Tue, 10 May 2011 00:33:21 -0700
6883834: ParNew: assert(!_g->to()->is_in_reserved(obj),"Scanning field twice?") with LargeObjects tests
Summary: Fixed process_chunk_boundaries(), used for parallel card scanning when using ParNew/CMS, so as to prevent double-scanning, or worse, non-scanning of imprecisely marked objects exceeding parallel chunk size. Made some sizing parameters for parallel card scanning diagnostic, disabled ParallelGCRetainPLAB, and elaborated and clarified some comments.
Reviewed-by: stefank, johnc
1 /*
2 * Copyright (c) 2001, 2011, 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 */
25 #include "precompiled.hpp"
26 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp"
27 #include "gc_implementation/parNew/parGCAllocBuffer.hpp"
28 #include "gc_implementation/parNew/parNewGeneration.hpp"
29 #include "gc_implementation/parNew/parOopClosures.inline.hpp"
30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp"
31 #include "gc_implementation/shared/ageTable.hpp"
32 #include "gc_implementation/shared/spaceDecorator.hpp"
33 #include "memory/defNewGeneration.inline.hpp"
34 #include "memory/genCollectedHeap.hpp"
35 #include "memory/genOopClosures.inline.hpp"
36 #include "memory/generation.hpp"
37 #include "memory/generation.inline.hpp"
38 #include "memory/referencePolicy.hpp"
39 #include "memory/resourceArea.hpp"
40 #include "memory/sharedHeap.hpp"
41 #include "memory/space.hpp"
42 #include "oops/objArrayOop.hpp"
43 #include "oops/oop.inline.hpp"
44 #include "oops/oop.pcgc.inline.hpp"
45 #include "runtime/handles.hpp"
46 #include "runtime/handles.inline.hpp"
47 #include "runtime/java.hpp"
48 #include "runtime/thread.hpp"
49 #include "utilities/copy.hpp"
50 #include "utilities/globalDefinitions.hpp"
51 #include "utilities/workgroup.hpp"
53 #ifdef _MSC_VER
54 #pragma warning( push )
55 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
56 #endif
57 ParScanThreadState::ParScanThreadState(Space* to_space_,
58 ParNewGeneration* gen_,
59 Generation* old_gen_,
60 int thread_num_,
61 ObjToScanQueueSet* work_queue_set_,
62 Stack<oop>* overflow_stacks_,
63 size_t desired_plab_sz_,
64 ParallelTaskTerminator& term_) :
65 _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_),
66 _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false),
67 _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
68 _ageTable(false), // false ==> not the global age table, no perf data.
69 _to_space_alloc_buffer(desired_plab_sz_),
70 _to_space_closure(gen_, this), _old_gen_closure(gen_, this),
71 _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this),
72 _older_gen_closure(gen_, this),
73 _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
74 &_to_space_root_closure, gen_, &_old_gen_root_closure,
75 work_queue_set_, &term_),
76 _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this),
77 _keep_alive_closure(&_scan_weak_ref_closure),
78 _promotion_failure_size(0),
79 _strong_roots_time(0.0), _term_time(0.0)
80 {
81 #if TASKQUEUE_STATS
82 _term_attempts = 0;
83 _overflow_refills = 0;
84 _overflow_refill_objs = 0;
85 #endif // TASKQUEUE_STATS
87 _survivor_chunk_array =
88 (ChunkArray*) old_gen()->get_data_recorder(thread_num());
89 _hash_seed = 17; // Might want to take time-based random value.
90 _start = os::elapsedTime();
91 _old_gen_closure.set_generation(old_gen_);
92 _old_gen_root_closure.set_generation(old_gen_);
93 }
94 #ifdef _MSC_VER
95 #pragma warning( pop )
96 #endif
98 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
99 size_t plab_word_size) {
100 ChunkArray* sca = survivor_chunk_array();
101 if (sca != NULL) {
102 // A non-null SCA implies that we want the PLAB data recorded.
103 sca->record_sample(plab_start, plab_word_size);
104 }
105 }
107 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
108 return new_obj->is_objArray() &&
109 arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
110 new_obj != old_obj;
111 }
113 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
114 assert(old->is_objArray(), "must be obj array");
115 assert(old->is_forwarded(), "must be forwarded");
116 assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
117 assert(!old_gen()->is_in(old), "must be in young generation.");
119 objArrayOop obj = objArrayOop(old->forwardee());
120 // Process ParGCArrayScanChunk elements now
121 // and push the remainder back onto queue
122 int start = arrayOop(old)->length();
123 int end = obj->length();
124 int remainder = end - start;
125 assert(start <= end, "just checking");
126 if (remainder > 2 * ParGCArrayScanChunk) {
127 // Test above combines last partial chunk with a full chunk
128 end = start + ParGCArrayScanChunk;
129 arrayOop(old)->set_length(end);
130 // Push remainder.
131 bool ok = work_queue()->push(old);
132 assert(ok, "just popped, push must be okay");
133 } else {
134 // Restore length so that it can be used if there
135 // is a promotion failure and forwarding pointers
136 // must be removed.
137 arrayOop(old)->set_length(end);
138 }
140 // process our set of indices (include header in first chunk)
141 // should make sure end is even (aligned to HeapWord in case of compressed oops)
142 if ((HeapWord *)obj < young_old_boundary()) {
143 // object is in to_space
144 obj->oop_iterate_range(&_to_space_closure, start, end);
145 } else {
146 // object is in old generation
147 obj->oop_iterate_range(&_old_gen_closure, start, end);
148 }
149 }
152 void ParScanThreadState::trim_queues(int max_size) {
153 ObjToScanQueue* queue = work_queue();
154 do {
155 while (queue->size() > (juint)max_size) {
156 oop obj_to_scan;
157 if (queue->pop_local(obj_to_scan)) {
158 if ((HeapWord *)obj_to_scan < young_old_boundary()) {
159 if (obj_to_scan->is_objArray() &&
160 obj_to_scan->is_forwarded() &&
161 obj_to_scan->forwardee() != obj_to_scan) {
162 scan_partial_array_and_push_remainder(obj_to_scan);
163 } else {
164 // object is in to_space
165 obj_to_scan->oop_iterate(&_to_space_closure);
166 }
167 } else {
168 // object is in old generation
169 obj_to_scan->oop_iterate(&_old_gen_closure);
170 }
171 }
172 }
173 // For the case of compressed oops, we have a private, non-shared
174 // overflow stack, so we eagerly drain it so as to more evenly
175 // distribute load early. Note: this may be good to do in
176 // general rather than delay for the final stealing phase.
177 // If applicable, we'll transfer a set of objects over to our
178 // work queue, allowing them to be stolen and draining our
179 // private overflow stack.
180 } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
181 }
183 bool ParScanThreadState::take_from_overflow_stack() {
184 assert(ParGCUseLocalOverflow, "Else should not call");
185 assert(young_gen()->overflow_list() == NULL, "Error");
186 ObjToScanQueue* queue = work_queue();
187 Stack<oop>* const of_stack = overflow_stack();
188 const size_t num_overflow_elems = of_stack->size();
189 const size_t space_available = queue->max_elems() - queue->size();
190 const size_t num_take_elems = MIN3(space_available / 4,
191 ParGCDesiredObjsFromOverflowList,
192 num_overflow_elems);
193 // Transfer the most recent num_take_elems from the overflow
194 // stack to our work queue.
195 for (size_t i = 0; i != num_take_elems; i++) {
196 oop cur = of_stack->pop();
197 oop obj_to_push = cur->forwardee();
198 assert(Universe::heap()->is_in_reserved(cur), "Should be in heap");
199 assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
200 assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap");
201 if (should_be_partially_scanned(obj_to_push, cur)) {
202 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
203 obj_to_push = cur;
204 }
205 bool ok = queue->push(obj_to_push);
206 assert(ok, "Should have succeeded");
207 }
208 assert(young_gen()->overflow_list() == NULL, "Error");
209 return num_take_elems > 0; // was something transferred?
210 }
212 void ParScanThreadState::push_on_overflow_stack(oop p) {
213 assert(ParGCUseLocalOverflow, "Else should not call");
214 overflow_stack()->push(p);
215 assert(young_gen()->overflow_list() == NULL, "Error");
216 }
218 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
220 // Otherwise, if the object is small enough, try to reallocate the
221 // buffer.
222 HeapWord* obj = NULL;
223 if (!_to_space_full) {
224 ParGCAllocBuffer* const plab = to_space_alloc_buffer();
225 Space* const sp = to_space();
226 if (word_sz * 100 <
227 ParallelGCBufferWastePct * plab->word_sz()) {
228 // Is small enough; abandon this buffer and start a new one.
229 plab->retire(false, false);
230 size_t buf_size = plab->word_sz();
231 HeapWord* buf_space = sp->par_allocate(buf_size);
232 if (buf_space == NULL) {
233 const size_t min_bytes =
234 ParGCAllocBuffer::min_size() << LogHeapWordSize;
235 size_t free_bytes = sp->free();
236 while(buf_space == NULL && free_bytes >= min_bytes) {
237 buf_size = free_bytes >> LogHeapWordSize;
238 assert(buf_size == (size_t)align_object_size(buf_size),
239 "Invariant");
240 buf_space = sp->par_allocate(buf_size);
241 free_bytes = sp->free();
242 }
243 }
244 if (buf_space != NULL) {
245 plab->set_word_size(buf_size);
246 plab->set_buf(buf_space);
247 record_survivor_plab(buf_space, buf_size);
248 obj = plab->allocate(word_sz);
249 // Note that we cannot compare buf_size < word_sz below
250 // because of AlignmentReserve (see ParGCAllocBuffer::allocate()).
251 assert(obj != NULL || plab->words_remaining() < word_sz,
252 "Else should have been able to allocate");
253 // It's conceivable that we may be able to use the
254 // buffer we just grabbed for subsequent small requests
255 // even if not for this one.
256 } else {
257 // We're used up.
258 _to_space_full = true;
259 }
261 } else {
262 // Too large; allocate the object individually.
263 obj = sp->par_allocate(word_sz);
264 }
265 }
266 return obj;
267 }
270 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj,
271 size_t word_sz) {
272 // Is the alloc in the current alloc buffer?
273 if (to_space_alloc_buffer()->contains(obj)) {
274 assert(to_space_alloc_buffer()->contains(obj + word_sz - 1),
275 "Should contain whole object.");
276 to_space_alloc_buffer()->undo_allocation(obj, word_sz);
277 } else {
278 CollectedHeap::fill_with_object(obj, word_sz);
279 }
280 }
282 void ParScanThreadState::print_and_clear_promotion_failure_size() {
283 if (_promotion_failure_size != 0) {
284 if (PrintPromotionFailure) {
285 gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ",
286 _thread_num, _promotion_failure_size);
287 }
288 _promotion_failure_size = 0;
289 }
290 }
292 class ParScanThreadStateSet: private ResourceArray {
293 public:
294 // Initializes states for the specified number of threads;
295 ParScanThreadStateSet(int num_threads,
296 Space& to_space,
297 ParNewGeneration& gen,
298 Generation& old_gen,
299 ObjToScanQueueSet& queue_set,
300 Stack<oop>* overflow_stacks_,
301 size_t desired_plab_sz,
302 ParallelTaskTerminator& term);
304 ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
306 inline ParScanThreadState& thread_state(int i);
308 void reset(bool promotion_failed);
309 void flush();
311 #if TASKQUEUE_STATS
312 static void
313 print_termination_stats_hdr(outputStream* const st = gclog_or_tty);
314 void print_termination_stats(outputStream* const st = gclog_or_tty);
315 static void
316 print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty);
317 void print_taskqueue_stats(outputStream* const st = gclog_or_tty);
318 void reset_stats();
319 #endif // TASKQUEUE_STATS
321 private:
322 ParallelTaskTerminator& _term;
323 ParNewGeneration& _gen;
324 Generation& _next_gen;
325 };
328 ParScanThreadStateSet::ParScanThreadStateSet(
329 int num_threads, Space& to_space, ParNewGeneration& gen,
330 Generation& old_gen, ObjToScanQueueSet& queue_set,
331 Stack<oop>* overflow_stacks,
332 size_t desired_plab_sz, ParallelTaskTerminator& term)
333 : ResourceArray(sizeof(ParScanThreadState), num_threads),
334 _gen(gen), _next_gen(old_gen), _term(term)
335 {
336 assert(num_threads > 0, "sanity check!");
337 assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
338 "overflow_stack allocation mismatch");
339 // Initialize states.
340 for (int i = 0; i < num_threads; ++i) {
341 new ((ParScanThreadState*)_data + i)
342 ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set,
343 overflow_stacks, desired_plab_sz, term);
344 }
345 }
347 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i)
348 {
349 assert(i >= 0 && i < length(), "sanity check!");
350 return ((ParScanThreadState*)_data)[i];
351 }
354 void ParScanThreadStateSet::reset(bool promotion_failed)
355 {
356 _term.reset_for_reuse();
357 if (promotion_failed) {
358 for (int i = 0; i < length(); ++i) {
359 thread_state(i).print_and_clear_promotion_failure_size();
360 }
361 }
362 }
364 #if TASKQUEUE_STATS
365 void
366 ParScanThreadState::reset_stats()
367 {
368 taskqueue_stats().reset();
369 _term_attempts = 0;
370 _overflow_refills = 0;
371 _overflow_refill_objs = 0;
372 }
374 void ParScanThreadStateSet::reset_stats()
375 {
376 for (int i = 0; i < length(); ++i) {
377 thread_state(i).reset_stats();
378 }
379 }
381 void
382 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st)
383 {
384 st->print_raw_cr("GC Termination Stats");
385 st->print_raw_cr(" elapsed --strong roots-- "
386 "-------termination-------");
387 st->print_raw_cr("thr ms ms % "
388 " ms % attempts");
389 st->print_raw_cr("--- --------- --------- ------ "
390 "--------- ------ --------");
391 }
393 void ParScanThreadStateSet::print_termination_stats(outputStream* const st)
394 {
395 print_termination_stats_hdr(st);
397 for (int i = 0; i < length(); ++i) {
398 const ParScanThreadState & pss = thread_state(i);
399 const double elapsed_ms = pss.elapsed_time() * 1000.0;
400 const double s_roots_ms = pss.strong_roots_time() * 1000.0;
401 const double term_ms = pss.term_time() * 1000.0;
402 st->print_cr("%3d %9.2f %9.2f %6.2f "
403 "%9.2f %6.2f " SIZE_FORMAT_W(8),
404 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
405 term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
406 }
407 }
409 // Print stats related to work queue activity.
410 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st)
411 {
412 st->print_raw_cr("GC Task Stats");
413 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
414 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
415 }
417 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st)
418 {
419 print_taskqueue_stats_hdr(st);
421 TaskQueueStats totals;
422 for (int i = 0; i < length(); ++i) {
423 const ParScanThreadState & pss = thread_state(i);
424 const TaskQueueStats & stats = pss.taskqueue_stats();
425 st->print("%3d ", i); stats.print(st); st->cr();
426 totals += stats;
428 if (pss.overflow_refills() > 0) {
429 st->print_cr(" " SIZE_FORMAT_W(10) " overflow refills "
430 SIZE_FORMAT_W(10) " overflow objects",
431 pss.overflow_refills(), pss.overflow_refill_objs());
432 }
433 }
434 st->print("tot "); totals.print(st); st->cr();
436 DEBUG_ONLY(totals.verify());
437 }
438 #endif // TASKQUEUE_STATS
440 void ParScanThreadStateSet::flush()
441 {
442 // Work in this loop should be kept as lightweight as
443 // possible since this might otherwise become a bottleneck
444 // to scaling. Should we add heavy-weight work into this
445 // loop, consider parallelizing the loop into the worker threads.
446 for (int i = 0; i < length(); ++i) {
447 ParScanThreadState& par_scan_state = thread_state(i);
449 // Flush stats related to To-space PLAB activity and
450 // retire the last buffer.
451 par_scan_state.to_space_alloc_buffer()->
452 flush_stats_and_retire(_gen.plab_stats(),
453 false /* !retain */);
455 // Every thread has its own age table. We need to merge
456 // them all into one.
457 ageTable *local_table = par_scan_state.age_table();
458 _gen.age_table()->merge(local_table);
460 // Inform old gen that we're done.
461 _next_gen.par_promote_alloc_done(i);
462 _next_gen.par_oop_since_save_marks_iterate_done(i);
463 }
465 if (UseConcMarkSweepGC && ParallelGCThreads > 0) {
466 // We need to call this even when ResizeOldPLAB is disabled
467 // so as to avoid breaking some asserts. While we may be able
468 // to avoid this by reorganizing the code a bit, I am loathe
469 // to do that unless we find cases where ergo leads to bad
470 // performance.
471 CFLS_LAB::compute_desired_plab_size();
472 }
473 }
475 ParScanClosure::ParScanClosure(ParNewGeneration* g,
476 ParScanThreadState* par_scan_state) :
477 OopsInGenClosure(g), _par_scan_state(par_scan_state), _g(g)
478 {
479 assert(_g->level() == 0, "Optimized for youngest generation");
480 _boundary = _g->reserved().end();
481 }
483 void ParScanWithBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, false); }
484 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
486 void ParScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, false); }
487 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
489 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, true, true); }
490 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
492 void ParRootScanWithoutBarrierClosure::do_oop(oop* p) { ParScanClosure::do_oop_work(p, false, true); }
493 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
495 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
496 ParScanThreadState* par_scan_state)
497 : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
498 {}
500 void ParScanWeakRefClosure::do_oop(oop* p) { ParScanWeakRefClosure::do_oop_work(p); }
501 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
503 #ifdef WIN32
504 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
505 #endif
507 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
508 ParScanThreadState* par_scan_state_,
509 ParScanWithoutBarrierClosure* to_space_closure_,
510 ParScanWithBarrierClosure* old_gen_closure_,
511 ParRootScanWithoutBarrierClosure* to_space_root_closure_,
512 ParNewGeneration* par_gen_,
513 ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
514 ObjToScanQueueSet* task_queues_,
515 ParallelTaskTerminator* terminator_) :
517 _par_scan_state(par_scan_state_),
518 _to_space_closure(to_space_closure_),
519 _old_gen_closure(old_gen_closure_),
520 _to_space_root_closure(to_space_root_closure_),
521 _old_gen_root_closure(old_gen_root_closure_),
522 _par_gen(par_gen_),
523 _task_queues(task_queues_),
524 _terminator(terminator_)
525 {}
527 void ParEvacuateFollowersClosure::do_void() {
528 ObjToScanQueue* work_q = par_scan_state()->work_queue();
530 while (true) {
532 // Scan to-space and old-gen objs until we run out of both.
533 oop obj_to_scan;
534 par_scan_state()->trim_queues(0);
536 // We have no local work, attempt to steal from other threads.
538 // attempt to steal work from promoted.
539 if (task_queues()->steal(par_scan_state()->thread_num(),
540 par_scan_state()->hash_seed(),
541 obj_to_scan)) {
542 bool res = work_q->push(obj_to_scan);
543 assert(res, "Empty queue should have room for a push.");
545 // if successful, goto Start.
546 continue;
548 // try global overflow list.
549 } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
550 continue;
551 }
553 // Otherwise, offer termination.
554 par_scan_state()->start_term_time();
555 if (terminator()->offer_termination()) break;
556 par_scan_state()->end_term_time();
557 }
558 assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
559 "Broken overflow list?");
560 // Finish the last termination pause.
561 par_scan_state()->end_term_time();
562 }
564 ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* next_gen,
565 HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) :
566 AbstractGangTask("ParNewGeneration collection"),
567 _gen(gen), _next_gen(next_gen),
568 _young_old_boundary(young_old_boundary),
569 _state_set(state_set)
570 {}
572 void ParNewGenTask::work(int i) {
573 GenCollectedHeap* gch = GenCollectedHeap::heap();
574 // Since this is being done in a separate thread, need new resource
575 // and handle marks.
576 ResourceMark rm;
577 HandleMark hm;
578 // We would need multiple old-gen queues otherwise.
579 assert(gch->n_gens() == 2, "Par young collection currently only works with one older gen.");
581 Generation* old_gen = gch->next_gen(_gen);
583 ParScanThreadState& par_scan_state = _state_set->thread_state(i);
584 par_scan_state.set_young_old_boundary(_young_old_boundary);
586 par_scan_state.start_strong_roots();
587 gch->gen_process_strong_roots(_gen->level(),
588 true, // Process younger gens, if any,
589 // as strong roots.
590 false, // no scope; this is parallel code
591 false, // not collecting perm generation.
592 SharedHeap::SO_AllClasses,
593 &par_scan_state.to_space_root_closure(),
594 true, // walk *all* scavengable nmethods
595 &par_scan_state.older_gen_closure());
596 par_scan_state.end_strong_roots();
598 // "evacuate followers".
599 par_scan_state.evacuate_followers_closure().do_void();
600 }
602 #ifdef _MSC_VER
603 #pragma warning( push )
604 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
605 #endif
606 ParNewGeneration::
607 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level)
608 : DefNewGeneration(rs, initial_byte_size, level, "PCopy"),
609 _overflow_list(NULL),
610 _is_alive_closure(this),
611 _plab_stats(YoungPLABSize, PLABWeight)
612 {
613 NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
614 NOT_PRODUCT(_num_par_pushes = 0;)
615 _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
616 guarantee(_task_queues != NULL, "task_queues allocation failure.");
618 for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
619 ObjToScanQueue *q = new ObjToScanQueue();
620 guarantee(q != NULL, "work_queue Allocation failure.");
621 _task_queues->register_queue(i1, q);
622 }
624 for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
625 _task_queues->queue(i2)->initialize();
627 _overflow_stacks = NULL;
628 if (ParGCUseLocalOverflow) {
629 _overflow_stacks = NEW_C_HEAP_ARRAY(Stack<oop>, ParallelGCThreads);
630 for (size_t i = 0; i < ParallelGCThreads; ++i) {
631 new (_overflow_stacks + i) Stack<oop>();
632 }
633 }
635 if (UsePerfData) {
636 EXCEPTION_MARK;
637 ResourceMark rm;
639 const char* cname =
640 PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
641 PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
642 ParallelGCThreads, CHECK);
643 }
644 }
645 #ifdef _MSC_VER
646 #pragma warning( pop )
647 #endif
649 // ParNewGeneration::
650 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
651 DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
653 template <class T>
654 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
655 #ifdef ASSERT
656 {
657 assert(!oopDesc::is_null(*p), "expected non-null ref");
658 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
659 // We never expect to see a null reference being processed
660 // as a weak reference.
661 assert(obj->is_oop(), "expected an oop while scanning weak refs");
662 }
663 #endif // ASSERT
665 _par_cl->do_oop_nv(p);
667 if (Universe::heap()->is_in_reserved(p)) {
668 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
669 _rs->write_ref_field_gc_par(p, obj);
670 }
671 }
673 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p) { ParKeepAliveClosure::do_oop_work(p); }
674 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
676 // ParNewGeneration::
677 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
678 DefNewGeneration::KeepAliveClosure(cl) {}
680 template <class T>
681 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
682 #ifdef ASSERT
683 {
684 assert(!oopDesc::is_null(*p), "expected non-null ref");
685 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
686 // We never expect to see a null reference being processed
687 // as a weak reference.
688 assert(obj->is_oop(), "expected an oop while scanning weak refs");
689 }
690 #endif // ASSERT
692 _cl->do_oop_nv(p);
694 if (Universe::heap()->is_in_reserved(p)) {
695 oop obj = oopDesc::load_decode_heap_oop_not_null(p);
696 _rs->write_ref_field_gc_par(p, obj);
697 }
698 }
700 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p) { KeepAliveClosure::do_oop_work(p); }
701 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
703 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
704 T heap_oop = oopDesc::load_heap_oop(p);
705 if (!oopDesc::is_null(heap_oop)) {
706 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
707 if ((HeapWord*)obj < _boundary) {
708 assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
709 oop new_obj = obj->is_forwarded()
710 ? obj->forwardee()
711 : _g->DefNewGeneration::copy_to_survivor_space(obj);
712 oopDesc::encode_store_heap_oop_not_null(p, new_obj);
713 }
714 if (_gc_barrier) {
715 // If p points to a younger generation, mark the card.
716 if ((HeapWord*)obj < _gen_boundary) {
717 _rs->write_ref_field_gc_par(p, obj);
718 }
719 }
720 }
721 }
723 void ScanClosureWithParBarrier::do_oop(oop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
724 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
726 class ParNewRefProcTaskProxy: public AbstractGangTask {
727 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
728 public:
729 ParNewRefProcTaskProxy(ProcessTask& task, ParNewGeneration& gen,
730 Generation& next_gen,
731 HeapWord* young_old_boundary,
732 ParScanThreadStateSet& state_set);
734 private:
735 virtual void work(int i);
737 private:
738 ParNewGeneration& _gen;
739 ProcessTask& _task;
740 Generation& _next_gen;
741 HeapWord* _young_old_boundary;
742 ParScanThreadStateSet& _state_set;
743 };
745 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(
746 ProcessTask& task, ParNewGeneration& gen,
747 Generation& next_gen,
748 HeapWord* young_old_boundary,
749 ParScanThreadStateSet& state_set)
750 : AbstractGangTask("ParNewGeneration parallel reference processing"),
751 _gen(gen),
752 _task(task),
753 _next_gen(next_gen),
754 _young_old_boundary(young_old_boundary),
755 _state_set(state_set)
756 {
757 }
759 void ParNewRefProcTaskProxy::work(int i)
760 {
761 ResourceMark rm;
762 HandleMark hm;
763 ParScanThreadState& par_scan_state = _state_set.thread_state(i);
764 par_scan_state.set_young_old_boundary(_young_old_boundary);
765 _task.work(i, par_scan_state.is_alive_closure(),
766 par_scan_state.keep_alive_closure(),
767 par_scan_state.evacuate_followers_closure());
768 }
770 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
771 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
772 EnqueueTask& _task;
774 public:
775 ParNewRefEnqueueTaskProxy(EnqueueTask& task)
776 : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
777 _task(task)
778 { }
780 virtual void work(int i)
781 {
782 _task.work(i);
783 }
784 };
787 void ParNewRefProcTaskExecutor::execute(ProcessTask& task)
788 {
789 GenCollectedHeap* gch = GenCollectedHeap::heap();
790 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
791 "not a generational heap");
792 WorkGang* workers = gch->workers();
793 assert(workers != NULL, "Need parallel worker threads.");
794 ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(),
795 _generation.reserved().end(), _state_set);
796 workers->run_task(&rp_task);
797 _state_set.reset(_generation.promotion_failed());
798 }
800 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task)
801 {
802 GenCollectedHeap* gch = GenCollectedHeap::heap();
803 WorkGang* workers = gch->workers();
804 assert(workers != NULL, "Need parallel worker threads.");
805 ParNewRefEnqueueTaskProxy enq_task(task);
806 workers->run_task(&enq_task);
807 }
809 void ParNewRefProcTaskExecutor::set_single_threaded_mode()
810 {
811 _state_set.flush();
812 GenCollectedHeap* gch = GenCollectedHeap::heap();
813 gch->set_par_threads(0); // 0 ==> non-parallel.
814 gch->save_marks();
815 }
817 ScanClosureWithParBarrier::
818 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
819 ScanClosure(g, gc_barrier) {}
821 EvacuateFollowersClosureGeneral::
822 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
823 OopsInGenClosure* cur,
824 OopsInGenClosure* older) :
825 _gch(gch), _level(level),
826 _scan_cur_or_nonheap(cur), _scan_older(older)
827 {}
829 void EvacuateFollowersClosureGeneral::do_void() {
830 do {
831 // Beware: this call will lead to closure applications via virtual
832 // calls.
833 _gch->oop_since_save_marks_iterate(_level,
834 _scan_cur_or_nonheap,
835 _scan_older);
836 } while (!_gch->no_allocs_since_save_marks(_level));
837 }
840 bool ParNewGeneration::_avoid_promotion_undo = false;
842 void ParNewGeneration::adjust_desired_tenuring_threshold() {
843 // Set the desired survivor size to half the real survivor space
844 _tenuring_threshold =
845 age_table()->compute_tenuring_threshold(to()->capacity()/HeapWordSize);
846 }
848 // A Generation that does parallel young-gen collection.
850 void ParNewGeneration::collect(bool full,
851 bool clear_all_soft_refs,
852 size_t size,
853 bool is_tlab) {
854 assert(full || size > 0, "otherwise we don't want to collect");
855 GenCollectedHeap* gch = GenCollectedHeap::heap();
856 assert(gch->kind() == CollectedHeap::GenCollectedHeap,
857 "not a CMS generational heap");
858 AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
859 WorkGang* workers = gch->workers();
860 _next_gen = gch->next_gen(this);
861 assert(_next_gen != NULL,
862 "This must be the youngest gen, and not the only gen");
863 assert(gch->n_gens() == 2,
864 "Par collection currently only works with single older gen.");
865 // Do we have to avoid promotion_undo?
866 if (gch->collector_policy()->is_concurrent_mark_sweep_policy()) {
867 set_avoid_promotion_undo(true);
868 }
870 // If the next generation is too full to accomodate worst-case promotion
871 // from this generation, pass on collection; let the next generation
872 // do it.
873 if (!collection_attempt_is_safe()) {
874 gch->set_incremental_collection_failed(); // slight lie, in that we did not even attempt one
875 return;
876 }
877 assert(to()->is_empty(), "Else not collection_attempt_is_safe");
879 init_assuming_no_promotion_failure();
881 if (UseAdaptiveSizePolicy) {
882 set_survivor_overflow(false);
883 size_policy->minor_collection_begin();
884 }
886 TraceTime t1("GC", PrintGC && !PrintGCDetails, true, gclog_or_tty);
887 // Capture heap used before collection (for printing).
888 size_t gch_prev_used = gch->used();
890 SpecializationStats::clear();
892 age_table()->clear();
893 to()->clear(SpaceDecorator::Mangle);
895 gch->save_marks();
896 assert(workers != NULL, "Need parallel worker threads.");
897 ParallelTaskTerminator _term(workers->total_workers(), task_queues());
898 ParScanThreadStateSet thread_state_set(workers->total_workers(),
899 *to(), *this, *_next_gen, *task_queues(),
900 _overflow_stacks, desired_plab_sz(), _term);
902 ParNewGenTask tsk(this, _next_gen, reserved().end(), &thread_state_set);
903 int n_workers = workers->total_workers();
904 gch->set_par_threads(n_workers);
905 gch->rem_set()->prepare_for_younger_refs_iterate(true);
906 // It turns out that even when we're using 1 thread, doing the work in a
907 // separate thread causes wide variance in run times. We can't help this
908 // in the multi-threaded case, but we special-case n=1 here to get
909 // repeatable measurements of the 1-thread overhead of the parallel code.
910 if (n_workers > 1) {
911 GenCollectedHeap::StrongRootsScope srs(gch);
912 workers->run_task(&tsk);
913 } else {
914 GenCollectedHeap::StrongRootsScope srs(gch);
915 tsk.work(0);
916 }
917 thread_state_set.reset(promotion_failed());
919 // Process (weak) reference objects found during scavenge.
920 ReferenceProcessor* rp = ref_processor();
921 IsAliveClosure is_alive(this);
922 ScanWeakRefClosure scan_weak_ref(this);
923 KeepAliveClosure keep_alive(&scan_weak_ref);
924 ScanClosure scan_without_gc_barrier(this, false);
925 ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
926 set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
927 EvacuateFollowersClosureGeneral evacuate_followers(gch, _level,
928 &scan_without_gc_barrier, &scan_with_gc_barrier);
929 rp->setup_policy(clear_all_soft_refs);
930 if (rp->processing_is_mt()) {
931 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
932 rp->process_discovered_references(&is_alive, &keep_alive,
933 &evacuate_followers, &task_executor);
934 } else {
935 thread_state_set.flush();
936 gch->set_par_threads(0); // 0 ==> non-parallel.
937 gch->save_marks();
938 rp->process_discovered_references(&is_alive, &keep_alive,
939 &evacuate_followers, NULL);
940 }
941 if (!promotion_failed()) {
942 // Swap the survivor spaces.
943 eden()->clear(SpaceDecorator::Mangle);
944 from()->clear(SpaceDecorator::Mangle);
945 if (ZapUnusedHeapArea) {
946 // This is now done here because of the piece-meal mangling which
947 // can check for valid mangling at intermediate points in the
948 // collection(s). When a minor collection fails to collect
949 // sufficient space resizing of the young generation can occur
950 // an redistribute the spaces in the young generation. Mangle
951 // here so that unzapped regions don't get distributed to
952 // other spaces.
953 to()->mangle_unused_area();
954 }
955 swap_spaces();
957 // A successful scavenge should restart the GC time limit count which is
958 // for full GC's.
959 size_policy->reset_gc_overhead_limit_count();
961 assert(to()->is_empty(), "to space should be empty now");
962 } else {
963 assert(_promo_failure_scan_stack.is_empty(), "post condition");
964 _promo_failure_scan_stack.clear(true); // Clear cached segments.
966 remove_forwarding_pointers();
967 if (PrintGCDetails) {
968 gclog_or_tty->print(" (promotion failed)");
969 }
970 // All the spaces are in play for mark-sweep.
971 swap_spaces(); // Make life simpler for CMS || rescan; see 6483690.
972 from()->set_next_compaction_space(to());
973 gch->set_incremental_collection_failed();
974 // Inform the next generation that a promotion failure occurred.
975 _next_gen->promotion_failure_occurred();
977 // Reset the PromotionFailureALot counters.
978 NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();)
979 }
980 // set new iteration safe limit for the survivor spaces
981 from()->set_concurrent_iteration_safe_limit(from()->top());
982 to()->set_concurrent_iteration_safe_limit(to()->top());
984 adjust_desired_tenuring_threshold();
985 if (ResizePLAB) {
986 plab_stats()->adjust_desired_plab_sz();
987 }
989 if (PrintGC && !PrintGCDetails) {
990 gch->print_heap_change(gch_prev_used);
991 }
993 if (PrintGCDetails && ParallelGCVerbose) {
994 TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
995 TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
996 }
998 if (UseAdaptiveSizePolicy) {
999 size_policy->minor_collection_end(gch->gc_cause());
1000 size_policy->avg_survived()->sample(from()->used());
1001 }
1003 update_time_of_last_gc(os::javaTimeMillis());
1005 SpecializationStats::print();
1007 rp->set_enqueuing_is_done(true);
1008 if (rp->processing_is_mt()) {
1009 ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1010 rp->enqueue_discovered_references(&task_executor);
1011 } else {
1012 rp->enqueue_discovered_references(NULL);
1013 }
1014 rp->verify_no_references_recorded();
1015 }
1017 static int sum;
1018 void ParNewGeneration::waste_some_time() {
1019 for (int i = 0; i < 100; i++) {
1020 sum += i;
1021 }
1022 }
1024 static const oop ClaimedForwardPtr = oop(0x4);
1026 // Because of concurrency, there are times where an object for which
1027 // "is_forwarded()" is true contains an "interim" forwarding pointer
1028 // value. Such a value will soon be overwritten with a real value.
1029 // This method requires "obj" to have a forwarding pointer, and waits, if
1030 // necessary for a real one to be inserted, and returns it.
1032 oop ParNewGeneration::real_forwardee(oop obj) {
1033 oop forward_ptr = obj->forwardee();
1034 if (forward_ptr != ClaimedForwardPtr) {
1035 return forward_ptr;
1036 } else {
1037 return real_forwardee_slow(obj);
1038 }
1039 }
1041 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1042 // Spin-read if it is claimed but not yet written by another thread.
1043 oop forward_ptr = obj->forwardee();
1044 while (forward_ptr == ClaimedForwardPtr) {
1045 waste_some_time();
1046 assert(obj->is_forwarded(), "precondition");
1047 forward_ptr = obj->forwardee();
1048 }
1049 return forward_ptr;
1050 }
1052 #ifdef ASSERT
1053 bool ParNewGeneration::is_legal_forward_ptr(oop p) {
1054 return
1055 (_avoid_promotion_undo && p == ClaimedForwardPtr)
1056 || Universe::heap()->is_in_reserved(p);
1057 }
1058 #endif
1060 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
1061 if (m->must_be_preserved_for_promotion_failure(obj)) {
1062 // We should really have separate per-worker stacks, rather
1063 // than use locking of a common pair of stacks.
1064 MutexLocker ml(ParGCRareEvent_lock);
1065 preserve_mark(obj, m);
1066 }
1067 }
1069 // Multiple GC threads may try to promote an object. If the object
1070 // is successfully promoted, a forwarding pointer will be installed in
1071 // the object in the young generation. This method claims the right
1072 // to install the forwarding pointer before it copies the object,
1073 // thus avoiding the need to undo the copy as in
1074 // copy_to_survivor_space_avoiding_with_undo.
1076 oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo(
1077 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1078 // In the sequential version, this assert also says that the object is
1079 // not forwarded. That might not be the case here. It is the case that
1080 // the caller observed it to be not forwarded at some time in the past.
1081 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1083 // The sequential code read "old->age()" below. That doesn't work here,
1084 // since the age is in the mark word, and that might be overwritten with
1085 // a forwarding pointer by a parallel thread. So we must save the mark
1086 // word in a local and then analyze it.
1087 oopDesc dummyOld;
1088 dummyOld.set_mark(m);
1089 assert(!dummyOld.is_forwarded(),
1090 "should not be called with forwarding pointer mark word.");
1092 oop new_obj = NULL;
1093 oop forward_ptr;
1095 // Try allocating obj in to-space (unless too old)
1096 if (dummyOld.age() < tenuring_threshold()) {
1097 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1098 if (new_obj == NULL) {
1099 set_survivor_overflow(true);
1100 }
1101 }
1103 if (new_obj == NULL) {
1104 // Either to-space is full or we decided to promote
1105 // try allocating obj tenured
1107 // Attempt to install a null forwarding pointer (atomically),
1108 // to claim the right to install the real forwarding pointer.
1109 forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1110 if (forward_ptr != NULL) {
1111 // someone else beat us to it.
1112 return real_forwardee(old);
1113 }
1115 new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
1116 old, m, sz);
1118 if (new_obj == NULL) {
1119 // promotion failed, forward to self
1120 _promotion_failed = true;
1121 new_obj = old;
1123 preserve_mark_if_necessary(old, m);
1124 // Log the size of the maiden promotion failure
1125 par_scan_state->log_promotion_failure(sz);
1126 }
1128 old->forward_to(new_obj);
1129 forward_ptr = NULL;
1130 } else {
1131 // Is in to-space; do copying ourselves.
1132 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1133 forward_ptr = old->forward_to_atomic(new_obj);
1134 // Restore the mark word copied above.
1135 new_obj->set_mark(m);
1136 // Increment age if obj still in new generation
1137 new_obj->incr_age();
1138 par_scan_state->age_table()->add(new_obj, sz);
1139 }
1140 assert(new_obj != NULL, "just checking");
1142 if (forward_ptr == NULL) {
1143 oop obj_to_push = new_obj;
1144 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1145 // Length field used as index of next element to be scanned.
1146 // Real length can be obtained from real_forwardee()
1147 arrayOop(old)->set_length(0);
1148 obj_to_push = old;
1149 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1150 "push forwarded object");
1151 }
1152 // Push it on one of the queues of to-be-scanned objects.
1153 bool simulate_overflow = false;
1154 NOT_PRODUCT(
1155 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1156 // simulate a stack overflow
1157 simulate_overflow = true;
1158 }
1159 )
1160 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1161 // Add stats for overflow pushes.
1162 if (Verbose && PrintGCDetails) {
1163 gclog_or_tty->print("queue overflow!\n");
1164 }
1165 push_on_overflow_list(old, par_scan_state);
1166 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1167 }
1169 return new_obj;
1170 }
1172 // Oops. Someone beat us to it. Undo the allocation. Where did we
1173 // allocate it?
1174 if (is_in_reserved(new_obj)) {
1175 // Must be in to_space.
1176 assert(to()->is_in_reserved(new_obj), "Checking");
1177 if (forward_ptr == ClaimedForwardPtr) {
1178 // Wait to get the real forwarding pointer value.
1179 forward_ptr = real_forwardee(old);
1180 }
1181 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1182 }
1184 return forward_ptr;
1185 }
1188 // Multiple GC threads may try to promote the same object. If two
1189 // or more GC threads copy the object, only one wins the race to install
1190 // the forwarding pointer. The other threads have to undo their copy.
1192 oop ParNewGeneration::copy_to_survivor_space_with_undo(
1193 ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1195 // In the sequential version, this assert also says that the object is
1196 // not forwarded. That might not be the case here. It is the case that
1197 // the caller observed it to be not forwarded at some time in the past.
1198 assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1200 // The sequential code read "old->age()" below. That doesn't work here,
1201 // since the age is in the mark word, and that might be overwritten with
1202 // a forwarding pointer by a parallel thread. So we must save the mark
1203 // word here, install it in a local oopDesc, and then analyze it.
1204 oopDesc dummyOld;
1205 dummyOld.set_mark(m);
1206 assert(!dummyOld.is_forwarded(),
1207 "should not be called with forwarding pointer mark word.");
1209 bool failed_to_promote = false;
1210 oop new_obj = NULL;
1211 oop forward_ptr;
1213 // Try allocating obj in to-space (unless too old)
1214 if (dummyOld.age() < tenuring_threshold()) {
1215 new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1216 if (new_obj == NULL) {
1217 set_survivor_overflow(true);
1218 }
1219 }
1221 if (new_obj == NULL) {
1222 // Either to-space is full or we decided to promote
1223 // try allocating obj tenured
1224 new_obj = _next_gen->par_promote(par_scan_state->thread_num(),
1225 old, m, sz);
1227 if (new_obj == NULL) {
1228 // promotion failed, forward to self
1229 forward_ptr = old->forward_to_atomic(old);
1230 new_obj = old;
1232 if (forward_ptr != NULL) {
1233 return forward_ptr; // someone else succeeded
1234 }
1236 _promotion_failed = true;
1237 failed_to_promote = true;
1239 preserve_mark_if_necessary(old, m);
1240 // Log the size of the maiden promotion failure
1241 par_scan_state->log_promotion_failure(sz);
1242 }
1243 } else {
1244 // Is in to-space; do copying ourselves.
1245 Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1246 // Restore the mark word copied above.
1247 new_obj->set_mark(m);
1248 // Increment age if new_obj still in new generation
1249 new_obj->incr_age();
1250 par_scan_state->age_table()->add(new_obj, sz);
1251 }
1252 assert(new_obj != NULL, "just checking");
1254 // Now attempt to install the forwarding pointer (atomically).
1255 // We have to copy the mark word before overwriting with forwarding
1256 // ptr, so we can restore it below in the copy.
1257 if (!failed_to_promote) {
1258 forward_ptr = old->forward_to_atomic(new_obj);
1259 }
1261 if (forward_ptr == NULL) {
1262 oop obj_to_push = new_obj;
1263 if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1264 // Length field used as index of next element to be scanned.
1265 // Real length can be obtained from real_forwardee()
1266 arrayOop(old)->set_length(0);
1267 obj_to_push = old;
1268 assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1269 "push forwarded object");
1270 }
1271 // Push it on one of the queues of to-be-scanned objects.
1272 bool simulate_overflow = false;
1273 NOT_PRODUCT(
1274 if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1275 // simulate a stack overflow
1276 simulate_overflow = true;
1277 }
1278 )
1279 if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1280 // Add stats for overflow pushes.
1281 push_on_overflow_list(old, par_scan_state);
1282 TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1283 }
1285 return new_obj;
1286 }
1288 // Oops. Someone beat us to it. Undo the allocation. Where did we
1289 // allocate it?
1290 if (is_in_reserved(new_obj)) {
1291 // Must be in to_space.
1292 assert(to()->is_in_reserved(new_obj), "Checking");
1293 par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1294 } else {
1295 assert(!_avoid_promotion_undo, "Should not be here if avoiding.");
1296 _next_gen->par_promote_alloc_undo(par_scan_state->thread_num(),
1297 (HeapWord*)new_obj, sz);
1298 }
1300 return forward_ptr;
1301 }
1303 #ifndef PRODUCT
1304 // It's OK to call this multi-threaded; the worst thing
1305 // that can happen is that we'll get a bunch of closely
1306 // spaced simulated oveflows, but that's OK, in fact
1307 // probably good as it would exercise the overflow code
1308 // under contention.
1309 bool ParNewGeneration::should_simulate_overflow() {
1310 if (_overflow_counter-- <= 0) { // just being defensive
1311 _overflow_counter = ParGCWorkQueueOverflowInterval;
1312 return true;
1313 } else {
1314 return false;
1315 }
1316 }
1317 #endif
1319 // In case we are using compressed oops, we need to be careful.
1320 // If the object being pushed is an object array, then its length
1321 // field keeps track of the "grey boundary" at which the next
1322 // incremental scan will be done (see ParGCArrayScanChunk).
1323 // When using compressed oops, this length field is kept in the
1324 // lower 32 bits of the erstwhile klass word and cannot be used
1325 // for the overflow chaining pointer (OCP below). As such the OCP
1326 // would itself need to be compressed into the top 32-bits in this
1327 // case. Unfortunately, see below, in the event that we have a
1328 // promotion failure, the node to be pushed on the list can be
1329 // outside of the Java heap, so the heap-based pointer compression
1330 // would not work (we would have potential aliasing between C-heap
1331 // and Java-heap pointers). For this reason, when using compressed
1332 // oops, we simply use a worker-thread-local, non-shared overflow
1333 // list in the form of a growable array, with a slightly different
1334 // overflow stack draining strategy. If/when we start using fat
1335 // stacks here, we can go back to using (fat) pointer chains
1336 // (although some performance comparisons would be useful since
1337 // single global lists have their own performance disadvantages
1338 // as we were made painfully aware not long ago, see 6786503).
1339 #define BUSY (oop(0x1aff1aff))
1340 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1341 assert(is_in_reserved(from_space_obj), "Should be from this generation");
1342 if (ParGCUseLocalOverflow) {
1343 // In the case of compressed oops, we use a private, not-shared
1344 // overflow stack.
1345 par_scan_state->push_on_overflow_stack(from_space_obj);
1346 } else {
1347 assert(!UseCompressedOops, "Error");
1348 // if the object has been forwarded to itself, then we cannot
1349 // use the klass pointer for the linked list. Instead we have
1350 // to allocate an oopDesc in the C-Heap and use that for the linked list.
1351 // XXX This is horribly inefficient when a promotion failure occurs
1352 // and should be fixed. XXX FIX ME !!!
1353 #ifndef PRODUCT
1354 Atomic::inc_ptr(&_num_par_pushes);
1355 assert(_num_par_pushes > 0, "Tautology");
1356 #endif
1357 if (from_space_obj->forwardee() == from_space_obj) {
1358 oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1);
1359 listhead->forward_to(from_space_obj);
1360 from_space_obj = listhead;
1361 }
1362 oop observed_overflow_list = _overflow_list;
1363 oop cur_overflow_list;
1364 do {
1365 cur_overflow_list = observed_overflow_list;
1366 if (cur_overflow_list != BUSY) {
1367 from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1368 } else {
1369 from_space_obj->set_klass_to_list_ptr(NULL);
1370 }
1371 observed_overflow_list =
1372 (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1373 } while (cur_overflow_list != observed_overflow_list);
1374 }
1375 }
1377 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1378 bool res;
1380 if (ParGCUseLocalOverflow) {
1381 res = par_scan_state->take_from_overflow_stack();
1382 } else {
1383 assert(!UseCompressedOops, "Error");
1384 res = take_from_overflow_list_work(par_scan_state);
1385 }
1386 return res;
1387 }
1390 // *NOTE*: The overflow list manipulation code here and
1391 // in CMSCollector:: are very similar in shape,
1392 // except that in the CMS case we thread the objects
1393 // directly into the list via their mark word, and do
1394 // not need to deal with special cases below related
1395 // to chunking of object arrays and promotion failure
1396 // handling.
1397 // CR 6797058 has been filed to attempt consolidation of
1398 // the common code.
1399 // Because of the common code, if you make any changes in
1400 // the code below, please check the CMS version to see if
1401 // similar changes might be needed.
1402 // See CMSCollector::par_take_from_overflow_list() for
1403 // more extensive documentation comments.
1404 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1405 ObjToScanQueue* work_q = par_scan_state->work_queue();
1406 // How many to take?
1407 size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1408 (size_t)ParGCDesiredObjsFromOverflowList);
1410 assert(!UseCompressedOops, "Error");
1411 assert(par_scan_state->overflow_stack() == NULL, "Error");
1412 if (_overflow_list == NULL) return false;
1414 // Otherwise, there was something there; try claiming the list.
1415 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
1416 // Trim off a prefix of at most objsFromOverflow items
1417 Thread* tid = Thread::current();
1418 size_t spin_count = (size_t)ParallelGCThreads;
1419 size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1420 for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1421 // someone grabbed it before we did ...
1422 // ... we spin for a short while...
1423 os::sleep(tid, sleep_time_millis, false);
1424 if (_overflow_list == NULL) {
1425 // nothing left to take
1426 return false;
1427 } else if (_overflow_list != BUSY) {
1428 // try and grab the prefix
1429 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
1430 }
1431 }
1432 if (prefix == NULL || prefix == BUSY) {
1433 // Nothing to take or waited long enough
1434 if (prefix == NULL) {
1435 // Write back the NULL in case we overwrote it with BUSY above
1436 // and it is still the same value.
1437 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1438 }
1439 return false;
1440 }
1441 assert(prefix != NULL && prefix != BUSY, "Error");
1442 size_t i = 1;
1443 oop cur = prefix;
1444 while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1445 i++; cur = oop(cur->klass());
1446 }
1448 // Reattach remaining (suffix) to overflow list
1449 if (cur->klass_or_null() == NULL) {
1450 // Write back the NULL in lieu of the BUSY we wrote
1451 // above and it is still the same value.
1452 if (_overflow_list == BUSY) {
1453 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1454 }
1455 } else {
1456 assert(cur->klass_or_null() != BUSY, "Error");
1457 oop suffix = oop(cur->klass()); // suffix will be put back on global list
1458 cur->set_klass_to_list_ptr(NULL); // break off suffix
1459 // It's possible that the list is still in the empty(busy) state
1460 // we left it in a short while ago; in that case we may be
1461 // able to place back the suffix.
1462 oop observed_overflow_list = _overflow_list;
1463 oop cur_overflow_list = observed_overflow_list;
1464 bool attached = false;
1465 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1466 observed_overflow_list =
1467 (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1468 if (cur_overflow_list == observed_overflow_list) {
1469 attached = true;
1470 break;
1471 } else cur_overflow_list = observed_overflow_list;
1472 }
1473 if (!attached) {
1474 // Too bad, someone else got in in between; we'll need to do a splice.
1475 // Find the last item of suffix list
1476 oop last = suffix;
1477 while (last->klass_or_null() != NULL) {
1478 last = oop(last->klass());
1479 }
1480 // Atomically prepend suffix to current overflow list
1481 observed_overflow_list = _overflow_list;
1482 do {
1483 cur_overflow_list = observed_overflow_list;
1484 if (cur_overflow_list != BUSY) {
1485 // Do the splice ...
1486 last->set_klass_to_list_ptr(cur_overflow_list);
1487 } else { // cur_overflow_list == BUSY
1488 last->set_klass_to_list_ptr(NULL);
1489 }
1490 observed_overflow_list =
1491 (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1492 } while (cur_overflow_list != observed_overflow_list);
1493 }
1494 }
1496 // Push objects on prefix list onto this thread's work queue
1497 assert(prefix != NULL && prefix != BUSY, "program logic");
1498 cur = prefix;
1499 ssize_t n = 0;
1500 while (cur != NULL) {
1501 oop obj_to_push = cur->forwardee();
1502 oop next = oop(cur->klass_or_null());
1503 cur->set_klass(obj_to_push->klass());
1504 // This may be an array object that is self-forwarded. In that case, the list pointer
1505 // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1506 if (!is_in_reserved(cur)) {
1507 // This can become a scaling bottleneck when there is work queue overflow coincident
1508 // with promotion failure.
1509 oopDesc* f = cur;
1510 FREE_C_HEAP_ARRAY(oopDesc, f);
1511 } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1512 assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1513 obj_to_push = cur;
1514 }
1515 bool ok = work_q->push(obj_to_push);
1516 assert(ok, "Should have succeeded");
1517 cur = next;
1518 n++;
1519 }
1520 TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1521 #ifndef PRODUCT
1522 assert(_num_par_pushes >= n, "Too many pops?");
1523 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1524 #endif
1525 return true;
1526 }
1527 #undef BUSY
1529 void ParNewGeneration::ref_processor_init()
1530 {
1531 if (_ref_processor == NULL) {
1532 // Allocate and initialize a reference processor
1533 _ref_processor =
1534 new ReferenceProcessor(_reserved, // span
1535 ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1536 (int) ParallelGCThreads, // mt processing degree
1537 refs_discovery_is_mt(), // mt discovery
1538 (int) ParallelGCThreads, // mt discovery degree
1539 refs_discovery_is_atomic(), // atomic_discovery
1540 NULL, // is_alive_non_header
1541 false); // write barrier for next field updates
1542 }
1543 }
1545 const char* ParNewGeneration::name() const {
1546 return "par new generation";
1547 }
1549 bool ParNewGeneration::in_use() {
1550 return UseParNewGC && ParallelGCThreads > 0;
1551 }