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/promotionFailedInfo.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/handles.hpp"
51 #include "runtime/handles.inline.hpp"
52 #include "runtime/java.hpp"
53 #include "runtime/thread.hpp"
54 #include "utilities/copy.hpp"
55 #include "utilities/globalDefinitions.hpp"
56 #include "utilities/workgroup.hpp"
57
58 #ifdef _MSC_VER
59 #pragma warning( push )
60 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
61 #endif
62 ParScanThreadState::ParScanThreadState(Space* to_space_,
63 ParNewGeneration* gen_,
64 Generation* old_gen_,
65 int thread_num_,
66 ObjToScanQueueSet* work_queue_set_,
67 Stack<oop, mtGC>* overflow_stacks_,
68 size_t desired_plab_sz_,
69 ParallelTaskTerminator& term_) :
70 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_),
71 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false),
72 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
73 _ageTable(false), // false ==> not the global age table, no perf data.
74 _to_space_alloc_buffer(desired_plab_sz_),
75 _to_space_closure(gen_, this), _old_gen_closure(gen_, this),
76 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this),
77 _older_gen_closure(gen_, this),
78 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
79 &_to_space_root_closure, gen_, &_old_gen_root_closure,
80 work_queue_set_, &term_),
81 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this),
82 _keep_alive_closure(&_scan_weak_ref_closure),
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_promotion_failure_size() {
287 if (_promotion_failed_info.promotion_failed() && PrintPromotionFailure) {
288 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ",
289 _thread_num, _promotion_failed_info.first_size());
290 }
291 }
292
293 class ParScanThreadStateSet: private ResourceArray {
294 public:
295 // Initializes states for the specified number of threads;
296 ParScanThreadStateSet(int num_threads,
297 Space& to_space,
298 ParNewGeneration& gen,
299 Generation& old_gen,
300 ObjToScanQueueSet& queue_set,
301 Stack<oop, mtGC>* overflow_stacks_,
302 size_t desired_plab_sz,
303 ParallelTaskTerminator& term);
304
305 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
306
307 inline ParScanThreadState& thread_state(int i);
308
309 void trace_promotion_failed(YoungGCTracer& gc_tracer);
310 void reset(int active_workers, bool promotion_failed);
311 void flush();
312
313 #if TASKQUEUE_STATS
314 static void
315 print_termination_stats_hdr(outputStream* const st = gclog_or_tty);
316 void print_termination_stats(outputStream* const st = gclog_or_tty);
317 static void
318 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty);
319 void print_taskqueue_stats(outputStream* const st = gclog_or_tty);
320 void reset_stats();
321 #endif // TASKQUEUE_STATS
322
323 private:
324 ParallelTaskTerminator& _term;
325 ParNewGeneration& _gen;
326 Generation& _next_gen;
327 public:
328 bool is_valid(int id) const { return id < length(); }
329 ParallelTaskTerminator* terminator() { return &_term; }
330 };
331
332
333 ParScanThreadStateSet::ParScanThreadStateSet(
334 int num_threads, Space& to_space, ParNewGeneration& gen,
335 Generation& old_gen, ObjToScanQueueSet& queue_set,
336 Stack<oop, mtGC>* overflow_stacks,
337 size_t desired_plab_sz, ParallelTaskTerminator& term)
338 : ResourceArray(sizeof(ParScanThreadState), num_threads),
339 _gen(gen), _next_gen(old_gen), _term(term)
340 {
341 assert(num_threads > 0, "sanity check!");
342 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
343 "overflow_stack allocation mismatch");
344 // Initialize states.
345 for (int i = 0; i < num_threads; ++i) {
346 new ((ParScanThreadState*)_data + i)
347 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set,
348 overflow_stacks, desired_plab_sz, term);
349 }
350 }
351
352 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i)
353 {
354 assert(i >= 0 && i < length(), "sanity check!");
355 return ((ParScanThreadState*)_data)[i];
356 }
357
358 void ParScanThreadStateSet::trace_promotion_failed(YoungGCTracer& gc_tracer) {
359 for (int i = 0; i < length(); ++i) {
360 if (thread_state(i).promotion_failed()) {
361 gc_tracer.report_promotion_failed(thread_state(i).promotion_failed_info());
362 thread_state(i).promotion_failed_info().reset();
363 }
364 }
365 }
366
367 void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed)
368 {
369 _term.reset_for_reuse(active_threads);
370 if (promotion_failed) {
371 for (int i = 0; i < length(); ++i) {
372 thread_state(i).print_promotion_failure_size();
373 }
374 }
375 }
376
377 #if TASKQUEUE_STATS
378 void
379 ParScanThreadState::reset_stats()
380 {
381 taskqueue_stats().reset();
382 _term_attempts = 0;
383 _overflow_refills = 0;
384 _overflow_refill_objs = 0;
385 }
386
387 void ParScanThreadStateSet::reset_stats()
388 {
389 for (int i = 0; i < length(); ++i) {
390 thread_state(i).reset_stats();
391 }
392 }
393
394 void
395 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st)
396 {
397 st->print_raw_cr("GC Termination Stats");
398 st->print_raw_cr(" elapsed --strong roots-- "
399 "-------termination-------");
400 st->print_raw_cr("thr ms ms % "
401 " ms % attempts");
402 st->print_raw_cr("--- --------- --------- ------ "
403 "--------- ------ --------");
404 }
405
406 void ParScanThreadStateSet::print_termination_stats(outputStream* const st)
407 {
408 print_termination_stats_hdr(st);
409
410 for (int i = 0; i < length(); ++i) {
411 const ParScanThreadState & pss = thread_state(i);
412 const double elapsed_ms = pss.elapsed_time() * 1000.0;
413 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
414 const double term_ms = pss.term_time() * 1000.0;
415 st->print_cr("%3d %9.2f %9.2f %6.2f "
416 "%9.2f %6.2f " SIZE_FORMAT_W(8),
417 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
418 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
419 }
420 }
421
422 // Print stats related to work queue activity.
423 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st)
424 {
425 st->print_raw_cr("GC Task Stats");
426 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
427 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
428 }
429
430 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st)
431 {
432 print_taskqueue_stats_hdr(st);
433
434 TaskQueueStats totals;
435 for (int i = 0; i < length(); ++i) {
436 const ParScanThreadState & pss = thread_state(i);
437 const TaskQueueStats & stats = pss.taskqueue_stats();
438 st->print("%3d ", i); stats.print(st); st->cr();
439 totals += stats;
440
441 if (pss.overflow_refills() > 0) {
442 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
443 SIZE_FORMAT_W(10) " overflow objects",
444 pss.overflow_refills(), pss.overflow_refill_objs());
445 }
446 }
447 st->print("tot "); totals.print(st); st->cr();
448
449 DEBUG_ONLY(totals.verify());
450 }
451 #endif // TASKQUEUE_STATS
452
453 void ParScanThreadStateSet::flush()
454 {
455 // Work in this loop should be kept as lightweight as
456 // possible since this might otherwise become a bottleneck
457 // to scaling. Should we add heavy-weight work into this
458 // loop, consider parallelizing the loop into the worker threads.
459 for (int i = 0; i < length(); ++i) {
460 ParScanThreadState& par_scan_state = thread_state(i);
461
462 // Flush stats related to To-space PLAB activity and
463 // retire the last buffer.
464 par_scan_state.to_space_alloc_buffer()->
465 flush_stats_and_retire(_gen.plab_stats(),
466 true /* end_of_gc */,
467 false /* retain */);
468
469 // Every thread has its own age table. We need to merge
470 // them all into one.
471 ageTable *local_table = par_scan_state.age_table();
472 _gen.age_table()->merge(local_table);
473
474 // Inform old gen that we're done.
475 _next_gen.par_promote_alloc_done(i);
476 _next_gen.par_oop_since_save_marks_iterate_done(i);
477 }
478
479 if (UseConcMarkSweepGC && ParallelGCThreads > 0) {
480 // We need to call this even when ResizeOldPLAB is disabled
481 // so as to avoid breaking some asserts. While we may be able
482 // to avoid this by reorganizing the code a bit, I am loathe
483 // to do that unless we find cases where ergo leads to bad
484 // performance.
485 CFLS_LAB::compute_desired_plab_size();
486 }
487 }
488
489 ParScanClosure::ParScanClosure(ParNewGeneration* g,
490 ParScanThreadState* par_scan_state) :
491 OopsInGenClosure(g), _par_scan_state(par_scan_state), _g(g)
492 {
493 assert(_g->level() == 0, "Optimized for youngest generation");
494 _boundary = _g->reserved().end();
495 }
496
497 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
498 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
499
500 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
501 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
502
503 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
504 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
505
506 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
507 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
508
509 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
510 ParScanThreadState* par_scan_state)
511 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
512 {}
513
514 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
515 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
516
517 #ifdef WIN32
518 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
519 #endif
520
521 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
522 ParScanThreadState* par_scan_state_,
523 ParScanWithoutBarrierClosure* to_space_closure_,
524 ParScanWithBarrierClosure* old_gen_closure_,
525 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
526 ParNewGeneration* par_gen_,
527 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
528 ObjToScanQueueSet* task_queues_,
529 ParallelTaskTerminator* terminator_) :
530
531 _par_scan_state(par_scan_state_),
532 _to_space_closure(to_space_closure_),
533 _old_gen_closure(old_gen_closure_),
534 _to_space_root_closure(to_space_root_closure_),
535 _old_gen_root_closure(old_gen_root_closure_),
536 _par_gen(par_gen_),
537 _task_queues(task_queues_),
538 _terminator(terminator_)
539 {}
540
541 void ParEvacuateFollowersClosure::do_void() {
542 ObjToScanQueue* work_q = par_scan_state()->work_queue();
543
544 while (true) {
545
546 // Scan to-space and old-gen objs until we run out of both.
547 oop obj_to_scan;
548 par_scan_state()->trim_queues(0);
549
550 // We have no local work, attempt to steal from other threads.
551
552 // attempt to steal work from promoted.
553 if (task_queues()->steal(par_scan_state()->thread_num(),
554 par_scan_state()->hash_seed(),
555 obj_to_scan)) {
556 bool res = work_q->push(obj_to_scan);
557 assert(res, "Empty queue should have room for a push.");
558
559 // if successful, goto Start.
560 continue;
561
562 // try global overflow list.
563 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
564 continue;
565 }
566
567 // Otherwise, offer termination.
568 par_scan_state()->start_term_time();
569 if (terminator()->offer_termination()) break;
570 par_scan_state()->end_term_time();
571 }
572 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
573 "Broken overflow list?");
574 // Finish the last termination pause.
575 par_scan_state()->end_term_time();
576 }
577
578 ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen,
579 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) :
580 AbstractGangTask("ParNewGeneration collection"),
581 _gen(gen), _next_gen(next_gen),
582 _young_old_boundary(young_old_boundary),
583 _state_set(state_set)
584 {}
585
586 // Reset the terminator for the given number of
587 // active threads.
588 void ParNewGenTask::set_for_termination(int active_workers) {
589 _state_set->reset(active_workers, _gen->promotion_failed());
590 // Should the heap be passed in? There's only 1 for now so
591 // grab it instead.
592 GenCollectedHeap* gch = GenCollectedHeap::heap();
593 gch->set_n_termination(active_workers);
594 }
595
596 // The "i" passed to this method is the part of the work for
597 // this thread. It is not the worker ID. The "i" is derived
598 // from _started_workers which is incremented in internal_note_start()
599 // called in GangWorker loop() and which is called under the
600 // which is called under the protection of the gang monitor and is
601 // called after a task is started. So "i" is based on
602 // first-come-first-served.
603
604 void ParNewGenTask::work(uint worker_id) {
605 GenCollectedHeap* gch = GenCollectedHeap::heap();
606 // Since this is being done in a separate thread, need new resource
607 // and handle marks.
608 ResourceMark rm;
609 HandleMark hm;
610 // We would need multiple old-gen queues otherwise.
611 assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen.");
612
613 Generation* old_gen = gch->next_gen(_gen);
614
615 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
616 assert(_state_set->is_valid(worker_id), "Should not have been called");
617
618 par_scan_state.set_young_old_boundary(_young_old_boundary);
619
620 par_scan_state.start_strong_roots();
621 gch->gen_process_strong_roots(_gen->level(),
622 true, // Process younger gens, if any,
623 // as strong roots.
624 false, // no scope; this is parallel code
625 false, // not collecting perm generation.
626 SharedHeap::SO_AllClasses,
627 &par_scan_state.to_space_root_closure(),
628 true, // walk *all* scavengable nmethods
629 &par_scan_state.older_gen_closure());
630 par_scan_state.end_strong_roots();
631
632 // "evacuate followers".
633 par_scan_state.evacuate_followers_closure().do_void();
634 }
635
636 #ifdef _MSC_VER
637 #pragma warning( push )
638 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
639 #endif
640 ParNewGeneration::
641 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level)
642 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"),
643 _overflow_list(NULL),
644 _is_alive_closure(this),
645 _plab_stats(YoungPLABSize, PLABWeight)
646 {
647 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
648 NOT_PRODUCT(_num_par_pushes = 0;)
649 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
650 guarantee(_task_queues != NULL, "task_queues allocation failure.");
651
652 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
653 ObjToScanQueue *q = new ObjToScanQueue();
654 guarantee(q != NULL, "work_queue Allocation failure.");
655 _task_queues->register_queue(i1, q);
656 }
657
658 for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
659 _task_queues->queue(i2)->initialize();
660
661 _overflow_stacks = NULL;
662 if (ParGCUseLocalOverflow) {
663
664 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal
665 // with ','
666 typedef Stack<oop, mtGC> GCOopStack;
667
668 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
669 for (size_t i = 0; i < ParallelGCThreads; ++i) {
670 new (_overflow_stacks + i) Stack<oop, mtGC>();
671 }
672 }
673
674 if (UsePerfData) {
675 EXCEPTION_MARK;
676 ResourceMark rm;
677
678 const char* cname =
679 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
680 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
681 ParallelGCThreads, CHECK);
682 }
683 }
684 #ifdef _MSC_VER
685 #pragma warning( pop )
686 #endif
687
688 // ParNewGeneration::
689 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
690 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
691
692 template <class T>
693 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
694 #ifdef ASSERT
695 {
696 assert(!oopDesc::is_null(*p), "expected non-null ref");
697 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
698 // We never expect to see a null reference being processed
699 // as a weak reference.
700 assert(obj->is_oop(), "expected an oop while scanning weak refs");
701 }
702 #endif // ASSERT
703
704 _par_cl->do_oop_nv(p);
705
706 if (Universe::heap()->is_in_reserved(p)) {
707 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
708 _rs->write_ref_field_gc_par(p, obj);
709 }
710 }
711
712 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
713 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
714
715 // ParNewGeneration::
716 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
717 DefNewGeneration::KeepAliveClosure(cl) {}
718
719 template <class T>
720 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
721 #ifdef ASSERT
722 {
723 assert(!oopDesc::is_null(*p), "expected non-null ref");
724 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
725 // We never expect to see a null reference being processed
726 // as a weak reference.
727 assert(obj->is_oop(), "expected an oop while scanning weak refs");
728 }
729 #endif // ASSERT
730
731 _cl->do_oop_nv(p);
732
733 if (Universe::heap()->is_in_reserved(p)) {
734 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
735 _rs->write_ref_field_gc_par(p, obj);
736 }
737 }
738
739 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
740 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
741
742 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
743 T heap_oop = oopDesc::load_heap_oop(p);
744 if (!oopDesc::is_null(heap_oop)) {
745 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
746 if ((HeapWord*)obj < _boundary) {
747 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
748 oop new_obj = obj->is_forwarded()
749 ? obj->forwardee()
750 : _g->DefNewGeneration::copy_to_survivor_space(obj);
751 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
752 }
753 if (_gc_barrier) {
754 // If p points to a younger generation, mark the card.
755 if ((HeapWord*)obj < _gen_boundary) {
756 _rs->write_ref_field_gc_par(p, obj);
757 }
758 }
759 }
760 }
761
762 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
763 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
764
765 class ParNewRefProcTaskProxy: public AbstractGangTask {
766 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
767 public:
768 ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen,
769 Generation& next_gen,
770 HeapWord* young_old_boundary,
771 ParScanThreadStateSet& state_set);
772
773 private:
774 virtual void work(uint worker_id);
775 virtual void set_for_termination(int active_workers) {
776 _state_set.terminator()->reset_for_reuse(active_workers);
777 }
778 private:
779 ParNewGeneration& _gen;
780 ProcessTask& _task;
781 Generation& _next_gen;
782 HeapWord* _young_old_boundary;
783 ParScanThreadStateSet& _state_set;
784 };
785
786 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(
787 ProcessTask& task, ParNewGeneration& gen,
788 Generation& next_gen,
789 HeapWord* young_old_boundary,
790 ParScanThreadStateSet& state_set)
791 : AbstractGangTask("ParNewGeneration parallel reference processing"),
792 _gen(gen),
793 _task(task),
794 _next_gen(next_gen),
795 _young_old_boundary(young_old_boundary),
796 _state_set(state_set)
797 {
798 }
799
800 void ParNewRefProcTaskProxy::work(uint worker_id)
801 {
802 ResourceMark rm;
803 HandleMark hm;
804 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
805 par_scan_state.set_young_old_boundary(_young_old_boundary);
806 _task.work(worker_id, par_scan_state.is_alive_closure(),
807 par_scan_state.keep_alive_closure(),
808 par_scan_state.evacuate_followers_closure());
809 }
810
811 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
812 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
813 EnqueueTask& _task;
814
815 public:
816 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
817 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
818 _task(task)
819 { }
820
821 virtual void work(uint worker_id)
822 {
823 _task.work(worker_id);
824 }
825 };
826
827
828 void ParNewRefProcTaskExecutor::execute(ProcessTask& task)
829 {
830 GenCollectedHeap* gch = GenCollectedHeap::heap();
831 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
832 "not a generational heap");
833 FlexibleWorkGang* workers = gch->workers();
834 assert(workers != NULL, "Need parallel worker threads.");
835 _state_set.reset(workers->active_workers(), _generation.promotion_failed());
836 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(),
837 _generation.reserved().end(), _state_set);
838 workers->run_task(&rp_task);
839 _state_set.reset(0 /* bad value in debug if not reset */,
840 _generation.promotion_failed());
841 }
842
843 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task)
844 {
845 GenCollectedHeap* gch = GenCollectedHeap::heap();
846 FlexibleWorkGang* workers = gch->workers();
847 assert(workers != NULL, "Need parallel worker threads.");
848 ParNewRefEnqueueTaskProxy enq_task(task);
849 workers->run_task(&enq_task);
850 }
851
852 void ParNewRefProcTaskExecutor::set_single_threaded_mode()
853 {
854 _state_set.flush();
855 GenCollectedHeap* gch = GenCollectedHeap::heap();
856 gch->set_par_threads(0); // 0 ==> non-parallel.
857 gch->save_marks();
858 }
859
860 ScanClosureWithParBarrier::
861 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
862 ScanClosure(g, gc_barrier) {}
863
864 EvacuateFollowersClosureGeneral::
865 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
866 OopsInGenClosure* cur,
867 OopsInGenClosure* older) :
868 _gch(gch), _level(level),
869 _scan_cur_or_nonheap(cur), _scan_older(older)
870 {}
871
872 void EvacuateFollowersClosureGeneral::do_void() {
873 do {
874 // Beware: this call will lead to closure applications via virtual
875 // calls.
876 _gch->oop_since_save_marks_iterate(_level,
877 _scan_cur_or_nonheap,
878 _scan_older);
879 } while (!_gch->no_allocs_since_save_marks(_level));
880 }
881
882
883 // A Generation that does parallel young-gen collection.
884
885 bool ParNewGeneration::_avoid_promotion_undo = false;
886
887 void ParNewGeneration::adjust_desired_tenuring_threshold() {
888 // Set the desired survivor size to half the real survivor space
889 _tenuring_threshold =
890 age_table()->compute_tenuring_threshold(to()->capacity()/HeapWordSize);
891 }
892
893 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set, ParNewTracer& gc_tracer) {
894 assert(_promo_failure_scan_stack.is_empty(), "post condition");
895 _promo_failure_scan_stack.clear(true); // Clear cached segments.
896
897 remove_forwarding_pointers();
898 if (PrintGCDetails) {
899 gclog_or_tty->print(" (promotion failed)");
900 }
901 // All the spaces are in play for mark-sweep.
902 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
903 from()->set_next_compaction_space(to());
904 gch->set_incremental_collection_failed();
905 // Inform the next generation that a promotion failure occurred.
906 _next_gen->promotion_failure_occurred();
907
908 // Trace promotion failure in the parallel GC threads
909 thread_state_set.trace_promotion_failed(gc_tracer);
910 // Single threaded code may have reported promotion failure to the global state
911 if (_promotion_failed_info.promotion_failed()) {
912 gc_tracer.report_promotion_failed(_promotion_failed_info);
913 }
914 // Reset the PromotionFailureALot counters.
915 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();)
916 }
917
918 void ParNewGeneration::collect(bool full,
919 bool clear_all_soft_refs,
920 size_t size,
921 bool is_tlab) {
922 assert(full || size > 0, "otherwise we don't want to collect");
923
924 GenCollectedHeap* gch = GenCollectedHeap::heap();
925
926 _gc_timer->register_gc_start(os::elapsed_counter());
927
928 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
929 "not a CMS generational heap");
930 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
931 FlexibleWorkGang* workers = gch->workers();
932 assert(workers != NULL, "Need workgang for parallel work");
933 int active_workers =
934 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
935 workers->active_workers(),
936 Threads::number_of_non_daemon_threads());
937 workers->set_active_workers(active_workers);
938 _next_gen = gch->next_gen(this);
939 assert(_next_gen != NULL,
940 "This must be the youngest gen, and not the only gen");
941 assert(gch->n_gens() == 2,
942 "Par collection currently only works with single older gen.");
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);
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(), _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);
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);
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 } else {
1058 handle_promotion_failed(gch, thread_state_set, gc_tracer);
1059 }
1060 // set new iteration safe limit for the survivor spaces
1061 from()->set_concurrent_iteration_safe_limit(from()->top());
1062 to()->set_concurrent_iteration_safe_limit(to()->top());
1063
1064 adjust_desired_tenuring_threshold();
1065 if (ResizePLAB) {
1066 plab_stats()->adjust_desired_plab_sz(n_workers);
1067 }
1068
1069 if (PrintGC && !PrintGCDetails) {
1070 gch->print_heap_change(gch_prev_used);
1071 }
1072
1073 if (PrintGCDetails && ParallelGCVerbose) {
1074 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
1075 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
1076 }
1077
1078 if (UseAdaptiveSizePolicy) {
1079 size_policy->minor_collection_end(gch->gc_cause());
1080 size_policy->avg_survived()->sample(from()->used());
1081 }
1082
1083 // We need to use a monotonically non-deccreasing time in ms
1084 // or we will see time-warp warnings and os::javaTimeMillis()
1085 // does not guarantee monotonicity.
1086 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1087 update_time_of_last_gc(now);
1088
1089 SpecializationStats::print();
1090
1091 rp->set_enqueuing_is_done(true);
1092 if (rp->processing_is_mt()) {
1093 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1094 rp->enqueue_discovered_references(&task_executor);
1095 } else {
1096 rp->enqueue_discovered_references(NULL);
1097 }
1098 rp->verify_no_references_recorded();
1099
1100 gch->trace_heap_after_gc(&gc_tracer);
1101
1102 _gc_timer->register_gc_end(os::elapsed_counter());
1103
1104 gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1105 }
1106
1107 static int sum;
1108 void ParNewGeneration::waste_some_time() {
1109 for (int i = 0; i < 100; i++) {
1110 sum += i;
1111 }
1112 }
1113
1114 static const oop ClaimedForwardPtr = oop(0x4);
1115
1116 // Because of concurrency, there are times where an object for which
1117 // "is_forwarded()" is true contains an "interim" forwarding pointer
1118 // value. Such a value will soon be overwritten with a real value.
1119 // This method requires "obj" to have a forwarding pointer, and waits, if
1120 // necessary for a real one to be inserted, and returns it.
1121
1122 oop ParNewGeneration::real_forwardee(oop obj) {
1123 oop forward_ptr = obj->forwardee();
1124 if (forward_ptr != ClaimedForwardPtr) {
1125 return forward_ptr;
1126 } else {
1127 return real_forwardee_slow(obj);
1128 }
1129 }
1130
1131 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1132 // Spin-read if it is claimed but not yet written by another thread.
1133 oop forward_ptr = obj->forwardee();
1134 while (forward_ptr == ClaimedForwardPtr) {
1135 waste_some_time();
1136 assert(obj->is_forwarded(), "precondition");
1137 forward_ptr = obj->forwardee();
1138 }
1139 return forward_ptr;
1140 }
1141
1142 #ifdef ASSERT
1143 bool ParNewGeneration::is_legal_forward_ptr(oop p) {
1144 return
1145 (_avoid_promotion_undo && p == ClaimedForwardPtr)
1146 || Universe::heap()->is_in_reserved(p);
1147 }
1148 #endif
1149
1150 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
1151 if (m->must_be_preserved_for_promotion_failure(obj)) {
1152 // We should really have separate per-worker stacks, rather
1153 // than use locking of a common pair of stacks.
1154 MutexLocker ml(ParGCRareEvent_lock);
1155 preserve_mark(obj, m);
1156 }
1157 }
1158
1159 // Multiple GC threads may try to promote an object. If the object
1160 // is successfully promoted, a forwarding pointer will be installed in
1161 // the object in the young generation. This method claims the right
1162 // to install the forwarding pointer before it copies the object,
1163 // thus avoiding the need to undo the copy as in
1164 // copy_to_survivor_space_avoiding_with_undo.
1165
1166 oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo(
1167 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1168 // In the sequential version, this assert also says that the object is
1169 // not forwarded. That might not be the case here. It is the case that
1170 // the caller observed it to be not forwarded at some time in the past.
1171 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1172
1173 // The sequential code read "old->age()" below. That doesn't work here,
1174 // since the age is in the mark word, and that might be overwritten with
1175 // a forwarding pointer by a parallel thread. So we must save the mark
1176 // word in a local and then analyze it.
1177 oopDesc dummyOld;
1178 dummyOld.set_mark(m);
1179 assert(!dummyOld.is_forwarded(),
1180 "should not be called with forwarding pointer mark word.");
1181
1182 oop new_obj = NULL;
1183 oop forward_ptr;
1184
1185 // Try allocating obj in to-space (unless too old)
1186 if (dummyOld.age() < tenuring_threshold()) {
1187 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1188 if (new_obj == NULL) {
1189 set_survivor_overflow(true);
1190 }
1191 }
1192
1193 if (new_obj == NULL) {
1194 // Either to-space is full or we decided to promote
1195 // try allocating obj tenured
1196
1197 // Attempt to install a null forwarding pointer (atomically),
1198 // to claim the right to install the real forwarding pointer.
1199 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1200 if (forward_ptr != NULL) {
1201 // someone else beat us to it.
1202 return real_forwardee(old);
1203 }
1204
1205 new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
1206 old, m, sz);
1207
1208 if (new_obj == NULL) {
1209 // promotion failed, forward to self
1210 _promotion_failed = true;
1211 new_obj = old;
1212
1213 preserve_mark_if_necessary(old, m);
1214 par_scan_state->register_promotion_failure(sz);
1215 }
1216
1217 old->forward_to(new_obj);
1218 forward_ptr = NULL;
1219 } else {
1220 // Is in to-space; do copying ourselves.
1221 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1222 forward_ptr = old->forward_to_atomic(new_obj);
1223 // Restore the mark word copied above.
1224 new_obj->set_mark(m);
1225 // Increment age if obj still in new generation
1226 new_obj->incr_age();
1227 par_scan_state->age_table()->add(new_obj, sz);
1228 }
1229 assert(new_obj != NULL, "just checking");
1230
1231 if (forward_ptr == NULL) {
1232 oop obj_to_push = new_obj;
1233 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1234 // Length field used as index of next element to be scanned.
1235 // Real length can be obtained from real_forwardee()
1236 arrayOop(old)->set_length(0);
1237 obj_to_push = old;
1238 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1239 "push forwarded object");
1240 }
1241 // Push it on one of the queues of to-be-scanned objects.
1242 bool simulate_overflow = false;
1243 NOT_PRODUCT(
1244 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1245 // simulate a stack overflow
1246 simulate_overflow = true;
1247 }
1248 )
1249 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1250 // Add stats for overflow pushes.
1251 if (Verbose && PrintGCDetails) {
1252 gclog_or_tty->print("queue overflow!\n");
1253 }
1254 push_on_overflow_list(old, par_scan_state);
1255 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1256 }
1257
1258 return new_obj;
1259 }
1260
1261 // Oops. Someone beat us to it. Undo the allocation. Where did we
1262 // allocate it?
1263 if (is_in_reserved(new_obj)) {
1264 // Must be in to_space.
1265 assert(to()->is_in_reserved(new_obj), "Checking");
1266 if (forward_ptr == ClaimedForwardPtr) {
1267 // Wait to get the real forwarding pointer value.
1268 forward_ptr = real_forwardee(old);
1269 }
1270 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1271 }
1272
1273 return forward_ptr;
1274 }
1275
1276
1277 // Multiple GC threads may try to promote the same object. If two
1278 // or more GC threads copy the object, only one wins the race to install
1279 // the forwarding pointer. The other threads have to undo their copy.
1280
1281 oop ParNewGeneration::copy_to_survivor_space_with_undo(
1282 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1283
1284 // In the sequential version, this assert also says that the object is
1285 // not forwarded. That might not be the case here. It is the case that
1286 // the caller observed it to be not forwarded at some time in the past.
1287 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1288
1289 // The sequential code read "old->age()" below. That doesn't work here,
1290 // since the age is in the mark word, and that might be overwritten with
1291 // a forwarding pointer by a parallel thread. So we must save the mark
1292 // word here, install it in a local oopDesc, and then analyze it.
1293 oopDesc dummyOld;
1294 dummyOld.set_mark(m);
1295 assert(!dummyOld.is_forwarded(),
1296 "should not be called with forwarding pointer mark word.");
1297
1298 bool failed_to_promote = false;
1299 oop new_obj = NULL;
1300 oop forward_ptr;
1301
1302 // Try allocating obj in to-space (unless too old)
1303 if (dummyOld.age() < tenuring_threshold()) {
1304 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1305 if (new_obj == NULL) {
1306 set_survivor_overflow(true);
1307 }
1308 }
1309
1310 if (new_obj == NULL) {
1311 // Either to-space is full or we decided to promote
1312 // try allocating obj tenured
1313 new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
1314 old, m, sz);
1315
1316 if (new_obj == NULL) {
1317 // promotion failed, forward to self
1318 forward_ptr = old->forward_to_atomic(old);
1319 new_obj = old;
1320
1321 if (forward_ptr != NULL) {
1322 return forward_ptr; // someone else succeeded
1323 }
1324
1325 _promotion_failed = true;
1326 failed_to_promote = true;
1327
1328 preserve_mark_if_necessary(old, m);
1329 par_scan_state->register_promotion_failure(sz);
1330 }
1331 } else {
1332 // Is in to-space; do copying ourselves.
1333 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1334 // Restore the mark word copied above.
1335 new_obj->set_mark(m);
1336 // Increment age if new_obj still in new generation
1337 new_obj->incr_age();
1338 par_scan_state->age_table()->add(new_obj, sz);
1339 }
1340 assert(new_obj != NULL, "just checking");
1341
1342 // Now attempt to install the forwarding pointer (atomically).
1343 // We have to copy the mark word before overwriting with forwarding
1344 // ptr, so we can restore it below in the copy.
1345 if (!failed_to_promote) {
1346 forward_ptr = old->forward_to_atomic(new_obj);
1347 }
1348
1349 if (forward_ptr == NULL) {
1350 oop obj_to_push = new_obj;
1351 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1352 // Length field used as index of next element to be scanned.
1353 // Real length can be obtained from real_forwardee()
1354 arrayOop(old)->set_length(0);
1355 obj_to_push = old;
1356 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1357 "push forwarded object");
1358 }
1359 // Push it on one of the queues of to-be-scanned objects.
1360 bool simulate_overflow = false;
1361 NOT_PRODUCT(
1362 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1363 // simulate a stack overflow
1364 simulate_overflow = true;
1365 }
1366 )
1367 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1368 // Add stats for overflow pushes.
1369 push_on_overflow_list(old, par_scan_state);
1370 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1371 }
1372
1373 return new_obj;
1374 }
1375
1376 // Oops. Someone beat us to it. Undo the allocation. Where did we
1377 // allocate it?
1378 if (is_in_reserved(new_obj)) {
1379 // Must be in to_space.
1380 assert(to()->is_in_reserved(new_obj), "Checking");
1381 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1382 } else {
1383 assert(!_avoid_promotion_undo, "Should not be here if avoiding.");
1384 _next_gen->par_promote_alloc_undo(par_scan_state->thread_num(),
1385 (HeapWord*)new_obj, sz);
1386 }
1387
1388 return forward_ptr;
1389 }
1390
1391 #ifndef PRODUCT
1392 // It's OK to call this multi-threaded; the worst thing
1393 // that can happen is that we'll get a bunch of closely
1394 // spaced simulated oveflows, but that's OK, in fact
1395 // probably good as it would exercise the overflow code
1396 // under contention.
1397 bool ParNewGeneration::should_simulate_overflow() {
1398 if (_overflow_counter-- <= 0) { // just being defensive
1399 _overflow_counter = ParGCWorkQueueOverflowInterval;
1400 return true;
1401 } else {
1402 return false;
1403 }
1404 }
1405 #endif
1406
1407 // In case we are using compressed oops, we need to be careful.
1408 // If the object being pushed is an object array, then its length
1409 // field keeps track of the "grey boundary" at which the next
1410 // incremental scan will be done (see ParGCArrayScanChunk).
1411 // When using compressed oops, this length field is kept in the
1412 // lower 32 bits of the erstwhile klass word and cannot be used
1413 // for the overflow chaining pointer (OCP below). As such the OCP
1414 // would itself need to be compressed into the top 32-bits in this
1415 // case. Unfortunately, see below, in the event that we have a
1416 // promotion failure, the node to be pushed on the list can be
1417 // outside of the Java heap, so the heap-based pointer compression
1418 // would not work (we would have potential aliasing between C-heap
1419 // and Java-heap pointers). For this reason, when using compressed
1420 // oops, we simply use a worker-thread-local, non-shared overflow
1421 // list in the form of a growable array, with a slightly different
1422 // overflow stack draining strategy. If/when we start using fat
1423 // stacks here, we can go back to using (fat) pointer chains
1424 // (although some performance comparisons would be useful since
1425 // single global lists have their own performance disadvantages
1426 // as we were made painfully aware not long ago, see 6786503).
1427 #define BUSY (oop(0x1aff1aff))
1428 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1429 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1430 if (ParGCUseLocalOverflow) {
1431 // In the case of compressed oops, we use a private, not-shared
1432 // overflow stack.
1433 par_scan_state->push_on_overflow_stack(from_space_obj);
1434 } else {
1435 assert(!UseCompressedOops, "Error");
1436 // if the object has been forwarded to itself, then we cannot
1437 // use the klass pointer for the linked list. Instead we have
1438 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1439 // XXX This is horribly inefficient when a promotion failure occurs
1440 // and should be fixed. XXX FIX ME !!!
1441 #ifndef PRODUCT
1442 Atomic::inc_ptr(&_num_par_pushes);
1443 assert(_num_par_pushes > 0, "Tautology");
1444 #endif
1445 if (from_space_obj->forwardee() == from_space_obj) {
1446 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1447 listhead->forward_to(from_space_obj);
1448 from_space_obj = listhead;
1449 }
1450 oop observed_overflow_list = _overflow_list;
1451 oop cur_overflow_list;
1452 do {
1453 cur_overflow_list = observed_overflow_list;
1454 if (cur_overflow_list != BUSY) {
1455 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1456 } else {
1457 from_space_obj->set_klass_to_list_ptr(NULL);
1458 }
1459 observed_overflow_list =
1460 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1461 } while (cur_overflow_list != observed_overflow_list);
1462 }
1463 }
1464
1465 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1466 bool res;
1467
1468 if (ParGCUseLocalOverflow) {
1469 res = par_scan_state->take_from_overflow_stack();
1470 } else {
1471 assert(!UseCompressedOops, "Error");
1472 res = take_from_overflow_list_work(par_scan_state);
1473 }
1474 return res;
1475 }
1476
1477
1478 // *NOTE*: The overflow list manipulation code here and
1479 // in CMSCollector:: are very similar in shape,
1480 // except that in the CMS case we thread the objects
1481 // directly into the list via their mark word, and do
1482 // not need to deal with special cases below related
1483 // to chunking of object arrays and promotion failure
1484 // handling.
1485 // CR 6797058 has been filed to attempt consolidation of
1486 // the common code.
1487 // Because of the common code, if you make any changes in
1488 // the code below, please check the CMS version to see if
1489 // similar changes might be needed.
1490 // See CMSCollector::par_take_from_overflow_list() for
1491 // more extensive documentation comments.
1492 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1493 ObjToScanQueue* work_q = par_scan_state->work_queue();
1494 // How many to take?
1495 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1496 (size_t)ParGCDesiredObjsFromOverflowList);
1497
1498 assert(!UseCompressedOops, "Error");
1499 assert(par_scan_state->overflow_stack() == NULL, "Error");
1500 if (_overflow_list == NULL) return false;
1501
1502 // Otherwise, there was something there; try claiming the list.
1503 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
1504 // Trim off a prefix of at most objsFromOverflow items
1505 Thread* tid = Thread::current();
1506 size_t spin_count = (size_t)ParallelGCThreads;
1507 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1508 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1509 // someone grabbed it before we did ...
1510 // ... we spin for a short while...
1511 os::sleep(tid, sleep_time_millis, false);
1512 if (_overflow_list == NULL) {
1513 // nothing left to take
1514 return false;
1515 } else if (_overflow_list != BUSY) {
1516 // try and grab the prefix
1517 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
1518 }
1519 }
1520 if (prefix == NULL || prefix == BUSY) {
1521 // Nothing to take or waited long enough
1522 if (prefix == NULL) {
1523 // Write back the NULL in case we overwrote it with BUSY above
1524 // and it is still the same value.
1525 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1526 }
1527 return false;
1528 }
1529 assert(prefix != NULL && prefix != BUSY, "Error");
1530 size_t i = 1;
1531 oop cur = prefix;
1532 while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1533 i++; cur = oop(cur->klass());
1534 }
1535
1536 // Reattach remaining (suffix) to overflow list
1537 if (cur->klass_or_null() == NULL) {
1538 // Write back the NULL in lieu of the BUSY we wrote
1539 // above and it is still the same value.
1540 if (_overflow_list == BUSY) {
1541 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1542 }
1543 } else {
1544 assert(cur->klass_or_null() != BUSY, "Error");
1545 oop suffix = oop(cur->klass()); // suffix will be put back on global list
1546 cur->set_klass_to_list_ptr(NULL); // break off suffix
1547 // It's possible that the list is still in the empty(busy) state
1548 // we left it in a short while ago; in that case we may be
1549 // able to place back the suffix.
1550 oop observed_overflow_list = _overflow_list;
1551 oop cur_overflow_list = observed_overflow_list;
1552 bool attached = false;
1553 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1554 observed_overflow_list =
1555 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1556 if (cur_overflow_list == observed_overflow_list) {
1557 attached = true;
1558 break;
1559 } else cur_overflow_list = observed_overflow_list;
1560 }
1561 if (!attached) {
1562 // Too bad, someone else got in in between; we'll need to do a splice.
1563 // Find the last item of suffix list
1564 oop last = suffix;
1565 while (last->klass_or_null() != NULL) {
1566 last = oop(last->klass());
1567 }
1568 // Atomically prepend suffix to current overflow list
1569 observed_overflow_list = _overflow_list;
1570 do {
1571 cur_overflow_list = observed_overflow_list;
1572 if (cur_overflow_list != BUSY) {
1573 // Do the splice ...
1574 last->set_klass_to_list_ptr(cur_overflow_list);
1575 } else { // cur_overflow_list == BUSY
1576 last->set_klass_to_list_ptr(NULL);
1577 }
1578 observed_overflow_list =
1579 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1580 } while (cur_overflow_list != observed_overflow_list);
1581 }
1582 }
1583
1584 // Push objects on prefix list onto this thread's work queue
1585 assert(prefix != NULL && prefix != BUSY, "program logic");
1586 cur = prefix;
1587 ssize_t n = 0;
1588 while (cur != NULL) {
1589 oop obj_to_push = cur->forwardee();
1590 oop next = oop(cur->klass_or_null());
1591 cur->set_klass(obj_to_push->klass());
1592 // This may be an array object that is self-forwarded. In that case, the list pointer
1593 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1594 if (!is_in_reserved(cur)) {
1595 // This can become a scaling bottleneck when there is work queue overflow coincident
1596 // with promotion failure.
1597 oopDesc* f = cur;
1598 FREE_C_HEAP_ARRAY(oopDesc, f, mtGC);
1599 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1600 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1601 obj_to_push = cur;
1602 }
1603 bool ok = work_q->push(obj_to_push);
1604 assert(ok, "Should have succeeded");
1605 cur = next;
1606 n++;
1607 }
1608 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1609 #ifndef PRODUCT
1610 assert(_num_par_pushes >= n, "Too many pops?");
1611 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1612 #endif
1613 return true;
1614 }
1615 #undef BUSY
1616
1617 void ParNewGeneration::ref_processor_init() {
1618 if (_ref_processor == NULL) {
1619 // Allocate and initialize a reference processor
1620 _ref_processor =
1621 new ReferenceProcessor(_reserved, // span
1622 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1623 (int) ParallelGCThreads, // mt processing degree
1624 refs_discovery_is_mt(), // mt discovery
1625 (int) ParallelGCThreads, // mt discovery degree
1626 refs_discovery_is_atomic(), // atomic_discovery
1627 NULL, // is_alive_non_header
1628 false); // write barrier for next field updates
1629 }
1630 }
1631
1632 const char* ParNewGeneration::name() const {
1633 return "par new generation";
1634 }
1635
1636 bool ParNewGeneration::in_use() {
1637 return UseParNewGC && ParallelGCThreads > 0;
1638 }