1 /*
2 * Copyright (c) 2001, 2018, 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 "classfile/stringTable.hpp"
27 #include "gc/cms/cmsHeap.inline.hpp"
28 #include "gc/cms/compactibleFreeListSpace.hpp"
29 #include "gc/cms/concurrentMarkSweepGeneration.hpp"
30 #include "gc/cms/parNewGeneration.inline.hpp"
31 #include "gc/cms/parOopClosures.inline.hpp"
32 #include "gc/serial/defNewGeneration.inline.hpp"
33 #include "gc/shared/adaptiveSizePolicy.hpp"
34 #include "gc/shared/ageTable.inline.hpp"
35 #include "gc/shared/copyFailedInfo.hpp"
36 #include "gc/shared/gcHeapSummary.hpp"
37 #include "gc/shared/gcTimer.hpp"
38 #include "gc/shared/gcTrace.hpp"
39 #include "gc/shared/gcTraceTime.inline.hpp"
40 #include "gc/shared/genOopClosures.inline.hpp"
41 #include "gc/shared/generation.hpp"
42 #include "gc/shared/plab.inline.hpp"
43 #include "gc/shared/preservedMarks.inline.hpp"
44 #include "gc/shared/referencePolicy.hpp"
45 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
46 #include "gc/shared/space.hpp"
47 #include "gc/shared/spaceDecorator.hpp"
48 #include "gc/shared/strongRootsScope.hpp"
49 #include "gc/shared/taskqueue.inline.hpp"
50 #include "gc/shared/weakProcessor.hpp"
51 #include "gc/shared/workgroup.hpp"
52 #include "logging/log.hpp"
53 #include "logging/logStream.hpp"
54 #include "memory/resourceArea.hpp"
55 #include "oops/access.inline.hpp"
56 #include "oops/compressedOops.inline.hpp"
57 #include "oops/objArrayOop.hpp"
58 #include "oops/oop.inline.hpp"
59 #include "runtime/atomic.hpp"
60 #include "runtime/handles.hpp"
61 #include "runtime/handles.inline.hpp"
62 #include "runtime/java.hpp"
63 #include "runtime/thread.inline.hpp"
64 #include "utilities/copy.hpp"
65 #include "utilities/globalDefinitions.hpp"
66 #include "utilities/stack.inline.hpp"
67
68 ParScanThreadState::ParScanThreadState(Space* to_space_,
69 ParNewGeneration* young_gen_,
70 Generation* old_gen_,
71 int thread_num_,
72 ObjToScanQueueSet* work_queue_set_,
73 Stack<oop, mtGC>* overflow_stacks_,
74 PreservedMarks* preserved_marks_,
75 size_t desired_plab_sz_,
76 ParallelTaskTerminator& term_) :
77 _to_space(to_space_),
78 _old_gen(old_gen_),
79 _young_gen(young_gen_),
80 _thread_num(thread_num_),
81 _work_queue(work_queue_set_->queue(thread_num_)),
82 _to_space_full(false),
83 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
84 _preserved_marks(preserved_marks_),
85 _ageTable(false), // false ==> not the global age table, no perf data.
86 _to_space_alloc_buffer(desired_plab_sz_),
87 _to_space_closure(young_gen_, this),
88 _old_gen_closure(young_gen_, this),
89 _to_space_root_closure(young_gen_, this),
90 _old_gen_root_closure(young_gen_, this),
91 _older_gen_closure(young_gen_, this),
92 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
93 &_to_space_root_closure, young_gen_, &_old_gen_root_closure,
94 work_queue_set_, &term_),
95 _is_alive_closure(young_gen_),
96 _scan_weak_ref_closure(young_gen_, this),
97 _keep_alive_closure(&_scan_weak_ref_closure),
98 _strong_roots_time(0.0),
99 _term_time(0.0)
100 {
101 #if TASKQUEUE_STATS
102 _term_attempts = 0;
103 _overflow_refills = 0;
104 _overflow_refill_objs = 0;
105 #endif // TASKQUEUE_STATS
106
107 _survivor_chunk_array = (ChunkArray*) old_gen()->get_data_recorder(thread_num());
108 _hash_seed = 17; // Might want to take time-based random value.
109 _start = os::elapsedTime();
110 _old_gen_closure.set_generation(old_gen_);
111 _old_gen_root_closure.set_generation(old_gen_);
112 }
113
114 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
115 size_t plab_word_size) {
116 ChunkArray* sca = survivor_chunk_array();
117 if (sca != NULL) {
118 // A non-null SCA implies that we want the PLAB data recorded.
119 sca->record_sample(plab_start, plab_word_size);
120 }
121 }
122
123 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
124 return new_obj->is_objArray() &&
125 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
126 new_obj != old_obj;
127 }
128
129 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
130 assert(old->is_objArray(), "must be obj array");
131 assert(old->is_forwarded(), "must be forwarded");
132 assert(CMSHeap::heap()->is_in_reserved(old), "must be in heap.");
133 assert(!old_gen()->is_in(old), "must be in young generation.");
134
135 objArrayOop obj = objArrayOop(old->forwardee());
136 // Process ParGCArrayScanChunk elements now
137 // and push the remainder back onto queue
138 int start = arrayOop(old)->length();
139 int end = obj->length();
140 int remainder = end - start;
141 assert(start <= end, "just checking");
142 if (remainder > 2 * ParGCArrayScanChunk) {
143 // Test above combines last partial chunk with a full chunk
144 end = start + ParGCArrayScanChunk;
145 arrayOop(old)->set_length(end);
146 // Push remainder.
147 bool ok = work_queue()->push(old);
148 assert(ok, "just popped, push must be okay");
149 } else {
150 // Restore length so that it can be used if there
151 // is a promotion failure and forwarding pointers
152 // must be removed.
153 arrayOop(old)->set_length(end);
154 }
155
156 // process our set of indices (include header in first chunk)
157 // should make sure end is even (aligned to HeapWord in case of compressed oops)
158 if ((HeapWord *)obj < young_old_boundary()) {
159 // object is in to_space
160 obj->oop_iterate_range(&_to_space_closure, start, end);
161 } else {
162 // object is in old generation
163 obj->oop_iterate_range(&_old_gen_closure, start, end);
164 }
165 }
166
167 void ParScanThreadState::trim_queues(int max_size) {
168 ObjToScanQueue* queue = work_queue();
169 do {
170 while (queue->size() > (juint)max_size) {
171 oop obj_to_scan;
172 if (queue->pop_local(obj_to_scan)) {
173 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
174 if (obj_to_scan->is_objArray() &&
175 obj_to_scan->is_forwarded() &&
176 obj_to_scan->forwardee() != obj_to_scan) {
177 scan_partial_array_and_push_remainder(obj_to_scan);
178 } else {
179 // object is in to_space
180 obj_to_scan->oop_iterate(&_to_space_closure);
181 }
182 } else {
183 // object is in old generation
184 obj_to_scan->oop_iterate(&_old_gen_closure);
185 }
186 }
187 }
188 // For the case of compressed oops, we have a private, non-shared
189 // overflow stack, so we eagerly drain it so as to more evenly
190 // distribute load early. Note: this may be good to do in
191 // general rather than delay for the final stealing phase.
192 // If applicable, we'll transfer a set of objects over to our
193 // work queue, allowing them to be stolen and draining our
194 // private overflow stack.
195 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
196 }
197
198 bool ParScanThreadState::take_from_overflow_stack() {
199 assert(ParGCUseLocalOverflow, "Else should not call");
200 assert(young_gen()->overflow_list() == NULL, "Error");
201 ObjToScanQueue* queue = work_queue();
202 Stack<oop, mtGC>* const of_stack = overflow_stack();
203 const size_t num_overflow_elems = of_stack->size();
204 const size_t space_available = queue->max_elems() - queue->size();
205 const size_t num_take_elems = MIN3(space_available / 4,
206 (size_t)ParGCDesiredObjsFromOverflowList,
207 num_overflow_elems);
208 // Transfer the most recent num_take_elems from the overflow
209 // stack to our work queue.
210 for (size_t i = 0; i != num_take_elems; i++) {
211 oop cur = of_stack->pop();
212 oop obj_to_push = cur->forwardee();
213 assert(CMSHeap::heap()->is_in_reserved(cur), "Should be in heap");
214 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
215 assert(CMSHeap::heap()->is_in_reserved(obj_to_push), "Should be in heap");
216 if (should_be_partially_scanned(obj_to_push, cur)) {
217 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
218 obj_to_push = cur;
219 }
220 bool ok = queue->push(obj_to_push);
221 assert(ok, "Should have succeeded");
222 }
223 assert(young_gen()->overflow_list() == NULL, "Error");
224 return num_take_elems > 0; // was something transferred?
225 }
226
227 void ParScanThreadState::push_on_overflow_stack(oop p) {
228 assert(ParGCUseLocalOverflow, "Else should not call");
229 overflow_stack()->push(p);
230 assert(young_gen()->overflow_list() == NULL, "Error");
231 }
232
233 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
234 // If the object is small enough, try to reallocate the buffer.
235 HeapWord* obj = NULL;
236 if (!_to_space_full) {
237 PLAB* const plab = to_space_alloc_buffer();
238 Space* const sp = to_space();
239 if (word_sz * 100 < ParallelGCBufferWastePct * plab->word_sz()) {
240 // Is small enough; abandon this buffer and start a new one.
241 plab->retire();
242 // The minimum size has to be twice SurvivorAlignmentInBytes to
243 // allow for padding used in the alignment of 1 word. A padding
244 // of 1 is too small for a filler word so the padding size will
245 // be increased by SurvivorAlignmentInBytes.
246 size_t min_usable_size = 2 * static_cast<size_t>(SurvivorAlignmentInBytes >> LogHeapWordSize);
247 size_t buf_size = MAX2(plab->word_sz(), min_usable_size);
248 HeapWord* buf_space = sp->par_allocate(buf_size);
249 if (buf_space == NULL) {
250 const size_t min_bytes = MAX2(PLAB::min_size(), min_usable_size) << LogHeapWordSize;
251 size_t free_bytes = sp->free();
252 while(buf_space == NULL && free_bytes >= min_bytes) {
253 buf_size = free_bytes >> LogHeapWordSize;
254 assert(buf_size == (size_t)align_object_size(buf_size), "Invariant");
255 buf_space = sp->par_allocate(buf_size);
256 free_bytes = sp->free();
257 }
258 }
259 if (buf_space != NULL) {
260 plab->set_buf(buf_space, buf_size);
261 record_survivor_plab(buf_space, buf_size);
262 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
263 // Note that we cannot compare buf_size < word_sz below
264 // because of AlignmentReserve (see PLAB::allocate()).
265 assert(obj != NULL || plab->words_remaining() < word_sz,
266 "Else should have been able to allocate requested object size "
267 SIZE_FORMAT ", PLAB size " SIZE_FORMAT ", SurvivorAlignmentInBytes "
268 SIZE_FORMAT ", words_remaining " SIZE_FORMAT,
269 word_sz, buf_size, SurvivorAlignmentInBytes, plab->words_remaining());
270 // It's conceivable that we may be able to use the
271 // buffer we just grabbed for subsequent small requests
272 // even if not for this one.
273 } else {
274 // We're used up.
275 _to_space_full = true;
276 }
277 } else {
278 // Too large; allocate the object individually.
279 obj = sp->par_allocate(word_sz);
280 }
281 }
282 return obj;
283 }
284
285 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) {
286 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
287 }
288
289 void ParScanThreadState::print_promotion_failure_size() {
290 if (_promotion_failed_info.has_failed()) {
291 log_trace(gc, promotion)(" (%d: promotion failure size = " SIZE_FORMAT ") ",
292 _thread_num, _promotion_failed_info.first_size());
293 }
294 }
295
296 class ParScanThreadStateSet: StackObj {
297 public:
298 // Initializes states for the specified number of threads;
299 ParScanThreadStateSet(int num_threads,
300 Space& to_space,
301 ParNewGeneration& young_gen,
302 Generation& old_gen,
303 ObjToScanQueueSet& queue_set,
304 Stack<oop, mtGC>* overflow_stacks_,
305 PreservedMarksSet& preserved_marks_set,
306 size_t desired_plab_sz,
307 ParallelTaskTerminator& term);
308
309 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
310
311 inline ParScanThreadState& thread_state(int i);
312
313 void trace_promotion_failed(const YoungGCTracer* gc_tracer);
314 void reset(uint active_workers, bool promotion_failed);
315 void flush();
316
317 #if TASKQUEUE_STATS
318 static void
319 print_termination_stats_hdr(outputStream* const st);
320 void print_termination_stats();
321 static void
322 print_taskqueue_stats_hdr(outputStream* const st);
323 void print_taskqueue_stats();
324 void reset_stats();
325 #endif // TASKQUEUE_STATS
326
327 private:
328 ParallelTaskTerminator& _term;
329 ParNewGeneration& _young_gen;
330 Generation& _old_gen;
331 ParScanThreadState* _per_thread_states;
332 const int _num_threads;
333 public:
334 bool is_valid(int id) const { return id < _num_threads; }
335 ParallelTaskTerminator* terminator() { return &_term; }
336 };
337
338 ParScanThreadStateSet::ParScanThreadStateSet(int num_threads,
339 Space& to_space,
340 ParNewGeneration& young_gen,
341 Generation& old_gen,
342 ObjToScanQueueSet& queue_set,
343 Stack<oop, mtGC>* overflow_stacks,
344 PreservedMarksSet& preserved_marks_set,
345 size_t desired_plab_sz,
346 ParallelTaskTerminator& term)
347 : _young_gen(young_gen),
348 _old_gen(old_gen),
349 _term(term),
350 _per_thread_states(NEW_RESOURCE_ARRAY(ParScanThreadState, num_threads)),
351 _num_threads(num_threads)
352 {
353 assert(num_threads > 0, "sanity check!");
354 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
355 "overflow_stack allocation mismatch");
356 // Initialize states.
357 for (int i = 0; i < num_threads; ++i) {
358 new(_per_thread_states + i)
359 ParScanThreadState(&to_space, &young_gen, &old_gen, i, &queue_set,
360 overflow_stacks, preserved_marks_set.get(i),
361 desired_plab_sz, term);
362 }
363 }
364
365 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) {
366 assert(i >= 0 && i < _num_threads, "sanity check!");
367 return _per_thread_states[i];
368 }
369
370 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) {
371 for (int i = 0; i < _num_threads; ++i) {
372 if (thread_state(i).promotion_failed()) {
373 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info());
374 thread_state(i).promotion_failed_info().reset();
375 }
376 }
377 }
378
379 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed) {
380 _term.reset_for_reuse(active_threads);
381 if (promotion_failed) {
382 for (int i = 0; i < _num_threads; ++i) {
383 thread_state(i).print_promotion_failure_size();
384 }
385 }
386 }
387
388 #if TASKQUEUE_STATS
389 void ParScanThreadState::reset_stats() {
390 taskqueue_stats().reset();
391 _term_attempts = 0;
392 _overflow_refills = 0;
393 _overflow_refill_objs = 0;
394 }
395
396 void ParScanThreadStateSet::reset_stats() {
397 for (int i = 0; i < _num_threads; ++i) {
398 thread_state(i).reset_stats();
399 }
400 }
401
402 void ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) {
403 st->print_raw_cr("GC Termination Stats");
404 st->print_raw_cr(" elapsed --strong roots-- -------termination-------");
405 st->print_raw_cr("thr ms ms % ms % attempts");
406 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------");
407 }
408
409 void ParScanThreadStateSet::print_termination_stats() {
410 Log(gc, task, stats) log;
411 if (!log.is_debug()) {
412 return;
413 }
414
415 ResourceMark rm;
416 LogStream ls(log.debug());
417 outputStream* st = &ls;
418
419 print_termination_stats_hdr(st);
420
421 for (int i = 0; i < _num_threads; ++i) {
422 const ParScanThreadState & pss = thread_state(i);
423 const double elapsed_ms = pss.elapsed_time() * 1000.0;
424 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
425 const double term_ms = pss.term_time() * 1000.0;
426 st->print_cr("%3d %9.2f %9.2f %6.2f %9.2f %6.2f " SIZE_FORMAT_W(8),
427 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
428 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
429 }
430 }
431
432 // Print stats related to work queue activity.
433 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) {
434 st->print_raw_cr("GC Task Stats");
435 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
436 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
437 }
438
439 void ParScanThreadStateSet::print_taskqueue_stats() {
440 if (!log_develop_is_enabled(Trace, gc, task, stats)) {
441 return;
442 }
443 Log(gc, task, stats) log;
444 ResourceMark rm;
445 LogStream ls(log.trace());
446 outputStream* st = &ls;
447 print_taskqueue_stats_hdr(st);
448
449 TaskQueueStats totals;
450 for (int i = 0; i < _num_threads; ++i) {
451 const ParScanThreadState & pss = thread_state(i);
452 const TaskQueueStats & stats = pss.taskqueue_stats();
453 st->print("%3d ", i); stats.print(st); st->cr();
454 totals += stats;
455
456 if (pss.overflow_refills() > 0) {
457 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
458 SIZE_FORMAT_W(10) " overflow objects",
459 pss.overflow_refills(), pss.overflow_refill_objs());
460 }
461 }
462 st->print("tot "); totals.print(st); st->cr();
463
464 DEBUG_ONLY(totals.verify());
465 }
466 #endif // TASKQUEUE_STATS
467
468 void ParScanThreadStateSet::flush() {
469 // Work in this loop should be kept as lightweight as
470 // possible since this might otherwise become a bottleneck
471 // to scaling. Should we add heavy-weight work into this
472 // loop, consider parallelizing the loop into the worker threads.
473 for (int i = 0; i < _num_threads; ++i) {
474 ParScanThreadState& par_scan_state = thread_state(i);
475
476 // Flush stats related to To-space PLAB activity and
477 // retire the last buffer.
478 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_young_gen.plab_stats());
479
480 // Every thread has its own age table. We need to merge
481 // them all into one.
482 AgeTable *local_table = par_scan_state.age_table();
483 _young_gen.age_table()->merge(local_table);
484
485 // Inform old gen that we're done.
486 _old_gen.par_promote_alloc_done(i);
487 }
488
489 if (UseConcMarkSweepGC) {
490 // We need to call this even when ResizeOldPLAB is disabled
491 // so as to avoid breaking some asserts. While we may be able
492 // to avoid this by reorganizing the code a bit, I am loathe
493 // to do that unless we find cases where ergo leads to bad
494 // performance.
495 CompactibleFreeListSpaceLAB::compute_desired_plab_size();
496 }
497 }
498
499 ParScanClosure::ParScanClosure(ParNewGeneration* g,
500 ParScanThreadState* par_scan_state) :
501 OopsInClassLoaderDataOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) {
502 _boundary = _g->reserved().end();
503 }
504
505 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
506 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
507
508 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
509 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
510
511 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
512 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
513
514 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
515 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
516
517 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
518 ParScanThreadState* par_scan_state)
519 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
520 {}
521
522 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
523 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
524
525 #ifdef WIN32
526 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
527 #endif
528
529 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
530 ParScanThreadState* par_scan_state_,
531 ParScanWithoutBarrierClosure* to_space_closure_,
532 ParScanWithBarrierClosure* old_gen_closure_,
533 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
534 ParNewGeneration* par_gen_,
535 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
536 ObjToScanQueueSet* task_queues_,
537 ParallelTaskTerminator* terminator_) :
538
539 _par_scan_state(par_scan_state_),
540 _to_space_closure(to_space_closure_),
541 _old_gen_closure(old_gen_closure_),
542 _to_space_root_closure(to_space_root_closure_),
543 _old_gen_root_closure(old_gen_root_closure_),
544 _par_gen(par_gen_),
545 _task_queues(task_queues_),
546 _terminator(terminator_)
547 {}
548
549 void ParEvacuateFollowersClosure::do_void() {
550 ObjToScanQueue* work_q = par_scan_state()->work_queue();
551
552 while (true) {
553 // Scan to-space and old-gen objs until we run out of both.
554 oop obj_to_scan;
555 par_scan_state()->trim_queues(0);
556
557 // We have no local work, attempt to steal from other threads.
558
559 // Attempt to steal work from promoted.
560 if (task_queues()->steal(par_scan_state()->thread_num(),
561 par_scan_state()->hash_seed(),
562 obj_to_scan)) {
563 bool res = work_q->push(obj_to_scan);
564 assert(res, "Empty queue should have room for a push.");
565
566 // If successful, goto Start.
567 continue;
568
569 // Try global overflow list.
570 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
571 continue;
572 }
573
574 // Otherwise, offer termination.
575 par_scan_state()->start_term_time();
576 if (terminator()->offer_termination()) break;
577 par_scan_state()->end_term_time();
578 }
579 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
580 "Broken overflow list?");
581 // Finish the last termination pause.
582 par_scan_state()->end_term_time();
583 }
584
585 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen,
586 Generation* old_gen,
587 HeapWord* young_old_boundary,
588 ParScanThreadStateSet* state_set,
589 StrongRootsScope* strong_roots_scope) :
590 AbstractGangTask("ParNewGeneration collection"),
591 _young_gen(young_gen), _old_gen(old_gen),
592 _young_old_boundary(young_old_boundary),
593 _state_set(state_set),
594 _strong_roots_scope(strong_roots_scope),
595 _par_state_string(StringTable::weak_storage())
596 {}
597
598 void ParNewGenTask::work(uint worker_id) {
599 CMSHeap* heap = CMSHeap::heap();
600 // Since this is being done in a separate thread, need new resource
601 // and handle marks.
602 ResourceMark rm;
603 HandleMark hm;
604
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 CLDScanClosure cld_scan_closure(&par_scan_state.to_space_root_closure(),
611 heap->rem_set()->cld_rem_set()->accumulate_modified_oops());
612
613 par_scan_state.start_strong_roots();
614 heap->young_process_roots(_strong_roots_scope,
615 &par_scan_state.to_space_root_closure(),
616 &par_scan_state.older_gen_closure(),
617 &cld_scan_closure,
618 &_par_state_string);
619
620 par_scan_state.end_strong_roots();
621
622 // "evacuate followers".
623 par_scan_state.evacuate_followers_closure().do_void();
624
625 // This will collapse this worker's promoted object list that's
626 // created during the main ParNew parallel phase of ParNew. This has
627 // to be called after all workers have finished promoting objects
628 // and scanning promoted objects. It should be safe calling it from
629 // here, given that we can only reach here after all thread have
630 // offered termination, i.e., after there is no more work to be
631 // done. It will also disable promotion tracking for the rest of
632 // this GC as it's not necessary to be on during reference processing.
633 _old_gen->par_oop_since_save_marks_iterate_done((int) worker_id);
634 }
635
636 ParNewGeneration::ParNewGeneration(ReservedSpace rs, size_t initial_byte_size)
637 : DefNewGeneration(rs, initial_byte_size, "PCopy"),
638 _overflow_list(NULL),
639 _is_alive_closure(this),
640 _plab_stats("Young", YoungPLABSize, PLABWeight)
641 {
642 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
643 NOT_PRODUCT(_num_par_pushes = 0;)
644 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
645 guarantee(_task_queues != NULL, "task_queues allocation failure.");
646
647 for (uint i = 0; i < ParallelGCThreads; i++) {
648 ObjToScanQueue *q = new ObjToScanQueue();
649 guarantee(q != NULL, "work_queue Allocation failure.");
650 _task_queues->register_queue(i, q);
651 }
652
653 for (uint i = 0; i < ParallelGCThreads; i++) {
654 _task_queues->queue(i)->initialize();
655 }
656
657 _overflow_stacks = NULL;
658 if (ParGCUseLocalOverflow) {
659 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal with ','
660 typedef Stack<oop, mtGC> GCOopStack;
661
662 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
663 for (size_t i = 0; i < ParallelGCThreads; ++i) {
664 new (_overflow_stacks + i) Stack<oop, mtGC>();
665 }
666 }
667
668 if (UsePerfData) {
669 EXCEPTION_MARK;
670 ResourceMark rm;
671
672 const char* cname =
673 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
674 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
675 ParallelGCThreads, CHECK);
676 }
677 }
678
679 // ParNewGeneration::
680 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
681 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
682
683 template <class T>
684 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
685 #ifdef ASSERT
686 {
687 oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);
688 // We never expect to see a null reference being processed
689 // as a weak reference.
690 assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs");
691 }
692 #endif // ASSERT
693
694 _par_cl->do_oop_nv(p);
695
696 if (CMSHeap::heap()->is_in_reserved(p)) {
697 oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);;
698 _rs->write_ref_field_gc_par(p, obj);
699 }
700 }
701
702 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
703 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
704
705 // ParNewGeneration::
706 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
707 DefNewGeneration::KeepAliveClosure(cl) {}
708
709 template <class T>
710 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
711 #ifdef ASSERT
712 {
713 oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);
714 // We never expect to see a null reference being processed
715 // as a weak reference.
716 assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs");
717 }
718 #endif // ASSERT
719
720 _cl->do_oop_nv(p);
721
722 if (CMSHeap::heap()->is_in_reserved(p)) {
723 oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);
724 _rs->write_ref_field_gc_par(p, obj);
725 }
726 }
727
728 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
729 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
730
731 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
732 T heap_oop = RawAccess<>::oop_load(p);
733 if (!CompressedOops::is_null(heap_oop)) {
734 oop obj = CompressedOops::decode_not_null(heap_oop);
735 if ((HeapWord*)obj < _boundary) {
736 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
737 oop new_obj = obj->is_forwarded()
738 ? obj->forwardee()
739 : _g->DefNewGeneration::copy_to_survivor_space(obj);
740 RawAccess<OOP_NOT_NULL>::oop_store(p, new_obj);
741 }
742 if (_gc_barrier) {
743 // If p points to a younger generation, mark the card.
744 if ((HeapWord*)obj < _gen_boundary) {
745 _rs->write_ref_field_gc_par(p, obj);
746 }
747 }
748 }
749 }
750
751 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
752 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
753
754 class ParNewRefProcTaskProxy: public AbstractGangTask {
755 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
756 public:
757 ParNewRefProcTaskProxy(ProcessTask& task,
758 ParNewGeneration& young_gen,
759 Generation& old_gen,
760 HeapWord* young_old_boundary,
761 ParScanThreadStateSet& state_set);
762
763 private:
764 virtual void work(uint worker_id);
765 private:
766 ParNewGeneration& _young_gen;
767 ProcessTask& _task;
768 Generation& _old_gen;
769 HeapWord* _young_old_boundary;
770 ParScanThreadStateSet& _state_set;
771 };
772
773 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
774 ParNewGeneration& young_gen,
775 Generation& old_gen,
776 HeapWord* young_old_boundary,
777 ParScanThreadStateSet& state_set)
778 : AbstractGangTask("ParNewGeneration parallel reference processing"),
779 _young_gen(young_gen),
780 _task(task),
781 _old_gen(old_gen),
782 _young_old_boundary(young_old_boundary),
783 _state_set(state_set)
784 { }
785
786 void ParNewRefProcTaskProxy::work(uint worker_id) {
787 ResourceMark rm;
788 HandleMark hm;
789 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
790 par_scan_state.set_young_old_boundary(_young_old_boundary);
791 _task.work(worker_id, par_scan_state.is_alive_closure(),
792 par_scan_state.keep_alive_closure(),
793 par_scan_state.evacuate_followers_closure());
794 }
795
796 void ParNewRefProcTaskExecutor::execute(ProcessTask& task, uint ergo_workers) {
797 CMSHeap* gch = CMSHeap::heap();
798 WorkGang* workers = gch->workers();
799 assert(workers != NULL, "Need parallel worker threads.");
800 assert(workers->active_workers() == ergo_workers,
801 "Ergonomically chosen workers (%u) must be equal to active workers (%u)",
802 ergo_workers, workers->active_workers());
803 _state_set.reset(workers->active_workers(), _young_gen.promotion_failed());
804 ParNewRefProcTaskProxy rp_task(task, _young_gen, _old_gen,
805 _young_gen.reserved().end(), _state_set);
806 workers->run_task(&rp_task, workers->active_workers());
807 _state_set.reset(0 /* bad value in debug if not reset */,
808 _young_gen.promotion_failed());
809 }
810
811 void ParNewRefProcTaskExecutor::set_single_threaded_mode() {
812 _state_set.flush();
813 CMSHeap* heap = CMSHeap::heap();
814 heap->save_marks();
815 }
816
817 ScanClosureWithParBarrier::
818 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
819 OopsInClassLoaderDataOrGenClosure(g), _g(g), _boundary(g->reserved().end()), _gc_barrier(gc_barrier)
820 { }
821
822 template <typename OopClosureType1, typename OopClosureType2>
823 EvacuateFollowersClosureGeneral<OopClosureType1, OopClosureType2>::
824 EvacuateFollowersClosureGeneral(CMSHeap* heap,
825 OopClosureType1* cur,
826 OopClosureType2* older) :
827 _heap(heap),
828 _scan_cur_or_nonheap(cur), _scan_older(older)
829 { }
830
831 template <typename OopClosureType1, typename OopClosureType2>
832 void EvacuateFollowersClosureGeneral<OopClosureType1, OopClosureType2>::do_void() {
833 do {
834 _heap->oop_since_save_marks_iterate(_scan_cur_or_nonheap,
835 _scan_older);
836 } while (!_heap->no_allocs_since_save_marks());
837 }
838
839 // A Generation that does parallel young-gen collection.
840
841 void ParNewGeneration::handle_promotion_failed(CMSHeap* gch, ParScanThreadStateSet& thread_state_set) {
842 assert(_promo_failure_scan_stack.is_empty(), "post condition");
843 _promo_failure_scan_stack.clear(true); // Clear cached segments.
844
845 remove_forwarding_pointers();
846 log_info(gc, promotion)("Promotion failed");
847 // All the spaces are in play for mark-sweep.
848 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
849 from()->set_next_compaction_space(to());
850 gch->set_incremental_collection_failed();
851 // Inform the next generation that a promotion failure occurred.
852 _old_gen->promotion_failure_occurred();
853
854 // Trace promotion failure in the parallel GC threads
855 thread_state_set.trace_promotion_failed(gc_tracer());
856 // Single threaded code may have reported promotion failure to the global state
857 if (_promotion_failed_info.has_failed()) {
858 _gc_tracer.report_promotion_failed(_promotion_failed_info);
859 }
860 // Reset the PromotionFailureALot counters.
861 NOT_PRODUCT(gch->reset_promotion_should_fail();)
862 }
863
864 void ParNewGeneration::collect(bool full,
865 bool clear_all_soft_refs,
866 size_t size,
867 bool is_tlab) {
868 assert(full || size > 0, "otherwise we don't want to collect");
869
870 CMSHeap* gch = CMSHeap::heap();
871
872 _gc_timer->register_gc_start();
873
874 AdaptiveSizePolicy* size_policy = gch->size_policy();
875 WorkGang* workers = gch->workers();
876 assert(workers != NULL, "Need workgang for parallel work");
877 uint active_workers =
878 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
879 workers->active_workers(),
880 Threads::number_of_non_daemon_threads());
881 active_workers = workers->update_active_workers(active_workers);
882 log_info(gc,task)("Using %u workers of %u for evacuation", active_workers, workers->total_workers());
883
884 _old_gen = gch->old_gen();
885
886 // If the next generation is too full to accommodate worst-case promotion
887 // from this generation, pass on collection; let the next generation
888 // do it.
889 if (!collection_attempt_is_safe()) {
890 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
891 return;
892 }
893 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
894
895 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
896 gch->trace_heap_before_gc(gc_tracer());
897
898 init_assuming_no_promotion_failure();
899
900 if (UseAdaptiveSizePolicy) {
901 set_survivor_overflow(false);
902 size_policy->minor_collection_begin();
903 }
904
905 GCTraceTime(Trace, gc, phases) t1("ParNew", NULL, gch->gc_cause());
906
907 age_table()->clear();
908 to()->clear(SpaceDecorator::Mangle);
909
910 gch->save_marks();
911
912 // Set the correct parallelism (number of queues) in the reference processor
913 ref_processor()->set_active_mt_degree(active_workers);
914
915 // Need to initialize the preserved marks before the ThreadStateSet c'tor.
916 _preserved_marks_set.init(active_workers);
917
918 // Always set the terminator for the active number of workers
919 // because only those workers go through the termination protocol.
920 ParallelTaskTerminator _term(active_workers, task_queues());
921 ParScanThreadStateSet thread_state_set(active_workers,
922 *to(), *this, *_old_gen, *task_queues(),
923 _overflow_stacks, _preserved_marks_set,
924 desired_plab_sz(), _term);
925
926 thread_state_set.reset(active_workers, promotion_failed());
927
928 {
929 StrongRootsScope srs(active_workers);
930
931 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs);
932 gch->rem_set()->prepare_for_younger_refs_iterate(true);
933 // It turns out that even when we're using 1 thread, doing the work in a
934 // separate thread causes wide variance in run times. We can't help this
935 // in the multi-threaded case, but we special-case n=1 here to get
936 // repeatable measurements of the 1-thread overhead of the parallel code.
937 // Might multiple workers ever be used? If yes, initialization
938 // has been done such that the single threaded path should not be used.
939 if (workers->total_workers() > 1) {
940 workers->run_task(&tsk);
941 } else {
942 tsk.work(0);
943 }
944 }
945
946 thread_state_set.reset(0 /* Bad value in debug if not reset */,
947 promotion_failed());
948
949 // Trace and reset failed promotion info.
950 if (promotion_failed()) {
951 thread_state_set.trace_promotion_failed(gc_tracer());
952 }
953
954 // Process (weak) reference objects found during scavenge.
955 ReferenceProcessor* rp = ref_processor();
956 IsAliveClosure is_alive(this);
957 ScanWeakRefClosure scan_weak_ref(this);
958 KeepAliveClosure keep_alive(&scan_weak_ref);
959 ScanClosure scan_without_gc_barrier(this, false);
960 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
961 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
962 EvacuateFollowersClosureGeneral<ScanClosure, ScanClosureWithParBarrier> evacuate_followers(
963 gch, &scan_without_gc_barrier, &scan_with_gc_barrier);
964 rp->setup_policy(clear_all_soft_refs);
965 // Can the mt_degree be set later (at run_task() time would be best)?
966 rp->set_active_mt_degree(active_workers);
967 ReferenceProcessorStats stats;
968 ReferenceProcessorPhaseTimes pt(_gc_timer, rp->max_num_queues());
969 if (rp->processing_is_mt()) {
970 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set);
971 stats = rp->process_discovered_references(&is_alive, &keep_alive,
972 &evacuate_followers, &task_executor,
973 &pt);
974 } else {
975 thread_state_set.flush();
976 gch->save_marks();
977 stats = rp->process_discovered_references(&is_alive, &keep_alive,
978 &evacuate_followers, NULL,
979 &pt);
980 }
981 _gc_tracer.report_gc_reference_stats(stats);
982 _gc_tracer.report_tenuring_threshold(tenuring_threshold());
983 pt.print_all_references();
984
985 assert(gch->no_allocs_since_save_marks(), "evacuation should be done at this point");
986
987 WeakProcessor::weak_oops_do(&is_alive, &keep_alive);
988
989 // Verify that the usage of keep_alive only forwarded
990 // the oops and did not find anything new to copy.
991 assert(gch->no_allocs_since_save_marks(), "unexpectedly copied objects");
992
993 if (!promotion_failed()) {
994 // Swap the survivor spaces.
995 eden()->clear(SpaceDecorator::Mangle);
996 from()->clear(SpaceDecorator::Mangle);
997 if (ZapUnusedHeapArea) {
998 // This is now done here because of the piece-meal mangling which
999 // can check for valid mangling at intermediate points in the
1000 // collection(s). When a young collection fails to collect
1001 // sufficient space resizing of the young generation can occur
1002 // and redistribute the spaces in the young generation. Mangle
1003 // here so that unzapped regions don't get distributed to
1004 // other spaces.
1005 to()->mangle_unused_area();
1006 }
1007 swap_spaces();
1008
1009 // A successful scavenge should restart the GC time limit count which is
1010 // for full GC's.
1011 size_policy->reset_gc_overhead_limit_count();
1012
1013 assert(to()->is_empty(), "to space should be empty now");
1014
1015 adjust_desired_tenuring_threshold();
1016 } else {
1017 handle_promotion_failed(gch, thread_state_set);
1018 }
1019 _preserved_marks_set.reclaim();
1020 // set new iteration safe limit for the survivor spaces
1021 from()->set_concurrent_iteration_safe_limit(from()->top());
1022 to()->set_concurrent_iteration_safe_limit(to()->top());
1023
1024 plab_stats()->adjust_desired_plab_sz();
1025
1026 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
1027 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
1028
1029 if (UseAdaptiveSizePolicy) {
1030 size_policy->minor_collection_end(gch->gc_cause());
1031 size_policy->avg_survived()->sample(from()->used());
1032 }
1033
1034 // We need to use a monotonically non-decreasing time in ms
1035 // or we will see time-warp warnings and os::javaTimeMillis()
1036 // does not guarantee monotonicity.
1037 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1038 update_time_of_last_gc(now);
1039
1040 rp->set_enqueuing_is_done(true);
1041 rp->verify_no_references_recorded();
1042
1043 gch->trace_heap_after_gc(gc_tracer());
1044
1045 _gc_timer->register_gc_end();
1046
1047 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1048 }
1049
1050 size_t ParNewGeneration::desired_plab_sz() {
1051 return _plab_stats.desired_plab_sz(CMSHeap::heap()->workers()->active_workers());
1052 }
1053
1054 static int sum;
1055 void ParNewGeneration::waste_some_time() {
1056 for (int i = 0; i < 100; i++) {
1057 sum += i;
1058 }
1059 }
1060
1061 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1062
1063 // Because of concurrency, there are times where an object for which
1064 // "is_forwarded()" is true contains an "interim" forwarding pointer
1065 // value. Such a value will soon be overwritten with a real value.
1066 // This method requires "obj" to have a forwarding pointer, and waits, if
1067 // necessary for a real one to be inserted, and returns it.
1068
1069 oop ParNewGeneration::real_forwardee(oop obj) {
1070 oop forward_ptr = obj->forwardee();
1071 if (forward_ptr != ClaimedForwardPtr) {
1072 return forward_ptr;
1073 } else {
1074 return real_forwardee_slow(obj);
1075 }
1076 }
1077
1078 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1079 // Spin-read if it is claimed but not yet written by another thread.
1080 oop forward_ptr = obj->forwardee();
1081 while (forward_ptr == ClaimedForwardPtr) {
1082 waste_some_time();
1083 assert(obj->is_forwarded(), "precondition");
1084 forward_ptr = obj->forwardee();
1085 }
1086 return forward_ptr;
1087 }
1088
1089 // Multiple GC threads may try to promote an object. If the object
1090 // is successfully promoted, a forwarding pointer will be installed in
1091 // the object in the young generation. This method claims the right
1092 // to install the forwarding pointer before it copies the object,
1093 // thus avoiding the need to undo the copy as in
1094 // copy_to_survivor_space_avoiding_with_undo.
1095
1096 oop ParNewGeneration::copy_to_survivor_space(ParScanThreadState* par_scan_state,
1097 oop old,
1098 size_t sz,
1099 markOop m) {
1100 // In the sequential version, this assert also says that the object is
1101 // not forwarded. That might not be the case here. It is the case that
1102 // the caller observed it to be not forwarded at some time in the past.
1103 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1104
1105 // The sequential code read "old->age()" below. That doesn't work here,
1106 // since the age is in the mark word, and that might be overwritten with
1107 // a forwarding pointer by a parallel thread. So we must save the mark
1108 // word in a local and then analyze it.
1109 oopDesc dummyOld;
1110 dummyOld.set_mark_raw(m);
1111 assert(!dummyOld.is_forwarded(),
1112 "should not be called with forwarding pointer mark word.");
1113
1114 oop new_obj = NULL;
1115 oop forward_ptr;
1116
1117 // Try allocating obj in to-space (unless too old)
1118 if (dummyOld.age() < tenuring_threshold()) {
1119 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1120 if (new_obj == NULL) {
1121 set_survivor_overflow(true);
1122 }
1123 }
1124
1125 if (new_obj == NULL) {
1126 // Either to-space is full or we decided to promote try allocating obj tenured
1127
1128 // Attempt to install a null forwarding pointer (atomically),
1129 // to claim the right to install the real forwarding pointer.
1130 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1131 if (forward_ptr != NULL) {
1132 // someone else beat us to it.
1133 return real_forwardee(old);
1134 }
1135
1136 if (!_promotion_failed) {
1137 new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1138 old, m, sz);
1139 }
1140
1141 if (new_obj == NULL) {
1142 // promotion failed, forward to self
1143 _promotion_failed = true;
1144 new_obj = old;
1145
1146 par_scan_state->preserved_marks()->push_if_necessary(old, m);
1147 par_scan_state->register_promotion_failure(sz);
1148 }
1149
1150 old->forward_to(new_obj);
1151 forward_ptr = NULL;
1152 } else {
1153 // Is in to-space; do copying ourselves.
1154 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1155 assert(CMSHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1156 forward_ptr = old->forward_to_atomic(new_obj);
1157 // Restore the mark word copied above.
1158 new_obj->set_mark_raw(m);
1159 // Increment age if obj still in new generation
1160 new_obj->incr_age();
1161 par_scan_state->age_table()->add(new_obj, sz);
1162 }
1163 assert(new_obj != NULL, "just checking");
1164
1165 // This code must come after the CAS test, or it will print incorrect
1166 // information.
1167 log_develop_trace(gc, scavenge)("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1168 is_in_reserved(new_obj) ? "copying" : "tenuring",
1169 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size());
1170
1171 if (forward_ptr == NULL) {
1172 oop obj_to_push = new_obj;
1173 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1174 // Length field used as index of next element to be scanned.
1175 // Real length can be obtained from real_forwardee()
1176 arrayOop(old)->set_length(0);
1177 obj_to_push = old;
1178 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1179 "push forwarded object");
1180 }
1181 // Push it on one of the queues of to-be-scanned objects.
1182 bool simulate_overflow = false;
1183 NOT_PRODUCT(
1184 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1185 // simulate a stack overflow
1186 simulate_overflow = true;
1187 }
1188 )
1189 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1190 // Add stats for overflow pushes.
1191 log_develop_trace(gc)("Queue Overflow");
1192 push_on_overflow_list(old, par_scan_state);
1193 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1194 }
1195
1196 return new_obj;
1197 }
1198
1199 // Oops. Someone beat us to it. Undo the allocation. Where did we
1200 // allocate it?
1201 if (is_in_reserved(new_obj)) {
1202 // Must be in to_space.
1203 assert(to()->is_in_reserved(new_obj), "Checking");
1204 if (forward_ptr == ClaimedForwardPtr) {
1205 // Wait to get the real forwarding pointer value.
1206 forward_ptr = real_forwardee(old);
1207 }
1208 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1209 }
1210
1211 return forward_ptr;
1212 }
1213
1214 #ifndef PRODUCT
1215 // It's OK to call this multi-threaded; the worst thing
1216 // that can happen is that we'll get a bunch of closely
1217 // spaced simulated overflows, but that's OK, in fact
1218 // probably good as it would exercise the overflow code
1219 // under contention.
1220 bool ParNewGeneration::should_simulate_overflow() {
1221 if (_overflow_counter-- <= 0) { // just being defensive
1222 _overflow_counter = ParGCWorkQueueOverflowInterval;
1223 return true;
1224 } else {
1225 return false;
1226 }
1227 }
1228 #endif
1229
1230 // In case we are using compressed oops, we need to be careful.
1231 // If the object being pushed is an object array, then its length
1232 // field keeps track of the "grey boundary" at which the next
1233 // incremental scan will be done (see ParGCArrayScanChunk).
1234 // When using compressed oops, this length field is kept in the
1235 // lower 32 bits of the erstwhile klass word and cannot be used
1236 // for the overflow chaining pointer (OCP below). As such the OCP
1237 // would itself need to be compressed into the top 32-bits in this
1238 // case. Unfortunately, see below, in the event that we have a
1239 // promotion failure, the node to be pushed on the list can be
1240 // outside of the Java heap, so the heap-based pointer compression
1241 // would not work (we would have potential aliasing between C-heap
1242 // and Java-heap pointers). For this reason, when using compressed
1243 // oops, we simply use a worker-thread-local, non-shared overflow
1244 // list in the form of a growable array, with a slightly different
1245 // overflow stack draining strategy. If/when we start using fat
1246 // stacks here, we can go back to using (fat) pointer chains
1247 // (although some performance comparisons would be useful since
1248 // single global lists have their own performance disadvantages
1249 // as we were made painfully aware not long ago, see 6786503).
1250 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1251 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1252 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1253 if (ParGCUseLocalOverflow) {
1254 // In the case of compressed oops, we use a private, not-shared
1255 // overflow stack.
1256 par_scan_state->push_on_overflow_stack(from_space_obj);
1257 } else {
1258 assert(!UseCompressedOops, "Error");
1259 // if the object has been forwarded to itself, then we cannot
1260 // use the klass pointer for the linked list. Instead we have
1261 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1262 // XXX This is horribly inefficient when a promotion failure occurs
1263 // and should be fixed. XXX FIX ME !!!
1264 #ifndef PRODUCT
1265 Atomic::inc(&_num_par_pushes);
1266 assert(_num_par_pushes > 0, "Tautology");
1267 #endif
1268 if (from_space_obj->forwardee() == from_space_obj) {
1269 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1270 listhead->forward_to(from_space_obj);
1271 from_space_obj = listhead;
1272 }
1273 oop observed_overflow_list = _overflow_list;
1274 oop cur_overflow_list;
1275 do {
1276 cur_overflow_list = observed_overflow_list;
1277 if (cur_overflow_list != BUSY) {
1278 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1279 } else {
1280 from_space_obj->set_klass_to_list_ptr(NULL);
1281 }
1282 observed_overflow_list =
1283 Atomic::cmpxchg((oopDesc*)from_space_obj, &_overflow_list, (oopDesc*)cur_overflow_list);
1284 } while (cur_overflow_list != observed_overflow_list);
1285 }
1286 }
1287
1288 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1289 bool res;
1290
1291 if (ParGCUseLocalOverflow) {
1292 res = par_scan_state->take_from_overflow_stack();
1293 } else {
1294 assert(!UseCompressedOops, "Error");
1295 res = take_from_overflow_list_work(par_scan_state);
1296 }
1297 return res;
1298 }
1299
1300
1301 // *NOTE*: The overflow list manipulation code here and
1302 // in CMSCollector:: are very similar in shape,
1303 // except that in the CMS case we thread the objects
1304 // directly into the list via their mark word, and do
1305 // not need to deal with special cases below related
1306 // to chunking of object arrays and promotion failure
1307 // handling.
1308 // CR 6797058 has been filed to attempt consolidation of
1309 // the common code.
1310 // Because of the common code, if you make any changes in
1311 // the code below, please check the CMS version to see if
1312 // similar changes might be needed.
1313 // See CMSCollector::par_take_from_overflow_list() for
1314 // more extensive documentation comments.
1315 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1316 ObjToScanQueue* work_q = par_scan_state->work_queue();
1317 // How many to take?
1318 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1319 (size_t)ParGCDesiredObjsFromOverflowList);
1320
1321 assert(!UseCompressedOops, "Error");
1322 assert(par_scan_state->overflow_stack() == NULL, "Error");
1323 if (_overflow_list == NULL) return false;
1324
1325 // Otherwise, there was something there; try claiming the list.
1326 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
1327 // Trim off a prefix of at most objsFromOverflow items
1328 Thread* tid = Thread::current();
1329 size_t spin_count = ParallelGCThreads;
1330 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1331 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1332 // someone grabbed it before we did ...
1333 // ... we spin for a short while...
1334 os::sleep(tid, sleep_time_millis, false);
1335 if (_overflow_list == NULL) {
1336 // nothing left to take
1337 return false;
1338 } else if (_overflow_list != BUSY) {
1339 // try and grab the prefix
1340 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
1341 }
1342 }
1343 if (prefix == NULL || prefix == BUSY) {
1344 // Nothing to take or waited long enough
1345 if (prefix == NULL) {
1346 // Write back the NULL in case we overwrote it with BUSY above
1347 // and it is still the same value.
1348 (void) Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
1349 }
1350 return false;
1351 }
1352 assert(prefix != NULL && prefix != BUSY, "Error");
1353 oop cur = prefix;
1354 for (size_t i = 1; i < objsFromOverflow; ++i) {
1355 oop next = cur->list_ptr_from_klass();
1356 if (next == NULL) break;
1357 cur = next;
1358 }
1359 assert(cur != NULL, "Loop postcondition");
1360
1361 // Reattach remaining (suffix) to overflow list
1362 oop suffix = cur->list_ptr_from_klass();
1363 if (suffix == NULL) {
1364 // Write back the NULL in lieu of the BUSY we wrote
1365 // above and it is still the same value.
1366 if (_overflow_list == BUSY) {
1367 (void) Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
1368 }
1369 } else {
1370 assert(suffix != BUSY, "Error");
1371 // suffix will be put back on global list
1372 cur->set_klass_to_list_ptr(NULL); // break off suffix
1373 // It's possible that the list is still in the empty(busy) state
1374 // we left it in a short while ago; in that case we may be
1375 // able to place back the suffix.
1376 oop observed_overflow_list = _overflow_list;
1377 oop cur_overflow_list = observed_overflow_list;
1378 bool attached = false;
1379 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1380 observed_overflow_list =
1381 Atomic::cmpxchg((oopDesc*)suffix, &_overflow_list, (oopDesc*)cur_overflow_list);
1382 if (cur_overflow_list == observed_overflow_list) {
1383 attached = true;
1384 break;
1385 } else cur_overflow_list = observed_overflow_list;
1386 }
1387 if (!attached) {
1388 // Too bad, someone else got in in between; we'll need to do a splice.
1389 // Find the last item of suffix list
1390 oop last = suffix;
1391 while (true) {
1392 oop next = last->list_ptr_from_klass();
1393 if (next == NULL) break;
1394 last = next;
1395 }
1396 // Atomically prepend suffix to current overflow list
1397 observed_overflow_list = _overflow_list;
1398 do {
1399 cur_overflow_list = observed_overflow_list;
1400 if (cur_overflow_list != BUSY) {
1401 // Do the splice ...
1402 last->set_klass_to_list_ptr(cur_overflow_list);
1403 } else { // cur_overflow_list == BUSY
1404 last->set_klass_to_list_ptr(NULL);
1405 }
1406 observed_overflow_list =
1407 Atomic::cmpxchg((oopDesc*)suffix, &_overflow_list, (oopDesc*)cur_overflow_list);
1408 } while (cur_overflow_list != observed_overflow_list);
1409 }
1410 }
1411
1412 // Push objects on prefix list onto this thread's work queue
1413 assert(prefix != NULL && prefix != BUSY, "program logic");
1414 cur = prefix;
1415 ssize_t n = 0;
1416 while (cur != NULL) {
1417 oop obj_to_push = cur->forwardee();
1418 oop next = cur->list_ptr_from_klass();
1419 cur->set_klass(obj_to_push->klass());
1420 // This may be an array object that is self-forwarded. In that case, the list pointer
1421 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1422 if (!is_in_reserved(cur)) {
1423 // This can become a scaling bottleneck when there is work queue overflow coincident
1424 // with promotion failure.
1425 oopDesc* f = cur;
1426 FREE_C_HEAP_ARRAY(oopDesc, f);
1427 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1428 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1429 obj_to_push = cur;
1430 }
1431 bool ok = work_q->push(obj_to_push);
1432 assert(ok, "Should have succeeded");
1433 cur = next;
1434 n++;
1435 }
1436 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1437 #ifndef PRODUCT
1438 assert(_num_par_pushes >= n, "Too many pops?");
1439 Atomic::sub(n, &_num_par_pushes);
1440 #endif
1441 return true;
1442 }
1443 #undef BUSY
1444
1445 void ParNewGeneration::ref_processor_init() {
1446 if (_ref_processor == NULL) {
1447 // Allocate and initialize a reference processor
1448 _span_based_discoverer.set_span(_reserved);
1449 _ref_processor =
1450 new ReferenceProcessor(&_span_based_discoverer, // span
1451 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1452 ParallelGCThreads, // mt processing degree
1453 refs_discovery_is_mt(), // mt discovery
1454 ParallelGCThreads, // mt discovery degree
1455 refs_discovery_is_atomic(), // atomic_discovery
1456 NULL, // is_alive_non_header
1457 false); // disable adjusting number of processing threads
1458 }
1459 }
1460
1461 const char* ParNewGeneration::name() const {
1462 return "par new generation";
1463 }
1464
1465 void ParNewGeneration::restore_preserved_marks() {
1466 SharedRestorePreservedMarksTaskExecutor task_executor(CMSHeap::heap()->workers());
1467 _preserved_marks_set.restore(&task_executor);
1468 }