203 // stack to our work queue.
204 for (size_t i = 0; i != num_take_elems; i++) {
205 oop cur = of_stack->pop();
206 oop obj_to_push = cur->forwardee();
207 assert(Universe::heap()->is_in_reserved(cur), "Should be in heap");
208 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
209 assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap");
210 if (should_be_partially_scanned(obj_to_push, cur)) {
211 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
212 obj_to_push = cur;
213 }
214 bool ok = queue->push(obj_to_push);
215 assert(ok, "Should have succeeded");
216 }
217 assert(young_gen()->overflow_list() == NULL, "Error");
218 return num_take_elems > 0; // was something transferred?
219 }
220
221 void ParScanThreadState::push_on_overflow_stack(oop p) {
222 assert(ParGCUseLocalOverflow, "Else should not call");
223 overflow_stack()->push(p);
224 assert(young_gen()->overflow_list() == NULL, "Error");
225 }
226
227 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz, oop const old, uint age) {
228
229 // Otherwise, if the object is small enough, try to reallocate the
230 // buffer.
231 HeapWord* obj = NULL;
232 if (!_to_space_full) {
233 ParGCAllocBuffer* const plab = to_space_alloc_buffer();
234 Space* const sp = to_space();
235 if (word_sz * 100 <
236 ParallelGCBufferWastePct * plab->word_sz()) {
237 // Is small enough; abandon this buffer and start a new one.
238 plab->retire(false, false);
239 size_t buf_size = plab->word_sz();
240 HeapWord* buf_space = sp->par_allocate(buf_size);
241 if (buf_space == NULL) {
242 const size_t min_bytes =
243 ParGCAllocBuffer::min_size() << LogHeapWordSize;
244 size_t free_bytes = sp->free();
245 while(buf_space == NULL && free_bytes >= min_bytes) {
246 buf_size = free_bytes >> LogHeapWordSize;
247 assert(buf_size == (size_t)align_object_size(buf_size),
248 "Invariant");
249 buf_space = sp->par_allocate(buf_size);
250 free_bytes = sp->free();
251 }
252 }
253 if (buf_space != NULL) {
254 plab->set_word_size(buf_size);
255 plab->set_buf(buf_space);
256 record_survivor_plab(buf_space, buf_size);
257 gc_tracer()->report_promotion_to_new_plab(old, age, false, buf_size);
258 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
259 // Note that we cannot compare buf_size < word_sz below
260 // because of AlignmentReserve (see ParGCAllocBuffer::allocate()).
261 assert(obj != NULL || plab->words_remaining() < word_sz,
262 "Else should have been able to allocate");
263 // It's conceivable that we may be able to use the
264 // buffer we just grabbed for subsequent small requests
265 // even if not for this one.
266 } else {
267 // We're used up.
268 _to_space_full = true;
269 }
270
271 } else {
272 // Too large; allocate the object individually.
273 gc_tracer()->report_promotion_to_new_plab(old, age, false, old->size());
274 obj = sp->par_allocate(word_sz);
275 }
276 }
277 return obj;
278 }
279
280
281 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj,
282 size_t word_sz) {
283 // Is the alloc in the current alloc buffer?
284 if (to_space_alloc_buffer()->contains(obj)) {
285 assert(to_space_alloc_buffer()->contains(obj + word_sz - 1),
286 "Should contain whole object.");
287 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
288 } else {
289 CollectedHeap::fill_with_object(obj, word_sz);
290 }
291 }
292
293 void ParScanThreadState::print_promotion_failure_size() {
294 if (_promotion_failed_info.has_failed() && PrintPromotionFailure) {
295 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ",
296 _thread_num, _promotion_failed_info.first_size());
297 }
298 }
299
300 class ParScanThreadStateSet: private ResourceArray {
301 public:
302 // Initializes states for the specified number of threads;
303 ParScanThreadStateSet(int num_threads,
304 Space& to_space,
305 ParNewGeneration& gen,
306 Generation& old_gen,
307 ObjToScanQueueSet& queue_set,
308 Stack<oop, mtGC>* overflow_stacks_,
309 size_t desired_plab_sz,
310 ParNewTracer* gc_tracer,
311 ParallelTaskTerminator& term);
312
313 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
314
315 inline ParScanThreadState& thread_state(int i);
316
317 void trace_promotion_failed(YoungGCTracer& gc_tracer);
318 void reset(int active_workers, bool promotion_failed);
319 void flush();
320
321 #if TASKQUEUE_STATS
322 static void
323 print_termination_stats_hdr(outputStream* const st = gclog_or_tty);
324 void print_termination_stats(outputStream* const st = gclog_or_tty);
325 static void
326 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty);
327 void print_taskqueue_stats(outputStream* const st = gclog_or_tty);
328 void reset_stats();
329 #endif // TASKQUEUE_STATS
330
331 private:
332 ParallelTaskTerminator& _term;
333 ParNewGeneration& _gen;
334 Generation& _next_gen;
335 public:
336 bool is_valid(int id) const { return id < length(); }
337 ParallelTaskTerminator* terminator() { return &_term; }
338 };
339
340
341 ParScanThreadStateSet::ParScanThreadStateSet(
342 int num_threads, Space& to_space, ParNewGeneration& gen,
343 Generation& old_gen, ObjToScanQueueSet& queue_set,
344 Stack<oop, mtGC>* overflow_stacks,
345 size_t desired_plab_sz,
346 ParNewTracer* gc_tracer, ParallelTaskTerminator& term)
347 : ResourceArray(sizeof(ParScanThreadState), num_threads),
348 _gen(gen), _next_gen(old_gen), _term(term)
349 {
350 assert(num_threads > 0, "sanity check!");
351 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
352 "overflow_stack allocation mismatch");
353 // Initialize states.
354 for (int i = 0; i < num_threads; ++i) {
355 new ((ParScanThreadState*)_data + i)
356 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set,
357 overflow_stacks, desired_plab_sz, gc_tracer, term);
358 }
359 }
360
361 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i)
362 {
363 assert(i >= 0 && i < length(), "sanity check!");
364 return ((ParScanThreadState*)_data)[i];
365 }
366
367 void ParScanThreadStateSet::trace_promotion_failed(YoungGCTracer& gc_tracer) {
368 for (int i = 0; i < length(); ++i) {
369 if (thread_state(i).promotion_failed()) {
370 gc_tracer.report_promotion_failed(thread_state(i).promotion_failed_info());
371 thread_state(i).promotion_failed_info().reset();
372 }
373 }
374 }
375
376 void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed)
377 {
378 _term.reset_for_reuse(active_threads);
379 if (promotion_failed) {
380 for (int i = 0; i < length(); ++i) {
381 thread_state(i).print_promotion_failure_size();
382 }
383 }
384 }
385
386 #if TASKQUEUE_STATS
387 void
388 ParScanThreadState::reset_stats()
389 {
390 taskqueue_stats().reset();
391 _term_attempts = 0;
392 _overflow_refills = 0;
393 _overflow_refill_objs = 0;
394 }
395
396 void ParScanThreadStateSet::reset_stats()
397 {
398 for (int i = 0; i < length(); ++i) {
399 thread_state(i).reset_stats();
400 }
401 }
402
403 void
404 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st)
405 {
406 st->print_raw_cr("GC Termination Stats");
407 st->print_raw_cr(" elapsed --strong roots-- "
408 "-------termination-------");
409 st->print_raw_cr("thr ms ms % "
410 " ms % attempts");
411 st->print_raw_cr("--- --------- --------- ------ "
412 "--------- ------ --------");
413 }
414
415 void ParScanThreadStateSet::print_termination_stats(outputStream* const st)
416 {
417 print_termination_stats_hdr(st);
418
419 for (int i = 0; i < length(); ++i) {
420 const ParScanThreadState & pss = thread_state(i);
421 const double elapsed_ms = pss.elapsed_time() * 1000.0;
422 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
423 const double term_ms = pss.term_time() * 1000.0;
424 st->print_cr("%3d %9.2f %9.2f %6.2f "
425 "%9.2f %6.2f " SIZE_FORMAT_W(8),
426 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
427 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
428 }
429 }
430
431 // Print stats related to work queue activity.
432 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st)
433 {
434 st->print_raw_cr("GC Task Stats");
435 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
436 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
437 }
438
439 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st)
440 {
441 print_taskqueue_stats_hdr(st);
442
443 TaskQueueStats totals;
444 for (int i = 0; i < length(); ++i) {
445 const ParScanThreadState & pss = thread_state(i);
446 const TaskQueueStats & stats = pss.taskqueue_stats();
447 st->print("%3d ", i); stats.print(st); st->cr();
448 totals += stats;
449
450 if (pss.overflow_refills() > 0) {
451 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
452 SIZE_FORMAT_W(10) " overflow objects",
453 pss.overflow_refills(), pss.overflow_refill_objs());
454 }
455 }
456 st->print("tot "); totals.print(st); st->cr();
457
458 DEBUG_ONLY(totals.verify());
459 }
460 #endif // TASKQUEUE_STATS
461
462 void ParScanThreadStateSet::flush()
463 {
464 // Work in this loop should be kept as lightweight as
465 // possible since this might otherwise become a bottleneck
466 // to scaling. Should we add heavy-weight work into this
467 // loop, consider parallelizing the loop into the worker threads.
468 for (int i = 0; i < length(); ++i) {
469 ParScanThreadState& par_scan_state = thread_state(i);
470
471 // Flush stats related to To-space PLAB activity and
472 // retire the last buffer.
473 par_scan_state.to_space_alloc_buffer()->
474 flush_stats_and_retire(_gen.plab_stats(),
475 true /* end_of_gc */,
476 false /* retain */);
477
478 // Every thread has its own age table. We need to merge
479 // them all into one.
480 ageTable *local_table = par_scan_state.age_table();
481 _gen.age_table()->merge(local_table);
482
483 // Inform old gen that we're done.
484 _next_gen.par_promote_alloc_done(i);
485 _next_gen.par_oop_since_save_marks_iterate_done(i);
486 }
487
488 if (UseConcMarkSweepGC && ParallelGCThreads > 0) {
489 // We need to call this even when ResizeOldPLAB is disabled
490 // so as to avoid breaking some asserts. While we may be able
491 // to avoid this by reorganizing the code a bit, I am loathe
492 // to do that unless we find cases where ergo leads to bad
493 // performance.
494 CFLS_LAB::compute_desired_plab_size();
495 }
496 }
497
498 ParScanClosure::ParScanClosure(ParNewGeneration* g,
499 ParScanThreadState* par_scan_state) :
500 OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g)
501 {
502 assert(_g->level() == 0, "Optimized for youngest generation");
503 _boundary = _g->reserved().end();
504 }
505
506 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
507 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
508
509 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
510 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
511
512 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
513 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
514
515 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
516 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
517
518 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
519 ParScanThreadState* par_scan_state)
520 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
521 {}
522
523 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
524 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
525
526 #ifdef WIN32
527 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
528 #endif
529
530 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
531 ParScanThreadState* par_scan_state_,
532 ParScanWithoutBarrierClosure* to_space_closure_,
533 ParScanWithBarrierClosure* old_gen_closure_,
534 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
535 ParNewGeneration* par_gen_,
536 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
537 ObjToScanQueueSet* task_queues_,
538 ParallelTaskTerminator* terminator_) :
539
540 _par_scan_state(par_scan_state_),
541 _to_space_closure(to_space_closure_),
542 _old_gen_closure(old_gen_closure_),
543 _to_space_root_closure(to_space_root_closure_),
544 _old_gen_root_closure(old_gen_root_closure_),
545 _par_gen(par_gen_),
546 _task_queues(task_queues_),
547 _terminator(terminator_)
548 {}
549
550 void ParEvacuateFollowersClosure::do_void() {
551 ObjToScanQueue* work_q = par_scan_state()->work_queue();
552
553 while (true) {
554
555 // Scan to-space and old-gen objs until we run out of both.
556 oop obj_to_scan;
557 par_scan_state()->trim_queues(0);
558
559 // We have no local work, attempt to steal from other threads.
560
561 // attempt to steal work from promoted.
562 if (task_queues()->steal(par_scan_state()->thread_num(),
563 par_scan_state()->hash_seed(),
564 obj_to_scan)) {
565 bool res = work_q->push(obj_to_scan);
566 assert(res, "Empty queue should have room for a push.");
567
568 // if successful, goto Start.
569 continue;
570
571 // try global overflow list.
572 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
573 continue;
574 }
575
576 // Otherwise, offer termination.
577 par_scan_state()->start_term_time();
578 if (terminator()->offer_termination()) break;
579 par_scan_state()->end_term_time();
580 }
581 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
582 "Broken overflow list?");
583 // Finish the last termination pause.
584 par_scan_state()->end_term_time();
585 }
586
587 ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen,
588 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) :
589 AbstractGangTask("ParNewGeneration collection"),
590 _gen(gen), _next_gen(next_gen),
591 _young_old_boundary(young_old_boundary),
592 _state_set(state_set)
593 {}
594
595 // Reset the terminator for the given number of
596 // active threads.
597 void ParNewGenTask::set_for_termination(int active_workers) {
598 _state_set->reset(active_workers, _gen->promotion_failed());
599 // Should the heap be passed in? There's only 1 for now so
600 // grab it instead.
601 GenCollectedHeap* gch = GenCollectedHeap::heap();
602 gch->set_n_termination(active_workers);
603 }
604
605 void ParNewGenTask::work(uint worker_id) {
606 GenCollectedHeap* gch = GenCollectedHeap::heap();
607 // Since this is being done in a separate thread, need new resource
608 // and handle marks.
609 ResourceMark rm;
610 HandleMark hm;
611 // We would need multiple old-gen queues otherwise.
612 assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen.");
613
614 Generation* old_gen = gch->next_gen(_gen);
615
616 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
617 assert(_state_set->is_valid(worker_id), "Should not have been called");
618
619 par_scan_state.set_young_old_boundary(_young_old_boundary);
620
621 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(),
622 gch->rem_set()->klass_rem_set());
623 CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure,
624 &par_scan_state.to_space_root_closure(),
625 false);
626
627 par_scan_state.start_strong_roots();
628 gch->gen_process_roots(_gen->level(),
629 true, // Process younger gens, if any,
630 // as strong roots.
631 false, // no scope; this is parallel code
632 SharedHeap::SO_ScavengeCodeCache,
633 GenCollectedHeap::StrongAndWeakRoots,
634 &par_scan_state.to_space_root_closure(),
635 &par_scan_state.older_gen_closure(),
636 &cld_scan_closure);
637
638 par_scan_state.end_strong_roots();
639
640 // "evacuate followers".
641 par_scan_state.evacuate_followers_closure().do_void();
642 }
643
644 #ifdef _MSC_VER
645 #pragma warning( push )
646 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
647 #endif
648 ParNewGeneration::
649 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level)
650 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"),
651 _overflow_list(NULL),
652 _is_alive_closure(this),
653 _plab_stats(YoungPLABSize, PLABWeight)
654 {
655 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
656 NOT_PRODUCT(_num_par_pushes = 0;)
657 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
658 guarantee(_task_queues != NULL, "task_queues allocation failure.");
659
660 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
661 ObjToScanQueue *q = new ObjToScanQueue();
662 guarantee(q != NULL, "work_queue Allocation failure.");
663 _task_queues->register_queue(i1, q);
664 }
665
666 for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
667 _task_queues->queue(i2)->initialize();
668
669 _overflow_stacks = NULL;
670 if (ParGCUseLocalOverflow) {
671
672 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal
673 // with ','
674 typedef Stack<oop, mtGC> GCOopStack;
675
676 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
677 for (size_t i = 0; i < ParallelGCThreads; ++i) {
678 new (_overflow_stacks + i) Stack<oop, mtGC>();
679 }
680 }
681
682 if (UsePerfData) {
683 EXCEPTION_MARK;
684 ResourceMark rm;
685
686 const char* cname =
687 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
688 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
689 ParallelGCThreads, CHECK);
690 }
691 }
692 #ifdef _MSC_VER
693 #pragma warning( pop )
694 #endif
695
696 // ParNewGeneration::
697 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
698 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
699
700 template <class T>
701 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
702 #ifdef ASSERT
703 {
704 assert(!oopDesc::is_null(*p), "expected non-null ref");
705 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
706 // We never expect to see a null reference being processed
707 // as a weak reference.
708 assert(obj->is_oop(), "expected an oop while scanning weak refs");
709 }
710 #endif // ASSERT
711
712 _par_cl->do_oop_nv(p);
713
714 if (Universe::heap()->is_in_reserved(p)) {
715 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
716 _rs->write_ref_field_gc_par(p, obj);
717 }
718 }
719
720 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
721 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
722
723 // ParNewGeneration::
724 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
725 DefNewGeneration::KeepAliveClosure(cl) {}
726
727 template <class T>
728 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
729 #ifdef ASSERT
730 {
731 assert(!oopDesc::is_null(*p), "expected non-null ref");
732 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
733 // We never expect to see a null reference being processed
734 // as a weak reference.
735 assert(obj->is_oop(), "expected an oop while scanning weak refs");
736 }
737 #endif // ASSERT
738
739 _cl->do_oop_nv(p);
740
741 if (Universe::heap()->is_in_reserved(p)) {
742 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
743 _rs->write_ref_field_gc_par(p, obj);
744 }
745 }
746
747 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
748 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
749
750 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
751 T heap_oop = oopDesc::load_heap_oop(p);
752 if (!oopDesc::is_null(heap_oop)) {
753 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
754 if ((HeapWord*)obj < _boundary) {
755 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
756 oop new_obj = obj->is_forwarded()
757 ? obj->forwardee()
758 : _g->DefNewGeneration::copy_to_survivor_space(obj);
759 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
760 }
761 if (_gc_barrier) {
762 // If p points to a younger generation, mark the card.
763 if ((HeapWord*)obj < _gen_boundary) {
764 _rs->write_ref_field_gc_par(p, obj);
765 }
766 }
767 }
768 }
769
770 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
771 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
772
773 class ParNewRefProcTaskProxy: public AbstractGangTask {
774 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
775 public:
776 ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen,
777 Generation& next_gen,
778 HeapWord* young_old_boundary,
779 ParScanThreadStateSet& state_set);
780
781 private:
782 virtual void work(uint worker_id);
783 virtual void set_for_termination(int active_workers) {
784 _state_set.terminator()->reset_for_reuse(active_workers);
785 }
786 private:
787 ParNewGeneration& _gen;
788 ProcessTask& _task;
789 Generation& _next_gen;
790 HeapWord* _young_old_boundary;
791 ParScanThreadStateSet& _state_set;
792 };
793
794 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(
795 ProcessTask& task, ParNewGeneration& gen,
796 Generation& next_gen,
797 HeapWord* young_old_boundary,
798 ParScanThreadStateSet& state_set)
799 : AbstractGangTask("ParNewGeneration parallel reference processing"),
800 _gen(gen),
801 _task(task),
802 _next_gen(next_gen),
803 _young_old_boundary(young_old_boundary),
804 _state_set(state_set)
805 {
806 }
807
808 void ParNewRefProcTaskProxy::work(uint worker_id)
809 {
810 ResourceMark rm;
811 HandleMark hm;
812 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
813 par_scan_state.set_young_old_boundary(_young_old_boundary);
814 _task.work(worker_id, par_scan_state.is_alive_closure(),
815 par_scan_state.keep_alive_closure(),
816 par_scan_state.evacuate_followers_closure());
817 }
818
819 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
820 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
821 EnqueueTask& _task;
822
823 public:
824 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
825 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
826 _task(task)
827 { }
828
829 virtual void work(uint worker_id)
830 {
831 _task.work(worker_id);
832 }
833 };
834
835
836 void ParNewRefProcTaskExecutor::execute(ProcessTask& task)
837 {
838 GenCollectedHeap* gch = GenCollectedHeap::heap();
839 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
840 "not a generational heap");
841 FlexibleWorkGang* workers = gch->workers();
842 assert(workers != NULL, "Need parallel worker threads.");
843 _state_set.reset(workers->active_workers(), _generation.promotion_failed());
844 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(),
845 _generation.reserved().end(), _state_set);
846 workers->run_task(&rp_task);
847 _state_set.reset(0 /* bad value in debug if not reset */,
848 _generation.promotion_failed());
849 }
850
851 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task)
852 {
853 GenCollectedHeap* gch = GenCollectedHeap::heap();
854 FlexibleWorkGang* workers = gch->workers();
855 assert(workers != NULL, "Need parallel worker threads.");
856 ParNewRefEnqueueTaskProxy enq_task(task);
857 workers->run_task(&enq_task);
858 }
859
860 void ParNewRefProcTaskExecutor::set_single_threaded_mode()
861 {
862 _state_set.flush();
863 GenCollectedHeap* gch = GenCollectedHeap::heap();
864 gch->set_par_threads(0); // 0 ==> non-parallel.
865 gch->save_marks();
866 }
867
868 ScanClosureWithParBarrier::
869 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
870 ScanClosure(g, gc_barrier) {}
871
872 EvacuateFollowersClosureGeneral::
873 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
874 OopsInGenClosure* cur,
875 OopsInGenClosure* older) :
876 _gch(gch), _level(level),
877 _scan_cur_or_nonheap(cur), _scan_older(older)
878 {}
879
880 void EvacuateFollowersClosureGeneral::do_void() {
881 do {
882 // Beware: this call will lead to closure applications via virtual
883 // calls.
884 _gch->oop_since_save_marks_iterate(_level,
885 _scan_cur_or_nonheap,
886 _scan_older);
887 } while (!_gch->no_allocs_since_save_marks(_level));
888 }
889
890
891 // A Generation that does parallel young-gen collection.
892
893 bool ParNewGeneration::_avoid_promotion_undo = false;
894
895 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set, ParNewTracer& gc_tracer) {
896 assert(_promo_failure_scan_stack.is_empty(), "post condition");
897 _promo_failure_scan_stack.clear(true); // Clear cached segments.
898
899 remove_forwarding_pointers();
900 if (PrintGCDetails) {
901 gclog_or_tty->print(" (promotion failed)");
902 }
903 // All the spaces are in play for mark-sweep.
904 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
905 from()->set_next_compaction_space(to());
906 gch->set_incremental_collection_failed();
907 // Inform the next generation that a promotion failure occurred.
908 _next_gen->promotion_failure_occurred();
909
910 // Trace promotion failure in the parallel GC threads
911 thread_state_set.trace_promotion_failed(gc_tracer);
912 // Single threaded code may have reported promotion failure to the global state
913 if (_promotion_failed_info.has_failed()) {
914 gc_tracer.report_promotion_failed(_promotion_failed_info);
915 }
916 // Reset the PromotionFailureALot counters.
917 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();)
918 }
919
920 void ParNewGeneration::collect(bool full,
921 bool clear_all_soft_refs,
922 size_t size,
923 bool is_tlab) {
924 assert(full || size > 0, "otherwise we don't want to collect");
925
926 GenCollectedHeap* gch = GenCollectedHeap::heap();
927
928 _gc_timer->register_gc_start();
929
930 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
931 "not a CMS generational heap");
932 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
933 FlexibleWorkGang* workers = gch->workers();
934 assert(workers != NULL, "Need workgang for parallel work");
935 int active_workers =
936 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
937 workers->active_workers(),
938 Threads::number_of_non_daemon_threads());
939 workers->set_active_workers(active_workers);
940 assert(gch->n_gens() == 2,
941 "Par collection currently only works with single older gen.");
942 _next_gen = gch->next_gen(this);
943 // Do we have to avoid promotion_undo?
944 if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) {
945 set_avoid_promotion_undo(true);
946 }
947
948 // If the next generation is too full to accommodate worst-case promotion
949 // from this generation, pass on collection; let the next generation
950 // do it.
951 if (!collection_attempt_is_safe()) {
952 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
953 return;
954 }
955 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
956
957 ParNewTracer gc_tracer;
958 gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
959 gch->trace_heap_before_gc(&gc_tracer);
960
961 init_assuming_no_promotion_failure();
962
963 if (UseAdaptiveSizePolicy) {
964 set_survivor_overflow(false);
965 size_policy->minor_collection_begin();
966 }
967
968 GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL, gc_tracer.gc_id());
969 // Capture heap used before collection (for printing).
970 size_t gch_prev_used = gch->used();
971
972 SpecializationStats::clear();
973
974 age_table()->clear();
975 to()->clear(SpaceDecorator::Mangle);
976
977 gch->save_marks();
978 assert(workers != NULL, "Need parallel worker threads.");
979 int n_workers = active_workers;
980
981 // Set the correct parallelism (number of queues) in the reference processor
982 ref_processor()->set_active_mt_degree(n_workers);
983
984 // Always set the terminator for the active number of workers
985 // because only those workers go through the termination protocol.
986 ParallelTaskTerminator _term(n_workers, task_queues());
987 ParScanThreadStateSet thread_state_set(workers->active_workers(),
988 *to(), *this, *_next_gen, *task_queues(),
989 _overflow_stacks, desired_plab_sz(), &gc_tracer, _term);
990
991 ParNewGenTask tsk(this, _next_gen, reserved().end(), &thread_state_set);
992 gch->set_par_threads(n_workers);
993 gch->rem_set()->prepare_for_younger_refs_iterate(true);
994 // It turns out that even when we're using 1 thread, doing the work in a
995 // separate thread causes wide variance in run times. We can't help this
996 // in the multi-threaded case, but we special-case n=1 here to get
997 // repeatable measurements of the 1-thread overhead of the parallel code.
998 if (n_workers > 1) {
999 GenCollectedHeap::StrongRootsScope srs(gch);
1000 workers->run_task(&tsk);
1001 } else {
1002 GenCollectedHeap::StrongRootsScope srs(gch);
1003 tsk.work(0);
1004 }
1005 thread_state_set.reset(0 /* Bad value in debug if not reset */,
1006 promotion_failed());
1007
1008 // Process (weak) reference objects found during scavenge.
1009 ReferenceProcessor* rp = ref_processor();
1010 IsAliveClosure is_alive(this);
1011 ScanWeakRefClosure scan_weak_ref(this);
1012 KeepAliveClosure keep_alive(&scan_weak_ref);
1013 ScanClosure scan_without_gc_barrier(this, false);
1014 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
1015 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
1016 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level,
1017 &scan_without_gc_barrier, &scan_with_gc_barrier);
1018 rp->setup_policy(clear_all_soft_refs);
1019 // Can the mt_degree be set later (at run_task() time would be best)?
1020 rp->set_active_mt_degree(active_workers);
1021 ReferenceProcessorStats stats;
1022 if (rp->processing_is_mt()) {
1023 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1024 stats = rp->process_discovered_references(&is_alive, &keep_alive,
1025 &evacuate_followers, &task_executor,
1026 _gc_timer, gc_tracer.gc_id());
1027 } else {
1028 thread_state_set.flush();
1029 gch->set_par_threads(0); // 0 ==> non-parallel.
1030 gch->save_marks();
1031 stats = rp->process_discovered_references(&is_alive, &keep_alive,
1032 &evacuate_followers, NULL,
1033 _gc_timer, gc_tracer.gc_id());
1034 }
1035 gc_tracer.report_gc_reference_stats(stats);
1036 if (!promotion_failed()) {
1037 // Swap the survivor spaces.
1038 eden()->clear(SpaceDecorator::Mangle);
1039 from()->clear(SpaceDecorator::Mangle);
1040 if (ZapUnusedHeapArea) {
1041 // This is now done here because of the piece-meal mangling which
1042 // can check for valid mangling at intermediate points in the
1043 // collection(s). When a minor collection fails to collect
1044 // sufficient space resizing of the young generation can occur
1045 // an redistribute the spaces in the young generation. Mangle
1046 // here so that unzapped regions don't get distributed to
1047 // other spaces.
1048 to()->mangle_unused_area();
1049 }
1050 swap_spaces();
1051
1052 // A successful scavenge should restart the GC time limit count which is
1053 // for full GC's.
1054 size_policy->reset_gc_overhead_limit_count();
1055
1056 assert(to()->is_empty(), "to space should be empty now");
1057
1058 adjust_desired_tenuring_threshold();
1059 } else {
1060 handle_promotion_failed(gch, thread_state_set, gc_tracer);
1061 }
1062 // set new iteration safe limit for the survivor spaces
1063 from()->set_concurrent_iteration_safe_limit(from()->top());
1064 to()->set_concurrent_iteration_safe_limit(to()->top());
1065
1066 if (ResizePLAB) {
1067 plab_stats()->adjust_desired_plab_sz(n_workers);
1068 }
1069
1070 if (PrintGC && !PrintGCDetails) {
1071 gch->print_heap_change(gch_prev_used);
1072 }
1073
1074 if (PrintGCDetails && ParallelGCVerbose) {
1075 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
1076 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
1077 }
1078
1079 if (UseAdaptiveSizePolicy) {
1080 size_policy->minor_collection_end(gch->gc_cause());
1081 size_policy->avg_survived()->sample(from()->used());
1082 }
1083
1084 // We need to use a monotonically non-decreasing time in ms
1085 // or we will see time-warp warnings and os::javaTimeMillis()
1086 // does not guarantee monotonicity.
1087 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1088 update_time_of_last_gc(now);
1089
1090 SpecializationStats::print();
1091
1092 rp->set_enqueuing_is_done(true);
1093 if (rp->processing_is_mt()) {
1094 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1095 rp->enqueue_discovered_references(&task_executor);
1096 } else {
1097 rp->enqueue_discovered_references(NULL);
1098 }
1099 rp->verify_no_references_recorded();
1100
1101 gch->trace_heap_after_gc(&gc_tracer);
1102 gc_tracer.report_tenuring_threshold(tenuring_threshold());
1103
1104 _gc_timer->register_gc_end();
1105
1106 gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1107 }
1108
1109 static int sum;
1110 void ParNewGeneration::waste_some_time() {
1111 for (int i = 0; i < 100; i++) {
1112 sum += i;
1113 }
1114 }
1115
1116 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1117
1118 // Because of concurrency, there are times where an object for which
1119 // "is_forwarded()" is true contains an "interim" forwarding pointer
1120 // value. Such a value will soon be overwritten with a real value.
1121 // This method requires "obj" to have a forwarding pointer, and waits, if
1122 // necessary for a real one to be inserted, and returns it.
1123
1124 oop ParNewGeneration::real_forwardee(oop obj) {
1125 oop forward_ptr = obj->forwardee();
1126 if (forward_ptr != ClaimedForwardPtr) {
1127 return forward_ptr;
1128 } else {
1129 return real_forwardee_slow(obj);
1130 }
1131 }
1132
1133 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1134 // Spin-read if it is claimed but not yet written by another thread.
1135 oop forward_ptr = obj->forwardee();
1136 while (forward_ptr == ClaimedForwardPtr) {
1137 waste_some_time();
1138 assert(obj->is_forwarded(), "precondition");
1139 forward_ptr = obj->forwardee();
1140 }
1141 return forward_ptr;
1142 }
1143
1144 #ifdef ASSERT
1145 bool ParNewGeneration::is_legal_forward_ptr(oop p) {
1146 return
1147 (_avoid_promotion_undo && p == ClaimedForwardPtr)
1148 || Universe::heap()->is_in_reserved(p);
1149 }
1150 #endif
1151
1152 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
1153 if (m->must_be_preserved_for_promotion_failure(obj)) {
1154 // We should really have separate per-worker stacks, rather
1155 // than use locking of a common pair of stacks.
1156 MutexLocker ml(ParGCRareEvent_lock);
1157 preserve_mark(obj, m);
1158 }
1159 }
1160
1161 // Multiple GC threads may try to promote an object. If the object
1162 // is successfully promoted, a forwarding pointer will be installed in
1163 // the object in the young generation. This method claims the right
1164 // to install the forwarding pointer before it copies the object,
1165 // thus avoiding the need to undo the copy as in
1166 // copy_to_survivor_space_avoiding_with_undo.
1167
1168 oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo(
1169 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1170 // In the sequential version, this assert also says that the object is
1171 // not forwarded. That might not be the case here. It is the case that
1172 // the caller observed it to be not forwarded at some time in the past.
1173 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1174
1175 // The sequential code read "old->age()" below. That doesn't work here,
1176 // since the age is in the mark word, and that might be overwritten with
1177 // a forwarding pointer by a parallel thread. So we must save the mark
1178 // word in a local and then analyze it.
1179 oopDesc dummyOld;
1180 dummyOld.set_mark(m);
1181 assert(!dummyOld.is_forwarded(),
1182 "should not be called with forwarding pointer mark word.");
1183
1184 oop new_obj = NULL;
1185 oop forward_ptr;
1186
1187 // Try allocating obj in to-space (unless too old)
1188 if (dummyOld.age() < tenuring_threshold()) {
1189 new_obj = (oop)par_scan_state->alloc_in_to_space(sz, old, dummyOld.age());
1190 if (new_obj == NULL) {
1191 set_survivor_overflow(true);
1192 }
1193 }
1194
1195 if (new_obj == NULL) {
1196 // Either to-space is full or we decided to promote
1197 // try allocating obj tenured
1198
1199 // Attempt to install a null forwarding pointer (atomically),
1200 // to claim the right to install the real forwarding pointer.
1201 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1202 if (forward_ptr != NULL) {
1203 // someone else beat us to it.
1204 return real_forwardee(old);
1205 }
1206
1207 new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
1208 old, m, sz);
1209
1210 if (new_obj == NULL) {
1211 // promotion failed, forward to self
1212 _promotion_failed = true;
1213 new_obj = old;
1214
1215 preserve_mark_if_necessary(old, m);
1216 par_scan_state->register_promotion_failure(sz);
1217 }
1218
1219 old->forward_to(new_obj);
1220 forward_ptr = NULL;
1221 } else {
1222 // Is in to-space; do copying ourselves.
1223 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1224 forward_ptr = old->forward_to_atomic(new_obj);
1225 // Restore the mark word copied above.
1226 new_obj->set_mark(m);
1227 // Increment age if obj still in new generation
1228 new_obj->incr_age();
1229 par_scan_state->age_table()->add(new_obj, sz);
1230 }
1231 assert(new_obj != NULL, "just checking");
1232
1233 #ifndef PRODUCT
1234 // This code must come after the CAS test, or it will print incorrect
1235 // information.
1236 if (TraceScavenge) {
1237 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1238 is_in_reserved(new_obj) ? "copying" : "tenuring",
1239 new_obj->klass()->internal_name(), (void *)old, (void *)new_obj, new_obj->size());
1240 }
1241 #endif
1242
1243 if (forward_ptr == NULL) {
1244 oop obj_to_push = new_obj;
1245 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1246 // Length field used as index of next element to be scanned.
1247 // Real length can be obtained from real_forwardee()
1248 arrayOop(old)->set_length(0);
1249 obj_to_push = old;
1250 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1251 "push forwarded object");
1252 }
1253 // Push it on one of the queues of to-be-scanned objects.
1254 bool simulate_overflow = false;
1255 NOT_PRODUCT(
1256 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1257 // simulate a stack overflow
1258 simulate_overflow = true;
1259 }
1260 )
1261 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1262 // Add stats for overflow pushes.
1263 if (Verbose && PrintGCDetails) {
1264 gclog_or_tty->print("queue overflow!\n");
1265 }
1266 push_on_overflow_list(old, par_scan_state);
1267 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1268 }
1269
1270 return new_obj;
1271 }
1272
1273 // Oops. Someone beat us to it. Undo the allocation. Where did we
1274 // allocate it?
1275 if (is_in_reserved(new_obj)) {
1276 // Must be in to_space.
1277 assert(to()->is_in_reserved(new_obj), "Checking");
1278 if (forward_ptr == ClaimedForwardPtr) {
1279 // Wait to get the real forwarding pointer value.
1280 forward_ptr = real_forwardee(old);
1281 }
1282 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1283 }
1284
1285 return forward_ptr;
1286 }
1287
1288
1289 // Multiple GC threads may try to promote the same object. If two
1290 // or more GC threads copy the object, only one wins the race to install
1291 // the forwarding pointer. The other threads have to undo their copy.
1292
1293 oop ParNewGeneration::copy_to_survivor_space_with_undo(
1294 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1295
1296 // In the sequential version, this assert also says that the object is
1297 // not forwarded. That might not be the case here. It is the case that
1298 // the caller observed it to be not forwarded at some time in the past.
1299 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1300
1301 // The sequential code read "old->age()" below. That doesn't work here,
1302 // since the age is in the mark word, and that might be overwritten with
1303 // a forwarding pointer by a parallel thread. So we must save the mark
1304 // word here, install it in a local oopDesc, and then analyze it.
1305 oopDesc dummyOld;
1306 dummyOld.set_mark(m);
1307 assert(!dummyOld.is_forwarded(),
1308 "should not be called with forwarding pointer mark word.");
1309
1310 bool failed_to_promote = false;
1311 oop new_obj = NULL;
1312 oop forward_ptr;
1313
1314 // Try allocating obj in to-space (unless too old)
1315 if (dummyOld.age() < tenuring_threshold()) {
1316 new_obj = (oop)par_scan_state->alloc_in_to_space(sz, old, dummyOld.age());
1317 if (new_obj == NULL) {
1318 set_survivor_overflow(true);
1319 }
1320 }
1321
1322 if (new_obj == NULL) {
1323 // Either to-space is full or we decided to promote
1324 // try allocating obj tenured
1325 new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
1326 old, m, sz);
1327
1328 if (new_obj == NULL) {
1329 // promotion failed, forward to self
1330 forward_ptr = old->forward_to_atomic(old);
1331 new_obj = old;
1332
1333 if (forward_ptr != NULL) {
1334 return forward_ptr; // someone else succeeded
1335 }
1336
1337 _promotion_failed = true;
1338 failed_to_promote = true;
1339
1340 preserve_mark_if_necessary(old, m);
1341 par_scan_state->register_promotion_failure(sz);
1342 }
1343 } else {
1344 // Is in to-space; do copying ourselves.
1345 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1346 // Restore the mark word copied above.
1347 new_obj->set_mark(m);
1348 // Increment age if new_obj still in new generation
1349 new_obj->incr_age();
1350 par_scan_state->age_table()->add(new_obj, sz);
1351 }
1352 assert(new_obj != NULL, "just checking");
1353
1354 #ifndef PRODUCT
1355 // This code must come after the CAS test, or it will print incorrect
1356 // information.
1357 if (TraceScavenge) {
1358 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1359 is_in_reserved(new_obj) ? "copying" : "tenuring",
1360 new_obj->klass()->internal_name(), (void *)old, (void *)new_obj, new_obj->size());
1361 }
1362 #endif
1363
1364 // Now attempt to install the forwarding pointer (atomically).
1365 // We have to copy the mark word before overwriting with forwarding
1366 // ptr, so we can restore it below in the copy.
1367 if (!failed_to_promote) {
1368 forward_ptr = old->forward_to_atomic(new_obj);
1369 }
1370
1371 if (forward_ptr == NULL) {
1372 oop obj_to_push = new_obj;
1373 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1374 // Length field used as index of next element to be scanned.
1375 // Real length can be obtained from real_forwardee()
1376 arrayOop(old)->set_length(0);
1377 obj_to_push = old;
1378 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1379 "push forwarded object");
1380 }
1381 // Push it on one of the queues of to-be-scanned objects.
1382 bool simulate_overflow = false;
1383 NOT_PRODUCT(
1384 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1385 // simulate a stack overflow
1386 simulate_overflow = true;
1387 }
1388 )
1389 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1390 // Add stats for overflow pushes.
1391 push_on_overflow_list(old, par_scan_state);
1392 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1393 }
1394
1395 return new_obj;
1396 }
1397
1398 // Oops. Someone beat us to it. Undo the allocation. Where did we
1399 // allocate it?
1400 if (is_in_reserved(new_obj)) {
1401 // Must be in to_space.
1402 assert(to()->is_in_reserved(new_obj), "Checking");
1403 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1404 } else {
1405 assert(!_avoid_promotion_undo, "Should not be here if avoiding.");
1406 _next_gen->par_promote_alloc_undo(par_scan_state->thread_num(),
1407 (HeapWord*)new_obj, sz);
1408 }
1409
1410 return forward_ptr;
1411 }
1412
1413 #ifndef PRODUCT
1414 // It's OK to call this multi-threaded; the worst thing
1415 // that can happen is that we'll get a bunch of closely
1416 // spaced simulated overflows, but that's OK, in fact
1417 // probably good as it would exercise the overflow code
1418 // under contention.
1419 bool ParNewGeneration::should_simulate_overflow() {
1420 if (_overflow_counter-- <= 0) { // just being defensive
1421 _overflow_counter = ParGCWorkQueueOverflowInterval;
1422 return true;
1423 } else {
1424 return false;
1425 }
1426 }
1427 #endif
1428
1429 // In case we are using compressed oops, we need to be careful.
1430 // If the object being pushed is an object array, then its length
1431 // field keeps track of the "grey boundary" at which the next
1432 // incremental scan will be done (see ParGCArrayScanChunk).
1433 // When using compressed oops, this length field is kept in the
1434 // lower 32 bits of the erstwhile klass word and cannot be used
1435 // for the overflow chaining pointer (OCP below). As such the OCP
1436 // would itself need to be compressed into the top 32-bits in this
1437 // case. Unfortunately, see below, in the event that we have a
1438 // promotion failure, the node to be pushed on the list can be
1439 // outside of the Java heap, so the heap-based pointer compression
1440 // would not work (we would have potential aliasing between C-heap
1441 // and Java-heap pointers). For this reason, when using compressed
1442 // oops, we simply use a worker-thread-local, non-shared overflow
1443 // list in the form of a growable array, with a slightly different
1444 // overflow stack draining strategy. If/when we start using fat
1445 // stacks here, we can go back to using (fat) pointer chains
1446 // (although some performance comparisons would be useful since
1447 // single global lists have their own performance disadvantages
1448 // as we were made painfully aware not long ago, see 6786503).
1449 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1450 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1451 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1452 if (ParGCUseLocalOverflow) {
1453 // In the case of compressed oops, we use a private, not-shared
1454 // overflow stack.
1455 par_scan_state->push_on_overflow_stack(from_space_obj);
1456 } else {
1457 assert(!UseCompressedOops, "Error");
1458 // if the object has been forwarded to itself, then we cannot
1459 // use the klass pointer for the linked list. Instead we have
1460 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1461 // XXX This is horribly inefficient when a promotion failure occurs
1462 // and should be fixed. XXX FIX ME !!!
1463 #ifndef PRODUCT
1464 Atomic::inc_ptr(&_num_par_pushes);
1465 assert(_num_par_pushes > 0, "Tautology");
1466 #endif
1467 if (from_space_obj->forwardee() == from_space_obj) {
1468 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1469 listhead->forward_to(from_space_obj);
1470 from_space_obj = listhead;
1471 }
1472 oop observed_overflow_list = _overflow_list;
1473 oop cur_overflow_list;
1474 do {
1475 cur_overflow_list = observed_overflow_list;
1476 if (cur_overflow_list != BUSY) {
1477 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1478 } else {
1479 from_space_obj->set_klass_to_list_ptr(NULL);
1480 }
1481 observed_overflow_list =
1482 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1483 } while (cur_overflow_list != observed_overflow_list);
1484 }
1485 }
1486
1487 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1488 bool res;
1489
1490 if (ParGCUseLocalOverflow) {
1491 res = par_scan_state->take_from_overflow_stack();
1492 } else {
1493 assert(!UseCompressedOops, "Error");
1494 res = take_from_overflow_list_work(par_scan_state);
1495 }
1496 return res;
1497 }
1498
1499
1500 // *NOTE*: The overflow list manipulation code here and
1501 // in CMSCollector:: are very similar in shape,
1502 // except that in the CMS case we thread the objects
1503 // directly into the list via their mark word, and do
1504 // not need to deal with special cases below related
1505 // to chunking of object arrays and promotion failure
1506 // handling.
1507 // CR 6797058 has been filed to attempt consolidation of
1508 // the common code.
1509 // Because of the common code, if you make any changes in
1510 // the code below, please check the CMS version to see if
1511 // similar changes might be needed.
1512 // See CMSCollector::par_take_from_overflow_list() for
1513 // more extensive documentation comments.
1514 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1515 ObjToScanQueue* work_q = par_scan_state->work_queue();
1516 // How many to take?
1517 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1518 (size_t)ParGCDesiredObjsFromOverflowList);
1519
1520 assert(!UseCompressedOops, "Error");
1521 assert(par_scan_state->overflow_stack() == NULL, "Error");
1522 if (_overflow_list == NULL) return false;
1523
1524 // Otherwise, there was something there; try claiming the list.
1525 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1526 // Trim off a prefix of at most objsFromOverflow items
1527 Thread* tid = Thread::current();
1528 size_t spin_count = (size_t)ParallelGCThreads;
1529 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1530 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1531 // someone grabbed it before we did ...
1532 // ... we spin for a short while...
1533 os::sleep(tid, sleep_time_millis, false);
1534 if (_overflow_list == NULL) {
1535 // nothing left to take
1536 return false;
1537 } else if (_overflow_list != BUSY) {
1538 // try and grab the prefix
1539 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1540 }
1541 }
1542 if (prefix == NULL || prefix == BUSY) {
1543 // Nothing to take or waited long enough
1544 if (prefix == NULL) {
1545 // Write back the NULL in case we overwrote it with BUSY above
1546 // and it is still the same value.
1547 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1548 }
1549 return false;
1550 }
1551 assert(prefix != NULL && prefix != BUSY, "Error");
1552 size_t i = 1;
1553 oop cur = prefix;
1554 while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1555 i++; cur = cur->list_ptr_from_klass();
1556 }
1557
1558 // Reattach remaining (suffix) to overflow list
1559 if (cur->klass_or_null() == NULL) {
1560 // Write back the NULL in lieu of the BUSY we wrote
1561 // above and it is still the same value.
1562 if (_overflow_list == BUSY) {
1563 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1564 }
1565 } else {
1566 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error");
1567 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list
1568 cur->set_klass_to_list_ptr(NULL); // break off suffix
1569 // It's possible that the list is still in the empty(busy) state
1570 // we left it in a short while ago; in that case we may be
1571 // able to place back the suffix.
1572 oop observed_overflow_list = _overflow_list;
1573 oop cur_overflow_list = observed_overflow_list;
1574 bool attached = false;
1575 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1576 observed_overflow_list =
1577 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1578 if (cur_overflow_list == observed_overflow_list) {
1579 attached = true;
1580 break;
1581 } else cur_overflow_list = observed_overflow_list;
1582 }
1583 if (!attached) {
1584 // Too bad, someone else got in in between; we'll need to do a splice.
1585 // Find the last item of suffix list
1586 oop last = suffix;
1587 while (last->klass_or_null() != NULL) {
1588 last = last->list_ptr_from_klass();
1589 }
1590 // Atomically prepend suffix to current overflow list
1591 observed_overflow_list = _overflow_list;
1592 do {
1593 cur_overflow_list = observed_overflow_list;
1594 if (cur_overflow_list != BUSY) {
1595 // Do the splice ...
1596 last->set_klass_to_list_ptr(cur_overflow_list);
1597 } else { // cur_overflow_list == BUSY
1598 last->set_klass_to_list_ptr(NULL);
1599 }
1600 observed_overflow_list =
1601 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1602 } while (cur_overflow_list != observed_overflow_list);
1603 }
1604 }
1605
1606 // Push objects on prefix list onto this thread's work queue
1607 assert(prefix != NULL && prefix != BUSY, "program logic");
1608 cur = prefix;
1609 ssize_t n = 0;
1610 while (cur != NULL) {
1611 oop obj_to_push = cur->forwardee();
1612 oop next = cur->list_ptr_from_klass();
1613 cur->set_klass(obj_to_push->klass());
1614 // This may be an array object that is self-forwarded. In that case, the list pointer
1615 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1616 if (!is_in_reserved(cur)) {
1617 // This can become a scaling bottleneck when there is work queue overflow coincident
1618 // with promotion failure.
1619 oopDesc* f = cur;
1620 FREE_C_HEAP_ARRAY(oopDesc, f, mtGC);
1621 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1622 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1623 obj_to_push = cur;
1624 }
1625 bool ok = work_q->push(obj_to_push);
1626 assert(ok, "Should have succeeded");
1627 cur = next;
1628 n++;
1629 }
1630 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1631 #ifndef PRODUCT
1632 assert(_num_par_pushes >= n, "Too many pops?");
1633 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1634 #endif
1635 return true;
1636 }
1637 #undef BUSY
1638
1639 void ParNewGeneration::ref_processor_init() {
1640 if (_ref_processor == NULL) {
1641 // Allocate and initialize a reference processor
1642 _ref_processor =
1643 new ReferenceProcessor(_reserved, // span
1644 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1645 (int) ParallelGCThreads, // mt processing degree
1646 refs_discovery_is_mt(), // mt discovery
1647 (int) ParallelGCThreads, // mt discovery degree
1648 refs_discovery_is_atomic(), // atomic_discovery
1649 NULL); // is_alive_non_header
1650 }
1651 }
1652
1653 const char* ParNewGeneration::name() const {
1654 return "par new generation";
1655 }