1 /*
2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
27 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp"
28 #include "gc_implementation/parNew/parNewGeneration.hpp"
29 #include "gc_implementation/parNew/parOopClosures.inline.hpp"
30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp"
31 #include "gc_implementation/shared/ageTable.hpp"
32 #include "gc_implementation/shared/copyFailedInfo.hpp"
33 #include "gc_implementation/shared/gcHeapSummary.hpp"
34 #include "gc_implementation/shared/gcTimer.hpp"
35 #include "gc_implementation/shared/gcTrace.hpp"
36 #include "gc_implementation/shared/gcTraceTime.hpp"
37 #include "gc_implementation/shared/parGCAllocBuffer.inline.hpp"
38 #include "gc_implementation/shared/spaceDecorator.hpp"
39 #include "memory/defNewGeneration.inline.hpp"
40 #include "memory/genCollectedHeap.hpp"
41 #include "memory/genOopClosures.inline.hpp"
42 #include "memory/generation.hpp"
43 #include "memory/referencePolicy.hpp"
44 #include "memory/resourceArea.hpp"
45 #include "memory/sharedHeap.hpp"
46 #include "memory/space.hpp"
47 #include "oops/objArrayOop.hpp"
48 #include "oops/oop.inline.hpp"
49 #include "oops/oop.pcgc.inline.hpp"
50 #include "runtime/atomic.inline.hpp"
51 #include "runtime/handles.hpp"
52 #include "runtime/handles.inline.hpp"
53 #include "runtime/java.hpp"
54 #include "runtime/thread.inline.hpp"
55 #include "utilities/copy.hpp"
56 #include "utilities/globalDefinitions.hpp"
57 #include "utilities/workgroup.hpp"
58
59 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
60
61 #ifdef _MSC_VER
62 #pragma warning( push )
63 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
64 #endif
65 ParScanThreadState::ParScanThreadState(Space* to_space_,
66 ParNewGeneration* gen_,
67 Generation* old_gen_,
68 int thread_num_,
69 ObjToScanQueueSet* work_queue_set_,
70 Stack<oop, mtGC>* overflow_stacks_,
71 size_t desired_plab_sz_,
72 ParallelTaskTerminator& term_) :
73 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_),
74 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false),
75 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
76 _ageTable(false), // false ==> not the global age table, no perf data.
77 _to_space_alloc_buffer(desired_plab_sz_),
78 _to_space_closure(gen_, this), _old_gen_closure(gen_, this),
79 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this),
80 _older_gen_closure(gen_, this),
81 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
82 &_to_space_root_closure, gen_, &_old_gen_root_closure,
83 work_queue_set_, &term_),
84 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this),
85 _keep_alive_closure(&_scan_weak_ref_closure),
86 _strong_roots_time(0.0), _term_time(0.0)
87 {
88 #if TASKQUEUE_STATS
89 _term_attempts = 0;
90 _overflow_refills = 0;
91 _overflow_refill_objs = 0;
92 #endif // TASKQUEUE_STATS
93
94 _survivor_chunk_array =
95 (ChunkArray*) old_gen()->get_data_recorder(thread_num());
96 _hash_seed = 17; // Might want to take time-based random value.
97 _start = os::elapsedTime();
98 _old_gen_closure.set_generation(old_gen_);
99 _old_gen_root_closure.set_generation(old_gen_);
100 }
101 #ifdef _MSC_VER
102 #pragma warning( pop )
103 #endif
104
105 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
106 size_t plab_word_size) {
107 ChunkArray* sca = survivor_chunk_array();
108 if (sca != NULL) {
109 // A non-null SCA implies that we want the PLAB data recorded.
110 sca->record_sample(plab_start, plab_word_size);
111 }
112 }
113
114 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
115 return new_obj->is_objArray() &&
116 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
117 new_obj != old_obj;
118 }
119
120 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
121 assert(old->is_objArray(), "must be obj array");
122 assert(old->is_forwarded(), "must be forwarded");
123 assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
124 assert(!old_gen()->is_in(old), "must be in young generation.");
125
126 objArrayOop obj = objArrayOop(old->forwardee());
127 // Process ParGCArrayScanChunk elements now
128 // and push the remainder back onto queue
129 int start = arrayOop(old)->length();
130 int end = obj->length();
131 int remainder = end - start;
132 assert(start <= end, "just checking");
133 if (remainder > 2 * ParGCArrayScanChunk) {
134 // Test above combines last partial chunk with a full chunk
135 end = start + ParGCArrayScanChunk;
136 arrayOop(old)->set_length(end);
137 // Push remainder.
138 bool ok = work_queue()->push(old);
139 assert(ok, "just popped, push must be okay");
140 } else {
141 // Restore length so that it can be used if there
142 // is a promotion failure and forwarding pointers
143 // must be removed.
144 arrayOop(old)->set_length(end);
145 }
146
147 // process our set of indices (include header in first chunk)
148 // should make sure end is even (aligned to HeapWord in case of compressed oops)
149 if ((HeapWord *)obj < young_old_boundary()) {
150 // object is in to_space
151 obj->oop_iterate_range(&_to_space_closure, start, end);
152 } else {
153 // object is in old generation
154 obj->oop_iterate_range(&_old_gen_closure, start, end);
155 }
156 }
157
158
159 void ParScanThreadState::trim_queues(int max_size) {
160 ObjToScanQueue* queue = work_queue();
161 do {
162 while (queue->size() > (juint)max_size) {
163 oop obj_to_scan;
164 if (queue->pop_local(obj_to_scan)) {
165 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
166 if (obj_to_scan->is_objArray() &&
167 obj_to_scan->is_forwarded() &&
168 obj_to_scan->forwardee() != obj_to_scan) {
169 scan_partial_array_and_push_remainder(obj_to_scan);
170 } else {
171 // object is in to_space
172 obj_to_scan->oop_iterate(&_to_space_closure);
173 }
174 } else {
175 // object is in old generation
176 obj_to_scan->oop_iterate(&_old_gen_closure);
177 }
178 }
179 }
180 // For the case of compressed oops, we have a private, non-shared
181 // overflow stack, so we eagerly drain it so as to more evenly
182 // distribute load early. Note: this may be good to do in
183 // general rather than delay for the final stealing phase.
184 // If applicable, we'll transfer a set of objects over to our
185 // work queue, allowing them to be stolen and draining our
186 // private overflow stack.
187 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
188 }
189
190 bool ParScanThreadState::take_from_overflow_stack() {
191 assert(ParGCUseLocalOverflow, "Else should not call");
192 assert(young_gen()->overflow_list() == NULL, "Error");
193 ObjToScanQueue* queue = work_queue();
194 Stack<oop, mtGC>* const of_stack = overflow_stack();
195 const size_t num_overflow_elems = of_stack->size();
196 const size_t space_available = queue->max_elems() - queue->size();
197 const size_t num_take_elems = MIN3(space_available / 4,
198 ParGCDesiredObjsFromOverflowList,
199 num_overflow_elems);
200 // Transfer the most recent num_take_elems from the overflow
201 // stack to our work queue.
202 for (size_t i = 0; i != num_take_elems; i++) {
203 oop cur = of_stack->pop();
204 oop obj_to_push = cur->forwardee();
205 assert(Universe::heap()->is_in_reserved(cur), "Should be in heap");
206 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
207 assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap");
208 if (should_be_partially_scanned(obj_to_push, cur)) {
209 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
210 obj_to_push = cur;
211 }
212 bool ok = queue->push(obj_to_push);
213 assert(ok, "Should have succeeded");
214 }
215 assert(young_gen()->overflow_list() == NULL, "Error");
216 return num_take_elems > 0; // was something transferred?
217 }
218
219 void ParScanThreadState::push_on_overflow_stack(oop p) {
220 assert(ParGCUseLocalOverflow, "Else should not call");
221 overflow_stack()->push(p);
222 assert(young_gen()->overflow_list() == NULL, "Error");
223 }
224
225 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
226
227 // Otherwise, if the object is small enough, try to reallocate the
228 // buffer.
229 HeapWord* obj = NULL;
230 if (!_to_space_full) {
231 ParGCAllocBuffer* const plab = to_space_alloc_buffer();
232 Space* const sp = to_space();
233 if (word_sz * 100 <
234 ParallelGCBufferWastePct * plab->word_sz()) {
235 // Is small enough; abandon this buffer and start a new one.
236 plab->retire();
237 size_t buf_size = plab->word_sz();
238 HeapWord* buf_space = sp->par_allocate(buf_size);
239 if (buf_space == NULL) {
240 const size_t min_bytes =
241 ParGCAllocBuffer::min_size() << LogHeapWordSize;
242 size_t free_bytes = sp->free();
243 while(buf_space == NULL && free_bytes >= min_bytes) {
244 buf_size = free_bytes >> LogHeapWordSize;
245 assert(buf_size == (size_t)align_object_size(buf_size),
246 "Invariant");
247 buf_space = sp->par_allocate(buf_size);
248 free_bytes = sp->free();
249 }
250 }
251 if (buf_space != NULL) {
252 plab->set_word_size(buf_size);
253 plab->set_buf(buf_space);
254 record_survivor_plab(buf_space, buf_size);
255 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
256 // Note that we cannot compare buf_size < word_sz below
257 // because of AlignmentReserve (see ParGCAllocBuffer::allocate()).
258 assert(obj != NULL || plab->words_remaining() < word_sz,
259 "Else should have been able to allocate");
260 // It's conceivable that we may be able to use the
261 // buffer we just grabbed for subsequent small requests
262 // even if not for this one.
263 } else {
264 // We're used up.
265 _to_space_full = true;
266 }
267
268 } else {
269 // Too large; allocate the object individually.
270 obj = sp->par_allocate(word_sz);
271 }
272 }
273 return obj;
274 }
275
276
277 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj,
278 size_t word_sz) {
279 // Is the alloc in the current alloc buffer?
280 if (to_space_alloc_buffer()->contains(obj)) {
281 assert(to_space_alloc_buffer()->contains(obj + word_sz - 1),
282 "Should contain whole object.");
283 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
284 } else {
285 CollectedHeap::fill_with_object(obj, word_sz);
286 }
287 }
288
289 void ParScanThreadState::print_promotion_failure_size() {
290 if (_promotion_failed_info.has_failed() && PrintPromotionFailure) {
291 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ",
292 _thread_num, _promotion_failed_info.first_size());
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 trace_promotion_failed(const YoungGCTracer* gc_tracer);
313 void reset(int active_workers, bool promotion_failed);
314 void flush();
315
316 #if TASKQUEUE_STATS
317 static void
318 print_termination_stats_hdr(outputStream* const st = gclog_or_tty);
319 void print_termination_stats(outputStream* const st = gclog_or_tty);
320 static void
321 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty);
322 void print_taskqueue_stats(outputStream* const st = gclog_or_tty);
323 void reset_stats();
324 #endif // TASKQUEUE_STATS
325
326 private:
327 ParallelTaskTerminator& _term;
328 ParNewGeneration& _gen;
329 Generation& _old_gen;
330 public:
331 bool is_valid(int id) const { return id < length(); }
332 ParallelTaskTerminator* terminator() { return &_term; }
333 };
334
335
336 ParScanThreadStateSet::ParScanThreadStateSet(
337 int num_threads, Space& to_space, ParNewGeneration& gen,
338 Generation& old_gen, ObjToScanQueueSet& queue_set,
339 Stack<oop, mtGC>* overflow_stacks,
340 size_t desired_plab_sz, ParallelTaskTerminator& term)
341 : ResourceArray(sizeof(ParScanThreadState), num_threads),
342 _gen(gen), _old_gen(old_gen), _term(term)
343 {
344 assert(num_threads > 0, "sanity check!");
345 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
346 "overflow_stack allocation mismatch");
347 // Initialize states.
348 for (int i = 0; i < num_threads; ++i) {
349 new ((ParScanThreadState*)_data + i)
350 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set,
351 overflow_stacks, desired_plab_sz, term);
352 }
353 }
354
355 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i)
356 {
357 assert(i >= 0 && i < length(), "sanity check!");
358 return ((ParScanThreadState*)_data)[i];
359 }
360
361 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) {
362 for (int i = 0; i < length(); ++i) {
363 if (thread_state(i).promotion_failed()) {
364 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info());
365 thread_state(i).promotion_failed_info().reset();
366 }
367 }
368 }
369
370 void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed)
371 {
372 _term.reset_for_reuse(active_threads);
373 if (promotion_failed) {
374 for (int i = 0; i < length(); ++i) {
375 thread_state(i).print_promotion_failure_size();
376 }
377 }
378 }
379
380 #if TASKQUEUE_STATS
381 void
382 ParScanThreadState::reset_stats()
383 {
384 taskqueue_stats().reset();
385 _term_attempts = 0;
386 _overflow_refills = 0;
387 _overflow_refill_objs = 0;
388 }
389
390 void ParScanThreadStateSet::reset_stats()
391 {
392 for (int i = 0; i < length(); ++i) {
393 thread_state(i).reset_stats();
394 }
395 }
396
397 void
398 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st)
399 {
400 st->print_raw_cr("GC Termination Stats");
401 st->print_raw_cr(" elapsed --strong roots-- "
402 "-------termination-------");
403 st->print_raw_cr("thr ms ms % "
404 " ms % attempts");
405 st->print_raw_cr("--- --------- --------- ------ "
406 "--------- ------ --------");
407 }
408
409 void ParScanThreadStateSet::print_termination_stats(outputStream* const st)
410 {
411 print_termination_stats_hdr(st);
412
413 for (int i = 0; i < length(); ++i) {
414 const ParScanThreadState & pss = thread_state(i);
415 const double elapsed_ms = pss.elapsed_time() * 1000.0;
416 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
417 const double term_ms = pss.term_time() * 1000.0;
418 st->print_cr("%3d %9.2f %9.2f %6.2f "
419 "%9.2f %6.2f " SIZE_FORMAT_W(8),
420 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
421 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
422 }
423 }
424
425 // Print stats related to work queue activity.
426 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st)
427 {
428 st->print_raw_cr("GC Task Stats");
429 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
430 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
431 }
432
433 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st)
434 {
435 print_taskqueue_stats_hdr(st);
436
437 TaskQueueStats totals;
438 for (int i = 0; i < length(); ++i) {
439 const ParScanThreadState & pss = thread_state(i);
440 const TaskQueueStats & stats = pss.taskqueue_stats();
441 st->print("%3d ", i); stats.print(st); st->cr();
442 totals += stats;
443
444 if (pss.overflow_refills() > 0) {
445 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
446 SIZE_FORMAT_W(10) " overflow objects",
447 pss.overflow_refills(), pss.overflow_refill_objs());
448 }
449 }
450 st->print("tot "); totals.print(st); st->cr();
451
452 DEBUG_ONLY(totals.verify());
453 }
454 #endif // TASKQUEUE_STATS
455
456 void ParScanThreadStateSet::flush()
457 {
458 // Work in this loop should be kept as lightweight as
459 // possible since this might otherwise become a bottleneck
460 // to scaling. Should we add heavy-weight work into this
461 // loop, consider parallelizing the loop into the worker threads.
462 for (int i = 0; i < length(); ++i) {
463 ParScanThreadState& par_scan_state = thread_state(i);
464
465 // Flush stats related to To-space PLAB activity and
466 // retire the last buffer.
467 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_gen.plab_stats());
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 _old_gen.par_promote_alloc_done(i);
476 _old_gen.par_oop_since_save_marks_iterate_done(i);
477 }
478
479 if (UseConcMarkSweepGC) {
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 OopsInKlassOrGenClosure(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* old_gen,
579 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) :
580 AbstractGangTask("ParNewGeneration collection"),
581 _gen(gen), _old_gen(old_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 void ParNewGenTask::work(uint worker_id) {
597 GenCollectedHeap* gch = GenCollectedHeap::heap();
598 // Since this is being done in a separate thread, need new resource
599 // and handle marks.
600 ResourceMark rm;
601 HandleMark hm;
602 // We would need multiple old-gen queues otherwise.
603 assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen.");
604
605 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
606 assert(_state_set->is_valid(worker_id), "Should not have been called");
607
608 par_scan_state.set_young_old_boundary(_young_old_boundary);
609
610 KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(),
611 gch->rem_set()->klass_rem_set());
612 CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure,
613 &par_scan_state.to_space_root_closure(),
614 false);
615
616 par_scan_state.start_strong_roots();
617 gch->gen_process_roots(_gen->level(),
618 true, // Process younger gens, if any,
619 // as strong roots.
620 false, // no scope; this is parallel code
621 GenCollectedHeap::SO_ScavengeCodeCache,
622 GenCollectedHeap::StrongAndWeakRoots,
623 &par_scan_state.to_space_root_closure(),
624 &par_scan_state.older_gen_closure(),
625 &cld_scan_closure);
626
627 par_scan_state.end_strong_roots();
628
629 // "evacuate followers".
630 par_scan_state.evacuate_followers_closure().do_void();
631 }
632
633 #ifdef _MSC_VER
634 #pragma warning( push )
635 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
636 #endif
637 ParNewGeneration::
638 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level)
639 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"),
640 _overflow_list(NULL),
641 _is_alive_closure(this),
642 _plab_stats(YoungPLABSize, PLABWeight)
643 {
644 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
645 NOT_PRODUCT(_num_par_pushes = 0;)
646 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
647 guarantee(_task_queues != NULL, "task_queues allocation failure.");
648
649 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
650 ObjToScanQueue *q = new ObjToScanQueue();
651 guarantee(q != NULL, "work_queue Allocation failure.");
652 _task_queues->register_queue(i1, q);
653 }
654
655 for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
656 _task_queues->queue(i2)->initialize();
657
658 _overflow_stacks = NULL;
659 if (ParGCUseLocalOverflow) {
660
661 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal
662 // with ','
663 typedef Stack<oop, mtGC> GCOopStack;
664
665 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
666 for (size_t i = 0; i < ParallelGCThreads; ++i) {
667 new (_overflow_stacks + i) Stack<oop, mtGC>();
668 }
669 }
670
671 if (UsePerfData) {
672 EXCEPTION_MARK;
673 ResourceMark rm;
674
675 const char* cname =
676 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
677 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
678 ParallelGCThreads, CHECK);
679 }
680 }
681 #ifdef _MSC_VER
682 #pragma warning( pop )
683 #endif
684
685 // ParNewGeneration::
686 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
687 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
688
689 template <class T>
690 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
691 #ifdef ASSERT
692 {
693 assert(!oopDesc::is_null(*p), "expected non-null ref");
694 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
695 // We never expect to see a null reference being processed
696 // as a weak reference.
697 assert(obj->is_oop(), "expected an oop while scanning weak refs");
698 }
699 #endif // ASSERT
700
701 _par_cl->do_oop_nv(p);
702
703 if (Universe::heap()->is_in_reserved(p)) {
704 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
705 _rs->write_ref_field_gc_par(p, obj);
706 }
707 }
708
709 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
710 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
711
712 // ParNewGeneration::
713 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
714 DefNewGeneration::KeepAliveClosure(cl) {}
715
716 template <class T>
717 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
718 #ifdef ASSERT
719 {
720 assert(!oopDesc::is_null(*p), "expected non-null ref");
721 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
722 // We never expect to see a null reference being processed
723 // as a weak reference.
724 assert(obj->is_oop(), "expected an oop while scanning weak refs");
725 }
726 #endif // ASSERT
727
728 _cl->do_oop_nv(p);
729
730 if (Universe::heap()->is_in_reserved(p)) {
731 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
732 _rs->write_ref_field_gc_par(p, obj);
733 }
734 }
735
736 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
737 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
738
739 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
740 T heap_oop = oopDesc::load_heap_oop(p);
741 if (!oopDesc::is_null(heap_oop)) {
742 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
743 if ((HeapWord*)obj < _boundary) {
744 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
745 oop new_obj = obj->is_forwarded()
746 ? obj->forwardee()
747 : _g->DefNewGeneration::copy_to_survivor_space(obj);
748 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
749 }
750 if (_gc_barrier) {
751 // If p points to a younger generation, mark the card.
752 if ((HeapWord*)obj < _gen_boundary) {
753 _rs->write_ref_field_gc_par(p, obj);
754 }
755 }
756 }
757 }
758
759 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
760 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
761
762 class ParNewRefProcTaskProxy: public AbstractGangTask {
763 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
764 public:
765 ParNewRefProcTaskProxy(ProcessTask& task,
766 ParNewGeneration& gen,
767 Generation& old_gen,
768 HeapWord* young_old_boundary,
769 ParScanThreadStateSet& state_set);
770
771 private:
772 virtual void work(uint worker_id);
773 virtual void set_for_termination(int active_workers) {
774 _state_set.terminator()->reset_for_reuse(active_workers);
775 }
776 private:
777 ParNewGeneration& _gen;
778 ProcessTask& _task;
779 Generation& _old_gen;
780 HeapWord* _young_old_boundary;
781 ParScanThreadStateSet& _state_set;
782 };
783
784 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
785 ParNewGeneration& gen,
786 Generation& old_gen,
787 HeapWord* young_old_boundary,
788 ParScanThreadStateSet& state_set)
789 : AbstractGangTask("ParNewGeneration parallel reference processing"),
790 _gen(gen),
791 _task(task),
792 _old_gen(old_gen),
793 _young_old_boundary(young_old_boundary),
794 _state_set(state_set)
795 {
796 }
797
798 void ParNewRefProcTaskProxy::work(uint worker_id)
799 {
800 ResourceMark rm;
801 HandleMark hm;
802 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
803 par_scan_state.set_young_old_boundary(_young_old_boundary);
804 _task.work(worker_id, par_scan_state.is_alive_closure(),
805 par_scan_state.keep_alive_closure(),
806 par_scan_state.evacuate_followers_closure());
807 }
808
809 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
810 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
811 EnqueueTask& _task;
812
813 public:
814 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
815 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
816 _task(task)
817 { }
818
819 virtual void work(uint worker_id)
820 {
821 _task.work(worker_id);
822 }
823 };
824
825
826 void ParNewRefProcTaskExecutor::execute(ProcessTask& task)
827 {
828 GenCollectedHeap* gch = GenCollectedHeap::heap();
829 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
830 "not a generational heap");
831 FlexibleWorkGang* workers = gch->workers();
832 assert(workers != NULL, "Need parallel worker threads.");
833 _state_set.reset(workers->active_workers(), _generation.promotion_failed());
834 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(),
835 _generation.reserved().end(), _state_set);
836 workers->run_task(&rp_task);
837 _state_set.reset(0 /* bad value in debug if not reset */,
838 _generation.promotion_failed());
839 }
840
841 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task)
842 {
843 GenCollectedHeap* gch = GenCollectedHeap::heap();
844 FlexibleWorkGang* workers = gch->workers();
845 assert(workers != NULL, "Need parallel worker threads.");
846 ParNewRefEnqueueTaskProxy enq_task(task);
847 workers->run_task(&enq_task);
848 }
849
850 void ParNewRefProcTaskExecutor::set_single_threaded_mode()
851 {
852 _state_set.flush();
853 GenCollectedHeap* gch = GenCollectedHeap::heap();
854 gch->set_par_threads(0); // 0 ==> non-parallel.
855 gch->save_marks();
856 }
857
858 ScanClosureWithParBarrier::
859 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
860 ScanClosure(g, gc_barrier) {}
861
862 EvacuateFollowersClosureGeneral::
863 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
864 OopsInGenClosure* cur,
865 OopsInGenClosure* older) :
866 _gch(gch), _level(level),
867 _scan_cur_or_nonheap(cur), _scan_older(older)
868 {}
869
870 void EvacuateFollowersClosureGeneral::do_void() {
871 do {
872 // Beware: this call will lead to closure applications via virtual
873 // calls.
874 _gch->oop_since_save_marks_iterate(_level,
875 _scan_cur_or_nonheap,
876 _scan_older);
877 } while (!_gch->no_allocs_since_save_marks(_level));
878 }
879
880
881 // A Generation that does parallel young-gen collection.
882
883 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set) {
884 assert(_promo_failure_scan_stack.is_empty(), "post condition");
885 _promo_failure_scan_stack.clear(true); // Clear cached segments.
886
887 remove_forwarding_pointers();
888 if (PrintGCDetails) {
889 gclog_or_tty->print(" (promotion failed)");
890 }
891 // All the spaces are in play for mark-sweep.
892 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
893 from()->set_next_compaction_space(to());
894 gch->set_incremental_collection_failed();
895 // Inform the next generation that a promotion failure occurred.
896 _old_gen->promotion_failure_occurred();
897
898 // Trace promotion failure in the parallel GC threads
899 thread_state_set.trace_promotion_failed(gc_tracer());
900 // Single threaded code may have reported promotion failure to the global state
901 if (_promotion_failed_info.has_failed()) {
902 _gc_tracer.report_promotion_failed(_promotion_failed_info);
903 }
904 // Reset the PromotionFailureALot counters.
905 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();)
906 }
907
908 void ParNewGeneration::collect(bool full,
909 bool clear_all_soft_refs,
910 size_t size,
911 bool is_tlab) {
912 assert(full || size > 0, "otherwise we don't want to collect");
913
914 GenCollectedHeap* gch = GenCollectedHeap::heap();
915
916 _gc_timer->register_gc_start();
917
918 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
919 "not a CMS generational heap");
920 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
921 FlexibleWorkGang* workers = gch->workers();
922 assert(workers != NULL, "Need workgang for parallel work");
923 int active_workers =
924 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
925 workers->active_workers(),
926 Threads::number_of_non_daemon_threads());
927 workers->set_active_workers(active_workers);
928 assert(gch->n_gens() == 2,
929 "Par collection currently only works with single older gen.");
930 _old_gen = gch->old_gen();
931
932 // If the next generation is too full to accommodate worst-case promotion
933 // from this generation, pass on collection; let the next generation
934 // do it.
935 if (!collection_attempt_is_safe()) {
936 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
937 return;
938 }
939 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
940
941 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
942 gch->trace_heap_before_gc(gc_tracer());
943
944 init_assuming_no_promotion_failure();
945
946 if (UseAdaptiveSizePolicy) {
947 set_survivor_overflow(false);
948 size_policy->minor_collection_begin();
949 }
950
951 GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL, _gc_tracer.gc_id());
952 // Capture heap used before collection (for printing).
953 size_t gch_prev_used = gch->used();
954
955 age_table()->clear();
956 to()->clear(SpaceDecorator::Mangle);
957
958 gch->save_marks();
959 assert(workers != NULL, "Need parallel worker threads.");
960 int n_workers = active_workers;
961
962 // Set the correct parallelism (number of queues) in the reference processor
963 ref_processor()->set_active_mt_degree(n_workers);
964
965 // Always set the terminator for the active number of workers
966 // because only those workers go through the termination protocol.
967 ParallelTaskTerminator _term(n_workers, task_queues());
968 ParScanThreadStateSet thread_state_set(workers->active_workers(),
969 *to(), *this, *_old_gen, *task_queues(),
970 _overflow_stacks, desired_plab_sz(), _term);
971
972 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set);
973 gch->set_par_threads(n_workers);
974 gch->rem_set()->prepare_for_younger_refs_iterate(true);
975 // It turns out that even when we're using 1 thread, doing the work in a
976 // separate thread causes wide variance in run times. We can't help this
977 // in the multi-threaded case, but we special-case n=1 here to get
978 // repeatable measurements of the 1-thread overhead of the parallel code.
979 if (n_workers > 1) {
980 GenCollectedHeap::StrongRootsScope srs(gch);
981 workers->run_task(&tsk);
982 } else {
983 GenCollectedHeap::StrongRootsScope srs(gch);
984 tsk.work(0);
985 }
986 thread_state_set.reset(0 /* Bad value in debug if not reset */,
987 promotion_failed());
988
989 // Trace and reset failed promotion info.
990 if (promotion_failed()) {
991 thread_state_set.trace_promotion_failed(gc_tracer());
992 }
993
994 // Process (weak) reference objects found during scavenge.
995 ReferenceProcessor* rp = ref_processor();
996 IsAliveClosure is_alive(this);
997 ScanWeakRefClosure scan_weak_ref(this);
998 KeepAliveClosure keep_alive(&scan_weak_ref);
999 ScanClosure scan_without_gc_barrier(this, false);
1000 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
1001 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
1002 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level,
1003 &scan_without_gc_barrier, &scan_with_gc_barrier);
1004 rp->setup_policy(clear_all_soft_refs);
1005 // Can the mt_degree be set later (at run_task() time would be best)?
1006 rp->set_active_mt_degree(active_workers);
1007 ReferenceProcessorStats stats;
1008 if (rp->processing_is_mt()) {
1009 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1010 stats = rp->process_discovered_references(&is_alive, &keep_alive,
1011 &evacuate_followers, &task_executor,
1012 _gc_timer, _gc_tracer.gc_id());
1013 } else {
1014 thread_state_set.flush();
1015 gch->set_par_threads(0); // 0 ==> non-parallel.
1016 gch->save_marks();
1017 stats = rp->process_discovered_references(&is_alive, &keep_alive,
1018 &evacuate_followers, NULL,
1019 _gc_timer, _gc_tracer.gc_id());
1020 }
1021 _gc_tracer.report_gc_reference_stats(stats);
1022 if (!promotion_failed()) {
1023 // Swap the survivor spaces.
1024 eden()->clear(SpaceDecorator::Mangle);
1025 from()->clear(SpaceDecorator::Mangle);
1026 if (ZapUnusedHeapArea) {
1027 // This is now done here because of the piece-meal mangling which
1028 // can check for valid mangling at intermediate points in the
1029 // collection(s). When a minor collection fails to collect
1030 // sufficient space resizing of the young generation can occur
1031 // an redistribute the spaces in the young generation. Mangle
1032 // here so that unzapped regions don't get distributed to
1033 // other spaces.
1034 to()->mangle_unused_area();
1035 }
1036 swap_spaces();
1037
1038 // A successful scavenge should restart the GC time limit count which is
1039 // for full GC's.
1040 size_policy->reset_gc_overhead_limit_count();
1041
1042 assert(to()->is_empty(), "to space should be empty now");
1043
1044 adjust_desired_tenuring_threshold();
1045 } else {
1046 handle_promotion_failed(gch, thread_state_set);
1047 }
1048 // set new iteration safe limit for the survivor spaces
1049 from()->set_concurrent_iteration_safe_limit(from()->top());
1050 to()->set_concurrent_iteration_safe_limit(to()->top());
1051
1052 if (ResizePLAB) {
1053 plab_stats()->adjust_desired_plab_sz(n_workers);
1054 }
1055
1056 if (PrintGC && !PrintGCDetails) {
1057 gch->print_heap_change(gch_prev_used);
1058 }
1059
1060 TASKQUEUE_STATS_ONLY(if (PrintTerminationStats) thread_state_set.print_termination_stats());
1061 TASKQUEUE_STATS_ONLY(if (PrintTaskqueue) thread_state_set.print_taskqueue_stats());
1062
1063 if (UseAdaptiveSizePolicy) {
1064 size_policy->minor_collection_end(gch->gc_cause());
1065 size_policy->avg_survived()->sample(from()->used());
1066 }
1067
1068 // We need to use a monotonically non-decreasing time in ms
1069 // or we will see time-warp warnings and os::javaTimeMillis()
1070 // does not guarantee monotonicity.
1071 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1072 update_time_of_last_gc(now);
1073
1074 rp->set_enqueuing_is_done(true);
1075 if (rp->processing_is_mt()) {
1076 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1077 rp->enqueue_discovered_references(&task_executor);
1078 } else {
1079 rp->enqueue_discovered_references(NULL);
1080 }
1081 rp->verify_no_references_recorded();
1082
1083 gch->trace_heap_after_gc(gc_tracer());
1084 _gc_tracer.report_tenuring_threshold(tenuring_threshold());
1085
1086 _gc_timer->register_gc_end();
1087
1088 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1089 }
1090
1091 static int sum;
1092 void ParNewGeneration::waste_some_time() {
1093 for (int i = 0; i < 100; i++) {
1094 sum += i;
1095 }
1096 }
1097
1098 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1099
1100 // Because of concurrency, there are times where an object for which
1101 // "is_forwarded()" is true contains an "interim" forwarding pointer
1102 // value. Such a value will soon be overwritten with a real value.
1103 // This method requires "obj" to have a forwarding pointer, and waits, if
1104 // necessary for a real one to be inserted, and returns it.
1105
1106 oop ParNewGeneration::real_forwardee(oop obj) {
1107 oop forward_ptr = obj->forwardee();
1108 if (forward_ptr != ClaimedForwardPtr) {
1109 return forward_ptr;
1110 } else {
1111 return real_forwardee_slow(obj);
1112 }
1113 }
1114
1115 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1116 // Spin-read if it is claimed but not yet written by another thread.
1117 oop forward_ptr = obj->forwardee();
1118 while (forward_ptr == ClaimedForwardPtr) {
1119 waste_some_time();
1120 assert(obj->is_forwarded(), "precondition");
1121 forward_ptr = obj->forwardee();
1122 }
1123 return forward_ptr;
1124 }
1125
1126 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
1127 if (m->must_be_preserved_for_promotion_failure(obj)) {
1128 // We should really have separate per-worker stacks, rather
1129 // than use locking of a common pair of stacks.
1130 MutexLocker ml(ParGCRareEvent_lock);
1131 preserve_mark(obj, m);
1132 }
1133 }
1134
1135 // Multiple GC threads may try to promote an object. If the object
1136 // is successfully promoted, a forwarding pointer will be installed in
1137 // the object in the young generation. This method claims the right
1138 // to install the forwarding pointer before it copies the object,
1139 // thus avoiding the need to undo the copy as in
1140 // copy_to_survivor_space_avoiding_with_undo.
1141
1142 oop ParNewGeneration::copy_to_survivor_space(
1143 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1144 // In the sequential version, this assert also says that the object is
1145 // not forwarded. That might not be the case here. It is the case that
1146 // the caller observed it to be not forwarded at some time in the past.
1147 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1148
1149 // The sequential code read "old->age()" below. That doesn't work here,
1150 // since the age is in the mark word, and that might be overwritten with
1151 // a forwarding pointer by a parallel thread. So we must save the mark
1152 // word in a local and then analyze it.
1153 oopDesc dummyOld;
1154 dummyOld.set_mark(m);
1155 assert(!dummyOld.is_forwarded(),
1156 "should not be called with forwarding pointer mark word.");
1157
1158 oop new_obj = NULL;
1159 oop forward_ptr;
1160
1161 // Try allocating obj in to-space (unless too old)
1162 if (dummyOld.age() < tenuring_threshold()) {
1163 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1164 if (new_obj == NULL) {
1165 set_survivor_overflow(true);
1166 }
1167 }
1168
1169 if (new_obj == NULL) {
1170 // Either to-space is full or we decided to promote
1171 // try allocating obj tenured
1172
1173 // Attempt to install a null forwarding pointer (atomically),
1174 // to claim the right to install the real forwarding pointer.
1175 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1176 if (forward_ptr != NULL) {
1177 // someone else beat us to it.
1178 return real_forwardee(old);
1179 }
1180
1181 if (!_promotion_failed) {
1182 new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1183 old, m, sz);
1184 }
1185
1186 if (new_obj == NULL) {
1187 // promotion failed, forward to self
1188 _promotion_failed = true;
1189 new_obj = old;
1190
1191 preserve_mark_if_necessary(old, m);
1192 par_scan_state->register_promotion_failure(sz);
1193 }
1194
1195 old->forward_to(new_obj);
1196 forward_ptr = NULL;
1197 } else {
1198 // Is in to-space; do copying ourselves.
1199 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1200 assert(Universe::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1201 forward_ptr = old->forward_to_atomic(new_obj);
1202 // Restore the mark word copied above.
1203 new_obj->set_mark(m);
1204 // Increment age if obj still in new generation
1205 new_obj->incr_age();
1206 par_scan_state->age_table()->add(new_obj, sz);
1207 }
1208 assert(new_obj != NULL, "just checking");
1209
1210 #ifndef PRODUCT
1211 // This code must come after the CAS test, or it will print incorrect
1212 // information.
1213 if (TraceScavenge) {
1214 gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1215 is_in_reserved(new_obj) ? "copying" : "tenuring",
1216 new_obj->klass()->internal_name(), (void *)old, (void *)new_obj, new_obj->size());
1217 }
1218 #endif
1219
1220 if (forward_ptr == NULL) {
1221 oop obj_to_push = new_obj;
1222 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1223 // Length field used as index of next element to be scanned.
1224 // Real length can be obtained from real_forwardee()
1225 arrayOop(old)->set_length(0);
1226 obj_to_push = old;
1227 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1228 "push forwarded object");
1229 }
1230 // Push it on one of the queues of to-be-scanned objects.
1231 bool simulate_overflow = false;
1232 NOT_PRODUCT(
1233 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1234 // simulate a stack overflow
1235 simulate_overflow = true;
1236 }
1237 )
1238 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1239 // Add stats for overflow pushes.
1240 if (Verbose && PrintGCDetails) {
1241 gclog_or_tty->print("queue overflow!\n");
1242 }
1243 push_on_overflow_list(old, par_scan_state);
1244 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1245 }
1246
1247 return new_obj;
1248 }
1249
1250 // Oops. Someone beat us to it. Undo the allocation. Where did we
1251 // allocate it?
1252 if (is_in_reserved(new_obj)) {
1253 // Must be in to_space.
1254 assert(to()->is_in_reserved(new_obj), "Checking");
1255 if (forward_ptr == ClaimedForwardPtr) {
1256 // Wait to get the real forwarding pointer value.
1257 forward_ptr = real_forwardee(old);
1258 }
1259 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1260 }
1261
1262 return forward_ptr;
1263 }
1264
1265 #ifndef PRODUCT
1266 // It's OK to call this multi-threaded; the worst thing
1267 // that can happen is that we'll get a bunch of closely
1268 // spaced simulated overflows, but that's OK, in fact
1269 // probably good as it would exercise the overflow code
1270 // under contention.
1271 bool ParNewGeneration::should_simulate_overflow() {
1272 if (_overflow_counter-- <= 0) { // just being defensive
1273 _overflow_counter = ParGCWorkQueueOverflowInterval;
1274 return true;
1275 } else {
1276 return false;
1277 }
1278 }
1279 #endif
1280
1281 // In case we are using compressed oops, we need to be careful.
1282 // If the object being pushed is an object array, then its length
1283 // field keeps track of the "grey boundary" at which the next
1284 // incremental scan will be done (see ParGCArrayScanChunk).
1285 // When using compressed oops, this length field is kept in the
1286 // lower 32 bits of the erstwhile klass word and cannot be used
1287 // for the overflow chaining pointer (OCP below). As such the OCP
1288 // would itself need to be compressed into the top 32-bits in this
1289 // case. Unfortunately, see below, in the event that we have a
1290 // promotion failure, the node to be pushed on the list can be
1291 // outside of the Java heap, so the heap-based pointer compression
1292 // would not work (we would have potential aliasing between C-heap
1293 // and Java-heap pointers). For this reason, when using compressed
1294 // oops, we simply use a worker-thread-local, non-shared overflow
1295 // list in the form of a growable array, with a slightly different
1296 // overflow stack draining strategy. If/when we start using fat
1297 // stacks here, we can go back to using (fat) pointer chains
1298 // (although some performance comparisons would be useful since
1299 // single global lists have their own performance disadvantages
1300 // as we were made painfully aware not long ago, see 6786503).
1301 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1302 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1303 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1304 if (ParGCUseLocalOverflow) {
1305 // In the case of compressed oops, we use a private, not-shared
1306 // overflow stack.
1307 par_scan_state->push_on_overflow_stack(from_space_obj);
1308 } else {
1309 assert(!UseCompressedOops, "Error");
1310 // if the object has been forwarded to itself, then we cannot
1311 // use the klass pointer for the linked list. Instead we have
1312 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1313 // XXX This is horribly inefficient when a promotion failure occurs
1314 // and should be fixed. XXX FIX ME !!!
1315 #ifndef PRODUCT
1316 Atomic::inc_ptr(&_num_par_pushes);
1317 assert(_num_par_pushes > 0, "Tautology");
1318 #endif
1319 if (from_space_obj->forwardee() == from_space_obj) {
1320 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1321 listhead->forward_to(from_space_obj);
1322 from_space_obj = listhead;
1323 }
1324 oop observed_overflow_list = _overflow_list;
1325 oop cur_overflow_list;
1326 do {
1327 cur_overflow_list = observed_overflow_list;
1328 if (cur_overflow_list != BUSY) {
1329 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1330 } else {
1331 from_space_obj->set_klass_to_list_ptr(NULL);
1332 }
1333 observed_overflow_list =
1334 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1335 } while (cur_overflow_list != observed_overflow_list);
1336 }
1337 }
1338
1339 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1340 bool res;
1341
1342 if (ParGCUseLocalOverflow) {
1343 res = par_scan_state->take_from_overflow_stack();
1344 } else {
1345 assert(!UseCompressedOops, "Error");
1346 res = take_from_overflow_list_work(par_scan_state);
1347 }
1348 return res;
1349 }
1350
1351
1352 // *NOTE*: The overflow list manipulation code here and
1353 // in CMSCollector:: are very similar in shape,
1354 // except that in the CMS case we thread the objects
1355 // directly into the list via their mark word, and do
1356 // not need to deal with special cases below related
1357 // to chunking of object arrays and promotion failure
1358 // handling.
1359 // CR 6797058 has been filed to attempt consolidation of
1360 // the common code.
1361 // Because of the common code, if you make any changes in
1362 // the code below, please check the CMS version to see if
1363 // similar changes might be needed.
1364 // See CMSCollector::par_take_from_overflow_list() for
1365 // more extensive documentation comments.
1366 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1367 ObjToScanQueue* work_q = par_scan_state->work_queue();
1368 // How many to take?
1369 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1370 (size_t)ParGCDesiredObjsFromOverflowList);
1371
1372 assert(!UseCompressedOops, "Error");
1373 assert(par_scan_state->overflow_stack() == NULL, "Error");
1374 if (_overflow_list == NULL) return false;
1375
1376 // Otherwise, there was something there; try claiming the list.
1377 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1378 // Trim off a prefix of at most objsFromOverflow items
1379 Thread* tid = Thread::current();
1380 size_t spin_count = (size_t)ParallelGCThreads;
1381 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1382 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1383 // someone grabbed it before we did ...
1384 // ... we spin for a short while...
1385 os::sleep(tid, sleep_time_millis, false);
1386 if (_overflow_list == NULL) {
1387 // nothing left to take
1388 return false;
1389 } else if (_overflow_list != BUSY) {
1390 // try and grab the prefix
1391 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1392 }
1393 }
1394 if (prefix == NULL || prefix == BUSY) {
1395 // Nothing to take or waited long enough
1396 if (prefix == NULL) {
1397 // Write back the NULL in case we overwrote it with BUSY above
1398 // and it is still the same value.
1399 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1400 }
1401 return false;
1402 }
1403 assert(prefix != NULL && prefix != BUSY, "Error");
1404 size_t i = 1;
1405 oop cur = prefix;
1406 while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1407 i++; cur = cur->list_ptr_from_klass();
1408 }
1409
1410 // Reattach remaining (suffix) to overflow list
1411 if (cur->klass_or_null() == NULL) {
1412 // Write back the NULL in lieu of the BUSY we wrote
1413 // above and it is still the same value.
1414 if (_overflow_list == BUSY) {
1415 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1416 }
1417 } else {
1418 assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error");
1419 oop suffix = cur->list_ptr_from_klass(); // suffix will be put back on global list
1420 cur->set_klass_to_list_ptr(NULL); // break off suffix
1421 // It's possible that the list is still in the empty(busy) state
1422 // we left it in a short while ago; in that case we may be
1423 // able to place back the suffix.
1424 oop observed_overflow_list = _overflow_list;
1425 oop cur_overflow_list = observed_overflow_list;
1426 bool attached = false;
1427 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1428 observed_overflow_list =
1429 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1430 if (cur_overflow_list == observed_overflow_list) {
1431 attached = true;
1432 break;
1433 } else cur_overflow_list = observed_overflow_list;
1434 }
1435 if (!attached) {
1436 // Too bad, someone else got in in between; we'll need to do a splice.
1437 // Find the last item of suffix list
1438 oop last = suffix;
1439 while (last->klass_or_null() != NULL) {
1440 last = last->list_ptr_from_klass();
1441 }
1442 // Atomically prepend suffix to current overflow list
1443 observed_overflow_list = _overflow_list;
1444 do {
1445 cur_overflow_list = observed_overflow_list;
1446 if (cur_overflow_list != BUSY) {
1447 // Do the splice ...
1448 last->set_klass_to_list_ptr(cur_overflow_list);
1449 } else { // cur_overflow_list == BUSY
1450 last->set_klass_to_list_ptr(NULL);
1451 }
1452 observed_overflow_list =
1453 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1454 } while (cur_overflow_list != observed_overflow_list);
1455 }
1456 }
1457
1458 // Push objects on prefix list onto this thread's work queue
1459 assert(prefix != NULL && prefix != BUSY, "program logic");
1460 cur = prefix;
1461 ssize_t n = 0;
1462 while (cur != NULL) {
1463 oop obj_to_push = cur->forwardee();
1464 oop next = cur->list_ptr_from_klass();
1465 cur->set_klass(obj_to_push->klass());
1466 // This may be an array object that is self-forwarded. In that case, the list pointer
1467 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1468 if (!is_in_reserved(cur)) {
1469 // This can become a scaling bottleneck when there is work queue overflow coincident
1470 // with promotion failure.
1471 oopDesc* f = cur;
1472 FREE_C_HEAP_ARRAY(oopDesc, f);
1473 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1474 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1475 obj_to_push = cur;
1476 }
1477 bool ok = work_q->push(obj_to_push);
1478 assert(ok, "Should have succeeded");
1479 cur = next;
1480 n++;
1481 }
1482 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1483 #ifndef PRODUCT
1484 assert(_num_par_pushes >= n, "Too many pops?");
1485 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1486 #endif
1487 return true;
1488 }
1489 #undef BUSY
1490
1491 void ParNewGeneration::ref_processor_init() {
1492 if (_ref_processor == NULL) {
1493 // Allocate and initialize a reference processor
1494 _ref_processor =
1495 new ReferenceProcessor(_reserved, // span
1496 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1497 (int) ParallelGCThreads, // mt processing degree
1498 refs_discovery_is_mt(), // mt discovery
1499 (int) ParallelGCThreads, // mt discovery degree
1500 refs_discovery_is_atomic(), // atomic_discovery
1501 NULL); // is_alive_non_header
1502 }
1503 }
1504
1505 const char* ParNewGeneration::name() const {
1506 return "par new generation";
1507 }