1 /*
2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp"
27 #include "gc_implementation/parNew/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/parGCAllocBuffer.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/spaceDecorator.hpp"
37 #include "memory/defNewGeneration.inline.hpp"
38 #include "memory/genCollectedHeap.hpp"
39 #include "memory/genOopClosures.inline.hpp"
40 #include "memory/generation.hpp"
41 #include "memory/generation.inline.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/handles.hpp"
50 #include "runtime/handles.inline.hpp"
51 #include "runtime/java.hpp"
52 #include "runtime/thread.hpp"
53 #include "utilities/copy.hpp"
54 #include "utilities/globalDefinitions.hpp"
55 #include "utilities/workgroup.hpp"
56
57 #ifdef _MSC_VER
58 #pragma warning( push )
59 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
60 #endif
61 ParScanThreadState::ParScanThreadState(Space* to_space_,
62 ParNewGeneration* gen_,
63 Generation* old_gen_,
64 int thread_num_,
65 ObjToScanQueueSet* work_queue_set_,
66 Stack<oop, mtGC>* overflow_stacks_,
67 size_t desired_plab_sz_,
68 ParallelTaskTerminator& term_) :
69 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_),
70 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false),
71 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
72 _ageTable(false), // false ==> not the global age table, no perf data.
73 _to_space_alloc_buffer(desired_plab_sz_),
74 _to_space_closure(gen_, this), _old_gen_closure(gen_, this),
75 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this),
76 _older_gen_closure(gen_, this),
77 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
78 &_to_space_root_closure, gen_, &_old_gen_root_closure,
79 work_queue_set_, &term_),
80 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this),
81 _keep_alive_closure(&_scan_weak_ref_closure),
82 _promotion_failure_size(0),
83 _strong_roots_time(0.0), _term_time(0.0)
84 {
85 #if TASKQUEUE_STATS
86 _term_attempts = 0;
87 _overflow_refills = 0;
88 _overflow_refill_objs = 0;
89 #endif // TASKQUEUE_STATS
90
91 _survivor_chunk_array =
92 (ChunkArray*) old_gen()->get_data_recorder(thread_num());
93 _hash_seed = 17; // Might want to take time-based random value.
94 _start = os::elapsedTime();
95 _old_gen_closure.set_generation(old_gen_);
96 _old_gen_root_closure.set_generation(old_gen_);
97 }
98 #ifdef _MSC_VER
99 #pragma warning( pop )
100 #endif
101
102 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
103 size_t plab_word_size) {
104 ChunkArray* sca = survivor_chunk_array();
105 if (sca != NULL) {
106 // A non-null SCA implies that we want the PLAB data recorded.
107 sca->record_sample(plab_start, plab_word_size);
108 }
109 }
110
111 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
112 return new_obj->is_objArray() &&
113 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
114 new_obj != old_obj;
115 }
116
117 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
118 assert(old->is_objArray(), "must be obj array");
119 assert(old->is_forwarded(), "must be forwarded");
120 assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
121 assert(!old_gen()->is_in(old), "must be in young generation.");
122
123 objArrayOop obj = objArrayOop(old->forwardee());
124 // Process ParGCArrayScanChunk elements now
125 // and push the remainder back onto queue
126 int start = arrayOop(old)->length();
127 int end = obj->length();
128 int remainder = end - start;
129 assert(start <= end, "just checking");
130 if (remainder > 2 * ParGCArrayScanChunk) {
131 // Test above combines last partial chunk with a full chunk
132 end = start + ParGCArrayScanChunk;
133 arrayOop(old)->set_length(end);
134 // Push remainder.
135 bool ok = work_queue()->push(old);
136 assert(ok, "just popped, push must be okay");
137 } else {
138 // Restore length so that it can be used if there
139 // is a promotion failure and forwarding pointers
140 // must be removed.
141 arrayOop(old)->set_length(end);
142 }
143
144 // process our set of indices (include header in first chunk)
145 // should make sure end is even (aligned to HeapWord in case of compressed oops)
146 if ((HeapWord *)obj < young_old_boundary()) {
147 // object is in to_space
148 obj->oop_iterate_range(&_to_space_closure, start, end);
149 } else {
150 // object is in old generation
151 obj->oop_iterate_range(&_old_gen_closure, start, end);
152 }
153 }
154
155
156 void ParScanThreadState::trim_queues(int max_size) {
157 ObjToScanQueue* queue = work_queue();
158 do {
159 while (queue->size() > (juint)max_size) {
160 oop obj_to_scan;
161 if (queue->pop_local(obj_to_scan)) {
162 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
163 if (obj_to_scan->is_objArray() &&
164 obj_to_scan->is_forwarded() &&
165 obj_to_scan->forwardee() != obj_to_scan) {
166 scan_partial_array_and_push_remainder(obj_to_scan);
167 } else {
168 // object is in to_space
169 obj_to_scan->oop_iterate(&_to_space_closure);
170 }
171 } else {
172 // object is in old generation
173 obj_to_scan->oop_iterate(&_old_gen_closure);
174 }
175 }
176 }
177 // For the case of compressed oops, we have a private, non-shared
178 // overflow stack, so we eagerly drain it so as to more evenly
179 // distribute load early. Note: this may be good to do in
180 // general rather than delay for the final stealing phase.
181 // If applicable, we'll transfer a set of objects over to our
182 // work queue, allowing them to be stolen and draining our
183 // private overflow stack.
184 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
185 }
186
187 bool ParScanThreadState::take_from_overflow_stack() {
188 assert(ParGCUseLocalOverflow, "Else should not call");
189 assert(young_gen()->overflow_list() == NULL, "Error");
190 ObjToScanQueue* queue = work_queue();
191 Stack<oop, mtGC>* const of_stack = overflow_stack();
192 const size_t num_overflow_elems = of_stack->size();
193 const size_t space_available = queue->max_elems() - queue->size();
194 const size_t num_take_elems = MIN3(space_available / 4,
195 ParGCDesiredObjsFromOverflowList,
196 num_overflow_elems);
197 // Transfer the most recent num_take_elems from the overflow
198 // stack to our work queue.
199 for (size_t i = 0; i != num_take_elems; i++) {
200 oop cur = of_stack->pop();
201 oop obj_to_push = cur->forwardee();
202 assert(Universe::heap()->is_in_reserved(cur), "Should be in heap");
203 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
204 assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap");
205 if (should_be_partially_scanned(obj_to_push, cur)) {
206 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
207 obj_to_push = cur;
208 }
209 bool ok = queue->push(obj_to_push);
210 assert(ok, "Should have succeeded");
211 }
212 assert(young_gen()->overflow_list() == NULL, "Error");
213 return num_take_elems > 0; // was something transferred?
214 }
215
216 void ParScanThreadState::push_on_overflow_stack(oop p) {
217 assert(ParGCUseLocalOverflow, "Else should not call");
218 overflow_stack()->push(p);
219 assert(young_gen()->overflow_list() == NULL, "Error");
220 }
221
222 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
223
224 // Otherwise, if the object is small enough, try to reallocate the
225 // buffer.
226 HeapWord* obj = NULL;
227 if (!_to_space_full) {
228 ParGCAllocBuffer* const plab = to_space_alloc_buffer();
229 Space* const sp = to_space();
230 if (word_sz * 100 <
231 ParallelGCBufferWastePct * plab->word_sz()) {
232 // Is small enough; abandon this buffer and start a new one.
233 plab->retire(false, false);
234 size_t buf_size = plab->word_sz();
235 HeapWord* buf_space = sp->par_allocate(buf_size);
236 if (buf_space == NULL) {
237 const size_t min_bytes =
238 ParGCAllocBuffer::min_size() << LogHeapWordSize;
239 size_t free_bytes = sp->free();
240 while(buf_space == NULL && free_bytes >= min_bytes) {
241 buf_size = free_bytes >> LogHeapWordSize;
242 assert(buf_size == (size_t)align_object_size(buf_size),
243 "Invariant");
244 buf_space = sp->par_allocate(buf_size);
245 free_bytes = sp->free();
246 }
247 }
248 if (buf_space != NULL) {
249 plab->set_word_size(buf_size);
250 plab->set_buf(buf_space);
251 record_survivor_plab(buf_space, buf_size);
252 obj = plab->allocate(word_sz);
253 // Note that we cannot compare buf_size < word_sz below
254 // because of AlignmentReserve (see ParGCAllocBuffer::allocate()).
255 assert(obj != NULL || plab->words_remaining() < word_sz,
256 "Else should have been able to allocate");
257 // It's conceivable that we may be able to use the
258 // buffer we just grabbed for subsequent small requests
259 // even if not for this one.
260 } else {
261 // We're used up.
262 _to_space_full = true;
263 }
264
265 } else {
266 // Too large; allocate the object individually.
267 obj = sp->par_allocate(word_sz);
268 }
269 }
270 return obj;
271 }
272
273
274 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj,
275 size_t word_sz) {
276 // Is the alloc in the current alloc buffer?
277 if (to_space_alloc_buffer()->contains(obj)) {
278 assert(to_space_alloc_buffer()->contains(obj + word_sz - 1),
279 "Should contain whole object.");
280 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
281 } else {
282 CollectedHeap::fill_with_object(obj, word_sz);
283 }
284 }
285
286 void ParScanThreadState::print_and_clear_promotion_failure_size() {
287 if (_promotion_failure_size != 0) {
288 if (PrintPromotionFailure) {
289 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ",
290 _thread_num, _promotion_failure_size);
291 }
292 _promotion_failure_size = 0;
293 }
294 }
295
296 class ParScanThreadStateSet: private ResourceArray {
297 public:
298 // Initializes states for the specified number of threads;
299 ParScanThreadStateSet(int num_threads,
300 Space& to_space,
301 ParNewGeneration& gen,
302 Generation& old_gen,
303 ObjToScanQueueSet& queue_set,
304 Stack<oop, mtGC>* overflow_stacks_,
305 size_t desired_plab_sz,
306 ParallelTaskTerminator& term);
307
308 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
309
310 inline ParScanThreadState& thread_state(int i);
311
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& _next_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), _next_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
361 void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed)
362 {
363 _term.reset_for_reuse(active_threads);
364 if (promotion_failed) {
365 for (int i = 0; i < length(); ++i) {
366 thread_state(i).print_and_clear_promotion_failure_size();
367 }
368 }
369 }
370
371 #if TASKQUEUE_STATS
372 void
373 ParScanThreadState::reset_stats()
374 {
375 taskqueue_stats().reset();
376 _term_attempts = 0;
377 _overflow_refills = 0;
378 _overflow_refill_objs = 0;
379 }
380
381 void ParScanThreadStateSet::reset_stats()
382 {
383 for (int i = 0; i < length(); ++i) {
384 thread_state(i).reset_stats();
385 }
386 }
387
388 void
389 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st)
390 {
391 st->print_raw_cr("GC Termination Stats");
392 st->print_raw_cr(" elapsed --strong roots-- "
393 "-------termination-------");
394 st->print_raw_cr("thr ms ms % "
395 " ms % attempts");
396 st->print_raw_cr("--- --------- --------- ------ "
397 "--------- ------ --------");
398 }
399
400 void ParScanThreadStateSet::print_termination_stats(outputStream* const st)
401 {
402 print_termination_stats_hdr(st);
403
404 for (int i = 0; i < length(); ++i) {
405 const ParScanThreadState & pss = thread_state(i);
406 const double elapsed_ms = pss.elapsed_time() * 1000.0;
407 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
408 const double term_ms = pss.term_time() * 1000.0;
409 st->print_cr("%3d %9.2f %9.2f %6.2f "
410 "%9.2f %6.2f " SIZE_FORMAT_W(8),
411 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
412 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
413 }
414 }
415
416 // Print stats related to work queue activity.
417 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st)
418 {
419 st->print_raw_cr("GC Task Stats");
420 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
421 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
422 }
423
424 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st)
425 {
426 print_taskqueue_stats_hdr(st);
427
428 TaskQueueStats totals;
429 for (int i = 0; i < length(); ++i) {
430 const ParScanThreadState & pss = thread_state(i);
431 const TaskQueueStats & stats = pss.taskqueue_stats();
432 st->print("%3d ", i); stats.print(st); st->cr();
433 totals += stats;
434
435 if (pss.overflow_refills() > 0) {
436 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
437 SIZE_FORMAT_W(10) " overflow objects",
438 pss.overflow_refills(), pss.overflow_refill_objs());
439 }
440 }
441 st->print("tot "); totals.print(st); st->cr();
442
443 DEBUG_ONLY(totals.verify());
444 }
445 #endif // TASKQUEUE_STATS
446
447 void ParScanThreadStateSet::flush()
448 {
449 // Work in this loop should be kept as lightweight as
450 // possible since this might otherwise become a bottleneck
451 // to scaling. Should we add heavy-weight work into this
452 // loop, consider parallelizing the loop into the worker threads.
453 for (int i = 0; i < length(); ++i) {
454 ParScanThreadState& par_scan_state = thread_state(i);
455
456 // Flush stats related to To-space PLAB activity and
457 // retire the last buffer.
458 par_scan_state.to_space_alloc_buffer()->
459 flush_stats_and_retire(_gen.plab_stats(),
460 true /* end_of_gc */,
461 false /* retain */);
462
463 // Every thread has its own age table. We need to merge
464 // them all into one.
465 ageTable *local_table = par_scan_state.age_table();
466 _gen.age_table()->merge(local_table);
467
468 // Inform old gen that we're done.
469 _next_gen.par_promote_alloc_done(i);
470 _next_gen.par_oop_since_save_marks_iterate_done(i);
471 }
472
473 if (UseConcMarkSweepGC && ParallelGCThreads > 0) {
474 // We need to call this even when ResizeOldPLAB is disabled
475 // so as to avoid breaking some asserts. While we may be able
476 // to avoid this by reorganizing the code a bit, I am loathe
477 // to do that unless we find cases where ergo leads to bad
478 // performance.
479 CFLS_LAB::compute_desired_plab_size();
480 }
481 }
482
483 ParScanClosure::ParScanClosure(ParNewGeneration* g,
484 ParScanThreadState* par_scan_state) :
485 OopsInGenClosure(g), _par_scan_state(par_scan_state), _g(g)
486 {
487 assert(_g->level() == 0, "Optimized for youngest generation");
488 _boundary = _g->reserved().end();
489 }
490
491 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
492 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
493
494 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
495 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
496
497 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
498 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
499
500 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
501 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
502
503 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
504 ParScanThreadState* par_scan_state)
505 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
506 {}
507
508 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
509 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
510
511 #ifdef WIN32
512 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
513 #endif
514
515 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
516 ParScanThreadState* par_scan_state_,
517 ParScanWithoutBarrierClosure* to_space_closure_,
518 ParScanWithBarrierClosure* old_gen_closure_,
519 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
520 ParNewGeneration* par_gen_,
521 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
522 ObjToScanQueueSet* task_queues_,
523 ParallelTaskTerminator* terminator_) :
524
525 _par_scan_state(par_scan_state_),
526 _to_space_closure(to_space_closure_),
527 _old_gen_closure(old_gen_closure_),
528 _to_space_root_closure(to_space_root_closure_),
529 _old_gen_root_closure(old_gen_root_closure_),
530 _par_gen(par_gen_),
531 _task_queues(task_queues_),
532 _terminator(terminator_)
533 {}
534
535 void ParEvacuateFollowersClosure::do_void() {
536 ObjToScanQueue* work_q = par_scan_state()->work_queue();
537
538 while (true) {
539
540 // Scan to-space and old-gen objs until we run out of both.
541 oop obj_to_scan;
542 par_scan_state()->trim_queues(0);
543
544 // We have no local work, attempt to steal from other threads.
545
546 // attempt to steal work from promoted.
547 if (task_queues()->steal(par_scan_state()->thread_num(),
548 par_scan_state()->hash_seed(),
549 obj_to_scan)) {
550 bool res = work_q->push(obj_to_scan);
551 assert(res, "Empty queue should have room for a push.");
552
553 // if successful, goto Start.
554 continue;
555
556 // try global overflow list.
557 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
558 continue;
559 }
560
561 // Otherwise, offer termination.
562 par_scan_state()->start_term_time();
563 if (terminator()->offer_termination()) break;
564 par_scan_state()->end_term_time();
565 }
566 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
567 "Broken overflow list?");
568 // Finish the last termination pause.
569 par_scan_state()->end_term_time();
570 }
571
572 ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen,
573 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) :
574 AbstractGangTask("ParNewGeneration collection"),
575 _gen(gen), _next_gen(next_gen),
576 _young_old_boundary(young_old_boundary),
577 _state_set(state_set)
578 {}
579
580 // Reset the terminator for the given number of
581 // active threads.
582 void ParNewGenTask::set_for_termination(int active_workers) {
583 _state_set->reset(active_workers, _gen->promotion_failed());
584 // Should the heap be passed in? There's only 1 for now so
585 // grab it instead.
586 GenCollectedHeap* gch = GenCollectedHeap::heap();
587 gch->set_n_termination(active_workers);
588 }
589
590 // The "i" passed to this method is the part of the work for
591 // this thread. It is not the worker ID. The "i" is derived
592 // from _started_workers which is incremented in internal_note_start()
593 // called in GangWorker loop() and which is called under the
594 // which is called under the protection of the gang monitor and is
595 // called after a task is started. So "i" is based on
596 // first-come-first-served.
597
598 void ParNewGenTask::work(uint worker_id) {
599 GenCollectedHeap* gch = GenCollectedHeap::heap();
600 // Since this is being done in a separate thread, need new resource
601 // and handle marks.
602 ResourceMark rm;
603 HandleMark hm;
604 // We would need multiple old-gen queues otherwise.
605 assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen.");
606
607 Generation* old_gen = gch->next_gen(_gen);
608
609 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
610 assert(_state_set->is_valid(worker_id), "Should not have been called");
611
612 par_scan_state.set_young_old_boundary(_young_old_boundary);
613
614 par_scan_state.start_strong_roots();
615 gch->gen_process_strong_roots(_gen->level(),
616 true, // Process younger gens, if any,
617 // as strong roots.
618 false, // no scope; this is parallel code
619 false, // not collecting perm generation.
620 SharedHeap::SO_AllClasses,
621 &par_scan_state.to_space_root_closure(),
622 true, // walk *all* scavengable nmethods
623 &par_scan_state.older_gen_closure());
624 par_scan_state.end_strong_roots();
625
626 // "evacuate followers".
627 par_scan_state.evacuate_followers_closure().do_void();
628 }
629
630 #ifdef _MSC_VER
631 #pragma warning( push )
632 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
633 #endif
634 ParNewGeneration::
635 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level)
636 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"),
637 _overflow_list(NULL),
638 _is_alive_closure(this),
639 _plab_stats(YoungPLABSize, PLABWeight)
640 {
641 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
642 NOT_PRODUCT(_num_par_pushes = 0;)
643 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
644 guarantee(_task_queues != NULL, "task_queues allocation failure.");
645
646 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
647 ObjToScanQueue *q = new ObjToScanQueue();
648 guarantee(q != NULL, "work_queue Allocation failure.");
649 _task_queues->register_queue(i1, q);
650 }
651
652 for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
653 _task_queues->queue(i2)->initialize();
654
655 _overflow_stacks = NULL;
656 if (ParGCUseLocalOverflow) {
657
658 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal
659 // with ','
660 typedef Stack<oop, mtGC> GCOopStack;
661
662 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
663 for (size_t i = 0; i < ParallelGCThreads; ++i) {
664 new (_overflow_stacks + i) Stack<oop, mtGC>();
665 }
666 }
667
668 if (UsePerfData) {
669 EXCEPTION_MARK;
670 ResourceMark rm;
671
672 const char* cname =
673 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
674 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
675 ParallelGCThreads, CHECK);
676 }
677 }
678 #ifdef _MSC_VER
679 #pragma warning( pop )
680 #endif
681
682 // ParNewGeneration::
683 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
684 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
685
686 template <class T>
687 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
688 #ifdef ASSERT
689 {
690 assert(!oopDesc::is_null(*p), "expected non-null ref");
691 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
692 // We never expect to see a null reference being processed
693 // as a weak reference.
694 assert(obj->is_oop(), "expected an oop while scanning weak refs");
695 }
696 #endif // ASSERT
697
698 _par_cl->do_oop_nv(p);
699
700 if (Universe::heap()->is_in_reserved(p)) {
701 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
702 _rs->write_ref_field_gc_par(p, obj);
703 }
704 }
705
706 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
707 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
708
709 // ParNewGeneration::
710 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
711 DefNewGeneration::KeepAliveClosure(cl) {}
712
713 template <class T>
714 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
715 #ifdef ASSERT
716 {
717 assert(!oopDesc::is_null(*p), "expected non-null ref");
718 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
719 // We never expect to see a null reference being processed
720 // as a weak reference.
721 assert(obj->is_oop(), "expected an oop while scanning weak refs");
722 }
723 #endif // ASSERT
724
725 _cl->do_oop_nv(p);
726
727 if (Universe::heap()->is_in_reserved(p)) {
728 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
729 _rs->write_ref_field_gc_par(p, obj);
730 }
731 }
732
733 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
734 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
735
736 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
737 T heap_oop = oopDesc::load_heap_oop(p);
738 if (!oopDesc::is_null(heap_oop)) {
739 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
740 if ((HeapWord*)obj < _boundary) {
741 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
742 oop new_obj = obj->is_forwarded()
743 ? obj->forwardee()
744 : _g->DefNewGeneration::copy_to_survivor_space(obj);
745 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
746 }
747 if (_gc_barrier) {
748 // If p points to a younger generation, mark the card.
749 if ((HeapWord*)obj < _gen_boundary) {
750 _rs->write_ref_field_gc_par(p, obj);
751 }
752 }
753 }
754 }
755
756 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
757 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
758
759 class ParNewRefProcTaskProxy: public AbstractGangTask {
760 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
761 public:
762 ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen,
763 Generation& next_gen,
764 HeapWord* young_old_boundary,
765 ParScanThreadStateSet& state_set);
766
767 private:
768 virtual void work(uint worker_id);
769 virtual void set_for_termination(int active_workers) {
770 _state_set.terminator()->reset_for_reuse(active_workers);
771 }
772 private:
773 ParNewGeneration& _gen;
774 ProcessTask& _task;
775 Generation& _next_gen;
776 HeapWord* _young_old_boundary;
777 ParScanThreadStateSet& _state_set;
778 };
779
780 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(
781 ProcessTask& task, ParNewGeneration& gen,
782 Generation& next_gen,
783 HeapWord* young_old_boundary,
784 ParScanThreadStateSet& state_set)
785 : AbstractGangTask("ParNewGeneration parallel reference processing"),
786 _gen(gen),
787 _task(task),
788 _next_gen(next_gen),
789 _young_old_boundary(young_old_boundary),
790 _state_set(state_set)
791 {
792 }
793
794 void ParNewRefProcTaskProxy::work(uint worker_id)
795 {
796 ResourceMark rm;
797 HandleMark hm;
798 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
799 par_scan_state.set_young_old_boundary(_young_old_boundary);
800 _task.work(worker_id, par_scan_state.is_alive_closure(),
801 par_scan_state.keep_alive_closure(),
802 par_scan_state.evacuate_followers_closure());
803 }
804
805 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
806 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
807 EnqueueTask& _task;
808
809 public:
810 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
811 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
812 _task(task)
813 { }
814
815 virtual void work(uint worker_id)
816 {
817 _task.work(worker_id);
818 }
819 };
820
821
822 void ParNewRefProcTaskExecutor::execute(ProcessTask& task)
823 {
824 GenCollectedHeap* gch = GenCollectedHeap::heap();
825 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
826 "not a generational heap");
827 FlexibleWorkGang* workers = gch->workers();
828 assert(workers != NULL, "Need parallel worker threads.");
829 _state_set.reset(workers->active_workers(), _generation.promotion_failed());
830 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(),
831 _generation.reserved().end(), _state_set);
832 workers->run_task(&rp_task);
833 _state_set.reset(0 /* bad value in debug if not reset */,
834 _generation.promotion_failed());
835 }
836
837 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task)
838 {
839 GenCollectedHeap* gch = GenCollectedHeap::heap();
840 FlexibleWorkGang* workers = gch->workers();
841 assert(workers != NULL, "Need parallel worker threads.");
842 ParNewRefEnqueueTaskProxy enq_task(task);
843 workers->run_task(&enq_task);
844 }
845
846 void ParNewRefProcTaskExecutor::set_single_threaded_mode()
847 {
848 _state_set.flush();
849 GenCollectedHeap* gch = GenCollectedHeap::heap();
850 gch->set_par_threads(0); // 0 ==> non-parallel.
851 gch->save_marks();
852 }
853
854 ScanClosureWithParBarrier::
855 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
856 ScanClosure(g, gc_barrier) {}
857
858 EvacuateFollowersClosureGeneral::
859 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
860 OopsInGenClosure* cur,
861 OopsInGenClosure* older) :
862 _gch(gch), _level(level),
863 _scan_cur_or_nonheap(cur), _scan_older(older)
864 {}
865
866 void EvacuateFollowersClosureGeneral::do_void() {
867 do {
868 // Beware: this call will lead to closure applications via virtual
869 // calls.
870 _gch->oop_since_save_marks_iterate(_level,
871 _scan_cur_or_nonheap,
872 _scan_older);
873 } while (!_gch->no_allocs_since_save_marks(_level));
874 }
875
876
877 bool ParNewGeneration::_avoid_promotion_undo = false;
878
879 void ParNewGeneration::adjust_desired_tenuring_threshold() {
880 // Set the desired survivor size to half the real survivor space
881 _tenuring_threshold =
882 age_table()->compute_tenuring_threshold(to()->capacity()/HeapWordSize);
883 }
884
885 // A Generation that does parallel young-gen collection.
886
887 void ParNewGeneration::collect(bool full,
888 bool clear_all_soft_refs,
889 size_t size,
890 bool is_tlab) {
891 assert(full || size > 0, "otherwise we don't want to collect");
892
893 GenCollectedHeap* gch = GenCollectedHeap::heap();
894
895 _gc_timer->register_gc_start(os::elapsed_counter());
896 ParNewTracer gc_tracer;
897 gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
898
899 gch->trace_heap_before_gc(&gc_tracer);
900
901 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
902 "not a CMS generational heap");
903 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
904 FlexibleWorkGang* workers = gch->workers();
905 assert(workers != NULL, "Need workgang for parallel work");
906 int active_workers =
907 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
908 workers->active_workers(),
909 Threads::number_of_non_daemon_threads());
910 workers->set_active_workers(active_workers);
911 _next_gen = gch->next_gen(this);
912 assert(_next_gen != NULL,
913 "This must be the youngest gen, and not the only gen");
914 assert(gch->n_gens() == 2,
915 "Par collection currently only works with single older gen.");
916 // Do we have to avoid promotion_undo?
917 if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) {
918 set_avoid_promotion_undo(true);
919 }
920
921 // If the next generation is too full to accomodate worst-case promotion
922 // from this generation, pass on collection; let the next generation
923 // do it.
924 if (!collection_attempt_is_safe()) {
925 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
926 return;
927 }
928 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
929
930 init_assuming_no_promotion_failure();
931
932 if (UseAdaptiveSizePolicy) {
933 set_survivor_overflow(false);
934 size_policy->minor_collection_begin();
935 }
936
937 GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL);
938 // Capture heap used before collection (for printing).
939 size_t gch_prev_used = gch->used();
940
941 SpecializationStats::clear();
942
943 age_table()->clear();
944 to()->clear(SpaceDecorator::Mangle);
945
946 gch->save_marks();
947 assert(workers != NULL, "Need parallel worker threads.");
948 int n_workers = active_workers;
949
950 // Set the correct parallelism (number of queues) in the reference processor
951 ref_processor()->set_active_mt_degree(n_workers);
952
953 // Always set the terminator for the active number of workers
954 // because only those workers go through the termination protocol.
955 ParallelTaskTerminator _term(n_workers, task_queues());
956 ParScanThreadStateSet thread_state_set(workers->active_workers(),
957 *to(), *this, *_next_gen, *task_queues(),
958 _overflow_stacks, desired_plab_sz(), _term);
959
960 ParNewGenTask tsk(this, _next_gen, reserved().end(), &thread_state_set);
961 gch->set_par_threads(n_workers);
962 gch->rem_set()->prepare_for_younger_refs_iterate(true);
963 // It turns out that even when we're using 1 thread, doing the work in a
964 // separate thread causes wide variance in run times. We can't help this
965 // in the multi-threaded case, but we special-case n=1 here to get
966 // repeatable measurements of the 1-thread overhead of the parallel code.
967 if (n_workers > 1) {
968 GenCollectedHeap::StrongRootsScope srs(gch);
969 workers->run_task(&tsk);
970 } else {
971 GenCollectedHeap::StrongRootsScope srs(gch);
972 tsk.work(0);
973 }
974 thread_state_set.reset(0 /* Bad value in debug if not reset */,
975 promotion_failed());
976
977 // Process (weak) reference objects found during scavenge.
978 ReferenceProcessor* rp = ref_processor();
979 IsAliveClosure is_alive(this);
980 ScanWeakRefClosure scan_weak_ref(this);
981 KeepAliveClosure keep_alive(&scan_weak_ref);
982 ScanClosure scan_without_gc_barrier(this, false);
983 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
984 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
985 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level,
986 &scan_without_gc_barrier, &scan_with_gc_barrier);
987 rp->setup_policy(clear_all_soft_refs);
988 // Can the mt_degree be set later (at run_task() time would be best)?
989 rp->set_active_mt_degree(active_workers);
990 if (rp->processing_is_mt()) {
991 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
992 rp->process_discovered_references(&is_alive, &keep_alive,
993 &evacuate_followers, &task_executor,
994 _gc_timer);
995 } else {
996 thread_state_set.flush();
997 gch->set_par_threads(0); // 0 ==> non-parallel.
998 gch->save_marks();
999 rp->process_discovered_references(&is_alive, &keep_alive,
1000 &evacuate_followers, NULL,
1001 _gc_timer);
1002 }
1003 if (!promotion_failed()) {
1004 // Swap the survivor spaces.
1005 eden()->clear(SpaceDecorator::Mangle);
1006 from()->clear(SpaceDecorator::Mangle);
1007 if (ZapUnusedHeapArea) {
1008 // This is now done here because of the piece-meal mangling which
1009 // can check for valid mangling at intermediate points in the
1010 // collection(s). When a minor collection fails to collect
1011 // sufficient space resizing of the young generation can occur
1012 // an redistribute the spaces in the young generation. Mangle
1013 // here so that unzapped regions don't get distributed to
1014 // other spaces.
1015 to()->mangle_unused_area();
1016 }
1017 swap_spaces();
1018
1019 // A successful scavenge should restart the GC time limit count which is
1020 // for full GC's.
1021 size_policy->reset_gc_overhead_limit_count();
1022
1023 assert(to()->is_empty(), "to space should be empty now");
1024 } else {
1025 assert(_promo_failure_scan_stack.is_empty(), "post condition");
1026 _promo_failure_scan_stack.clear(true); // Clear cached segments.
1027
1028 remove_forwarding_pointers();
1029 if (PrintGCDetails) {
1030 gclog_or_tty->print(" (promotion failed)");
1031 }
1032 // All the spaces are in play for mark-sweep.
1033 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
1034 from()->set_next_compaction_space(to());
1035 gch->set_incremental_collection_failed();
1036 // Inform the next generation that a promotion failure occurred.
1037 _next_gen->promotion_failure_occurred();
1038
1039 // Reset the PromotionFailureALot counters.
1040 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();)
1041 }
1042 // set new iteration safe limit for the survivor spaces
1043 from()->set_concurrent_iteration_safe_limit(from()->top());
1044 to()->set_concurrent_iteration_safe_limit(to()->top());
1045
1046 adjust_desired_tenuring_threshold();
1047 if (ResizePLAB) {
1048 plab_stats()->adjust_desired_plab_sz(n_workers);
1049 }
1050
1051 if (PrintGC && !PrintGCDetails) {
1052 gch->print_heap_change(gch_prev_used);
1053 }
1054
1055 if (PrintGCDetails && ParallelGCVerbose) {
1056 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
1057 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
1058 }
1059
1060 if (UseAdaptiveSizePolicy) {
1061 size_policy->minor_collection_end(gch->gc_cause());
1062 size_policy->avg_survived()->sample(from()->used());
1063 }
1064
1065 // We need to use a monotonically non-deccreasing time in ms
1066 // or we will see time-warp warnings and os::javaTimeMillis()
1067 // does not guarantee monotonicity.
1068 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1069 update_time_of_last_gc(now);
1070
1071 SpecializationStats::print();
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
1084 _gc_timer->register_gc_end(os::elapsed_counter());
1085
1086 gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1087 }
1088
1089 static int sum;
1090 void ParNewGeneration::waste_some_time() {
1091 for (int i = 0; i < 100; i++) {
1092 sum += i;
1093 }
1094 }
1095
1096 static const oop ClaimedForwardPtr = oop(0x4);
1097
1098 // Because of concurrency, there are times where an object for which
1099 // "is_forwarded()" is true contains an "interim" forwarding pointer
1100 // value. Such a value will soon be overwritten with a real value.
1101 // This method requires "obj" to have a forwarding pointer, and waits, if
1102 // necessary for a real one to be inserted, and returns it.
1103
1104 oop ParNewGeneration::real_forwardee(oop obj) {
1105 oop forward_ptr = obj->forwardee();
1106 if (forward_ptr != ClaimedForwardPtr) {
1107 return forward_ptr;
1108 } else {
1109 return real_forwardee_slow(obj);
1110 }
1111 }
1112
1113 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1114 // Spin-read if it is claimed but not yet written by another thread.
1115 oop forward_ptr = obj->forwardee();
1116 while (forward_ptr == ClaimedForwardPtr) {
1117 waste_some_time();
1118 assert(obj->is_forwarded(), "precondition");
1119 forward_ptr = obj->forwardee();
1120 }
1121 return forward_ptr;
1122 }
1123
1124 #ifdef ASSERT
1125 bool ParNewGeneration::is_legal_forward_ptr(oop p) {
1126 return
1127 (_avoid_promotion_undo && p == ClaimedForwardPtr)
1128 || Universe::heap()->is_in_reserved(p);
1129 }
1130 #endif
1131
1132 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
1133 if (m->must_be_preserved_for_promotion_failure(obj)) {
1134 // We should really have separate per-worker stacks, rather
1135 // than use locking of a common pair of stacks.
1136 MutexLocker ml(ParGCRareEvent_lock);
1137 preserve_mark(obj, m);
1138 }
1139 }
1140
1141 // Multiple GC threads may try to promote an object. If the object
1142 // is successfully promoted, a forwarding pointer will be installed in
1143 // the object in the young generation. This method claims the right
1144 // to install the forwarding pointer before it copies the object,
1145 // thus avoiding the need to undo the copy as in
1146 // copy_to_survivor_space_avoiding_with_undo.
1147
1148 oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo(
1149 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1150 // In the sequential version, this assert also says that the object is
1151 // not forwarded. That might not be the case here. It is the case that
1152 // the caller observed it to be not forwarded at some time in the past.
1153 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1154
1155 // The sequential code read "old->age()" below. That doesn't work here,
1156 // since the age is in the mark word, and that might be overwritten with
1157 // a forwarding pointer by a parallel thread. So we must save the mark
1158 // word in a local and then analyze it.
1159 oopDesc dummyOld;
1160 dummyOld.set_mark(m);
1161 assert(!dummyOld.is_forwarded(),
1162 "should not be called with forwarding pointer mark word.");
1163
1164 oop new_obj = NULL;
1165 oop forward_ptr;
1166
1167 // Try allocating obj in to-space (unless too old)
1168 if (dummyOld.age() < tenuring_threshold()) {
1169 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1170 if (new_obj == NULL) {
1171 set_survivor_overflow(true);
1172 }
1173 }
1174
1175 if (new_obj == NULL) {
1176 // Either to-space is full or we decided to promote
1177 // try allocating obj tenured
1178
1179 // Attempt to install a null forwarding pointer (atomically),
1180 // to claim the right to install the real forwarding pointer.
1181 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1182 if (forward_ptr != NULL) {
1183 // someone else beat us to it.
1184 return real_forwardee(old);
1185 }
1186
1187 new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
1188 old, m, sz);
1189
1190 if (new_obj == NULL) {
1191 // promotion failed, forward to self
1192 _promotion_failed = true;
1193 new_obj = old;
1194
1195 preserve_mark_if_necessary(old, m);
1196 // Log the size of the maiden promotion failure
1197 par_scan_state->log_promotion_failure(sz);
1198 }
1199
1200 old->forward_to(new_obj);
1201 forward_ptr = NULL;
1202 } else {
1203 // Is in to-space; do copying ourselves.
1204 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1205 forward_ptr = old->forward_to_atomic(new_obj);
1206 // Restore the mark word copied above.
1207 new_obj->set_mark(m);
1208 // Increment age if obj still in new generation
1209 new_obj->incr_age();
1210 par_scan_state->age_table()->add(new_obj, sz);
1211 }
1212 assert(new_obj != NULL, "just checking");
1213
1214 if (forward_ptr == NULL) {
1215 oop obj_to_push = new_obj;
1216 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1217 // Length field used as index of next element to be scanned.
1218 // Real length can be obtained from real_forwardee()
1219 arrayOop(old)->set_length(0);
1220 obj_to_push = old;
1221 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1222 "push forwarded object");
1223 }
1224 // Push it on one of the queues of to-be-scanned objects.
1225 bool simulate_overflow = false;
1226 NOT_PRODUCT(
1227 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1228 // simulate a stack overflow
1229 simulate_overflow = true;
1230 }
1231 )
1232 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1233 // Add stats for overflow pushes.
1234 if (Verbose && PrintGCDetails) {
1235 gclog_or_tty->print("queue overflow!\n");
1236 }
1237 push_on_overflow_list(old, par_scan_state);
1238 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1239 }
1240
1241 return new_obj;
1242 }
1243
1244 // Oops. Someone beat us to it. Undo the allocation. Where did we
1245 // allocate it?
1246 if (is_in_reserved(new_obj)) {
1247 // Must be in to_space.
1248 assert(to()->is_in_reserved(new_obj), "Checking");
1249 if (forward_ptr == ClaimedForwardPtr) {
1250 // Wait to get the real forwarding pointer value.
1251 forward_ptr = real_forwardee(old);
1252 }
1253 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1254 }
1255
1256 return forward_ptr;
1257 }
1258
1259
1260 // Multiple GC threads may try to promote the same object. If two
1261 // or more GC threads copy the object, only one wins the race to install
1262 // the forwarding pointer. The other threads have to undo their copy.
1263
1264 oop ParNewGeneration::copy_to_survivor_space_with_undo(
1265 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1266
1267 // In the sequential version, this assert also says that the object is
1268 // not forwarded. That might not be the case here. It is the case that
1269 // the caller observed it to be not forwarded at some time in the past.
1270 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1271
1272 // The sequential code read "old->age()" below. That doesn't work here,
1273 // since the age is in the mark word, and that might be overwritten with
1274 // a forwarding pointer by a parallel thread. So we must save the mark
1275 // word here, install it in a local oopDesc, and then analyze it.
1276 oopDesc dummyOld;
1277 dummyOld.set_mark(m);
1278 assert(!dummyOld.is_forwarded(),
1279 "should not be called with forwarding pointer mark word.");
1280
1281 bool failed_to_promote = false;
1282 oop new_obj = NULL;
1283 oop forward_ptr;
1284
1285 // Try allocating obj in to-space (unless too old)
1286 if (dummyOld.age() < tenuring_threshold()) {
1287 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1288 if (new_obj == NULL) {
1289 set_survivor_overflow(true);
1290 }
1291 }
1292
1293 if (new_obj == NULL) {
1294 // Either to-space is full or we decided to promote
1295 // try allocating obj tenured
1296 new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
1297 old, m, sz);
1298
1299 if (new_obj == NULL) {
1300 // promotion failed, forward to self
1301 forward_ptr = old->forward_to_atomic(old);
1302 new_obj = old;
1303
1304 if (forward_ptr != NULL) {
1305 return forward_ptr; // someone else succeeded
1306 }
1307
1308 _promotion_failed = true;
1309 failed_to_promote = true;
1310
1311 preserve_mark_if_necessary(old, m);
1312 // Log the size of the maiden promotion failure
1313 par_scan_state->log_promotion_failure(sz);
1314 }
1315 } else {
1316 // Is in to-space; do copying ourselves.
1317 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1318 // Restore the mark word copied above.
1319 new_obj->set_mark(m);
1320 // Increment age if new_obj still in new generation
1321 new_obj->incr_age();
1322 par_scan_state->age_table()->add(new_obj, sz);
1323 }
1324 assert(new_obj != NULL, "just checking");
1325
1326 // Now attempt to install the forwarding pointer (atomically).
1327 // We have to copy the mark word before overwriting with forwarding
1328 // ptr, so we can restore it below in the copy.
1329 if (!failed_to_promote) {
1330 forward_ptr = old->forward_to_atomic(new_obj);
1331 }
1332
1333 if (forward_ptr == NULL) {
1334 oop obj_to_push = new_obj;
1335 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1336 // Length field used as index of next element to be scanned.
1337 // Real length can be obtained from real_forwardee()
1338 arrayOop(old)->set_length(0);
1339 obj_to_push = old;
1340 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1341 "push forwarded object");
1342 }
1343 // Push it on one of the queues of to-be-scanned objects.
1344 bool simulate_overflow = false;
1345 NOT_PRODUCT(
1346 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1347 // simulate a stack overflow
1348 simulate_overflow = true;
1349 }
1350 )
1351 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1352 // Add stats for overflow pushes.
1353 push_on_overflow_list(old, par_scan_state);
1354 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1355 }
1356
1357 return new_obj;
1358 }
1359
1360 // Oops. Someone beat us to it. Undo the allocation. Where did we
1361 // allocate it?
1362 if (is_in_reserved(new_obj)) {
1363 // Must be in to_space.
1364 assert(to()->is_in_reserved(new_obj), "Checking");
1365 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1366 } else {
1367 assert(!_avoid_promotion_undo, "Should not be here if avoiding.");
1368 _next_gen->par_promote_alloc_undo(par_scan_state->thread_num(),
1369 (HeapWord*)new_obj, sz);
1370 }
1371
1372 return forward_ptr;
1373 }
1374
1375 #ifndef PRODUCT
1376 // It's OK to call this multi-threaded; the worst thing
1377 // that can happen is that we'll get a bunch of closely
1378 // spaced simulated oveflows, but that's OK, in fact
1379 // probably good as it would exercise the overflow code
1380 // under contention.
1381 bool ParNewGeneration::should_simulate_overflow() {
1382 if (_overflow_counter-- <= 0) { // just being defensive
1383 _overflow_counter = ParGCWorkQueueOverflowInterval;
1384 return true;
1385 } else {
1386 return false;
1387 }
1388 }
1389 #endif
1390
1391 // In case we are using compressed oops, we need to be careful.
1392 // If the object being pushed is an object array, then its length
1393 // field keeps track of the "grey boundary" at which the next
1394 // incremental scan will be done (see ParGCArrayScanChunk).
1395 // When using compressed oops, this length field is kept in the
1396 // lower 32 bits of the erstwhile klass word and cannot be used
1397 // for the overflow chaining pointer (OCP below). As such the OCP
1398 // would itself need to be compressed into the top 32-bits in this
1399 // case. Unfortunately, see below, in the event that we have a
1400 // promotion failure, the node to be pushed on the list can be
1401 // outside of the Java heap, so the heap-based pointer compression
1402 // would not work (we would have potential aliasing between C-heap
1403 // and Java-heap pointers). For this reason, when using compressed
1404 // oops, we simply use a worker-thread-local, non-shared overflow
1405 // list in the form of a growable array, with a slightly different
1406 // overflow stack draining strategy. If/when we start using fat
1407 // stacks here, we can go back to using (fat) pointer chains
1408 // (although some performance comparisons would be useful since
1409 // single global lists have their own performance disadvantages
1410 // as we were made painfully aware not long ago, see 6786503).
1411 #define BUSY (oop(0x1aff1aff))
1412 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1413 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1414 if (ParGCUseLocalOverflow) {
1415 // In the case of compressed oops, we use a private, not-shared
1416 // overflow stack.
1417 par_scan_state->push_on_overflow_stack(from_space_obj);
1418 } else {
1419 assert(!UseCompressedOops, "Error");
1420 // if the object has been forwarded to itself, then we cannot
1421 // use the klass pointer for the linked list. Instead we have
1422 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1423 // XXX This is horribly inefficient when a promotion failure occurs
1424 // and should be fixed. XXX FIX ME !!!
1425 #ifndef PRODUCT
1426 Atomic::inc_ptr(&_num_par_pushes);
1427 assert(_num_par_pushes > 0, "Tautology");
1428 #endif
1429 if (from_space_obj->forwardee() == from_space_obj) {
1430 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1431 listhead->forward_to(from_space_obj);
1432 from_space_obj = listhead;
1433 }
1434 oop observed_overflow_list = _overflow_list;
1435 oop cur_overflow_list;
1436 do {
1437 cur_overflow_list = observed_overflow_list;
1438 if (cur_overflow_list != BUSY) {
1439 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1440 } else {
1441 from_space_obj->set_klass_to_list_ptr(NULL);
1442 }
1443 observed_overflow_list =
1444 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1445 } while (cur_overflow_list != observed_overflow_list);
1446 }
1447 }
1448
1449 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1450 bool res;
1451
1452 if (ParGCUseLocalOverflow) {
1453 res = par_scan_state->take_from_overflow_stack();
1454 } else {
1455 assert(!UseCompressedOops, "Error");
1456 res = take_from_overflow_list_work(par_scan_state);
1457 }
1458 return res;
1459 }
1460
1461
1462 // *NOTE*: The overflow list manipulation code here and
1463 // in CMSCollector:: are very similar in shape,
1464 // except that in the CMS case we thread the objects
1465 // directly into the list via their mark word, and do
1466 // not need to deal with special cases below related
1467 // to chunking of object arrays and promotion failure
1468 // handling.
1469 // CR 6797058 has been filed to attempt consolidation of
1470 // the common code.
1471 // Because of the common code, if you make any changes in
1472 // the code below, please check the CMS version to see if
1473 // similar changes might be needed.
1474 // See CMSCollector::par_take_from_overflow_list() for
1475 // more extensive documentation comments.
1476 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1477 ObjToScanQueue* work_q = par_scan_state->work_queue();
1478 // How many to take?
1479 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1480 (size_t)ParGCDesiredObjsFromOverflowList);
1481
1482 assert(!UseCompressedOops, "Error");
1483 assert(par_scan_state->overflow_stack() == NULL, "Error");
1484 if (_overflow_list == NULL) return false;
1485
1486 // Otherwise, there was something there; try claiming the list.
1487 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
1488 // Trim off a prefix of at most objsFromOverflow items
1489 Thread* tid = Thread::current();
1490 size_t spin_count = (size_t)ParallelGCThreads;
1491 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1492 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1493 // someone grabbed it before we did ...
1494 // ... we spin for a short while...
1495 os::sleep(tid, sleep_time_millis, false);
1496 if (_overflow_list == NULL) {
1497 // nothing left to take
1498 return false;
1499 } else if (_overflow_list != BUSY) {
1500 // try and grab the prefix
1501 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
1502 }
1503 }
1504 if (prefix == NULL || prefix == BUSY) {
1505 // Nothing to take or waited long enough
1506 if (prefix == NULL) {
1507 // Write back the NULL in case we overwrote it with BUSY above
1508 // and it is still the same value.
1509 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1510 }
1511 return false;
1512 }
1513 assert(prefix != NULL && prefix != BUSY, "Error");
1514 size_t i = 1;
1515 oop cur = prefix;
1516 while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1517 i++; cur = oop(cur->klass());
1518 }
1519
1520 // Reattach remaining (suffix) to overflow list
1521 if (cur->klass_or_null() == NULL) {
1522 // Write back the NULL in lieu of the BUSY we wrote
1523 // above and it is still the same value.
1524 if (_overflow_list == BUSY) {
1525 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1526 }
1527 } else {
1528 assert(cur->klass_or_null() != BUSY, "Error");
1529 oop suffix = oop(cur->klass()); // suffix will be put back on global list
1530 cur->set_klass_to_list_ptr(NULL); // break off suffix
1531 // It's possible that the list is still in the empty(busy) state
1532 // we left it in a short while ago; in that case we may be
1533 // able to place back the suffix.
1534 oop observed_overflow_list = _overflow_list;
1535 oop cur_overflow_list = observed_overflow_list;
1536 bool attached = false;
1537 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1538 observed_overflow_list =
1539 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1540 if (cur_overflow_list == observed_overflow_list) {
1541 attached = true;
1542 break;
1543 } else cur_overflow_list = observed_overflow_list;
1544 }
1545 if (!attached) {
1546 // Too bad, someone else got in in between; we'll need to do a splice.
1547 // Find the last item of suffix list
1548 oop last = suffix;
1549 while (last->klass_or_null() != NULL) {
1550 last = oop(last->klass());
1551 }
1552 // Atomically prepend suffix to current overflow list
1553 observed_overflow_list = _overflow_list;
1554 do {
1555 cur_overflow_list = observed_overflow_list;
1556 if (cur_overflow_list != BUSY) {
1557 // Do the splice ...
1558 last->set_klass_to_list_ptr(cur_overflow_list);
1559 } else { // cur_overflow_list == BUSY
1560 last->set_klass_to_list_ptr(NULL);
1561 }
1562 observed_overflow_list =
1563 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1564 } while (cur_overflow_list != observed_overflow_list);
1565 }
1566 }
1567
1568 // Push objects on prefix list onto this thread's work queue
1569 assert(prefix != NULL && prefix != BUSY, "program logic");
1570 cur = prefix;
1571 ssize_t n = 0;
1572 while (cur != NULL) {
1573 oop obj_to_push = cur->forwardee();
1574 oop next = oop(cur->klass_or_null());
1575 cur->set_klass(obj_to_push->klass());
1576 // This may be an array object that is self-forwarded. In that case, the list pointer
1577 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1578 if (!is_in_reserved(cur)) {
1579 // This can become a scaling bottleneck when there is work queue overflow coincident
1580 // with promotion failure.
1581 oopDesc* f = cur;
1582 FREE_C_HEAP_ARRAY(oopDesc, f, mtGC);
1583 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1584 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1585 obj_to_push = cur;
1586 }
1587 bool ok = work_q->push(obj_to_push);
1588 assert(ok, "Should have succeeded");
1589 cur = next;
1590 n++;
1591 }
1592 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1593 #ifndef PRODUCT
1594 assert(_num_par_pushes >= n, "Too many pops?");
1595 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1596 #endif
1597 return true;
1598 }
1599 #undef BUSY
1600
1601 void ParNewGeneration::ref_processor_init() {
1602 if (_ref_processor == NULL) {
1603 // Allocate and initialize a reference processor
1604 _ref_processor =
1605 new ReferenceProcessor(_reserved, // span
1606 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1607 (int) ParallelGCThreads, // mt processing degree
1608 refs_discovery_is_mt(), // mt discovery
1609 (int) ParallelGCThreads, // mt discovery degree
1610 refs_discovery_is_atomic(), // atomic_discovery
1611 NULL, // is_alive_non_header
1612 false); // write barrier for next field updates
1613 }
1614 }
1615
1616 const char* ParNewGeneration::name() const {
1617 return "par new generation";
1618 }
1619
1620 bool ParNewGeneration::in_use() {
1621 return UseParNewGC && ParallelGCThreads > 0;
1622 }