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 "gc/cms/cmsHeap.inline.hpp"
27 #include "gc/cms/compactibleFreeListSpace.hpp"
28 #include "gc/cms/concurrentMarkSweepGeneration.hpp"
29 #include "gc/cms/parNewGeneration.inline.hpp"
30 #include "gc/cms/parOopClosures.inline.hpp"
31 #include "gc/serial/defNewGeneration.inline.hpp"
32 #include "gc/shared/adaptiveSizePolicy.hpp"
33 #include "gc/shared/ageTable.inline.hpp"
34 #include "gc/shared/copyFailedInfo.hpp"
35 #include "gc/shared/gcHeapSummary.hpp"
36 #include "gc/shared/gcTimer.hpp"
37 #include "gc/shared/gcTrace.hpp"
38 #include "gc/shared/gcTraceTime.inline.hpp"
39 #include "gc/shared/genOopClosures.inline.hpp"
40 #include "gc/shared/generation.hpp"
41 #include "gc/shared/plab.inline.hpp"
42 #include "gc/shared/preservedMarks.inline.hpp"
43 #include "gc/shared/referencePolicy.hpp"
44 #include "gc/shared/space.hpp"
45 #include "gc/shared/spaceDecorator.hpp"
46 #include "gc/shared/strongRootsScope.hpp"
47 #include "gc/shared/taskqueue.inline.hpp"
48 #include "gc/shared/weakProcessor.hpp"
49 #include "gc/shared/workgroup.hpp"
50 #include "logging/log.hpp"
51 #include "logging/logStream.hpp"
52 #include "memory/resourceArea.hpp"
53 #include "oops/access.inline.hpp"
54 #include "oops/compressedOops.inline.hpp"
55 #include "oops/objArrayOop.hpp"
56 #include "oops/oop.inline.hpp"
57 #include "runtime/atomic.hpp"
58 #include "runtime/handles.hpp"
59 #include "runtime/handles.inline.hpp"
60 #include "runtime/java.hpp"
61 #include "runtime/thread.inline.hpp"
62 #include "utilities/copy.hpp"
63 #include "utilities/globalDefinitions.hpp"
64 #include "utilities/stack.inline.hpp"
65
66 ParScanThreadState::ParScanThreadState(Space* to_space_,
67 ParNewGeneration* young_gen_,
68 Generation* old_gen_,
69 int thread_num_,
70 ObjToScanQueueSet* work_queue_set_,
71 Stack<oop, mtGC>* overflow_stacks_,
72 PreservedMarks* preserved_marks_,
73 size_t desired_plab_sz_,
74 ParallelTaskTerminator& term_) :
75 _to_space(to_space_),
76 _old_gen(old_gen_),
77 _young_gen(young_gen_),
78 _thread_num(thread_num_),
79 _work_queue(work_queue_set_->queue(thread_num_)),
80 _to_space_full(false),
81 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
82 _preserved_marks(preserved_marks_),
83 _ageTable(false), // false ==> not the global age table, no perf data.
84 _to_space_alloc_buffer(desired_plab_sz_),
85 _to_space_closure(young_gen_, this),
86 _old_gen_closure(young_gen_, this),
87 _to_space_root_closure(young_gen_, this),
88 _old_gen_root_closure(young_gen_, this),
89 _older_gen_closure(young_gen_, this),
90 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
91 &_to_space_root_closure, young_gen_, &_old_gen_root_closure,
92 work_queue_set_, &term_),
93 _is_alive_closure(young_gen_),
94 _scan_weak_ref_closure(young_gen_, this),
95 _keep_alive_closure(&_scan_weak_ref_closure),
96 _strong_roots_time(0.0),
97 _term_time(0.0)
98 {
99 #if TASKQUEUE_STATS
100 _term_attempts = 0;
101 _overflow_refills = 0;
102 _overflow_refill_objs = 0;
103 #endif // TASKQUEUE_STATS
104
105 _survivor_chunk_array = (ChunkArray*) old_gen()->get_data_recorder(thread_num());
106 _hash_seed = 17; // Might want to take time-based random value.
107 _start = os::elapsedTime();
108 _old_gen_closure.set_generation(old_gen_);
109 _old_gen_root_closure.set_generation(old_gen_);
110 }
111
112 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
113 size_t plab_word_size) {
114 ChunkArray* sca = survivor_chunk_array();
115 if (sca != NULL) {
116 // A non-null SCA implies that we want the PLAB data recorded.
117 sca->record_sample(plab_start, plab_word_size);
118 }
119 }
120
121 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
122 return new_obj->is_objArray() &&
123 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
124 new_obj != old_obj;
125 }
126
127 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
128 assert(old->is_objArray(), "must be obj array");
129 assert(old->is_forwarded(), "must be forwarded");
130 assert(CMSHeap::heap()->is_in_reserved(old), "must be in heap.");
131 assert(!old_gen()->is_in(old), "must be in young generation.");
132
133 objArrayOop obj = objArrayOop(old->forwardee());
134 // Process ParGCArrayScanChunk elements now
135 // and push the remainder back onto queue
136 int start = arrayOop(old)->length();
137 int end = obj->length();
138 int remainder = end - start;
139 assert(start <= end, "just checking");
140 if (remainder > 2 * ParGCArrayScanChunk) {
141 // Test above combines last partial chunk with a full chunk
142 end = start + ParGCArrayScanChunk;
143 arrayOop(old)->set_length(end);
144 // Push remainder.
145 bool ok = work_queue()->push(old);
146 assert(ok, "just popped, push must be okay");
147 } else {
148 // Restore length so that it can be used if there
149 // is a promotion failure and forwarding pointers
150 // must be removed.
151 arrayOop(old)->set_length(end);
152 }
153
154 // process our set of indices (include header in first chunk)
155 // should make sure end is even (aligned to HeapWord in case of compressed oops)
156 if ((HeapWord *)obj < young_old_boundary()) {
157 // object is in to_space
158 obj->oop_iterate_range(&_to_space_closure, start, end);
159 } else {
160 // object is in old generation
161 obj->oop_iterate_range(&_old_gen_closure, start, end);
162 }
163 }
164
165 void ParScanThreadState::trim_queues(int max_size) {
166 ObjToScanQueue* queue = work_queue();
167 do {
168 while (queue->size() > (juint)max_size) {
169 oop obj_to_scan;
170 if (queue->pop_local(obj_to_scan)) {
171 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
172 if (obj_to_scan->is_objArray() &&
173 obj_to_scan->is_forwarded() &&
174 obj_to_scan->forwardee() != obj_to_scan) {
175 scan_partial_array_and_push_remainder(obj_to_scan);
176 } else {
177 // object is in to_space
178 obj_to_scan->oop_iterate(&_to_space_closure);
179 }
180 } else {
181 // object is in old generation
182 obj_to_scan->oop_iterate(&_old_gen_closure);
183 }
184 }
185 }
186 // For the case of compressed oops, we have a private, non-shared
187 // overflow stack, so we eagerly drain it so as to more evenly
188 // distribute load early. Note: this may be good to do in
189 // general rather than delay for the final stealing phase.
190 // If applicable, we'll transfer a set of objects over to our
191 // work queue, allowing them to be stolen and draining our
192 // private overflow stack.
193 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
194 }
195
196 bool ParScanThreadState::take_from_overflow_stack() {
197 assert(ParGCUseLocalOverflow, "Else should not call");
198 assert(young_gen()->overflow_list() == NULL, "Error");
199 ObjToScanQueue* queue = work_queue();
200 Stack<oop, mtGC>* const of_stack = overflow_stack();
201 const size_t num_overflow_elems = of_stack->size();
202 const size_t space_available = queue->max_elems() - queue->size();
203 const size_t num_take_elems = MIN3(space_available / 4,
204 ParGCDesiredObjsFromOverflowList,
205 num_overflow_elems);
206 // Transfer the most recent num_take_elems from the overflow
207 // stack to our work queue.
208 for (size_t i = 0; i != num_take_elems; i++) {
209 oop cur = of_stack->pop();
210 oop obj_to_push = cur->forwardee();
211 assert(CMSHeap::heap()->is_in_reserved(cur), "Should be in heap");
212 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
213 assert(CMSHeap::heap()->is_in_reserved(obj_to_push), "Should be in heap");
214 if (should_be_partially_scanned(obj_to_push, cur)) {
215 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
216 obj_to_push = cur;
217 }
218 bool ok = queue->push(obj_to_push);
219 assert(ok, "Should have succeeded");
220 }
221 assert(young_gen()->overflow_list() == NULL, "Error");
222 return num_take_elems > 0; // was something transferred?
223 }
224
225 void ParScanThreadState::push_on_overflow_stack(oop p) {
226 assert(ParGCUseLocalOverflow, "Else should not call");
227 overflow_stack()->push(p);
228 assert(young_gen()->overflow_list() == NULL, "Error");
229 }
230
231 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
232 // If the object is small enough, try to reallocate the buffer.
233 HeapWord* obj = NULL;
234 if (!_to_space_full) {
235 PLAB* const plab = to_space_alloc_buffer();
236 Space* const sp = to_space();
237 if (word_sz * 100 < ParallelGCBufferWastePct * plab->word_sz()) {
238 // Is small enough; abandon this buffer and start a new one.
239 plab->retire();
240 // The minimum size has to be twice SurvivorAlignmentInBytes to
241 // allow for padding used in the alignment of 1 word. A padding
242 // of 1 is too small for a filler word so the padding size will
243 // be increased by SurvivorAlignmentInBytes.
244 size_t min_usable_size = 2 * static_cast<size_t>(SurvivorAlignmentInBytes >> LogHeapWordSize);
245 size_t buf_size = MAX2(plab->word_sz(), min_usable_size);
246 HeapWord* buf_space = sp->par_allocate(buf_size);
247 if (buf_space == NULL) {
248 const size_t min_bytes = MAX2(PLAB::min_size(), min_usable_size) << LogHeapWordSize;
249 size_t free_bytes = sp->free();
250 while(buf_space == NULL && free_bytes >= min_bytes) {
251 buf_size = free_bytes >> LogHeapWordSize;
252 assert(buf_size == (size_t)align_object_size(buf_size), "Invariant");
253 buf_space = sp->par_allocate(buf_size);
254 free_bytes = sp->free();
255 }
256 }
257 if (buf_space != NULL) {
258 plab->set_buf(buf_space, buf_size);
259 record_survivor_plab(buf_space, buf_size);
260 obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
261 // Note that we cannot compare buf_size < word_sz below
262 // because of AlignmentReserve (see PLAB::allocate()).
263 assert(obj != NULL || plab->words_remaining() < word_sz,
264 "Else should have been able to allocate requested object size "
265 SIZE_FORMAT ", PLAB size " SIZE_FORMAT ", SurvivorAlignmentInBytes "
266 SIZE_FORMAT ", words_remaining " SIZE_FORMAT,
267 word_sz, buf_size, SurvivorAlignmentInBytes, plab->words_remaining());
268 // It's conceivable that we may be able to use the
269 // buffer we just grabbed for subsequent small requests
270 // even if not for this one.
271 } else {
272 // We're used up.
273 _to_space_full = true;
274 }
275 } else {
276 // Too large; allocate the object individually.
277 obj = sp->par_allocate(word_sz);
278 }
279 }
280 return obj;
281 }
282
283 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) {
284 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
285 }
286
287 void ParScanThreadState::print_promotion_failure_size() {
288 if (_promotion_failed_info.has_failed()) {
289 log_trace(gc, promotion)(" (%d: promotion failure size = " SIZE_FORMAT ") ",
290 _thread_num, _promotion_failed_info.first_size());
291 }
292 }
293
294 class ParScanThreadStateSet: StackObj {
295 public:
296 // Initializes states for the specified number of threads;
297 ParScanThreadStateSet(int num_threads,
298 Space& to_space,
299 ParNewGeneration& young_gen,
300 Generation& old_gen,
301 ObjToScanQueueSet& queue_set,
302 Stack<oop, mtGC>* overflow_stacks_,
303 PreservedMarksSet& preserved_marks_set,
304 size_t desired_plab_sz,
305 ParallelTaskTerminator& term);
306
307 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
308
309 inline ParScanThreadState& thread_state(int i);
310
311 void trace_promotion_failed(const YoungGCTracer* gc_tracer);
312 void reset(uint active_workers, bool promotion_failed);
313 void flush();
314
315 #if TASKQUEUE_STATS
316 static void
317 print_termination_stats_hdr(outputStream* const st);
318 void print_termination_stats();
319 static void
320 print_taskqueue_stats_hdr(outputStream* const st);
321 void print_taskqueue_stats();
322 void reset_stats();
323 #endif // TASKQUEUE_STATS
324
325 private:
326 ParallelTaskTerminator& _term;
327 ParNewGeneration& _young_gen;
328 Generation& _old_gen;
329 ParScanThreadState* _per_thread_states;
330 const int _num_threads;
331 public:
332 bool is_valid(int id) const { return id < _num_threads; }
333 ParallelTaskTerminator* terminator() { return &_term; }
334 };
335
336 ParScanThreadStateSet::ParScanThreadStateSet(int num_threads,
337 Space& to_space,
338 ParNewGeneration& young_gen,
339 Generation& old_gen,
340 ObjToScanQueueSet& queue_set,
341 Stack<oop, mtGC>* overflow_stacks,
342 PreservedMarksSet& preserved_marks_set,
343 size_t desired_plab_sz,
344 ParallelTaskTerminator& term)
345 : _young_gen(young_gen),
346 _old_gen(old_gen),
347 _term(term),
348 _per_thread_states(NEW_RESOURCE_ARRAY(ParScanThreadState, num_threads)),
349 _num_threads(num_threads)
350 {
351 assert(num_threads > 0, "sanity check!");
352 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
353 "overflow_stack allocation mismatch");
354 // Initialize states.
355 for (int i = 0; i < num_threads; ++i) {
356 new(_per_thread_states + i)
357 ParScanThreadState(&to_space, &young_gen, &old_gen, i, &queue_set,
358 overflow_stacks, preserved_marks_set.get(i),
359 desired_plab_sz, term);
360 }
361 }
362
363 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) {
364 assert(i >= 0 && i < _num_threads, "sanity check!");
365 return _per_thread_states[i];
366 }
367
368 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) {
369 for (int i = 0; i < _num_threads; ++i) {
370 if (thread_state(i).promotion_failed()) {
371 gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info());
372 thread_state(i).promotion_failed_info().reset();
373 }
374 }
375 }
376
377 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed) {
378 _term.reset_for_reuse(active_threads);
379 if (promotion_failed) {
380 for (int i = 0; i < _num_threads; ++i) {
381 thread_state(i).print_promotion_failure_size();
382 }
383 }
384 }
385
386 #if TASKQUEUE_STATS
387 void ParScanThreadState::reset_stats() {
388 taskqueue_stats().reset();
389 _term_attempts = 0;
390 _overflow_refills = 0;
391 _overflow_refill_objs = 0;
392 }
393
394 void ParScanThreadStateSet::reset_stats() {
395 for (int i = 0; i < _num_threads; ++i) {
396 thread_state(i).reset_stats();
397 }
398 }
399
400 void ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) {
401 st->print_raw_cr("GC Termination Stats");
402 st->print_raw_cr(" elapsed --strong roots-- -------termination-------");
403 st->print_raw_cr("thr ms ms % ms % attempts");
404 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------");
405 }
406
407 void ParScanThreadStateSet::print_termination_stats() {
408 Log(gc, task, stats) log;
409 if (!log.is_debug()) {
410 return;
411 }
412
413 ResourceMark rm;
414 LogStream ls(log.debug());
415 outputStream* st = &ls;
416
417 print_termination_stats_hdr(st);
418
419 for (int i = 0; i < _num_threads; ++i) {
420 const ParScanThreadState & pss = thread_state(i);
421 const double elapsed_ms = pss.elapsed_time() * 1000.0;
422 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
423 const double term_ms = pss.term_time() * 1000.0;
424 st->print_cr("%3d %9.2f %9.2f %6.2f %9.2f %6.2f " SIZE_FORMAT_W(8),
425 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
426 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
427 }
428 }
429
430 // Print stats related to work queue activity.
431 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) {
432 st->print_raw_cr("GC Task Stats");
433 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
434 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
435 }
436
437 void ParScanThreadStateSet::print_taskqueue_stats() {
438 if (!log_develop_is_enabled(Trace, gc, task, stats)) {
439 return;
440 }
441 Log(gc, task, stats) log;
442 ResourceMark rm;
443 LogStream ls(log.trace());
444 outputStream* st = &ls;
445 print_taskqueue_stats_hdr(st);
446
447 TaskQueueStats totals;
448 for (int i = 0; i < _num_threads; ++i) {
449 const ParScanThreadState & pss = thread_state(i);
450 const TaskQueueStats & stats = pss.taskqueue_stats();
451 st->print("%3d ", i); stats.print(st); st->cr();
452 totals += stats;
453
454 if (pss.overflow_refills() > 0) {
455 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
456 SIZE_FORMAT_W(10) " overflow objects",
457 pss.overflow_refills(), pss.overflow_refill_objs());
458 }
459 }
460 st->print("tot "); totals.print(st); st->cr();
461
462 DEBUG_ONLY(totals.verify());
463 }
464 #endif // TASKQUEUE_STATS
465
466 void ParScanThreadStateSet::flush() {
467 // Work in this loop should be kept as lightweight as
468 // possible since this might otherwise become a bottleneck
469 // to scaling. Should we add heavy-weight work into this
470 // loop, consider parallelizing the loop into the worker threads.
471 for (int i = 0; i < _num_threads; ++i) {
472 ParScanThreadState& par_scan_state = thread_state(i);
473
474 // Flush stats related to To-space PLAB activity and
475 // retire the last buffer.
476 par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_young_gen.plab_stats());
477
478 // Every thread has its own age table. We need to merge
479 // them all into one.
480 AgeTable *local_table = par_scan_state.age_table();
481 _young_gen.age_table()->merge(local_table);
482
483 // Inform old gen that we're done.
484 _old_gen.par_promote_alloc_done(i);
485 }
486
487 if (UseConcMarkSweepGC) {
488 // We need to call this even when ResizeOldPLAB is disabled
489 // so as to avoid breaking some asserts. While we may be able
490 // to avoid this by reorganizing the code a bit, I am loathe
491 // to do that unless we find cases where ergo leads to bad
492 // performance.
493 CompactibleFreeListSpaceLAB::compute_desired_plab_size();
494 }
495 }
496
497 ParScanClosure::ParScanClosure(ParNewGeneration* g,
498 ParScanThreadState* par_scan_state) :
499 OopsInClassLoaderDataOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) {
500 _boundary = _g->reserved().end();
501 }
502
503 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
504 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
505
506 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
507 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
508
509 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
510 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
511
512 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
513 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
514
515 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
516 ParScanThreadState* par_scan_state)
517 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
518 {}
519
520 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
521 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
522
523 #ifdef WIN32
524 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
525 #endif
526
527 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
528 ParScanThreadState* par_scan_state_,
529 ParScanWithoutBarrierClosure* to_space_closure_,
530 ParScanWithBarrierClosure* old_gen_closure_,
531 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
532 ParNewGeneration* par_gen_,
533 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
534 ObjToScanQueueSet* task_queues_,
535 ParallelTaskTerminator* terminator_) :
536
537 _par_scan_state(par_scan_state_),
538 _to_space_closure(to_space_closure_),
539 _old_gen_closure(old_gen_closure_),
540 _to_space_root_closure(to_space_root_closure_),
541 _old_gen_root_closure(old_gen_root_closure_),
542 _par_gen(par_gen_),
543 _task_queues(task_queues_),
544 _terminator(terminator_)
545 {}
546
547 void ParEvacuateFollowersClosure::do_void() {
548 ObjToScanQueue* work_q = par_scan_state()->work_queue();
549
550 while (true) {
551 // Scan to-space and old-gen objs until we run out of both.
552 oop obj_to_scan;
553 par_scan_state()->trim_queues(0);
554
555 // We have no local work, attempt to steal from other threads.
556
557 // Attempt to steal work from promoted.
558 if (task_queues()->steal(par_scan_state()->thread_num(),
559 par_scan_state()->hash_seed(),
560 obj_to_scan)) {
561 bool res = work_q->push(obj_to_scan);
562 assert(res, "Empty queue should have room for a push.");
563
564 // If successful, goto Start.
565 continue;
566
567 // Try global overflow list.
568 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
569 continue;
570 }
571
572 // Otherwise, offer termination.
573 par_scan_state()->start_term_time();
574 if (terminator()->offer_termination()) break;
575 par_scan_state()->end_term_time();
576 }
577 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
578 "Broken overflow list?");
579 // Finish the last termination pause.
580 par_scan_state()->end_term_time();
581 }
582
583 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen,
584 Generation* old_gen,
585 HeapWord* young_old_boundary,
586 ParScanThreadStateSet* state_set,
587 StrongRootsScope* strong_roots_scope) :
588 AbstractGangTask("ParNewGeneration collection"),
589 _young_gen(young_gen), _old_gen(old_gen),
590 _young_old_boundary(young_old_boundary),
591 _state_set(state_set),
592 _strong_roots_scope(strong_roots_scope)
593 {}
594
595 void ParNewGenTask::work(uint worker_id) {
596 CMSHeap* heap = CMSHeap::heap();
597 // Since this is being done in a separate thread, need new resource
598 // and handle marks.
599 ResourceMark rm;
600 HandleMark hm;
601
602 ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
603 assert(_state_set->is_valid(worker_id), "Should not have been called");
604
605 par_scan_state.set_young_old_boundary(_young_old_boundary);
606
607 CLDScanClosure cld_scan_closure(&par_scan_state.to_space_root_closure(),
608 heap->rem_set()->cld_rem_set()->accumulate_modified_oops());
609
610 par_scan_state.start_strong_roots();
611 heap->young_process_roots(_strong_roots_scope,
612 &par_scan_state.to_space_root_closure(),
613 &par_scan_state.older_gen_closure(),
614 &cld_scan_closure);
615
616 par_scan_state.end_strong_roots();
617
618 // "evacuate followers".
619 par_scan_state.evacuate_followers_closure().do_void();
620
621 // This will collapse this worker's promoted object list that's
622 // created during the main ParNew parallel phase of ParNew. This has
623 // to be called after all workers have finished promoting objects
624 // and scanning promoted objects. It should be safe calling it from
625 // here, given that we can only reach here after all thread have
626 // offered termination, i.e., after there is no more work to be
627 // done. It will also disable promotion tracking for the rest of
628 // this GC as it's not necessary to be on during reference processing.
629 _old_gen->par_oop_since_save_marks_iterate_done((int) worker_id);
630 }
631
632 ParNewGeneration::ParNewGeneration(ReservedSpace rs, size_t initial_byte_size)
633 : DefNewGeneration(rs, initial_byte_size, "PCopy"),
634 _overflow_list(NULL),
635 _is_alive_closure(this),
636 _plab_stats("Young", YoungPLABSize, PLABWeight)
637 {
638 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
639 NOT_PRODUCT(_num_par_pushes = 0;)
640 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
641 guarantee(_task_queues != NULL, "task_queues allocation failure.");
642
643 for (uint i = 0; i < ParallelGCThreads; i++) {
644 ObjToScanQueue *q = new ObjToScanQueue();
645 guarantee(q != NULL, "work_queue Allocation failure.");
646 _task_queues->register_queue(i, q);
647 }
648
649 for (uint i = 0; i < ParallelGCThreads; i++) {
650 _task_queues->queue(i)->initialize();
651 }
652
653 _overflow_stacks = NULL;
654 if (ParGCUseLocalOverflow) {
655 // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal with ','
656 typedef Stack<oop, mtGC> GCOopStack;
657
658 _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
659 for (size_t i = 0; i < ParallelGCThreads; ++i) {
660 new (_overflow_stacks + i) Stack<oop, mtGC>();
661 }
662 }
663
664 if (UsePerfData) {
665 EXCEPTION_MARK;
666 ResourceMark rm;
667
668 const char* cname =
669 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
670 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
671 ParallelGCThreads, CHECK);
672 }
673 }
674
675 // ParNewGeneration::
676 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
677 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
678
679 template <class T>
680 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
681 #ifdef ASSERT
682 {
683 oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);
684 // We never expect to see a null reference being processed
685 // as a weak reference.
686 assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs");
687 }
688 #endif // ASSERT
689
690 _par_cl->do_oop_nv(p);
691
692 if (CMSHeap::heap()->is_in_reserved(p)) {
693 oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);;
694 _rs->write_ref_field_gc_par(p, obj);
695 }
696 }
697
698 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
699 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
700
701 // ParNewGeneration::
702 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
703 DefNewGeneration::KeepAliveClosure(cl) {}
704
705 template <class T>
706 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
707 #ifdef ASSERT
708 {
709 oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);
710 // We never expect to see a null reference being processed
711 // as a weak reference.
712 assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs");
713 }
714 #endif // ASSERT
715
716 _cl->do_oop_nv(p);
717
718 if (CMSHeap::heap()->is_in_reserved(p)) {
719 oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);
720 _rs->write_ref_field_gc_par(p, obj);
721 }
722 }
723
724 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
725 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
726
727 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
728 T heap_oop = RawAccess<>::oop_load(p);
729 if (!CompressedOops::is_null(heap_oop)) {
730 oop obj = CompressedOops::decode_not_null(heap_oop);
731 if ((HeapWord*)obj < _boundary) {
732 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
733 oop new_obj = obj->is_forwarded()
734 ? obj->forwardee()
735 : _g->DefNewGeneration::copy_to_survivor_space(obj);
736 RawAccess<OOP_NOT_NULL>::oop_store(p, new_obj);
737 }
738 if (_gc_barrier) {
739 // If p points to a younger generation, mark the card.
740 if ((HeapWord*)obj < _gen_boundary) {
741 _rs->write_ref_field_gc_par(p, obj);
742 }
743 }
744 }
745 }
746
747 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
748 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
749
750 class ParNewRefProcTaskProxy: public AbstractGangTask {
751 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
752 public:
753 ParNewRefProcTaskProxy(ProcessTask& task,
754 ParNewGeneration& young_gen,
755 Generation& old_gen,
756 HeapWord* young_old_boundary,
757 ParScanThreadStateSet& state_set);
758
759 private:
760 virtual void work(uint worker_id);
761 private:
762 ParNewGeneration& _young_gen;
763 ProcessTask& _task;
764 Generation& _old_gen;
765 HeapWord* _young_old_boundary;
766 ParScanThreadStateSet& _state_set;
767 };
768
769 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
770 ParNewGeneration& young_gen,
771 Generation& old_gen,
772 HeapWord* young_old_boundary,
773 ParScanThreadStateSet& state_set)
774 : AbstractGangTask("ParNewGeneration parallel reference processing"),
775 _young_gen(young_gen),
776 _task(task),
777 _old_gen(old_gen),
778 _young_old_boundary(young_old_boundary),
779 _state_set(state_set)
780 { }
781
782 void ParNewRefProcTaskProxy::work(uint worker_id) {
783 ResourceMark rm;
784 HandleMark hm;
785 ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
786 par_scan_state.set_young_old_boundary(_young_old_boundary);
787 _task.work(worker_id, par_scan_state.is_alive_closure(),
788 par_scan_state.keep_alive_closure(),
789 par_scan_state.evacuate_followers_closure());
790 }
791
792 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) {
793 CMSHeap* gch = CMSHeap::heap();
794 WorkGang* workers = gch->workers();
795 assert(workers != NULL, "Need parallel worker threads.");
796 _state_set.reset(workers->active_workers(), _young_gen.promotion_failed());
797 ParNewRefProcTaskProxy rp_task(task, _young_gen, _old_gen,
798 _young_gen.reserved().end(), _state_set);
799 workers->run_task(&rp_task);
800 _state_set.reset(0 /* bad value in debug if not reset */,
801 _young_gen.promotion_failed());
802 }
803
804 void ParNewRefProcTaskExecutor::set_single_threaded_mode() {
805 _state_set.flush();
806 CMSHeap* heap = CMSHeap::heap();
807 heap->save_marks();
808 }
809
810 ScanClosureWithParBarrier::
811 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
812 ScanClosure(g, gc_barrier)
813 { }
814
815 template <typename OopClosureType1, typename OopClosureType2>
816 EvacuateFollowersClosureGeneral<OopClosureType1, OopClosureType2>::
817 EvacuateFollowersClosureGeneral(CMSHeap* heap,
818 OopClosureType1* cur,
819 OopClosureType2* older) :
820 _heap(heap),
821 _scan_cur_or_nonheap(cur), _scan_older(older)
822 { }
823
824 template <typename OopClosureType1, typename OopClosureType2>
825 void EvacuateFollowersClosureGeneral<OopClosureType1, OopClosureType2>::do_void() {
826 do {
827 _heap->oop_since_save_marks_iterate(_scan_cur_or_nonheap,
828 _scan_older);
829 } while (!_heap->no_allocs_since_save_marks());
830 }
831
832 // A Generation that does parallel young-gen collection.
833
834 void ParNewGeneration::handle_promotion_failed(CMSHeap* gch, ParScanThreadStateSet& thread_state_set) {
835 assert(_promo_failure_scan_stack.is_empty(), "post condition");
836 _promo_failure_scan_stack.clear(true); // Clear cached segments.
837
838 remove_forwarding_pointers();
839 log_info(gc, promotion)("Promotion failed");
840 // All the spaces are in play for mark-sweep.
841 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
842 from()->set_next_compaction_space(to());
843 gch->set_incremental_collection_failed();
844 // Inform the next generation that a promotion failure occurred.
845 _old_gen->promotion_failure_occurred();
846
847 // Trace promotion failure in the parallel GC threads
848 thread_state_set.trace_promotion_failed(gc_tracer());
849 // Single threaded code may have reported promotion failure to the global state
850 if (_promotion_failed_info.has_failed()) {
851 _gc_tracer.report_promotion_failed(_promotion_failed_info);
852 }
853 // Reset the PromotionFailureALot counters.
854 NOT_PRODUCT(gch->reset_promotion_should_fail();)
855 }
856
857 void ParNewGeneration::collect(bool full,
858 bool clear_all_soft_refs,
859 size_t size,
860 bool is_tlab) {
861 assert(full || size > 0, "otherwise we don't want to collect");
862
863 CMSHeap* gch = CMSHeap::heap();
864
865 _gc_timer->register_gc_start();
866
867 AdaptiveSizePolicy* size_policy = gch->size_policy();
868 WorkGang* workers = gch->workers();
869 assert(workers != NULL, "Need workgang for parallel work");
870 uint active_workers =
871 AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
872 workers->active_workers(),
873 Threads::number_of_non_daemon_threads());
874 active_workers = workers->update_active_workers(active_workers);
875 log_info(gc,task)("Using %u workers of %u for evacuation", active_workers, workers->total_workers());
876
877 _old_gen = gch->old_gen();
878
879 // If the next generation is too full to accommodate worst-case promotion
880 // from this generation, pass on collection; let the next generation
881 // do it.
882 if (!collection_attempt_is_safe()) {
883 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
884 return;
885 }
886 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
887
888 _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
889 gch->trace_heap_before_gc(gc_tracer());
890
891 init_assuming_no_promotion_failure();
892
893 if (UseAdaptiveSizePolicy) {
894 set_survivor_overflow(false);
895 size_policy->minor_collection_begin();
896 }
897
898 GCTraceTime(Trace, gc, phases) t1("ParNew", NULL, gch->gc_cause());
899
900 age_table()->clear();
901 to()->clear(SpaceDecorator::Mangle);
902
903 gch->save_marks();
904
905 // Set the correct parallelism (number of queues) in the reference processor
906 ref_processor()->set_active_mt_degree(active_workers);
907
908 // Need to initialize the preserved marks before the ThreadStateSet c'tor.
909 _preserved_marks_set.init(active_workers);
910
911 // Always set the terminator for the active number of workers
912 // because only those workers go through the termination protocol.
913 ParallelTaskTerminator _term(active_workers, task_queues());
914 ParScanThreadStateSet thread_state_set(active_workers,
915 *to(), *this, *_old_gen, *task_queues(),
916 _overflow_stacks, _preserved_marks_set,
917 desired_plab_sz(), _term);
918
919 thread_state_set.reset(active_workers, promotion_failed());
920
921 {
922 StrongRootsScope srs(active_workers);
923
924 ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs);
925 gch->rem_set()->prepare_for_younger_refs_iterate(true);
926 // It turns out that even when we're using 1 thread, doing the work in a
927 // separate thread causes wide variance in run times. We can't help this
928 // in the multi-threaded case, but we special-case n=1 here to get
929 // repeatable measurements of the 1-thread overhead of the parallel code.
930 // Might multiple workers ever be used? If yes, initialization
931 // has been done such that the single threaded path should not be used.
932 if (workers->total_workers() > 1) {
933 workers->run_task(&tsk);
934 } else {
935 tsk.work(0);
936 }
937 }
938
939 thread_state_set.reset(0 /* Bad value in debug if not reset */,
940 promotion_failed());
941
942 // Trace and reset failed promotion info.
943 if (promotion_failed()) {
944 thread_state_set.trace_promotion_failed(gc_tracer());
945 }
946
947 // Process (weak) reference objects found during scavenge.
948 ReferenceProcessor* rp = ref_processor();
949 IsAliveClosure is_alive(this);
950 ScanWeakRefClosure scan_weak_ref(this);
951 KeepAliveClosure keep_alive(&scan_weak_ref);
952 ScanClosure scan_without_gc_barrier(this, false);
953 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
954 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
955 EvacuateFollowersClosureGeneral<ScanClosure, ScanClosureWithParBarrier> evacuate_followers(
956 gch, &scan_without_gc_barrier, &scan_with_gc_barrier);
957 rp->setup_policy(clear_all_soft_refs);
958 // Can the mt_degree be set later (at run_task() time would be best)?
959 rp->set_active_mt_degree(active_workers);
960 ReferenceProcessorStats stats;
961 ReferenceProcessorPhaseTimes pt(_gc_timer, rp->num_queues());
962 if (rp->processing_is_mt()) {
963 ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set);
964 stats = rp->process_discovered_references(&is_alive, &keep_alive,
965 &evacuate_followers, &task_executor,
966 &pt);
967 } else {
968 thread_state_set.flush();
969 gch->save_marks();
970 stats = rp->process_discovered_references(&is_alive, &keep_alive,
971 &evacuate_followers, NULL,
972 &pt);
973 }
974 _gc_tracer.report_gc_reference_stats(stats);
975 _gc_tracer.report_tenuring_threshold(tenuring_threshold());
976 pt.print_all_references();
977
978 assert(gch->no_allocs_since_save_marks(), "evacuation should be done at this point");
979
980 WeakProcessor::weak_oops_do(&is_alive, &keep_alive);
981
982 // Verify that the usage of keep_alive only forwarded
983 // the oops and did not find anything new to copy.
984 assert(gch->no_allocs_since_save_marks(), "unexpectedly copied objects");
985
986 if (!promotion_failed()) {
987 // Swap the survivor spaces.
988 eden()->clear(SpaceDecorator::Mangle);
989 from()->clear(SpaceDecorator::Mangle);
990 if (ZapUnusedHeapArea) {
991 // This is now done here because of the piece-meal mangling which
992 // can check for valid mangling at intermediate points in the
993 // collection(s). When a young collection fails to collect
994 // sufficient space resizing of the young generation can occur
995 // and redistribute the spaces in the young generation. Mangle
996 // here so that unzapped regions don't get distributed to
997 // other spaces.
998 to()->mangle_unused_area();
999 }
1000 swap_spaces();
1001
1002 // A successful scavenge should restart the GC time limit count which is
1003 // for full GC's.
1004 size_policy->reset_gc_overhead_limit_count();
1005
1006 assert(to()->is_empty(), "to space should be empty now");
1007
1008 adjust_desired_tenuring_threshold();
1009 } else {
1010 handle_promotion_failed(gch, thread_state_set);
1011 }
1012 _preserved_marks_set.reclaim();
1013 // set new iteration safe limit for the survivor spaces
1014 from()->set_concurrent_iteration_safe_limit(from()->top());
1015 to()->set_concurrent_iteration_safe_limit(to()->top());
1016
1017 plab_stats()->adjust_desired_plab_sz();
1018
1019 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
1020 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
1021
1022 if (UseAdaptiveSizePolicy) {
1023 size_policy->minor_collection_end(gch->gc_cause());
1024 size_policy->avg_survived()->sample(from()->used());
1025 }
1026
1027 // We need to use a monotonically non-decreasing time in ms
1028 // or we will see time-warp warnings and os::javaTimeMillis()
1029 // does not guarantee monotonicity.
1030 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1031 update_time_of_last_gc(now);
1032
1033 rp->set_enqueuing_is_done(true);
1034 rp->verify_no_references_recorded();
1035
1036 gch->trace_heap_after_gc(gc_tracer());
1037
1038 _gc_timer->register_gc_end();
1039
1040 _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1041 }
1042
1043 size_t ParNewGeneration::desired_plab_sz() {
1044 return _plab_stats.desired_plab_sz(CMSHeap::heap()->workers()->active_workers());
1045 }
1046
1047 static int sum;
1048 void ParNewGeneration::waste_some_time() {
1049 for (int i = 0; i < 100; i++) {
1050 sum += i;
1051 }
1052 }
1053
1054 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1055
1056 // Because of concurrency, there are times where an object for which
1057 // "is_forwarded()" is true contains an "interim" forwarding pointer
1058 // value. Such a value will soon be overwritten with a real value.
1059 // This method requires "obj" to have a forwarding pointer, and waits, if
1060 // necessary for a real one to be inserted, and returns it.
1061
1062 oop ParNewGeneration::real_forwardee(oop obj) {
1063 oop forward_ptr = obj->forwardee();
1064 if (forward_ptr != ClaimedForwardPtr) {
1065 return forward_ptr;
1066 } else {
1067 return real_forwardee_slow(obj);
1068 }
1069 }
1070
1071 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1072 // Spin-read if it is claimed but not yet written by another thread.
1073 oop forward_ptr = obj->forwardee();
1074 while (forward_ptr == ClaimedForwardPtr) {
1075 waste_some_time();
1076 assert(obj->is_forwarded(), "precondition");
1077 forward_ptr = obj->forwardee();
1078 }
1079 return forward_ptr;
1080 }
1081
1082 // Multiple GC threads may try to promote an object. If the object
1083 // is successfully promoted, a forwarding pointer will be installed in
1084 // the object in the young generation. This method claims the right
1085 // to install the forwarding pointer before it copies the object,
1086 // thus avoiding the need to undo the copy as in
1087 // copy_to_survivor_space_avoiding_with_undo.
1088
1089 oop ParNewGeneration::copy_to_survivor_space(ParScanThreadState* par_scan_state,
1090 oop old,
1091 size_t sz,
1092 markOop m) {
1093 // In the sequential version, this assert also says that the object is
1094 // not forwarded. That might not be the case here. It is the case that
1095 // the caller observed it to be not forwarded at some time in the past.
1096 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1097
1098 // The sequential code read "old->age()" below. That doesn't work here,
1099 // since the age is in the mark word, and that might be overwritten with
1100 // a forwarding pointer by a parallel thread. So we must save the mark
1101 // word in a local and then analyze it.
1102 oopDesc dummyOld;
1103 dummyOld.set_mark_raw(m);
1104 assert(!dummyOld.is_forwarded(),
1105 "should not be called with forwarding pointer mark word.");
1106
1107 oop new_obj = NULL;
1108 oop forward_ptr;
1109
1110 // Try allocating obj in to-space (unless too old)
1111 if (dummyOld.age() < tenuring_threshold()) {
1112 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1113 if (new_obj == NULL) {
1114 set_survivor_overflow(true);
1115 }
1116 }
1117
1118 if (new_obj == NULL) {
1119 // Either to-space is full or we decided to promote try allocating obj tenured
1120
1121 // Attempt to install a null forwarding pointer (atomically),
1122 // to claim the right to install the real forwarding pointer.
1123 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1124 if (forward_ptr != NULL) {
1125 // someone else beat us to it.
1126 return real_forwardee(old);
1127 }
1128
1129 if (!_promotion_failed) {
1130 new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1131 old, m, sz);
1132 }
1133
1134 if (new_obj == NULL) {
1135 // promotion failed, forward to self
1136 _promotion_failed = true;
1137 new_obj = old;
1138
1139 par_scan_state->preserved_marks()->push_if_necessary(old, m);
1140 par_scan_state->register_promotion_failure(sz);
1141 }
1142
1143 old->forward_to(new_obj);
1144 forward_ptr = NULL;
1145 } else {
1146 // Is in to-space; do copying ourselves.
1147 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1148 assert(CMSHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1149 forward_ptr = old->forward_to_atomic(new_obj);
1150 // Restore the mark word copied above.
1151 new_obj->set_mark_raw(m);
1152 // Increment age if obj still in new generation
1153 new_obj->incr_age();
1154 par_scan_state->age_table()->add(new_obj, sz);
1155 }
1156 assert(new_obj != NULL, "just checking");
1157
1158 // This code must come after the CAS test, or it will print incorrect
1159 // information.
1160 log_develop_trace(gc, scavenge)("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1161 is_in_reserved(new_obj) ? "copying" : "tenuring",
1162 new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size());
1163
1164 if (forward_ptr == NULL) {
1165 oop obj_to_push = new_obj;
1166 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1167 // Length field used as index of next element to be scanned.
1168 // Real length can be obtained from real_forwardee()
1169 arrayOop(old)->set_length(0);
1170 obj_to_push = old;
1171 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1172 "push forwarded object");
1173 }
1174 // Push it on one of the queues of to-be-scanned objects.
1175 bool simulate_overflow = false;
1176 NOT_PRODUCT(
1177 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1178 // simulate a stack overflow
1179 simulate_overflow = true;
1180 }
1181 )
1182 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1183 // Add stats for overflow pushes.
1184 log_develop_trace(gc)("Queue Overflow");
1185 push_on_overflow_list(old, par_scan_state);
1186 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1187 }
1188
1189 return new_obj;
1190 }
1191
1192 // Oops. Someone beat us to it. Undo the allocation. Where did we
1193 // allocate it?
1194 if (is_in_reserved(new_obj)) {
1195 // Must be in to_space.
1196 assert(to()->is_in_reserved(new_obj), "Checking");
1197 if (forward_ptr == ClaimedForwardPtr) {
1198 // Wait to get the real forwarding pointer value.
1199 forward_ptr = real_forwardee(old);
1200 }
1201 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1202 }
1203
1204 return forward_ptr;
1205 }
1206
1207 #ifndef PRODUCT
1208 // It's OK to call this multi-threaded; the worst thing
1209 // that can happen is that we'll get a bunch of closely
1210 // spaced simulated overflows, but that's OK, in fact
1211 // probably good as it would exercise the overflow code
1212 // under contention.
1213 bool ParNewGeneration::should_simulate_overflow() {
1214 if (_overflow_counter-- <= 0) { // just being defensive
1215 _overflow_counter = ParGCWorkQueueOverflowInterval;
1216 return true;
1217 } else {
1218 return false;
1219 }
1220 }
1221 #endif
1222
1223 // In case we are using compressed oops, we need to be careful.
1224 // If the object being pushed is an object array, then its length
1225 // field keeps track of the "grey boundary" at which the next
1226 // incremental scan will be done (see ParGCArrayScanChunk).
1227 // When using compressed oops, this length field is kept in the
1228 // lower 32 bits of the erstwhile klass word and cannot be used
1229 // for the overflow chaining pointer (OCP below). As such the OCP
1230 // would itself need to be compressed into the top 32-bits in this
1231 // case. Unfortunately, see below, in the event that we have a
1232 // promotion failure, the node to be pushed on the list can be
1233 // outside of the Java heap, so the heap-based pointer compression
1234 // would not work (we would have potential aliasing between C-heap
1235 // and Java-heap pointers). For this reason, when using compressed
1236 // oops, we simply use a worker-thread-local, non-shared overflow
1237 // list in the form of a growable array, with a slightly different
1238 // overflow stack draining strategy. If/when we start using fat
1239 // stacks here, we can go back to using (fat) pointer chains
1240 // (although some performance comparisons would be useful since
1241 // single global lists have their own performance disadvantages
1242 // as we were made painfully aware not long ago, see 6786503).
1243 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1244 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1245 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1246 if (ParGCUseLocalOverflow) {
1247 // In the case of compressed oops, we use a private, not-shared
1248 // overflow stack.
1249 par_scan_state->push_on_overflow_stack(from_space_obj);
1250 } else {
1251 assert(!UseCompressedOops, "Error");
1252 // if the object has been forwarded to itself, then we cannot
1253 // use the klass pointer for the linked list. Instead we have
1254 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1255 // XXX This is horribly inefficient when a promotion failure occurs
1256 // and should be fixed. XXX FIX ME !!!
1257 #ifndef PRODUCT
1258 Atomic::inc(&_num_par_pushes);
1259 assert(_num_par_pushes > 0, "Tautology");
1260 #endif
1261 if (from_space_obj->forwardee() == from_space_obj) {
1262 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1263 listhead->forward_to(from_space_obj);
1264 from_space_obj = listhead;
1265 }
1266 oop observed_overflow_list = _overflow_list;
1267 oop cur_overflow_list;
1268 do {
1269 cur_overflow_list = observed_overflow_list;
1270 if (cur_overflow_list != BUSY) {
1271 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1272 } else {
1273 from_space_obj->set_klass_to_list_ptr(NULL);
1274 }
1275 observed_overflow_list =
1276 Atomic::cmpxchg((oopDesc*)from_space_obj, &_overflow_list, (oopDesc*)cur_overflow_list);
1277 } while (cur_overflow_list != observed_overflow_list);
1278 }
1279 }
1280
1281 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1282 bool res;
1283
1284 if (ParGCUseLocalOverflow) {
1285 res = par_scan_state->take_from_overflow_stack();
1286 } else {
1287 assert(!UseCompressedOops, "Error");
1288 res = take_from_overflow_list_work(par_scan_state);
1289 }
1290 return res;
1291 }
1292
1293
1294 // *NOTE*: The overflow list manipulation code here and
1295 // in CMSCollector:: are very similar in shape,
1296 // except that in the CMS case we thread the objects
1297 // directly into the list via their mark word, and do
1298 // not need to deal with special cases below related
1299 // to chunking of object arrays and promotion failure
1300 // handling.
1301 // CR 6797058 has been filed to attempt consolidation of
1302 // the common code.
1303 // Because of the common code, if you make any changes in
1304 // the code below, please check the CMS version to see if
1305 // similar changes might be needed.
1306 // See CMSCollector::par_take_from_overflow_list() for
1307 // more extensive documentation comments.
1308 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1309 ObjToScanQueue* work_q = par_scan_state->work_queue();
1310 // How many to take?
1311 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1312 (size_t)ParGCDesiredObjsFromOverflowList);
1313
1314 assert(!UseCompressedOops, "Error");
1315 assert(par_scan_state->overflow_stack() == NULL, "Error");
1316 if (_overflow_list == NULL) return false;
1317
1318 // Otherwise, there was something there; try claiming the list.
1319 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
1320 // Trim off a prefix of at most objsFromOverflow items
1321 Thread* tid = Thread::current();
1322 size_t spin_count = ParallelGCThreads;
1323 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1324 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1325 // someone grabbed it before we did ...
1326 // ... we spin for a short while...
1327 os::sleep(tid, sleep_time_millis, false);
1328 if (_overflow_list == NULL) {
1329 // nothing left to take
1330 return false;
1331 } else if (_overflow_list != BUSY) {
1332 // try and grab the prefix
1333 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
1334 }
1335 }
1336 if (prefix == NULL || prefix == BUSY) {
1337 // Nothing to take or waited long enough
1338 if (prefix == NULL) {
1339 // Write back the NULL in case we overwrote it with BUSY above
1340 // and it is still the same value.
1341 (void) Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
1342 }
1343 return false;
1344 }
1345 assert(prefix != NULL && prefix != BUSY, "Error");
1346 oop cur = prefix;
1347 for (size_t i = 1; i < objsFromOverflow; ++i) {
1348 oop next = cur->list_ptr_from_klass();
1349 if (next == NULL) break;
1350 cur = next;
1351 }
1352 assert(cur != NULL, "Loop postcondition");
1353
1354 // Reattach remaining (suffix) to overflow list
1355 oop suffix = cur->list_ptr_from_klass();
1356 if (suffix == NULL) {
1357 // Write back the NULL in lieu of the BUSY we wrote
1358 // above and it is still the same value.
1359 if (_overflow_list == BUSY) {
1360 (void) Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
1361 }
1362 } else {
1363 assert(suffix != BUSY, "Error");
1364 // suffix will be put back on global list
1365 cur->set_klass_to_list_ptr(NULL); // break off suffix
1366 // It's possible that the list is still in the empty(busy) state
1367 // we left it in a short while ago; in that case we may be
1368 // able to place back the suffix.
1369 oop observed_overflow_list = _overflow_list;
1370 oop cur_overflow_list = observed_overflow_list;
1371 bool attached = false;
1372 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1373 observed_overflow_list =
1374 Atomic::cmpxchg((oopDesc*)suffix, &_overflow_list, (oopDesc*)cur_overflow_list);
1375 if (cur_overflow_list == observed_overflow_list) {
1376 attached = true;
1377 break;
1378 } else cur_overflow_list = observed_overflow_list;
1379 }
1380 if (!attached) {
1381 // Too bad, someone else got in in between; we'll need to do a splice.
1382 // Find the last item of suffix list
1383 oop last = suffix;
1384 while (true) {
1385 oop next = last->list_ptr_from_klass();
1386 if (next == NULL) break;
1387 last = next;
1388 }
1389 // Atomically prepend suffix to current overflow list
1390 observed_overflow_list = _overflow_list;
1391 do {
1392 cur_overflow_list = observed_overflow_list;
1393 if (cur_overflow_list != BUSY) {
1394 // Do the splice ...
1395 last->set_klass_to_list_ptr(cur_overflow_list);
1396 } else { // cur_overflow_list == BUSY
1397 last->set_klass_to_list_ptr(NULL);
1398 }
1399 observed_overflow_list =
1400 Atomic::cmpxchg((oopDesc*)suffix, &_overflow_list, (oopDesc*)cur_overflow_list);
1401 } while (cur_overflow_list != observed_overflow_list);
1402 }
1403 }
1404
1405 // Push objects on prefix list onto this thread's work queue
1406 assert(prefix != NULL && prefix != BUSY, "program logic");
1407 cur = prefix;
1408 ssize_t n = 0;
1409 while (cur != NULL) {
1410 oop obj_to_push = cur->forwardee();
1411 oop next = cur->list_ptr_from_klass();
1412 cur->set_klass(obj_to_push->klass());
1413 // This may be an array object that is self-forwarded. In that case, the list pointer
1414 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1415 if (!is_in_reserved(cur)) {
1416 // This can become a scaling bottleneck when there is work queue overflow coincident
1417 // with promotion failure.
1418 oopDesc* f = cur;
1419 FREE_C_HEAP_ARRAY(oopDesc, f);
1420 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1421 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1422 obj_to_push = cur;
1423 }
1424 bool ok = work_q->push(obj_to_push);
1425 assert(ok, "Should have succeeded");
1426 cur = next;
1427 n++;
1428 }
1429 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1430 #ifndef PRODUCT
1431 assert(_num_par_pushes >= n, "Too many pops?");
1432 Atomic::sub(n, &_num_par_pushes);
1433 #endif
1434 return true;
1435 }
1436 #undef BUSY
1437
1438 void ParNewGeneration::ref_processor_init() {
1439 if (_ref_processor == NULL) {
1440 // Allocate and initialize a reference processor
1441 _span_based_discoverer.set_span(_reserved);
1442 _ref_processor =
1443 new ReferenceProcessor(&_span_based_discoverer, // span
1444 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1445 ParallelGCThreads, // mt processing degree
1446 refs_discovery_is_mt(), // mt discovery
1447 ParallelGCThreads, // mt discovery degree
1448 refs_discovery_is_atomic(), // atomic_discovery
1449 NULL); // is_alive_non_header
1450 }
1451 }
1452
1453 const char* ParNewGeneration::name() const {
1454 return "par new generation";
1455 }
1456
1457 void ParNewGeneration::restore_preserved_marks() {
1458 SharedRestorePreservedMarksTaskExecutor task_executor(CMSHeap::heap()->workers());
1459 _preserved_marks_set.restore(&task_executor);
1460 }