1 /*
2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "code/icBuffer.hpp"
27 #include "gc_implementation/g1/bufferingOopClosure.hpp"
28 #include "gc_implementation/g1/concurrentG1Refine.hpp"
29 #include "gc_implementation/g1/concurrentG1RefineThread.hpp"
30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
31 #include "gc_implementation/g1/g1AllocRegion.inline.hpp"
32 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
33 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
34 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
35 #include "gc_implementation/g1/g1EvacFailure.hpp"
36 #include "gc_implementation/g1/g1GCPhaseTimes.hpp"
37 #include "gc_implementation/g1/g1Log.hpp"
38 #include "gc_implementation/g1/g1MarkSweep.hpp"
39 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
40 #include "gc_implementation/g1/g1PreserveMarkQueue.hpp"
41 #include "gc_implementation/g1/g1RemSet.inline.hpp"
42 #include "gc_implementation/g1/heapRegion.inline.hpp"
43 #include "gc_implementation/g1/heapRegionRemSet.hpp"
44 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
45 #include "gc_implementation/g1/vm_operations_g1.hpp"
46 #include "gc_implementation/shared/isGCActiveMark.hpp"
47 #include "memory/gcLocker.inline.hpp"
48 #include "memory/genOopClosures.inline.hpp"
49 #include "memory/generationSpec.hpp"
50 #include "memory/referenceProcessor.hpp"
51 #include "oops/oop.inline.hpp"
52 #include "oops/oop.pcgc.inline.hpp"
53 #include "runtime/aprofiler.hpp"
54 #include "runtime/vmThread.hpp"
55
56 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0;
57
58 // turn it on so that the contents of the young list (scan-only /
59 // to-be-collected) are printed at "strategic" points before / during
60 // / after the collection --- this is useful for debugging
61 #define YOUNG_LIST_VERBOSE 0
62 // CURRENT STATUS
63 // This file is under construction. Search for "FIXME".
64
65 // INVARIANTS/NOTES
66 //
67 // All allocation activity covered by the G1CollectedHeap interface is
68 // serialized by acquiring the HeapLock. This happens in mem_allocate
69 // and allocate_new_tlab, which are the "entry" points to the
70 // allocation code from the rest of the JVM. (Note that this does not
71 // apply to TLAB allocation, which is not part of this interface: it
72 // is done by clients of this interface.)
73
74 // Notes on implementation of parallelism in different tasks.
75 //
76 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism.
77 // The number of GC workers is passed to heap_region_par_iterate_chunked().
78 // It does use run_task() which sets _n_workers in the task.
79 // G1ParTask executes g1_process_strong_roots() ->
80 // SharedHeap::process_strong_roots() which calls eventuall to
81 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses
82 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also
83 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap).
84 //
85
86 // Local to this file.
87
88 class RefineCardTableEntryClosure: public CardTableEntryClosure {
89 SuspendibleThreadSet* _sts;
90 G1RemSet* _g1rs;
91 ConcurrentG1Refine* _cg1r;
92 bool _concurrent;
93 public:
94 RefineCardTableEntryClosure(SuspendibleThreadSet* sts,
95 G1RemSet* g1rs,
96 ConcurrentG1Refine* cg1r) :
97 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true)
98 {}
99 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
100 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false);
101 // This path is executed by the concurrent refine or mutator threads,
102 // concurrently, and so we do not care if card_ptr contains references
103 // that point into the collection set.
104 assert(!oops_into_cset, "should be");
105
106 if (_concurrent && _sts->should_yield()) {
107 // Caller will actually yield.
108 return false;
109 }
110 // Otherwise, we finished successfully; return true.
111 return true;
112 }
113 void set_concurrent(bool b) { _concurrent = b; }
114 };
115
116
117 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure {
118 int _calls;
119 G1CollectedHeap* _g1h;
120 CardTableModRefBS* _ctbs;
121 int _histo[256];
122 public:
123 ClearLoggedCardTableEntryClosure() :
124 _calls(0)
125 {
126 _g1h = G1CollectedHeap::heap();
127 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
128 for (int i = 0; i < 256; i++) _histo[i] = 0;
129 }
130 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
131 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
132 _calls++;
133 unsigned char* ujb = (unsigned char*)card_ptr;
134 int ind = (int)(*ujb);
135 _histo[ind]++;
136 *card_ptr = -1;
137 }
138 return true;
139 }
140 int calls() { return _calls; }
141 void print_histo() {
142 gclog_or_tty->print_cr("Card table value histogram:");
143 for (int i = 0; i < 256; i++) {
144 if (_histo[i] != 0) {
145 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]);
146 }
147 }
148 }
149 };
150
151 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure {
152 int _calls;
153 G1CollectedHeap* _g1h;
154 CardTableModRefBS* _ctbs;
155 public:
156 RedirtyLoggedCardTableEntryClosure() :
157 _calls(0)
158 {
159 _g1h = G1CollectedHeap::heap();
160 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
161 }
162 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
163 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
164 _calls++;
165 *card_ptr = 0;
166 }
167 return true;
168 }
169 int calls() { return _calls; }
170 };
171
172 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure {
173 public:
174 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
175 *card_ptr = CardTableModRefBS::dirty_card_val();
176 return true;
177 }
178 };
179
180 YoungList::YoungList(G1CollectedHeap* g1h) :
181 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0),
182 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) {
183 guarantee(check_list_empty(false), "just making sure...");
184 }
185
186 void YoungList::push_region(HeapRegion *hr) {
187 assert(!hr->is_young(), "should not already be young");
188 assert(hr->get_next_young_region() == NULL, "cause it should!");
189
190 hr->set_next_young_region(_head);
191 _head = hr;
192
193 _g1h->g1_policy()->set_region_eden(hr, (int) _length);
194 ++_length;
195 }
196
197 void YoungList::add_survivor_region(HeapRegion* hr) {
198 assert(hr->is_survivor(), "should be flagged as survivor region");
199 assert(hr->get_next_young_region() == NULL, "cause it should!");
200
201 hr->set_next_young_region(_survivor_head);
202 if (_survivor_head == NULL) {
203 _survivor_tail = hr;
204 }
205 _survivor_head = hr;
206 ++_survivor_length;
207 }
208
209 void YoungList::empty_list(HeapRegion* list) {
210 while (list != NULL) {
211 HeapRegion* next = list->get_next_young_region();
212 list->set_next_young_region(NULL);
213 list->uninstall_surv_rate_group();
214 list->set_not_young();
215 list = next;
216 }
217 }
218
219 void YoungList::empty_list() {
220 assert(check_list_well_formed(), "young list should be well formed");
221
222 empty_list(_head);
223 _head = NULL;
224 _length = 0;
225
226 empty_list(_survivor_head);
227 _survivor_head = NULL;
228 _survivor_tail = NULL;
229 _survivor_length = 0;
230
231 _last_sampled_rs_lengths = 0;
232
233 assert(check_list_empty(false), "just making sure...");
234 }
235
236 bool YoungList::check_list_well_formed() {
237 bool ret = true;
238
239 uint length = 0;
240 HeapRegion* curr = _head;
241 HeapRegion* last = NULL;
242 while (curr != NULL) {
243 if (!curr->is_young()) {
244 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" "
245 "incorrectly tagged (y: %d, surv: %d)",
246 curr->bottom(), curr->end(),
247 curr->is_young(), curr->is_survivor());
248 ret = false;
249 }
250 ++length;
251 last = curr;
252 curr = curr->get_next_young_region();
253 }
254 ret = ret && (length == _length);
255
256 if (!ret) {
257 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!");
258 gclog_or_tty->print_cr("### list has %u entries, _length is %u",
259 length, _length);
260 }
261
262 return ret;
263 }
264
265 bool YoungList::check_list_empty(bool check_sample) {
266 bool ret = true;
267
268 if (_length != 0) {
269 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u",
270 _length);
271 ret = false;
272 }
273 if (check_sample && _last_sampled_rs_lengths != 0) {
274 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths");
275 ret = false;
276 }
277 if (_head != NULL) {
278 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head");
279 ret = false;
280 }
281 if (!ret) {
282 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty");
283 }
284
285 return ret;
286 }
287
288 void
289 YoungList::rs_length_sampling_init() {
290 _sampled_rs_lengths = 0;
291 _curr = _head;
292 }
293
294 bool
295 YoungList::rs_length_sampling_more() {
296 return _curr != NULL;
297 }
298
299 void
300 YoungList::rs_length_sampling_next() {
301 assert( _curr != NULL, "invariant" );
302 size_t rs_length = _curr->rem_set()->occupied();
303
304 _sampled_rs_lengths += rs_length;
305
306 // The current region may not yet have been added to the
307 // incremental collection set (it gets added when it is
308 // retired as the current allocation region).
309 if (_curr->in_collection_set()) {
310 // Update the collection set policy information for this region
311 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length);
312 }
313
314 _curr = _curr->get_next_young_region();
315 if (_curr == NULL) {
316 _last_sampled_rs_lengths = _sampled_rs_lengths;
317 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths);
318 }
319 }
320
321 void
322 YoungList::reset_auxilary_lists() {
323 guarantee( is_empty(), "young list should be empty" );
324 assert(check_list_well_formed(), "young list should be well formed");
325
326 // Add survivor regions to SurvRateGroup.
327 _g1h->g1_policy()->note_start_adding_survivor_regions();
328 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */);
329
330 int young_index_in_cset = 0;
331 for (HeapRegion* curr = _survivor_head;
332 curr != NULL;
333 curr = curr->get_next_young_region()) {
334 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset);
335
336 // The region is a non-empty survivor so let's add it to
337 // the incremental collection set for the next evacuation
338 // pause.
339 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr);
340 young_index_in_cset += 1;
341 }
342 assert((uint) young_index_in_cset == _survivor_length, "post-condition");
343 _g1h->g1_policy()->note_stop_adding_survivor_regions();
344
345 _head = _survivor_head;
346 _length = _survivor_length;
347 if (_survivor_head != NULL) {
348 assert(_survivor_tail != NULL, "cause it shouldn't be");
349 assert(_survivor_length > 0, "invariant");
350 _survivor_tail->set_next_young_region(NULL);
351 }
352
353 // Don't clear the survivor list handles until the start of
354 // the next evacuation pause - we need it in order to re-tag
355 // the survivor regions from this evacuation pause as 'young'
356 // at the start of the next.
357
358 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */);
359
360 assert(check_list_well_formed(), "young list should be well formed");
361 }
362
363 void YoungList::print() {
364 HeapRegion* lists[] = {_head, _survivor_head};
365 const char* names[] = {"YOUNG", "SURVIVOR"};
366
367 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) {
368 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]);
369 HeapRegion *curr = lists[list];
370 if (curr == NULL)
371 gclog_or_tty->print_cr(" empty");
372 while (curr != NULL) {
373 gclog_or_tty->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d",
374 HR_FORMAT_PARAMS(curr),
375 curr->prev_top_at_mark_start(),
376 curr->next_top_at_mark_start(),
377 curr->age_in_surv_rate_group_cond());
378 curr = curr->get_next_young_region();
379 }
380 }
381
382 gclog_or_tty->print_cr("");
383 }
384
385 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr)
386 {
387 // Claim the right to put the region on the dirty cards region list
388 // by installing a self pointer.
389 HeapRegion* next = hr->get_next_dirty_cards_region();
390 if (next == NULL) {
391 HeapRegion* res = (HeapRegion*)
392 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(),
393 NULL);
394 if (res == NULL) {
395 HeapRegion* head;
396 do {
397 // Put the region to the dirty cards region list.
398 head = _dirty_cards_region_list;
399 next = (HeapRegion*)
400 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head);
401 if (next == head) {
402 assert(hr->get_next_dirty_cards_region() == hr,
403 "hr->get_next_dirty_cards_region() != hr");
404 if (next == NULL) {
405 // The last region in the list points to itself.
406 hr->set_next_dirty_cards_region(hr);
407 } else {
408 hr->set_next_dirty_cards_region(next);
409 }
410 }
411 } while (next != head);
412 }
413 }
414 }
415
416 HeapRegion* G1CollectedHeap::pop_dirty_cards_region()
417 {
418 HeapRegion* head;
419 HeapRegion* hr;
420 do {
421 head = _dirty_cards_region_list;
422 if (head == NULL) {
423 return NULL;
424 }
425 HeapRegion* new_head = head->get_next_dirty_cards_region();
426 if (head == new_head) {
427 // The last region.
428 new_head = NULL;
429 }
430 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list,
431 head);
432 } while (hr != head);
433 assert(hr != NULL, "invariant");
434 hr->set_next_dirty_cards_region(NULL);
435 return hr;
436 }
437
438 void G1CollectedHeap::stop_conc_gc_threads() {
439 _cg1r->stop();
440 _cmThread->stop();
441 }
442
443 #ifdef ASSERT
444 // A region is added to the collection set as it is retired
445 // so an address p can point to a region which will be in the
446 // collection set but has not yet been retired. This method
447 // therefore is only accurate during a GC pause after all
448 // regions have been retired. It is used for debugging
449 // to check if an nmethod has references to objects that can
450 // be move during a partial collection. Though it can be
451 // inaccurate, it is sufficient for G1 because the conservative
452 // implementation of is_scavengable() for G1 will indicate that
453 // all nmethods must be scanned during a partial collection.
454 bool G1CollectedHeap::is_in_partial_collection(const void* p) {
455 HeapRegion* hr = heap_region_containing(p);
456 return hr != NULL && hr->in_collection_set();
457 }
458 #endif
459
460 // Returns true if the reference points to an object that
461 // can move in an incremental collecction.
462 bool G1CollectedHeap::is_scavengable(const void* p) {
463 G1CollectedHeap* g1h = G1CollectedHeap::heap();
464 G1CollectorPolicy* g1p = g1h->g1_policy();
465 HeapRegion* hr = heap_region_containing(p);
466 if (hr == NULL) {
467 // null
468 assert(p == NULL, err_msg("Not NULL " PTR_FORMAT ,p));
469 return false;
470 } else {
471 return !hr->isHumongous();
472 }
473 }
474
475 void G1CollectedHeap::check_ct_logs_at_safepoint() {
476 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
477 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
478
479 // Count the dirty cards at the start.
480 CountNonCleanMemRegionClosure count1(this);
481 ct_bs->mod_card_iterate(&count1);
482 int orig_count = count1.n();
483
484 // First clear the logged cards.
485 ClearLoggedCardTableEntryClosure clear;
486 dcqs.set_closure(&clear);
487 dcqs.apply_closure_to_all_completed_buffers();
488 dcqs.iterate_closure_all_threads(false);
489 clear.print_histo();
490
491 // Now ensure that there's no dirty cards.
492 CountNonCleanMemRegionClosure count2(this);
493 ct_bs->mod_card_iterate(&count2);
494 if (count2.n() != 0) {
495 gclog_or_tty->print_cr("Card table has %d entries; %d originally",
496 count2.n(), orig_count);
497 }
498 guarantee(count2.n() == 0, "Card table should be clean.");
499
500 RedirtyLoggedCardTableEntryClosure redirty;
501 JavaThread::dirty_card_queue_set().set_closure(&redirty);
502 dcqs.apply_closure_to_all_completed_buffers();
503 dcqs.iterate_closure_all_threads(false);
504 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.",
505 clear.calls(), orig_count);
506 guarantee(redirty.calls() == clear.calls(),
507 "Or else mechanism is broken.");
508
509 CountNonCleanMemRegionClosure count3(this);
510 ct_bs->mod_card_iterate(&count3);
511 if (count3.n() != orig_count) {
512 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.",
513 orig_count, count3.n());
514 guarantee(count3.n() >= orig_count, "Should have restored them all.");
515 }
516
517 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
518 }
519
520 // Private class members.
521
522 G1CollectedHeap* G1CollectedHeap::_g1h;
523
524 // Private methods.
525
526 HeapRegion*
527 G1CollectedHeap::new_region_try_secondary_free_list() {
528 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
529 while (!_secondary_free_list.is_empty() || free_regions_coming()) {
530 if (!_secondary_free_list.is_empty()) {
531 if (G1ConcRegionFreeingVerbose) {
532 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
533 "secondary_free_list has %u entries",
534 _secondary_free_list.length());
535 }
536 // It looks as if there are free regions available on the
537 // secondary_free_list. Let's move them to the free_list and try
538 // again to allocate from it.
539 append_secondary_free_list();
540
541 assert(!_free_list.is_empty(), "if the secondary_free_list was not "
542 "empty we should have moved at least one entry to the free_list");
543 HeapRegion* res = _free_list.remove_head();
544 if (G1ConcRegionFreeingVerbose) {
545 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
546 "allocated "HR_FORMAT" from secondary_free_list",
547 HR_FORMAT_PARAMS(res));
548 }
549 return res;
550 }
551
552 // Wait here until we get notifed either when (a) there are no
553 // more free regions coming or (b) some regions have been moved on
554 // the secondary_free_list.
555 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
556 }
557
558 if (G1ConcRegionFreeingVerbose) {
559 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
560 "could not allocate from secondary_free_list");
561 }
562 return NULL;
563 }
564
565 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) {
566 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords,
567 "the only time we use this to allocate a humongous region is "
568 "when we are allocating a single humongous region");
569
570 HeapRegion* res;
571 if (G1StressConcRegionFreeing) {
572 if (!_secondary_free_list.is_empty()) {
573 if (G1ConcRegionFreeingVerbose) {
574 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
575 "forced to look at the secondary_free_list");
576 }
577 res = new_region_try_secondary_free_list();
578 if (res != NULL) {
579 return res;
580 }
581 }
582 }
583 res = _free_list.remove_head_or_null();
584 if (res == NULL) {
585 if (G1ConcRegionFreeingVerbose) {
586 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
587 "res == NULL, trying the secondary_free_list");
588 }
589 res = new_region_try_secondary_free_list();
590 }
591 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) {
592 // Currently, only attempts to allocate GC alloc regions set
593 // do_expand to true. So, we should only reach here during a
594 // safepoint. If this assumption changes we might have to
595 // reconsider the use of _expand_heap_after_alloc_failure.
596 assert(SafepointSynchronize::is_at_safepoint(), "invariant");
597
598 ergo_verbose1(ErgoHeapSizing,
599 "attempt heap expansion",
600 ergo_format_reason("region allocation request failed")
601 ergo_format_byte("allocation request"),
602 word_size * HeapWordSize);
603 if (expand(word_size * HeapWordSize)) {
604 // Given that expand() succeeded in expanding the heap, and we
605 // always expand the heap by an amount aligned to the heap
606 // region size, the free list should in theory not be empty. So
607 // it would probably be OK to use remove_head(). But the extra
608 // check for NULL is unlikely to be a performance issue here (we
609 // just expanded the heap!) so let's just be conservative and
610 // use remove_head_or_null().
611 res = _free_list.remove_head_or_null();
612 } else {
613 _expand_heap_after_alloc_failure = false;
614 }
615 }
616 return res;
617 }
618
619 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions,
620 size_t word_size) {
621 assert(isHumongous(word_size), "word_size should be humongous");
622 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
623
624 uint first = G1_NULL_HRS_INDEX;
625 if (num_regions == 1) {
626 // Only one region to allocate, no need to go through the slower
627 // path. The caller will attempt the expasion if this fails, so
628 // let's not try to expand here too.
629 HeapRegion* hr = new_region(word_size, false /* do_expand */);
630 if (hr != NULL) {
631 first = hr->hrs_index();
632 } else {
633 first = G1_NULL_HRS_INDEX;
634 }
635 } else {
636 // We can't allocate humongous regions while cleanupComplete() is
637 // running, since some of the regions we find to be empty might not
638 // yet be added to the free list and it is not straightforward to
639 // know which list they are on so that we can remove them. Note
640 // that we only need to do this if we need to allocate more than
641 // one region to satisfy the current humongous allocation
642 // request. If we are only allocating one region we use the common
643 // region allocation code (see above).
644 wait_while_free_regions_coming();
645 append_secondary_free_list_if_not_empty_with_lock();
646
647 if (free_regions() >= num_regions) {
648 first = _hrs.find_contiguous(num_regions);
649 if (first != G1_NULL_HRS_INDEX) {
650 for (uint i = first; i < first + num_regions; ++i) {
651 HeapRegion* hr = region_at(i);
652 assert(hr->is_empty(), "sanity");
653 assert(is_on_master_free_list(hr), "sanity");
654 hr->set_pending_removal(true);
655 }
656 _free_list.remove_all_pending(num_regions);
657 }
658 }
659 }
660 return first;
661 }
662
663 HeapWord*
664 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first,
665 uint num_regions,
666 size_t word_size) {
667 assert(first != G1_NULL_HRS_INDEX, "pre-condition");
668 assert(isHumongous(word_size), "word_size should be humongous");
669 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
670
671 // Index of last region in the series + 1.
672 uint last = first + num_regions;
673
674 // We need to initialize the region(s) we just discovered. This is
675 // a bit tricky given that it can happen concurrently with
676 // refinement threads refining cards on these regions and
677 // potentially wanting to refine the BOT as they are scanning
678 // those cards (this can happen shortly after a cleanup; see CR
679 // 6991377). So we have to set up the region(s) carefully and in
680 // a specific order.
681
682 // The word size sum of all the regions we will allocate.
683 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords;
684 assert(word_size <= word_size_sum, "sanity");
685
686 // This will be the "starts humongous" region.
687 HeapRegion* first_hr = region_at(first);
688 // The header of the new object will be placed at the bottom of
689 // the first region.
690 HeapWord* new_obj = first_hr->bottom();
691 // This will be the new end of the first region in the series that
692 // should also match the end of the last region in the seriers.
693 HeapWord* new_end = new_obj + word_size_sum;
694 // This will be the new top of the first region that will reflect
695 // this allocation.
696 HeapWord* new_top = new_obj + word_size;
697
698 // First, we need to zero the header of the space that we will be
699 // allocating. When we update top further down, some refinement
700 // threads might try to scan the region. By zeroing the header we
701 // ensure that any thread that will try to scan the region will
702 // come across the zero klass word and bail out.
703 //
704 // NOTE: It would not have been correct to have used
705 // CollectedHeap::fill_with_object() and make the space look like
706 // an int array. The thread that is doing the allocation will
707 // later update the object header to a potentially different array
708 // type and, for a very short period of time, the klass and length
709 // fields will be inconsistent. This could cause a refinement
710 // thread to calculate the object size incorrectly.
711 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0);
712
713 // We will set up the first region as "starts humongous". This
714 // will also update the BOT covering all the regions to reflect
715 // that there is a single object that starts at the bottom of the
716 // first region.
717 first_hr->set_startsHumongous(new_top, new_end);
718
719 // Then, if there are any, we will set up the "continues
720 // humongous" regions.
721 HeapRegion* hr = NULL;
722 for (uint i = first + 1; i < last; ++i) {
723 hr = region_at(i);
724 hr->set_continuesHumongous(first_hr);
725 }
726 // If we have "continues humongous" regions (hr != NULL), then the
727 // end of the last one should match new_end.
728 assert(hr == NULL || hr->end() == new_end, "sanity");
729
730 // Up to this point no concurrent thread would have been able to
731 // do any scanning on any region in this series. All the top
732 // fields still point to bottom, so the intersection between
733 // [bottom,top] and [card_start,card_end] will be empty. Before we
734 // update the top fields, we'll do a storestore to make sure that
735 // no thread sees the update to top before the zeroing of the
736 // object header and the BOT initialization.
737 OrderAccess::storestore();
738
739 // Now that the BOT and the object header have been initialized,
740 // we can update top of the "starts humongous" region.
741 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(),
742 "new_top should be in this region");
743 first_hr->set_top(new_top);
744 if (_hr_printer.is_active()) {
745 HeapWord* bottom = first_hr->bottom();
746 HeapWord* end = first_hr->orig_end();
747 if ((first + 1) == last) {
748 // the series has a single humongous region
749 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top);
750 } else {
751 // the series has more than one humongous regions
752 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end);
753 }
754 }
755
756 // Now, we will update the top fields of the "continues humongous"
757 // regions. The reason we need to do this is that, otherwise,
758 // these regions would look empty and this will confuse parts of
759 // G1. For example, the code that looks for a consecutive number
760 // of empty regions will consider them empty and try to
761 // re-allocate them. We can extend is_empty() to also include
762 // !continuesHumongous(), but it is easier to just update the top
763 // fields here. The way we set top for all regions (i.e., top ==
764 // end for all regions but the last one, top == new_top for the
765 // last one) is actually used when we will free up the humongous
766 // region in free_humongous_region().
767 hr = NULL;
768 for (uint i = first + 1; i < last; ++i) {
769 hr = region_at(i);
770 if ((i + 1) == last) {
771 // last continues humongous region
772 assert(hr->bottom() < new_top && new_top <= hr->end(),
773 "new_top should fall on this region");
774 hr->set_top(new_top);
775 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top);
776 } else {
777 // not last one
778 assert(new_top > hr->end(), "new_top should be above this region");
779 hr->set_top(hr->end());
780 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end());
781 }
782 }
783 // If we have continues humongous regions (hr != NULL), then the
784 // end of the last one should match new_end and its top should
785 // match new_top.
786 assert(hr == NULL ||
787 (hr->end() == new_end && hr->top() == new_top), "sanity");
788
789 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
790 _summary_bytes_used += first_hr->used();
791 _humongous_set.add(first_hr);
792
793 return new_obj;
794 }
795
796 // If could fit into free regions w/o expansion, try.
797 // Otherwise, if can expand, do so.
798 // Otherwise, if using ex regions might help, try with ex given back.
799 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) {
800 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
801
802 verify_region_sets_optional();
803
804 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords);
805 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords);
806 uint x_num = expansion_regions();
807 uint fs = _hrs.free_suffix();
808 uint first = humongous_obj_allocate_find_first(num_regions, word_size);
809 if (first == G1_NULL_HRS_INDEX) {
810 // The only thing we can do now is attempt expansion.
811 if (fs + x_num >= num_regions) {
812 // If the number of regions we're trying to allocate for this
813 // object is at most the number of regions in the free suffix,
814 // then the call to humongous_obj_allocate_find_first() above
815 // should have succeeded and we wouldn't be here.
816 //
817 // We should only be trying to expand when the free suffix is
818 // not sufficient for the object _and_ we have some expansion
819 // room available.
820 assert(num_regions > fs, "earlier allocation should have succeeded");
821
822 ergo_verbose1(ErgoHeapSizing,
823 "attempt heap expansion",
824 ergo_format_reason("humongous allocation request failed")
825 ergo_format_byte("allocation request"),
826 word_size * HeapWordSize);
827 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) {
828 // Even though the heap was expanded, it might not have
829 // reached the desired size. So, we cannot assume that the
830 // allocation will succeed.
831 first = humongous_obj_allocate_find_first(num_regions, word_size);
832 }
833 }
834 }
835
836 HeapWord* result = NULL;
837 if (first != G1_NULL_HRS_INDEX) {
838 result =
839 humongous_obj_allocate_initialize_regions(first, num_regions, word_size);
840 assert(result != NULL, "it should always return a valid result");
841
842 // A successful humongous object allocation changes the used space
843 // information of the old generation so we need to recalculate the
844 // sizes and update the jstat counters here.
845 g1mm()->update_sizes();
846 }
847
848 verify_region_sets_optional();
849
850 return result;
851 }
852
853 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) {
854 assert_heap_not_locked_and_not_at_safepoint();
855 assert(!isHumongous(word_size), "we do not allow humongous TLABs");
856
857 unsigned int dummy_gc_count_before;
858 return attempt_allocation(word_size, &dummy_gc_count_before);
859 }
860
861 HeapWord*
862 G1CollectedHeap::mem_allocate(size_t word_size,
863 bool* gc_overhead_limit_was_exceeded) {
864 assert_heap_not_locked_and_not_at_safepoint();
865
866 // Loop until the allocation is satisified, or unsatisfied after GC.
867 for (int try_count = 1; /* we'll return */; try_count += 1) {
868 unsigned int gc_count_before;
869
870 HeapWord* result = NULL;
871 if (!isHumongous(word_size)) {
872 result = attempt_allocation(word_size, &gc_count_before);
873 } else {
874 result = attempt_allocation_humongous(word_size, &gc_count_before);
875 }
876 if (result != NULL) {
877 return result;
878 }
879
880 // Create the garbage collection operation...
881 VM_G1CollectForAllocation op(gc_count_before, word_size);
882 // ...and get the VM thread to execute it.
883 VMThread::execute(&op);
884
885 if (op.prologue_succeeded() && op.pause_succeeded()) {
886 // If the operation was successful we'll return the result even
887 // if it is NULL. If the allocation attempt failed immediately
888 // after a Full GC, it's unlikely we'll be able to allocate now.
889 HeapWord* result = op.result();
890 if (result != NULL && !isHumongous(word_size)) {
891 // Allocations that take place on VM operations do not do any
892 // card dirtying and we have to do it here. We only have to do
893 // this for non-humongous allocations, though.
894 dirty_young_block(result, word_size);
895 }
896 return result;
897 } else {
898 assert(op.result() == NULL,
899 "the result should be NULL if the VM op did not succeed");
900 }
901
902 // Give a warning if we seem to be looping forever.
903 if ((QueuedAllocationWarningCount > 0) &&
904 (try_count % QueuedAllocationWarningCount == 0)) {
905 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
906 }
907 }
908
909 ShouldNotReachHere();
910 return NULL;
911 }
912
913 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
914 unsigned int *gc_count_before_ret) {
915 // Make sure you read the note in attempt_allocation_humongous().
916
917 assert_heap_not_locked_and_not_at_safepoint();
918 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
919 "be called for humongous allocation requests");
920
921 // We should only get here after the first-level allocation attempt
922 // (attempt_allocation()) failed to allocate.
923
924 // We will loop until a) we manage to successfully perform the
925 // allocation or b) we successfully schedule a collection which
926 // fails to perform the allocation. b) is the only case when we'll
927 // return NULL.
928 HeapWord* result = NULL;
929 for (int try_count = 1; /* we'll return */; try_count += 1) {
930 bool should_try_gc;
931 unsigned int gc_count_before;
932
933 {
934 MutexLockerEx x(Heap_lock);
935
936 result = _mutator_alloc_region.attempt_allocation_locked(word_size,
937 false /* bot_updates */);
938 if (result != NULL) {
939 return result;
940 }
941
942 // If we reach here, attempt_allocation_locked() above failed to
943 // allocate a new region. So the mutator alloc region should be NULL.
944 assert(_mutator_alloc_region.get() == NULL, "only way to get here");
945
946 if (GC_locker::is_active_and_needs_gc()) {
947 if (g1_policy()->can_expand_young_list()) {
948 // No need for an ergo verbose message here,
949 // can_expand_young_list() does this when it returns true.
950 result = _mutator_alloc_region.attempt_allocation_force(word_size,
951 false /* bot_updates */);
952 if (result != NULL) {
953 return result;
954 }
955 }
956 should_try_gc = false;
957 } else {
958 // The GCLocker may not be active but the GCLocker initiated
959 // GC may not yet have been performed (GCLocker::needs_gc()
960 // returns true). In this case we do not try this GC and
961 // wait until the GCLocker initiated GC is performed, and
962 // then retry the allocation.
963 if (GC_locker::needs_gc()) {
964 should_try_gc = false;
965 } else {
966 // Read the GC count while still holding the Heap_lock.
967 gc_count_before = total_collections();
968 should_try_gc = true;
969 }
970 }
971 }
972
973 if (should_try_gc) {
974 bool succeeded;
975 result = do_collection_pause(word_size, gc_count_before, &succeeded);
976 if (result != NULL) {
977 assert(succeeded, "only way to get back a non-NULL result");
978 return result;
979 }
980
981 if (succeeded) {
982 // If we get here we successfully scheduled a collection which
983 // failed to allocate. No point in trying to allocate
984 // further. We'll just return NULL.
985 MutexLockerEx x(Heap_lock);
986 *gc_count_before_ret = total_collections();
987 return NULL;
988 }
989 } else {
990 // The GCLocker is either active or the GCLocker initiated
991 // GC has not yet been performed. Stall until it is and
992 // then retry the allocation.
993 GC_locker::stall_until_clear();
994 }
995
996 // We can reach here if we were unsuccessul in scheduling a
997 // collection (because another thread beat us to it) or if we were
998 // stalled due to the GC locker. In either can we should retry the
999 // allocation attempt in case another thread successfully
1000 // performed a collection and reclaimed enough space. We do the
1001 // first attempt (without holding the Heap_lock) here and the
1002 // follow-on attempt will be at the start of the next loop
1003 // iteration (after taking the Heap_lock).
1004 result = _mutator_alloc_region.attempt_allocation(word_size,
1005 false /* bot_updates */);
1006 if (result != NULL) {
1007 return result;
1008 }
1009
1010 // Give a warning if we seem to be looping forever.
1011 if ((QueuedAllocationWarningCount > 0) &&
1012 (try_count % QueuedAllocationWarningCount == 0)) {
1013 warning("G1CollectedHeap::attempt_allocation_slow() "
1014 "retries %d times", try_count);
1015 }
1016 }
1017
1018 ShouldNotReachHere();
1019 return NULL;
1020 }
1021
1022 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
1023 unsigned int * gc_count_before_ret) {
1024 // The structure of this method has a lot of similarities to
1025 // attempt_allocation_slow(). The reason these two were not merged
1026 // into a single one is that such a method would require several "if
1027 // allocation is not humongous do this, otherwise do that"
1028 // conditional paths which would obscure its flow. In fact, an early
1029 // version of this code did use a unified method which was harder to
1030 // follow and, as a result, it had subtle bugs that were hard to
1031 // track down. So keeping these two methods separate allows each to
1032 // be more readable. It will be good to keep these two in sync as
1033 // much as possible.
1034
1035 assert_heap_not_locked_and_not_at_safepoint();
1036 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1037 "should only be called for humongous allocations");
1038
1039 // Humongous objects can exhaust the heap quickly, so we should check if we
1040 // need to start a marking cycle at each humongous object allocation. We do
1041 // the check before we do the actual allocation. The reason for doing it
1042 // before the allocation is that we avoid having to keep track of the newly
1043 // allocated memory while we do a GC.
1044 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation",
1045 word_size)) {
1046 collect(GCCause::_g1_humongous_allocation);
1047 }
1048
1049 // We will loop until a) we manage to successfully perform the
1050 // allocation or b) we successfully schedule a collection which
1051 // fails to perform the allocation. b) is the only case when we'll
1052 // return NULL.
1053 HeapWord* result = NULL;
1054 for (int try_count = 1; /* we'll return */; try_count += 1) {
1055 bool should_try_gc;
1056 unsigned int gc_count_before;
1057
1058 {
1059 MutexLockerEx x(Heap_lock);
1060
1061 // Given that humongous objects are not allocated in young
1062 // regions, we'll first try to do the allocation without doing a
1063 // collection hoping that there's enough space in the heap.
1064 result = humongous_obj_allocate(word_size);
1065 if (result != NULL) {
1066 return result;
1067 }
1068
1069 if (GC_locker::is_active_and_needs_gc()) {
1070 should_try_gc = false;
1071 } else {
1072 // The GCLocker may not be active but the GCLocker initiated
1073 // GC may not yet have been performed (GCLocker::needs_gc()
1074 // returns true). In this case we do not try this GC and
1075 // wait until the GCLocker initiated GC is performed, and
1076 // then retry the allocation.
1077 if (GC_locker::needs_gc()) {
1078 should_try_gc = false;
1079 } else {
1080 // Read the GC count while still holding the Heap_lock.
1081 gc_count_before = total_collections();
1082 should_try_gc = true;
1083 }
1084 }
1085 }
1086
1087 if (should_try_gc) {
1088 // If we failed to allocate the humongous object, we should try to
1089 // do a collection pause (if we're allowed) in case it reclaims
1090 // enough space for the allocation to succeed after the pause.
1091
1092 bool succeeded;
1093 result = do_collection_pause(word_size, gc_count_before, &succeeded);
1094 if (result != NULL) {
1095 assert(succeeded, "only way to get back a non-NULL result");
1096 return result;
1097 }
1098
1099 if (succeeded) {
1100 // If we get here we successfully scheduled a collection which
1101 // failed to allocate. No point in trying to allocate
1102 // further. We'll just return NULL.
1103 MutexLockerEx x(Heap_lock);
1104 *gc_count_before_ret = total_collections();
1105 return NULL;
1106 }
1107 } else {
1108 // The GCLocker is either active or the GCLocker initiated
1109 // GC has not yet been performed. Stall until it is and
1110 // then retry the allocation.
1111 GC_locker::stall_until_clear();
1112 }
1113
1114 // We can reach here if we were unsuccessul in scheduling a
1115 // collection (because another thread beat us to it) or if we were
1116 // stalled due to the GC locker. In either can we should retry the
1117 // allocation attempt in case another thread successfully
1118 // performed a collection and reclaimed enough space. Give a
1119 // warning if we seem to be looping forever.
1120
1121 if ((QueuedAllocationWarningCount > 0) &&
1122 (try_count % QueuedAllocationWarningCount == 0)) {
1123 warning("G1CollectedHeap::attempt_allocation_humongous() "
1124 "retries %d times", try_count);
1125 }
1126 }
1127
1128 ShouldNotReachHere();
1129 return NULL;
1130 }
1131
1132 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1133 bool expect_null_mutator_alloc_region) {
1134 assert_at_safepoint(true /* should_be_vm_thread */);
1135 assert(_mutator_alloc_region.get() == NULL ||
1136 !expect_null_mutator_alloc_region,
1137 "the current alloc region was unexpectedly found to be non-NULL");
1138
1139 if (!isHumongous(word_size)) {
1140 return _mutator_alloc_region.attempt_allocation_locked(word_size,
1141 false /* bot_updates */);
1142 } else {
1143 HeapWord* result = humongous_obj_allocate(word_size);
1144 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) {
1145 g1_policy()->set_initiate_conc_mark_if_possible();
1146 }
1147 return result;
1148 }
1149
1150 ShouldNotReachHere();
1151 }
1152
1153 class PostMCRemSetClearClosure: public HeapRegionClosure {
1154 G1CollectedHeap* _g1h;
1155 ModRefBarrierSet* _mr_bs;
1156 public:
1157 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) :
1158 _g1h(g1h), _mr_bs(mr_bs) { }
1159 bool doHeapRegion(HeapRegion* r) {
1160 if (r->continuesHumongous()) {
1161 return false;
1162 }
1163 _g1h->reset_gc_time_stamps(r);
1164 HeapRegionRemSet* hrrs = r->rem_set();
1165 if (hrrs != NULL) hrrs->clear();
1166 // You might think here that we could clear just the cards
1167 // corresponding to the used region. But no: if we leave a dirty card
1168 // in a region we might allocate into, then it would prevent that card
1169 // from being enqueued, and cause it to be missed.
1170 // Re: the performance cost: we shouldn't be doing full GC anyway!
1171 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1172 return false;
1173 }
1174 };
1175
1176 void G1CollectedHeap::clear_rsets_post_compaction() {
1177 PostMCRemSetClearClosure rs_clear(this, mr_bs());
1178 heap_region_iterate(&rs_clear);
1179 }
1180
1181 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1182 G1CollectedHeap* _g1h;
1183 UpdateRSOopClosure _cl;
1184 int _worker_i;
1185 public:
1186 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1187 _cl(g1->g1_rem_set(), worker_i),
1188 _worker_i(worker_i),
1189 _g1h(g1)
1190 { }
1191
1192 bool doHeapRegion(HeapRegion* r) {
1193 if (!r->continuesHumongous()) {
1194 _cl.set_from(r);
1195 r->oop_iterate(&_cl);
1196 }
1197 return false;
1198 }
1199 };
1200
1201 class ParRebuildRSTask: public AbstractGangTask {
1202 G1CollectedHeap* _g1;
1203 public:
1204 ParRebuildRSTask(G1CollectedHeap* g1)
1205 : AbstractGangTask("ParRebuildRSTask"),
1206 _g1(g1)
1207 { }
1208
1209 void work(uint worker_id) {
1210 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id);
1211 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id,
1212 _g1->workers()->active_workers(),
1213 HeapRegion::RebuildRSClaimValue);
1214 }
1215 };
1216
1217 class PostCompactionPrinterClosure: public HeapRegionClosure {
1218 private:
1219 G1HRPrinter* _hr_printer;
1220 public:
1221 bool doHeapRegion(HeapRegion* hr) {
1222 assert(!hr->is_young(), "not expecting to find young regions");
1223 // We only generate output for non-empty regions.
1224 if (!hr->is_empty()) {
1225 if (!hr->isHumongous()) {
1226 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1227 } else if (hr->startsHumongous()) {
1228 if (hr->region_num() == 1) {
1229 // single humongous region
1230 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1231 } else {
1232 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1233 }
1234 } else {
1235 assert(hr->continuesHumongous(), "only way to get here");
1236 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1237 }
1238 }
1239 return false;
1240 }
1241
1242 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1243 : _hr_printer(hr_printer) { }
1244 };
1245
1246 void G1CollectedHeap::print_hrs_post_compaction() {
1247 PostCompactionPrinterClosure cl(hr_printer());
1248 heap_region_iterate(&cl);
1249 }
1250
1251 double G1CollectedHeap::verify(bool guard, const char* msg) {
1252 double verify_time_ms = 0.0;
1253
1254 if (guard && total_collections() >= VerifyGCStartAt) {
1255 double verify_start = os::elapsedTime();
1256 HandleMark hm; // Discard invalid handles created during verification
1257 gclog_or_tty->print(msg);
1258 prepare_for_verify();
1259 Universe::verify(false /* silent */, VerifyOption_G1UsePrevMarking);
1260 verify_time_ms = (os::elapsedTime() - verify_start) * 1000;
1261 }
1262
1263 return verify_time_ms;
1264 }
1265
1266 void G1CollectedHeap::verify_before_gc() {
1267 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:");
1268 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms);
1269 }
1270
1271 void G1CollectedHeap::verify_after_gc() {
1272 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:");
1273 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms);
1274 }
1275
1276 bool G1CollectedHeap::do_collection(bool explicit_gc,
1277 bool clear_all_soft_refs,
1278 size_t word_size) {
1279 assert_at_safepoint(true /* should_be_vm_thread */);
1280
1281 if (GC_locker::check_active_before_gc()) {
1282 return false;
1283 }
1284
1285 SvcGCMarker sgcm(SvcGCMarker::FULL);
1286 ResourceMark rm;
1287
1288 print_heap_before_gc();
1289
1290 size_t metadata_prev_used = MetaspaceAux::used_in_bytes();
1291
1292 HRSPhaseSetter x(HRSPhaseFullGC);
1293 verify_region_sets_optional();
1294
1295 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1296 collector_policy()->should_clear_all_soft_refs();
1297
1298 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1299
1300 {
1301 IsGCActiveMark x;
1302
1303 // Timing
1304 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant");
1305 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps);
1306 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
1307
1308 TraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, gclog_or_tty);
1309 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1310 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1311
1312 double start = os::elapsedTime();
1313 g1_policy()->record_full_collection_start();
1314
1315 // Note: When we have a more flexible GC logging framework that
1316 // allows us to add optional attributes to a GC log record we
1317 // could consider timing and reporting how long we wait in the
1318 // following two methods.
1319 wait_while_free_regions_coming();
1320 // If we start the compaction before the CM threads finish
1321 // scanning the root regions we might trip them over as we'll
1322 // be moving objects / updating references. So let's wait until
1323 // they are done. By telling them to abort, they should complete
1324 // early.
1325 _cm->root_regions()->abort();
1326 _cm->root_regions()->wait_until_scan_finished();
1327 append_secondary_free_list_if_not_empty_with_lock();
1328
1329 gc_prologue(true);
1330 increment_total_collections(true /* full gc */);
1331 increment_old_marking_cycles_started();
1332
1333 size_t g1h_prev_used = used();
1334 assert(used() == recalculate_used(), "Should be equal");
1335
1336 verify_before_gc();
1337
1338 pre_full_gc_dump();
1339
1340 COMPILER2_PRESENT(DerivedPointerTable::clear());
1341
1342 // Disable discovery and empty the discovered lists
1343 // for the CM ref processor.
1344 ref_processor_cm()->disable_discovery();
1345 ref_processor_cm()->abandon_partial_discovery();
1346 ref_processor_cm()->verify_no_references_recorded();
1347
1348 // Abandon current iterations of concurrent marking and concurrent
1349 // refinement, if any are in progress. We have to do this before
1350 // wait_until_scan_finished() below.
1351 concurrent_mark()->abort();
1352
1353 // Make sure we'll choose a new allocation region afterwards.
1354 release_mutator_alloc_region();
1355 abandon_gc_alloc_regions();
1356 g1_rem_set()->cleanupHRRS();
1357
1358 // We should call this after we retire any currently active alloc
1359 // regions so that all the ALLOC / RETIRE events are generated
1360 // before the start GC event.
1361 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1362
1363 // We may have added regions to the current incremental collection
1364 // set between the last GC or pause and now. We need to clear the
1365 // incremental collection set and then start rebuilding it afresh
1366 // after this full GC.
1367 abandon_collection_set(g1_policy()->inc_cset_head());
1368 g1_policy()->clear_incremental_cset();
1369 g1_policy()->stop_incremental_cset_building();
1370
1371 tear_down_region_sets(false /* free_list_only */);
1372 g1_policy()->set_gcs_are_young(true);
1373
1374 // See the comments in g1CollectedHeap.hpp and
1375 // G1CollectedHeap::ref_processing_init() about
1376 // how reference processing currently works in G1.
1377
1378 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1379 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1380
1381 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1382 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1383
1384 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1385 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1386
1387 // Do collection work
1388 {
1389 HandleMark hm; // Discard invalid handles created during gc
1390 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1391 }
1392
1393 assert(free_regions() == 0, "we should not have added any free regions");
1394 rebuild_region_sets(false /* free_list_only */);
1395
1396 // Enqueue any discovered reference objects that have
1397 // not been removed from the discovered lists.
1398 ref_processor_stw()->enqueue_discovered_references();
1399
1400 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1401
1402 MemoryService::track_memory_usage();
1403
1404 verify_after_gc();
1405
1406 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1407 ref_processor_stw()->verify_no_references_recorded();
1408
1409 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1410 ClassLoaderDataGraph::purge();
1411
1412 // Note: since we've just done a full GC, concurrent
1413 // marking is no longer active. Therefore we need not
1414 // re-enable reference discovery for the CM ref processor.
1415 // That will be done at the start of the next marking cycle.
1416 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1417 ref_processor_cm()->verify_no_references_recorded();
1418
1419 reset_gc_time_stamp();
1420 // Since everything potentially moved, we will clear all remembered
1421 // sets, and clear all cards. Later we will rebuild remebered
1422 // sets. We will also reset the GC time stamps of the regions.
1423 clear_rsets_post_compaction();
1424 check_gc_time_stamps();
1425
1426 // Resize the heap if necessary.
1427 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1428
1429 if (_hr_printer.is_active()) {
1430 // We should do this after we potentially resize the heap so
1431 // that all the COMMIT / UNCOMMIT events are generated before
1432 // the end GC event.
1433
1434 print_hrs_post_compaction();
1435 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1436 }
1437
1438 if (_cg1r->use_cache()) {
1439 _cg1r->clear_and_record_card_counts();
1440 _cg1r->clear_hot_cache();
1441 }
1442
1443 // Rebuild remembered sets of all regions.
1444 if (G1CollectedHeap::use_parallel_gc_threads()) {
1445 uint n_workers =
1446 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
1447 workers()->active_workers(),
1448 Threads::number_of_non_daemon_threads());
1449 assert(UseDynamicNumberOfGCThreads ||
1450 n_workers == workers()->total_workers(),
1451 "If not dynamic should be using all the workers");
1452 workers()->set_active_workers(n_workers);
1453 // Set parallel threads in the heap (_n_par_threads) only
1454 // before a parallel phase and always reset it to 0 after
1455 // the phase so that the number of parallel threads does
1456 // no get carried forward to a serial phase where there
1457 // may be code that is "possibly_parallel".
1458 set_par_threads(n_workers);
1459
1460 ParRebuildRSTask rebuild_rs_task(this);
1461 assert(check_heap_region_claim_values(
1462 HeapRegion::InitialClaimValue), "sanity check");
1463 assert(UseDynamicNumberOfGCThreads ||
1464 workers()->active_workers() == workers()->total_workers(),
1465 "Unless dynamic should use total workers");
1466 // Use the most recent number of active workers
1467 assert(workers()->active_workers() > 0,
1468 "Active workers not properly set");
1469 set_par_threads(workers()->active_workers());
1470 workers()->run_task(&rebuild_rs_task);
1471 set_par_threads(0);
1472 assert(check_heap_region_claim_values(
1473 HeapRegion::RebuildRSClaimValue), "sanity check");
1474 reset_heap_region_claim_values();
1475 } else {
1476 RebuildRSOutOfRegionClosure rebuild_rs(this);
1477 heap_region_iterate(&rebuild_rs);
1478 }
1479
1480 if (G1Log::fine()) {
1481 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity());
1482 }
1483
1484 if (true) { // FIXME
1485 MetaspaceGC::compute_new_size();
1486 }
1487
1488 // Start a new incremental collection set for the next pause
1489 assert(g1_policy()->collection_set() == NULL, "must be");
1490 g1_policy()->start_incremental_cset_building();
1491
1492 // Clear the _cset_fast_test bitmap in anticipation of adding
1493 // regions to the incremental collection set for the next
1494 // evacuation pause.
1495 clear_cset_fast_test();
1496
1497 init_mutator_alloc_region();
1498
1499 double end = os::elapsedTime();
1500 g1_policy()->record_full_collection_end();
1501
1502 #ifdef TRACESPINNING
1503 ParallelTaskTerminator::print_termination_counts();
1504 #endif
1505
1506 gc_epilogue(true);
1507
1508 // Discard all rset updates
1509 JavaThread::dirty_card_queue_set().abandon_logs();
1510 assert(!G1DeferredRSUpdate
1511 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
1512
1513 _young_list->reset_sampled_info();
1514 // At this point there should be no regions in the
1515 // entire heap tagged as young.
1516 assert( check_young_list_empty(true /* check_heap */),
1517 "young list should be empty at this point");
1518
1519 // Update the number of full collections that have been completed.
1520 increment_old_marking_cycles_completed(false /* concurrent */);
1521
1522 _hrs.verify_optional();
1523 verify_region_sets_optional();
1524
1525 print_heap_after_gc();
1526
1527 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
1528 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
1529 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
1530 // before any GC notifications are raised.
1531 g1mm()->update_sizes();
1532 }
1533
1534 post_full_gc_dump();
1535
1536 return true;
1537 }
1538
1539 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1540 // do_collection() will return whether it succeeded in performing
1541 // the GC. Currently, there is no facility on the
1542 // do_full_collection() API to notify the caller than the collection
1543 // did not succeed (e.g., because it was locked out by the GC
1544 // locker). So, right now, we'll ignore the return value.
1545 bool dummy = do_collection(true, /* explicit_gc */
1546 clear_all_soft_refs,
1547 0 /* word_size */);
1548 }
1549
1550 // This code is mostly copied from TenuredGeneration.
1551 void
1552 G1CollectedHeap::
1553 resize_if_necessary_after_full_collection(size_t word_size) {
1554 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check");
1555
1556 // Include the current allocation, if any, and bytes that will be
1557 // pre-allocated to support collections, as "used".
1558 const size_t used_after_gc = used();
1559 const size_t capacity_after_gc = capacity();
1560 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1561
1562 // This is enforced in arguments.cpp.
1563 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1564 "otherwise the code below doesn't make sense");
1565
1566 // We don't have floating point command-line arguments
1567 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1568 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1569 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1570 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1571
1572 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1573 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1574
1575 // We have to be careful here as these two calculations can overflow
1576 // 32-bit size_t's.
1577 double used_after_gc_d = (double) used_after_gc;
1578 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1579 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1580
1581 // Let's make sure that they are both under the max heap size, which
1582 // by default will make them fit into a size_t.
1583 double desired_capacity_upper_bound = (double) max_heap_size;
1584 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1585 desired_capacity_upper_bound);
1586 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1587 desired_capacity_upper_bound);
1588
1589 // We can now safely turn them into size_t's.
1590 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1591 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1592
1593 // This assert only makes sense here, before we adjust them
1594 // with respect to the min and max heap size.
1595 assert(minimum_desired_capacity <= maximum_desired_capacity,
1596 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1597 "maximum_desired_capacity = "SIZE_FORMAT,
1598 minimum_desired_capacity, maximum_desired_capacity));
1599
1600 // Should not be greater than the heap max size. No need to adjust
1601 // it with respect to the heap min size as it's a lower bound (i.e.,
1602 // we'll try to make the capacity larger than it, not smaller).
1603 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1604 // Should not be less than the heap min size. No need to adjust it
1605 // with respect to the heap max size as it's an upper bound (i.e.,
1606 // we'll try to make the capacity smaller than it, not greater).
1607 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1608
1609 if (capacity_after_gc < minimum_desired_capacity) {
1610 // Don't expand unless it's significant
1611 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1612 ergo_verbose4(ErgoHeapSizing,
1613 "attempt heap expansion",
1614 ergo_format_reason("capacity lower than "
1615 "min desired capacity after Full GC")
1616 ergo_format_byte("capacity")
1617 ergo_format_byte("occupancy")
1618 ergo_format_byte_perc("min desired capacity"),
1619 capacity_after_gc, used_after_gc,
1620 minimum_desired_capacity, (double) MinHeapFreeRatio);
1621 expand(expand_bytes);
1622
1623 // No expansion, now see if we want to shrink
1624 } else if (capacity_after_gc > maximum_desired_capacity) {
1625 // Capacity too large, compute shrinking size
1626 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1627 ergo_verbose4(ErgoHeapSizing,
1628 "attempt heap shrinking",
1629 ergo_format_reason("capacity higher than "
1630 "max desired capacity after Full GC")
1631 ergo_format_byte("capacity")
1632 ergo_format_byte("occupancy")
1633 ergo_format_byte_perc("max desired capacity"),
1634 capacity_after_gc, used_after_gc,
1635 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1636 shrink(shrink_bytes);
1637 }
1638 }
1639
1640
1641 HeapWord*
1642 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1643 bool* succeeded) {
1644 assert_at_safepoint(true /* should_be_vm_thread */);
1645
1646 *succeeded = true;
1647 // Let's attempt the allocation first.
1648 HeapWord* result =
1649 attempt_allocation_at_safepoint(word_size,
1650 false /* expect_null_mutator_alloc_region */);
1651 if (result != NULL) {
1652 assert(*succeeded, "sanity");
1653 return result;
1654 }
1655
1656 // In a G1 heap, we're supposed to keep allocation from failing by
1657 // incremental pauses. Therefore, at least for now, we'll favor
1658 // expansion over collection. (This might change in the future if we can
1659 // do something smarter than full collection to satisfy a failed alloc.)
1660 result = expand_and_allocate(word_size);
1661 if (result != NULL) {
1662 assert(*succeeded, "sanity");
1663 return result;
1664 }
1665
1666 // Expansion didn't work, we'll try to do a Full GC.
1667 bool gc_succeeded = do_collection(false, /* explicit_gc */
1668 false, /* clear_all_soft_refs */
1669 word_size);
1670 if (!gc_succeeded) {
1671 *succeeded = false;
1672 return NULL;
1673 }
1674
1675 // Retry the allocation
1676 result = attempt_allocation_at_safepoint(word_size,
1677 true /* expect_null_mutator_alloc_region */);
1678 if (result != NULL) {
1679 assert(*succeeded, "sanity");
1680 return result;
1681 }
1682
1683 // Then, try a Full GC that will collect all soft references.
1684 gc_succeeded = do_collection(false, /* explicit_gc */
1685 true, /* clear_all_soft_refs */
1686 word_size);
1687 if (!gc_succeeded) {
1688 *succeeded = false;
1689 return NULL;
1690 }
1691
1692 // Retry the allocation once more
1693 result = attempt_allocation_at_safepoint(word_size,
1694 true /* expect_null_mutator_alloc_region */);
1695 if (result != NULL) {
1696 assert(*succeeded, "sanity");
1697 return result;
1698 }
1699
1700 assert(!collector_policy()->should_clear_all_soft_refs(),
1701 "Flag should have been handled and cleared prior to this point");
1702
1703 // What else? We might try synchronous finalization later. If the total
1704 // space available is large enough for the allocation, then a more
1705 // complete compaction phase than we've tried so far might be
1706 // appropriate.
1707 assert(*succeeded, "sanity");
1708 return NULL;
1709 }
1710
1711 // Attempting to expand the heap sufficiently
1712 // to support an allocation of the given "word_size". If
1713 // successful, perform the allocation and return the address of the
1714 // allocated block, or else "NULL".
1715
1716 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) {
1717 assert_at_safepoint(true /* should_be_vm_thread */);
1718
1719 verify_region_sets_optional();
1720
1721 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1722 ergo_verbose1(ErgoHeapSizing,
1723 "attempt heap expansion",
1724 ergo_format_reason("allocation request failed")
1725 ergo_format_byte("allocation request"),
1726 word_size * HeapWordSize);
1727 if (expand(expand_bytes)) {
1728 _hrs.verify_optional();
1729 verify_region_sets_optional();
1730 return attempt_allocation_at_safepoint(word_size,
1731 false /* expect_null_mutator_alloc_region */);
1732 }
1733 return NULL;
1734 }
1735
1736 void G1CollectedHeap::update_committed_space(HeapWord* old_end,
1737 HeapWord* new_end) {
1738 assert(old_end != new_end, "don't call this otherwise");
1739 assert((HeapWord*) _g1_storage.high() == new_end, "invariant");
1740
1741 // Update the committed mem region.
1742 _g1_committed.set_end(new_end);
1743 // Tell the card table about the update.
1744 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed);
1745 // Tell the BOT about the update.
1746 _bot_shared->resize(_g1_committed.word_size());
1747 }
1748
1749 bool G1CollectedHeap::expand(size_t expand_bytes) {
1750 size_t old_mem_size = _g1_storage.committed_size();
1751 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1752 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1753 HeapRegion::GrainBytes);
1754 ergo_verbose2(ErgoHeapSizing,
1755 "expand the heap",
1756 ergo_format_byte("requested expansion amount")
1757 ergo_format_byte("attempted expansion amount"),
1758 expand_bytes, aligned_expand_bytes);
1759
1760 // First commit the memory.
1761 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1762 bool successful = _g1_storage.expand_by(aligned_expand_bytes);
1763 if (successful) {
1764 // Then propagate this update to the necessary data structures.
1765 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1766 update_committed_space(old_end, new_end);
1767
1768 FreeRegionList expansion_list("Local Expansion List");
1769 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list);
1770 assert(mr.start() == old_end, "post-condition");
1771 // mr might be a smaller region than what was requested if
1772 // expand_by() was unable to allocate the HeapRegion instances
1773 assert(mr.end() <= new_end, "post-condition");
1774
1775 size_t actual_expand_bytes = mr.byte_size();
1776 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1777 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(),
1778 "post-condition");
1779 if (actual_expand_bytes < aligned_expand_bytes) {
1780 // We could not expand _hrs to the desired size. In this case we
1781 // need to shrink the committed space accordingly.
1782 assert(mr.end() < new_end, "invariant");
1783
1784 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes;
1785 // First uncommit the memory.
1786 _g1_storage.shrink_by(diff_bytes);
1787 // Then propagate this update to the necessary data structures.
1788 update_committed_space(new_end, mr.end());
1789 }
1790 _free_list.add_as_tail(&expansion_list);
1791
1792 if (_hr_printer.is_active()) {
1793 HeapWord* curr = mr.start();
1794 while (curr < mr.end()) {
1795 HeapWord* curr_end = curr + HeapRegion::GrainWords;
1796 _hr_printer.commit(curr, curr_end);
1797 curr = curr_end;
1798 }
1799 assert(curr == mr.end(), "post-condition");
1800 }
1801 g1_policy()->record_new_heap_size(n_regions());
1802 } else {
1803 ergo_verbose0(ErgoHeapSizing,
1804 "did not expand the heap",
1805 ergo_format_reason("heap expansion operation failed"));
1806 // The expansion of the virtual storage space was unsuccessful.
1807 // Let's see if it was because we ran out of swap.
1808 if (G1ExitOnExpansionFailure &&
1809 _g1_storage.uncommitted_size() >= aligned_expand_bytes) {
1810 // We had head room...
1811 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion");
1812 }
1813 }
1814 return successful;
1815 }
1816
1817 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1818 size_t old_mem_size = _g1_storage.committed_size();
1819 size_t aligned_shrink_bytes =
1820 ReservedSpace::page_align_size_down(shrink_bytes);
1821 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1822 HeapRegion::GrainBytes);
1823 uint num_regions_deleted = 0;
1824 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted);
1825 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1826 assert(mr.end() == old_end, "post-condition");
1827
1828 ergo_verbose3(ErgoHeapSizing,
1829 "shrink the heap",
1830 ergo_format_byte("requested shrinking amount")
1831 ergo_format_byte("aligned shrinking amount")
1832 ergo_format_byte("attempted shrinking amount"),
1833 shrink_bytes, aligned_shrink_bytes, mr.byte_size());
1834 if (mr.byte_size() > 0) {
1835 if (_hr_printer.is_active()) {
1836 HeapWord* curr = mr.end();
1837 while (curr > mr.start()) {
1838 HeapWord* curr_end = curr;
1839 curr -= HeapRegion::GrainWords;
1840 _hr_printer.uncommit(curr, curr_end);
1841 }
1842 assert(curr == mr.start(), "post-condition");
1843 }
1844
1845 _g1_storage.shrink_by(mr.byte_size());
1846 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1847 assert(mr.start() == new_end, "post-condition");
1848
1849 _expansion_regions += num_regions_deleted;
1850 update_committed_space(old_end, new_end);
1851 HeapRegionRemSet::shrink_heap(n_regions());
1852 g1_policy()->record_new_heap_size(n_regions());
1853 } else {
1854 ergo_verbose0(ErgoHeapSizing,
1855 "did not shrink the heap",
1856 ergo_format_reason("heap shrinking operation failed"));
1857 }
1858 }
1859
1860 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1861 verify_region_sets_optional();
1862
1863 // We should only reach here at the end of a Full GC which means we
1864 // should not not be holding to any GC alloc regions. The method
1865 // below will make sure of that and do any remaining clean up.
1866 abandon_gc_alloc_regions();
1867
1868 // Instead of tearing down / rebuilding the free lists here, we
1869 // could instead use the remove_all_pending() method on free_list to
1870 // remove only the ones that we need to remove.
1871 tear_down_region_sets(true /* free_list_only */);
1872 shrink_helper(shrink_bytes);
1873 rebuild_region_sets(true /* free_list_only */);
1874
1875 _hrs.verify_optional();
1876 verify_region_sets_optional();
1877 }
1878
1879 // Public methods.
1880
1881 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1882 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1883 #endif // _MSC_VER
1884
1885
1886 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1887 SharedHeap(policy_),
1888 _g1_policy(policy_),
1889 _dirty_card_queue_set(false),
1890 _into_cset_dirty_card_queue_set(false),
1891 _is_alive_closure_cm(this),
1892 _is_alive_closure_stw(this),
1893 _ref_processor_cm(NULL),
1894 _ref_processor_stw(NULL),
1895 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1896 _bot_shared(NULL),
1897 _preserved_marks(40, 10000),
1898 _evac_failure_scan_stack(NULL) ,
1899 _mark_in_progress(false),
1900 _cg1r(NULL), _summary_bytes_used(0),
1901 _g1mm(NULL),
1902 _refine_cte_cl(NULL),
1903 _full_collection(false),
1904 _free_list("Master Free List"),
1905 _secondary_free_list("Secondary Free List"),
1906 _old_set("Old Set"),
1907 _humongous_set("Master Humongous Set"),
1908 _free_regions_coming(false),
1909 _young_list(new YoungList(this)),
1910 _gc_time_stamp(0),
1911 _retained_old_gc_alloc_region(NULL),
1912 _survivor_plab_stats(YoungPLABSize, PLABWeight),
1913 _old_plab_stats(OldPLABSize, PLABWeight),
1914 _expand_heap_after_alloc_failure(true),
1915 _surviving_young_words(NULL),
1916 _old_marking_cycles_started(0),
1917 _old_marking_cycles_completed(0),
1918 _in_cset_fast_test(NULL),
1919 _in_cset_fast_test_base(NULL),
1920 _dirty_cards_region_list(NULL),
1921 _worker_cset_start_region(NULL),
1922 _worker_cset_start_region_time_stamp(NULL) {
1923 _g1h = this; // To catch bugs.
1924 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1925 vm_exit_during_initialization("Failed necessary allocation.");
1926 }
1927
1928 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1929
1930 int n_queues = MAX2((int)ParallelGCThreads, 1);
1931 _task_queues = new RefToScanQueueSet(n_queues);
1932
1933 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1934 assert(n_rem_sets > 0, "Invariant.");
1935
1936 HeapRegionRemSetIterator** iter_arr =
1937 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues, mtGC);
1938 for (int i = 0; i < n_queues; i++) {
1939 iter_arr[i] = new HeapRegionRemSetIterator();
1940 }
1941 _rem_set_iterator = iter_arr;
1942
1943 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC);
1944 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC);
1945
1946 for (int i = 0; i < n_queues; i++) {
1947 RefToScanQueue* q = new RefToScanQueue();
1948 q->initialize();
1949 _task_queues->register_queue(i, q);
1950 }
1951
1952 clear_cset_start_regions();
1953
1954 // Initialize the G1EvacuationFailureALot counters and flags.
1955 NOT_PRODUCT(reset_evacuation_should_fail();)
1956
1957 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1958 }
1959
1960 jint G1CollectedHeap::initialize() {
1961 CollectedHeap::pre_initialize();
1962 os::enable_vtime();
1963
1964 G1Log::init();
1965
1966 // Necessary to satisfy locking discipline assertions.
1967
1968 MutexLocker x(Heap_lock);
1969
1970 // We have to initialize the printer before committing the heap, as
1971 // it will be used then.
1972 _hr_printer.set_active(G1PrintHeapRegions);
1973
1974 // While there are no constraints in the GC code that HeapWordSize
1975 // be any particular value, there are multiple other areas in the
1976 // system which believe this to be true (e.g. oop->object_size in some
1977 // cases incorrectly returns the size in wordSize units rather than
1978 // HeapWordSize).
1979 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1980
1981 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1982 size_t max_byte_size = collector_policy()->max_heap_byte_size();
1983
1984 // Ensure that the sizes are properly aligned.
1985 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
1986 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
1987
1988 _cg1r = new ConcurrentG1Refine();
1989
1990 // Reserve the maximum.
1991
1992 // When compressed oops are enabled, the preferred heap base
1993 // is calculated by subtracting the requested size from the
1994 // 32Gb boundary and using the result as the base address for
1995 // heap reservation. If the requested size is not aligned to
1996 // HeapRegion::GrainBytes (i.e. the alignment that is passed
1997 // into the ReservedHeapSpace constructor) then the actual
1998 // base of the reserved heap may end up differing from the
1999 // address that was requested (i.e. the preferred heap base).
2000 // If this happens then we could end up using a non-optimal
2001 // compressed oops mode.
2002
2003 // Since max_byte_size is aligned to the size of a heap region (checked
2004 // above).
2005 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
2006
2007 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size,
2008 HeapRegion::GrainBytes);
2009
2010 // It is important to do this in a way such that concurrent readers can't
2011 // temporarily think somethings in the heap. (I've actually seen this
2012 // happen in asserts: DLD.)
2013 _reserved.set_word_size(0);
2014 _reserved.set_start((HeapWord*)heap_rs.base());
2015 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
2016
2017 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes);
2018
2019 // Create the gen rem set (and barrier set) for the entire reserved region.
2020 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
2021 set_barrier_set(rem_set()->bs());
2022 if (barrier_set()->is_a(BarrierSet::ModRef)) {
2023 _mr_bs = (ModRefBarrierSet*)_barrier_set;
2024 } else {
2025 vm_exit_during_initialization("G1 requires a mod ref bs.");
2026 return JNI_ENOMEM;
2027 }
2028
2029 // Also create a G1 rem set.
2030 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
2031 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
2032 } else {
2033 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
2034 return JNI_ENOMEM;
2035 }
2036
2037 // Carve out the G1 part of the heap.
2038
2039 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
2040 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
2041 g1_rs.size()/HeapWordSize);
2042
2043 _g1_storage.initialize(g1_rs, 0);
2044 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
2045 _hrs.initialize((HeapWord*) _g1_reserved.start(),
2046 (HeapWord*) _g1_reserved.end(),
2047 _expansion_regions);
2048
2049 // 6843694 - ensure that the maximum region index can fit
2050 // in the remembered set structures.
2051 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
2052 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
2053
2054 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
2055 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
2056 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
2057 "too many cards per region");
2058
2059 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
2060
2061 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
2062 heap_word_size(init_byte_size));
2063
2064 _g1h = this;
2065
2066 _in_cset_fast_test_length = max_regions();
2067 _in_cset_fast_test_base =
2068 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC);
2069
2070 // We're biasing _in_cset_fast_test to avoid subtracting the
2071 // beginning of the heap every time we want to index; basically
2072 // it's the same with what we do with the card table.
2073 _in_cset_fast_test = _in_cset_fast_test_base -
2074 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
2075
2076 // Clear the _cset_fast_test bitmap in anticipation of adding
2077 // regions to the incremental collection set for the first
2078 // evacuation pause.
2079 clear_cset_fast_test();
2080
2081 // Create the ConcurrentMark data structure and thread.
2082 // (Must do this late, so that "max_regions" is defined.)
2083 _cm = new ConcurrentMark(this, heap_rs);
2084 if (_cm == NULL || !_cm->completed_initialization()) {
2085 vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark");
2086 return JNI_ENOMEM;
2087 }
2088 _cmThread = _cm->cmThread();
2089
2090 // Initialize the from_card cache structure of HeapRegionRemSet.
2091 HeapRegionRemSet::init_heap(max_regions());
2092
2093 // Now expand into the initial heap size.
2094 if (!expand(init_byte_size)) {
2095 vm_shutdown_during_initialization("Failed to allocate initial heap.");
2096 return JNI_ENOMEM;
2097 }
2098
2099 // Perform any initialization actions delegated to the policy.
2100 g1_policy()->init();
2101
2102 _refine_cte_cl =
2103 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2104 g1_rem_set(),
2105 concurrent_g1_refine());
2106 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2107
2108 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2109 SATB_Q_FL_lock,
2110 G1SATBProcessCompletedThreshold,
2111 Shared_SATB_Q_lock);
2112
2113 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2114 DirtyCardQ_FL_lock,
2115 concurrent_g1_refine()->yellow_zone(),
2116 concurrent_g1_refine()->red_zone(),
2117 Shared_DirtyCardQ_lock);
2118
2119 if (G1DeferredRSUpdate) {
2120 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2121 DirtyCardQ_FL_lock,
2122 -1, // never trigger processing
2123 -1, // no limit on length
2124 Shared_DirtyCardQ_lock,
2125 &JavaThread::dirty_card_queue_set());
2126 }
2127
2128 // Initialize the card queue set used to hold cards containing
2129 // references into the collection set.
2130 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2131 DirtyCardQ_FL_lock,
2132 -1, // never trigger processing
2133 -1, // no limit on length
2134 Shared_DirtyCardQ_lock,
2135 &JavaThread::dirty_card_queue_set());
2136
2137 // In case we're keeping closure specialization stats, initialize those
2138 // counts and that mechanism.
2139 SpecializationStats::clear();
2140
2141 // Do later initialization work for concurrent refinement.
2142 _cg1r->init();
2143
2144 // Here we allocate the dummy full region that is required by the
2145 // G1AllocRegion class. If we don't pass an address in the reserved
2146 // space here, lots of asserts fire.
2147
2148 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2149 _g1_reserved.start());
2150 // We'll re-use the same region whether the alloc region will
2151 // require BOT updates or not and, if it doesn't, then a non-young
2152 // region will complain that it cannot support allocations without
2153 // BOT updates. So we'll tag the dummy region as young to avoid that.
2154 dummy_region->set_young();
2155 // Make sure it's full.
2156 dummy_region->set_top(dummy_region->end());
2157 G1AllocRegion::setup(this, dummy_region);
2158
2159 init_mutator_alloc_region();
2160
2161 // Do create of the monitoring and management support so that
2162 // values in the heap have been properly initialized.
2163 _g1mm = new G1MonitoringSupport(this);
2164
2165 return JNI_OK;
2166 }
2167
2168 void G1CollectedHeap::ref_processing_init() {
2169 // Reference processing in G1 currently works as follows:
2170 //
2171 // * There are two reference processor instances. One is
2172 // used to record and process discovered references
2173 // during concurrent marking; the other is used to
2174 // record and process references during STW pauses
2175 // (both full and incremental).
2176 // * Both ref processors need to 'span' the entire heap as
2177 // the regions in the collection set may be dotted around.
2178 //
2179 // * For the concurrent marking ref processor:
2180 // * Reference discovery is enabled at initial marking.
2181 // * Reference discovery is disabled and the discovered
2182 // references processed etc during remarking.
2183 // * Reference discovery is MT (see below).
2184 // * Reference discovery requires a barrier (see below).
2185 // * Reference processing may or may not be MT
2186 // (depending on the value of ParallelRefProcEnabled
2187 // and ParallelGCThreads).
2188 // * A full GC disables reference discovery by the CM
2189 // ref processor and abandons any entries on it's
2190 // discovered lists.
2191 //
2192 // * For the STW processor:
2193 // * Non MT discovery is enabled at the start of a full GC.
2194 // * Processing and enqueueing during a full GC is non-MT.
2195 // * During a full GC, references are processed after marking.
2196 //
2197 // * Discovery (may or may not be MT) is enabled at the start
2198 // of an incremental evacuation pause.
2199 // * References are processed near the end of a STW evacuation pause.
2200 // * For both types of GC:
2201 // * Discovery is atomic - i.e. not concurrent.
2202 // * Reference discovery will not need a barrier.
2203
2204 SharedHeap::ref_processing_init();
2205 MemRegion mr = reserved_region();
2206
2207 // Concurrent Mark ref processor
2208 _ref_processor_cm =
2209 new ReferenceProcessor(mr, // span
2210 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2211 // mt processing
2212 (int) ParallelGCThreads,
2213 // degree of mt processing
2214 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2215 // mt discovery
2216 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2217 // degree of mt discovery
2218 false,
2219 // Reference discovery is not atomic
2220 &_is_alive_closure_cm,
2221 // is alive closure
2222 // (for efficiency/performance)
2223 true);
2224 // Setting next fields of discovered
2225 // lists requires a barrier.
2226
2227 // STW ref processor
2228 _ref_processor_stw =
2229 new ReferenceProcessor(mr, // span
2230 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2231 // mt processing
2232 MAX2((int)ParallelGCThreads, 1),
2233 // degree of mt processing
2234 (ParallelGCThreads > 1),
2235 // mt discovery
2236 MAX2((int)ParallelGCThreads, 1),
2237 // degree of mt discovery
2238 true,
2239 // Reference discovery is atomic
2240 &_is_alive_closure_stw,
2241 // is alive closure
2242 // (for efficiency/performance)
2243 false);
2244 // Setting next fields of discovered
2245 // lists requires a barrier.
2246 }
2247
2248 size_t G1CollectedHeap::capacity() const {
2249 return _g1_committed.byte_size();
2250 }
2251
2252 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) {
2253 assert(!hr->continuesHumongous(), "pre-condition");
2254 hr->reset_gc_time_stamp();
2255 if (hr->startsHumongous()) {
2256 uint first_index = hr->hrs_index() + 1;
2257 uint last_index = hr->last_hc_index();
2258 for (uint i = first_index; i < last_index; i += 1) {
2259 HeapRegion* chr = region_at(i);
2260 assert(chr->continuesHumongous(), "sanity");
2261 chr->reset_gc_time_stamp();
2262 }
2263 }
2264 }
2265
2266 #ifndef PRODUCT
2267 class CheckGCTimeStampsHRClosure : public HeapRegionClosure {
2268 private:
2269 unsigned _gc_time_stamp;
2270 bool _failures;
2271
2272 public:
2273 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) :
2274 _gc_time_stamp(gc_time_stamp), _failures(false) { }
2275
2276 virtual bool doHeapRegion(HeapRegion* hr) {
2277 unsigned region_gc_time_stamp = hr->get_gc_time_stamp();
2278 if (_gc_time_stamp != region_gc_time_stamp) {
2279 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, "
2280 "expected %d", HR_FORMAT_PARAMS(hr),
2281 region_gc_time_stamp, _gc_time_stamp);
2282 _failures = true;
2283 }
2284 return false;
2285 }
2286
2287 bool failures() { return _failures; }
2288 };
2289
2290 void G1CollectedHeap::check_gc_time_stamps() {
2291 CheckGCTimeStampsHRClosure cl(_gc_time_stamp);
2292 heap_region_iterate(&cl);
2293 guarantee(!cl.failures(), "all GC time stamps should have been reset");
2294 }
2295 #endif // PRODUCT
2296
2297 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2298 DirtyCardQueue* into_cset_dcq,
2299 bool concurrent,
2300 int worker_i) {
2301 // Clean cards in the hot card cache
2302 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq);
2303
2304 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2305 int n_completed_buffers = 0;
2306 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2307 n_completed_buffers++;
2308 }
2309 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers);
2310 dcqs.clear_n_completed_buffers();
2311 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2312 }
2313
2314
2315 // Computes the sum of the storage used by the various regions.
2316
2317 size_t G1CollectedHeap::used() const {
2318 assert(Heap_lock->owner() != NULL,
2319 "Should be owned on this thread's behalf.");
2320 size_t result = _summary_bytes_used;
2321 // Read only once in case it is set to NULL concurrently
2322 HeapRegion* hr = _mutator_alloc_region.get();
2323 if (hr != NULL)
2324 result += hr->used();
2325 return result;
2326 }
2327
2328 size_t G1CollectedHeap::used_unlocked() const {
2329 size_t result = _summary_bytes_used;
2330 return result;
2331 }
2332
2333 class SumUsedClosure: public HeapRegionClosure {
2334 size_t _used;
2335 public:
2336 SumUsedClosure() : _used(0) {}
2337 bool doHeapRegion(HeapRegion* r) {
2338 if (!r->continuesHumongous()) {
2339 _used += r->used();
2340 }
2341 return false;
2342 }
2343 size_t result() { return _used; }
2344 };
2345
2346 size_t G1CollectedHeap::recalculate_used() const {
2347 SumUsedClosure blk;
2348 heap_region_iterate(&blk);
2349 return blk.result();
2350 }
2351
2352 size_t G1CollectedHeap::unsafe_max_alloc() {
2353 if (free_regions() > 0) return HeapRegion::GrainBytes;
2354 // otherwise, is there space in the current allocation region?
2355
2356 // We need to store the current allocation region in a local variable
2357 // here. The problem is that this method doesn't take any locks and
2358 // there may be other threads which overwrite the current allocation
2359 // region field. attempt_allocation(), for example, sets it to NULL
2360 // and this can happen *after* the NULL check here but before the call
2361 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2362 // to be a problem in the optimized build, since the two loads of the
2363 // current allocation region field are optimized away.
2364 HeapRegion* hr = _mutator_alloc_region.get();
2365 if (hr == NULL) {
2366 return 0;
2367 }
2368 return hr->free();
2369 }
2370
2371 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2372 switch (cause) {
2373 case GCCause::_gc_locker: return GCLockerInvokesConcurrent;
2374 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent;
2375 case GCCause::_g1_humongous_allocation: return true;
2376 default: return false;
2377 }
2378 }
2379
2380 #ifndef PRODUCT
2381 void G1CollectedHeap::allocate_dummy_regions() {
2382 // Let's fill up most of the region
2383 size_t word_size = HeapRegion::GrainWords - 1024;
2384 // And as a result the region we'll allocate will be humongous.
2385 guarantee(isHumongous(word_size), "sanity");
2386
2387 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2388 // Let's use the existing mechanism for the allocation
2389 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2390 if (dummy_obj != NULL) {
2391 MemRegion mr(dummy_obj, word_size);
2392 CollectedHeap::fill_with_object(mr);
2393 } else {
2394 // If we can't allocate once, we probably cannot allocate
2395 // again. Let's get out of the loop.
2396 break;
2397 }
2398 }
2399 }
2400 #endif // !PRODUCT
2401
2402 void G1CollectedHeap::increment_old_marking_cycles_started() {
2403 assert(_old_marking_cycles_started == _old_marking_cycles_completed ||
2404 _old_marking_cycles_started == _old_marking_cycles_completed + 1,
2405 err_msg("Wrong marking cycle count (started: %d, completed: %d)",
2406 _old_marking_cycles_started, _old_marking_cycles_completed));
2407
2408 _old_marking_cycles_started++;
2409 }
2410
2411 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) {
2412 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2413
2414 // We assume that if concurrent == true, then the caller is a
2415 // concurrent thread that was joined the Suspendible Thread
2416 // Set. If there's ever a cheap way to check this, we should add an
2417 // assert here.
2418
2419 // Given that this method is called at the end of a Full GC or of a
2420 // concurrent cycle, and those can be nested (i.e., a Full GC can
2421 // interrupt a concurrent cycle), the number of full collections
2422 // completed should be either one (in the case where there was no
2423 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2424 // behind the number of full collections started.
2425
2426 // This is the case for the inner caller, i.e. a Full GC.
2427 assert(concurrent ||
2428 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) ||
2429 (_old_marking_cycles_started == _old_marking_cycles_completed + 2),
2430 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u "
2431 "is inconsistent with _old_marking_cycles_completed = %u",
2432 _old_marking_cycles_started, _old_marking_cycles_completed));
2433
2434 // This is the case for the outer caller, i.e. the concurrent cycle.
2435 assert(!concurrent ||
2436 (_old_marking_cycles_started == _old_marking_cycles_completed + 1),
2437 err_msg("for outer caller (concurrent cycle): "
2438 "_old_marking_cycles_started = %u "
2439 "is inconsistent with _old_marking_cycles_completed = %u",
2440 _old_marking_cycles_started, _old_marking_cycles_completed));
2441
2442 _old_marking_cycles_completed += 1;
2443
2444 // We need to clear the "in_progress" flag in the CM thread before
2445 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2446 // is set) so that if a waiter requests another System.gc() it doesn't
2447 // incorrectly see that a marking cyle is still in progress.
2448 if (concurrent) {
2449 _cmThread->clear_in_progress();
2450 }
2451
2452 // This notify_all() will ensure that a thread that called
2453 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2454 // and it's waiting for a full GC to finish will be woken up. It is
2455 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2456 FullGCCount_lock->notify_all();
2457 }
2458
2459 void G1CollectedHeap::collect(GCCause::Cause cause) {
2460 assert_heap_not_locked();
2461
2462 unsigned int gc_count_before;
2463 unsigned int old_marking_count_before;
2464 bool retry_gc;
2465
2466 do {
2467 retry_gc = false;
2468
2469 {
2470 MutexLocker ml(Heap_lock);
2471
2472 // Read the GC count while holding the Heap_lock
2473 gc_count_before = total_collections();
2474 old_marking_count_before = _old_marking_cycles_started;
2475 }
2476
2477 if (should_do_concurrent_full_gc(cause)) {
2478 // Schedule an initial-mark evacuation pause that will start a
2479 // concurrent cycle. We're setting word_size to 0 which means that
2480 // we are not requesting a post-GC allocation.
2481 VM_G1IncCollectionPause op(gc_count_before,
2482 0, /* word_size */
2483 true, /* should_initiate_conc_mark */
2484 g1_policy()->max_pause_time_ms(),
2485 cause);
2486
2487 VMThread::execute(&op);
2488 if (!op.pause_succeeded()) {
2489 if (old_marking_count_before == _old_marking_cycles_started) {
2490 retry_gc = op.should_retry_gc();
2491 } else {
2492 // A Full GC happened while we were trying to schedule the
2493 // initial-mark GC. No point in starting a new cycle given
2494 // that the whole heap was collected anyway.
2495 }
2496
2497 if (retry_gc) {
2498 if (GC_locker::is_active_and_needs_gc()) {
2499 GC_locker::stall_until_clear();
2500 }
2501 }
2502 }
2503 } else {
2504 if (cause == GCCause::_gc_locker
2505 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2506
2507 // Schedule a standard evacuation pause. We're setting word_size
2508 // to 0 which means that we are not requesting a post-GC allocation.
2509 VM_G1IncCollectionPause op(gc_count_before,
2510 0, /* word_size */
2511 false, /* should_initiate_conc_mark */
2512 g1_policy()->max_pause_time_ms(),
2513 cause);
2514 VMThread::execute(&op);
2515 } else {
2516 // Schedule a Full GC.
2517 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause);
2518 VMThread::execute(&op);
2519 }
2520 }
2521 } while (retry_gc);
2522 }
2523
2524 bool G1CollectedHeap::is_in(const void* p) const {
2525 if (_g1_committed.contains(p)) {
2526 // Given that we know that p is in the committed space,
2527 // heap_region_containing_raw() should successfully
2528 // return the containing region.
2529 HeapRegion* hr = heap_region_containing_raw(p);
2530 return hr->is_in(p);
2531 } else {
2532 return false;
2533 }
2534 }
2535
2536 // Iteration functions.
2537
2538 // Iterates an OopClosure over all ref-containing fields of objects
2539 // within a HeapRegion.
2540
2541 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2542 MemRegion _mr;
2543 ExtendedOopClosure* _cl;
2544 public:
2545 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl)
2546 : _mr(mr), _cl(cl) {}
2547 bool doHeapRegion(HeapRegion* r) {
2548 if (!r->continuesHumongous()) {
2549 r->oop_iterate(_cl);
2550 }
2551 return false;
2552 }
2553 };
2554
2555 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) {
2556 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2557 heap_region_iterate(&blk);
2558 }
2559
2560 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
2561 IterateOopClosureRegionClosure blk(mr, cl);
2562 heap_region_iterate(&blk);
2563 }
2564
2565 // Iterates an ObjectClosure over all objects within a HeapRegion.
2566
2567 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2568 ObjectClosure* _cl;
2569 public:
2570 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2571 bool doHeapRegion(HeapRegion* r) {
2572 if (! r->continuesHumongous()) {
2573 r->object_iterate(_cl);
2574 }
2575 return false;
2576 }
2577 };
2578
2579 void G1CollectedHeap::object_iterate(ObjectClosure* cl) {
2580 IterateObjectClosureRegionClosure blk(cl);
2581 heap_region_iterate(&blk);
2582 }
2583
2584 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2585 // FIXME: is this right?
2586 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2587 }
2588
2589 // Calls a SpaceClosure on a HeapRegion.
2590
2591 class SpaceClosureRegionClosure: public HeapRegionClosure {
2592 SpaceClosure* _cl;
2593 public:
2594 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2595 bool doHeapRegion(HeapRegion* r) {
2596 _cl->do_space(r);
2597 return false;
2598 }
2599 };
2600
2601 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2602 SpaceClosureRegionClosure blk(cl);
2603 heap_region_iterate(&blk);
2604 }
2605
2606 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2607 _hrs.iterate(cl);
2608 }
2609
2610 void
2611 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2612 uint worker_id,
2613 uint no_of_par_workers,
2614 jint claim_value) {
2615 const uint regions = n_regions();
2616 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
2617 no_of_par_workers :
2618 1);
2619 assert(UseDynamicNumberOfGCThreads ||
2620 no_of_par_workers == workers()->total_workers(),
2621 "Non dynamic should use fixed number of workers");
2622 // try to spread out the starting points of the workers
2623 const HeapRegion* start_hr =
2624 start_region_for_worker(worker_id, no_of_par_workers);
2625 const uint start_index = start_hr->hrs_index();
2626
2627 // each worker will actually look at all regions
2628 for (uint count = 0; count < regions; ++count) {
2629 const uint index = (start_index + count) % regions;
2630 assert(0 <= index && index < regions, "sanity");
2631 HeapRegion* r = region_at(index);
2632 // we'll ignore "continues humongous" regions (we'll process them
2633 // when we come across their corresponding "start humongous"
2634 // region) and regions already claimed
2635 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2636 continue;
2637 }
2638 // OK, try to claim it
2639 if (r->claimHeapRegion(claim_value)) {
2640 // success!
2641 assert(!r->continuesHumongous(), "sanity");
2642 if (r->startsHumongous()) {
2643 // If the region is "starts humongous" we'll iterate over its
2644 // "continues humongous" first; in fact we'll do them
2645 // first. The order is important. In on case, calling the
2646 // closure on the "starts humongous" region might de-allocate
2647 // and clear all its "continues humongous" regions and, as a
2648 // result, we might end up processing them twice. So, we'll do
2649 // them first (notice: most closures will ignore them anyway) and
2650 // then we'll do the "starts humongous" region.
2651 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) {
2652 HeapRegion* chr = region_at(ch_index);
2653
2654 // if the region has already been claimed or it's not
2655 // "continues humongous" we're done
2656 if (chr->claim_value() == claim_value ||
2657 !chr->continuesHumongous()) {
2658 break;
2659 }
2660
2661 // Noone should have claimed it directly. We can given
2662 // that we claimed its "starts humongous" region.
2663 assert(chr->claim_value() != claim_value, "sanity");
2664 assert(chr->humongous_start_region() == r, "sanity");
2665
2666 if (chr->claimHeapRegion(claim_value)) {
2667 // we should always be able to claim it; noone else should
2668 // be trying to claim this region
2669
2670 bool res2 = cl->doHeapRegion(chr);
2671 assert(!res2, "Should not abort");
2672
2673 // Right now, this holds (i.e., no closure that actually
2674 // does something with "continues humongous" regions
2675 // clears them). We might have to weaken it in the future,
2676 // but let's leave these two asserts here for extra safety.
2677 assert(chr->continuesHumongous(), "should still be the case");
2678 assert(chr->humongous_start_region() == r, "sanity");
2679 } else {
2680 guarantee(false, "we should not reach here");
2681 }
2682 }
2683 }
2684
2685 assert(!r->continuesHumongous(), "sanity");
2686 bool res = cl->doHeapRegion(r);
2687 assert(!res, "Should not abort");
2688 }
2689 }
2690 }
2691
2692 class ResetClaimValuesClosure: public HeapRegionClosure {
2693 public:
2694 bool doHeapRegion(HeapRegion* r) {
2695 r->set_claim_value(HeapRegion::InitialClaimValue);
2696 return false;
2697 }
2698 };
2699
2700 void G1CollectedHeap::reset_heap_region_claim_values() {
2701 ResetClaimValuesClosure blk;
2702 heap_region_iterate(&blk);
2703 }
2704
2705 void G1CollectedHeap::reset_cset_heap_region_claim_values() {
2706 ResetClaimValuesClosure blk;
2707 collection_set_iterate(&blk);
2708 }
2709
2710 #ifdef ASSERT
2711 // This checks whether all regions in the heap have the correct claim
2712 // value. I also piggy-backed on this a check to ensure that the
2713 // humongous_start_region() information on "continues humongous"
2714 // regions is correct.
2715
2716 class CheckClaimValuesClosure : public HeapRegionClosure {
2717 private:
2718 jint _claim_value;
2719 uint _failures;
2720 HeapRegion* _sh_region;
2721
2722 public:
2723 CheckClaimValuesClosure(jint claim_value) :
2724 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2725 bool doHeapRegion(HeapRegion* r) {
2726 if (r->claim_value() != _claim_value) {
2727 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2728 "claim value = %d, should be %d",
2729 HR_FORMAT_PARAMS(r),
2730 r->claim_value(), _claim_value);
2731 ++_failures;
2732 }
2733 if (!r->isHumongous()) {
2734 _sh_region = NULL;
2735 } else if (r->startsHumongous()) {
2736 _sh_region = r;
2737 } else if (r->continuesHumongous()) {
2738 if (r->humongous_start_region() != _sh_region) {
2739 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2740 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2741 HR_FORMAT_PARAMS(r),
2742 r->humongous_start_region(),
2743 _sh_region);
2744 ++_failures;
2745 }
2746 }
2747 return false;
2748 }
2749 uint failures() { return _failures; }
2750 };
2751
2752 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2753 CheckClaimValuesClosure cl(claim_value);
2754 heap_region_iterate(&cl);
2755 return cl.failures() == 0;
2756 }
2757
2758 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure {
2759 private:
2760 jint _claim_value;
2761 uint _failures;
2762
2763 public:
2764 CheckClaimValuesInCSetHRClosure(jint claim_value) :
2765 _claim_value(claim_value), _failures(0) { }
2766
2767 uint failures() { return _failures; }
2768
2769 bool doHeapRegion(HeapRegion* hr) {
2770 assert(hr->in_collection_set(), "how?");
2771 assert(!hr->isHumongous(), "H-region in CSet");
2772 if (hr->claim_value() != _claim_value) {
2773 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", "
2774 "claim value = %d, should be %d",
2775 HR_FORMAT_PARAMS(hr),
2776 hr->claim_value(), _claim_value);
2777 _failures += 1;
2778 }
2779 return false;
2780 }
2781 };
2782
2783 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) {
2784 CheckClaimValuesInCSetHRClosure cl(claim_value);
2785 collection_set_iterate(&cl);
2786 return cl.failures() == 0;
2787 }
2788 #endif // ASSERT
2789
2790 // Clear the cached CSet starting regions and (more importantly)
2791 // the time stamps. Called when we reset the GC time stamp.
2792 void G1CollectedHeap::clear_cset_start_regions() {
2793 assert(_worker_cset_start_region != NULL, "sanity");
2794 assert(_worker_cset_start_region_time_stamp != NULL, "sanity");
2795
2796 int n_queues = MAX2((int)ParallelGCThreads, 1);
2797 for (int i = 0; i < n_queues; i++) {
2798 _worker_cset_start_region[i] = NULL;
2799 _worker_cset_start_region_time_stamp[i] = 0;
2800 }
2801 }
2802
2803 // Given the id of a worker, obtain or calculate a suitable
2804 // starting region for iterating over the current collection set.
2805 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) {
2806 assert(get_gc_time_stamp() > 0, "should have been updated by now");
2807
2808 HeapRegion* result = NULL;
2809 unsigned gc_time_stamp = get_gc_time_stamp();
2810
2811 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) {
2812 // Cached starting region for current worker was set
2813 // during the current pause - so it's valid.
2814 // Note: the cached starting heap region may be NULL
2815 // (when the collection set is empty).
2816 result = _worker_cset_start_region[worker_i];
2817 assert(result == NULL || result->in_collection_set(), "sanity");
2818 return result;
2819 }
2820
2821 // The cached entry was not valid so let's calculate
2822 // a suitable starting heap region for this worker.
2823
2824 // We want the parallel threads to start their collection
2825 // set iteration at different collection set regions to
2826 // avoid contention.
2827 // If we have:
2828 // n collection set regions
2829 // p threads
2830 // Then thread t will start at region floor ((t * n) / p)
2831
2832 result = g1_policy()->collection_set();
2833 if (G1CollectedHeap::use_parallel_gc_threads()) {
2834 uint cs_size = g1_policy()->cset_region_length();
2835 uint active_workers = workers()->active_workers();
2836 assert(UseDynamicNumberOfGCThreads ||
2837 active_workers == workers()->total_workers(),
2838 "Unless dynamic should use total workers");
2839
2840 uint end_ind = (cs_size * worker_i) / active_workers;
2841 uint start_ind = 0;
2842
2843 if (worker_i > 0 &&
2844 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) {
2845 // Previous workers starting region is valid
2846 // so let's iterate from there
2847 start_ind = (cs_size * (worker_i - 1)) / active_workers;
2848 result = _worker_cset_start_region[worker_i - 1];
2849 }
2850
2851 for (uint i = start_ind; i < end_ind; i++) {
2852 result = result->next_in_collection_set();
2853 }
2854 }
2855
2856 // Note: the calculated starting heap region may be NULL
2857 // (when the collection set is empty).
2858 assert(result == NULL || result->in_collection_set(), "sanity");
2859 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp,
2860 "should be updated only once per pause");
2861 _worker_cset_start_region[worker_i] = result;
2862 OrderAccess::storestore();
2863 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp;
2864 return result;
2865 }
2866
2867 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i,
2868 uint no_of_par_workers) {
2869 uint worker_num =
2870 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U;
2871 assert(UseDynamicNumberOfGCThreads ||
2872 no_of_par_workers == workers()->total_workers(),
2873 "Non dynamic should use fixed number of workers");
2874 const uint start_index = n_regions() * worker_i / worker_num;
2875 return region_at(start_index);
2876 }
2877
2878 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2879 HeapRegion* r = g1_policy()->collection_set();
2880 while (r != NULL) {
2881 HeapRegion* next = r->next_in_collection_set();
2882 if (cl->doHeapRegion(r)) {
2883 cl->incomplete();
2884 return;
2885 }
2886 r = next;
2887 }
2888 }
2889
2890 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2891 HeapRegionClosure *cl) {
2892 if (r == NULL) {
2893 // The CSet is empty so there's nothing to do.
2894 return;
2895 }
2896
2897 assert(r->in_collection_set(),
2898 "Start region must be a member of the collection set.");
2899 HeapRegion* cur = r;
2900 while (cur != NULL) {
2901 HeapRegion* next = cur->next_in_collection_set();
2902 if (cl->doHeapRegion(cur) && false) {
2903 cl->incomplete();
2904 return;
2905 }
2906 cur = next;
2907 }
2908 cur = g1_policy()->collection_set();
2909 while (cur != r) {
2910 HeapRegion* next = cur->next_in_collection_set();
2911 if (cl->doHeapRegion(cur) && false) {
2912 cl->incomplete();
2913 return;
2914 }
2915 cur = next;
2916 }
2917 }
2918
2919 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2920 return n_regions() > 0 ? region_at(0) : NULL;
2921 }
2922
2923
2924 Space* G1CollectedHeap::space_containing(const void* addr) const {
2925 Space* res = heap_region_containing(addr);
2926 return res;
2927 }
2928
2929 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2930 Space* sp = space_containing(addr);
2931 if (sp != NULL) {
2932 return sp->block_start(addr);
2933 }
2934 return NULL;
2935 }
2936
2937 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2938 Space* sp = space_containing(addr);
2939 assert(sp != NULL, "block_size of address outside of heap");
2940 return sp->block_size(addr);
2941 }
2942
2943 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2944 Space* sp = space_containing(addr);
2945 return sp->block_is_obj(addr);
2946 }
2947
2948 bool G1CollectedHeap::supports_tlab_allocation() const {
2949 return true;
2950 }
2951
2952 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2953 return HeapRegion::GrainBytes;
2954 }
2955
2956 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2957 // Return the remaining space in the cur alloc region, but not less than
2958 // the min TLAB size.
2959
2960 // Also, this value can be at most the humongous object threshold,
2961 // since we can't allow tlabs to grow big enough to accomodate
2962 // humongous objects.
2963
2964 HeapRegion* hr = _mutator_alloc_region.get();
2965 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2966 if (hr == NULL) {
2967 return max_tlab_size;
2968 } else {
2969 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2970 }
2971 }
2972
2973 size_t G1CollectedHeap::max_capacity() const {
2974 return _g1_reserved.byte_size();
2975 }
2976
2977 jlong G1CollectedHeap::millis_since_last_gc() {
2978 // assert(false, "NYI");
2979 return 0;
2980 }
2981
2982 void G1CollectedHeap::prepare_for_verify() {
2983 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2984 ensure_parsability(false);
2985 }
2986 g1_rem_set()->prepare_for_verify();
2987 }
2988
2989 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr,
2990 VerifyOption vo) {
2991 switch (vo) {
2992 case VerifyOption_G1UsePrevMarking:
2993 return hr->obj_allocated_since_prev_marking(obj);
2994 case VerifyOption_G1UseNextMarking:
2995 return hr->obj_allocated_since_next_marking(obj);
2996 case VerifyOption_G1UseMarkWord:
2997 return false;
2998 default:
2999 ShouldNotReachHere();
3000 }
3001 return false; // keep some compilers happy
3002 }
3003
3004 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) {
3005 switch (vo) {
3006 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start();
3007 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start();
3008 case VerifyOption_G1UseMarkWord: return NULL;
3009 default: ShouldNotReachHere();
3010 }
3011 return NULL; // keep some compilers happy
3012 }
3013
3014 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) {
3015 switch (vo) {
3016 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj);
3017 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj);
3018 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked();
3019 default: ShouldNotReachHere();
3020 }
3021 return false; // keep some compilers happy
3022 }
3023
3024 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) {
3025 switch (vo) {
3026 case VerifyOption_G1UsePrevMarking: return "PTAMS";
3027 case VerifyOption_G1UseNextMarking: return "NTAMS";
3028 case VerifyOption_G1UseMarkWord: return "NONE";
3029 default: ShouldNotReachHere();
3030 }
3031 return NULL; // keep some compilers happy
3032 }
3033
3034 class VerifyLivenessOopClosure: public OopClosure {
3035 G1CollectedHeap* _g1h;
3036 VerifyOption _vo;
3037 public:
3038 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
3039 _g1h(g1h), _vo(vo)
3040 { }
3041 void do_oop(narrowOop *p) { do_oop_work(p); }
3042 void do_oop( oop *p) { do_oop_work(p); }
3043
3044 template <class T> void do_oop_work(T *p) {
3045 oop obj = oopDesc::load_decode_heap_oop(p);
3046 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
3047 "Dead object referenced by a not dead object");
3048 }
3049 };
3050
3051 class VerifyObjsInRegionClosure: public ObjectClosure {
3052 private:
3053 G1CollectedHeap* _g1h;
3054 size_t _live_bytes;
3055 HeapRegion *_hr;
3056 VerifyOption _vo;
3057 public:
3058 // _vo == UsePrevMarking -> use "prev" marking information,
3059 // _vo == UseNextMarking -> use "next" marking information,
3060 // _vo == UseMarkWord -> use mark word from object header.
3061 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
3062 : _live_bytes(0), _hr(hr), _vo(vo) {
3063 _g1h = G1CollectedHeap::heap();
3064 }
3065 void do_object(oop o) {
3066 VerifyLivenessOopClosure isLive(_g1h, _vo);
3067 assert(o != NULL, "Huh?");
3068 if (!_g1h->is_obj_dead_cond(o, _vo)) {
3069 // If the object is alive according to the mark word,
3070 // then verify that the marking information agrees.
3071 // Note we can't verify the contra-positive of the
3072 // above: if the object is dead (according to the mark
3073 // word), it may not be marked, or may have been marked
3074 // but has since became dead, or may have been allocated
3075 // since the last marking.
3076 if (_vo == VerifyOption_G1UseMarkWord) {
3077 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
3078 }
3079
3080 o->oop_iterate_no_header(&isLive);
3081 if (!_hr->obj_allocated_since_prev_marking(o)) {
3082 size_t obj_size = o->size(); // Make sure we don't overflow
3083 _live_bytes += (obj_size * HeapWordSize);
3084 }
3085 }
3086 }
3087 size_t live_bytes() { return _live_bytes; }
3088 };
3089
3090 class PrintObjsInRegionClosure : public ObjectClosure {
3091 HeapRegion *_hr;
3092 G1CollectedHeap *_g1;
3093 public:
3094 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
3095 _g1 = G1CollectedHeap::heap();
3096 };
3097
3098 void do_object(oop o) {
3099 if (o != NULL) {
3100 HeapWord *start = (HeapWord *) o;
3101 size_t word_sz = o->size();
3102 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
3103 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
3104 (void*) o, word_sz,
3105 _g1->isMarkedPrev(o),
3106 _g1->isMarkedNext(o),
3107 _hr->obj_allocated_since_prev_marking(o));
3108 HeapWord *end = start + word_sz;
3109 HeapWord *cur;
3110 int *val;
3111 for (cur = start; cur < end; cur++) {
3112 val = (int *) cur;
3113 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
3114 }
3115 }
3116 }
3117 };
3118
3119 class VerifyRegionClosure: public HeapRegionClosure {
3120 private:
3121 bool _par;
3122 VerifyOption _vo;
3123 bool _failures;
3124 public:
3125 // _vo == UsePrevMarking -> use "prev" marking information,
3126 // _vo == UseNextMarking -> use "next" marking information,
3127 // _vo == UseMarkWord -> use mark word from object header.
3128 VerifyRegionClosure(bool par, VerifyOption vo)
3129 : _par(par),
3130 _vo(vo),
3131 _failures(false) {}
3132
3133 bool failures() {
3134 return _failures;
3135 }
3136
3137 bool doHeapRegion(HeapRegion* r) {
3138 if (!r->continuesHumongous()) {
3139 bool failures = false;
3140 r->verify(_vo, &failures);
3141 if (failures) {
3142 _failures = true;
3143 } else {
3144 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
3145 r->object_iterate(¬_dead_yet_cl);
3146 if (_vo != VerifyOption_G1UseNextMarking) {
3147 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
3148 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
3149 "max_live_bytes "SIZE_FORMAT" "
3150 "< calculated "SIZE_FORMAT,
3151 r->bottom(), r->end(),
3152 r->max_live_bytes(),
3153 not_dead_yet_cl.live_bytes());
3154 _failures = true;
3155 }
3156 } else {
3157 // When vo == UseNextMarking we cannot currently do a sanity
3158 // check on the live bytes as the calculation has not been
3159 // finalized yet.
3160 }
3161 }
3162 }
3163 return false; // stop the region iteration if we hit a failure
3164 }
3165 };
3166
3167 class YoungRefCounterClosure : public OopClosure {
3168 G1CollectedHeap* _g1h;
3169 int _count;
3170 public:
3171 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {}
3172 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } }
3173 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3174
3175 int count() { return _count; }
3176 void reset_count() { _count = 0; };
3177 };
3178
3179 class VerifyKlassClosure: public KlassClosure {
3180 YoungRefCounterClosure _young_ref_counter_closure;
3181 OopClosure *_oop_closure;
3182 public:
3183 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {}
3184 void do_klass(Klass* k) {
3185 k->oops_do(_oop_closure);
3186
3187 _young_ref_counter_closure.reset_count();
3188 k->oops_do(&_young_ref_counter_closure);
3189 if (_young_ref_counter_closure.count() > 0) {
3190 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k));
3191 }
3192 }
3193 };
3194
3195 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can
3196 // pass it as the perm_blk to SharedHeap::process_strong_roots.
3197 // When process_strong_roots stop calling perm_blk->younger_refs_iterate
3198 // we can change this closure to extend the simpler OopClosure.
3199 class VerifyRootsClosure: public OopsInGenClosure {
3200 private:
3201 G1CollectedHeap* _g1h;
3202 VerifyOption _vo;
3203 bool _failures;
3204 public:
3205 // _vo == UsePrevMarking -> use "prev" marking information,
3206 // _vo == UseNextMarking -> use "next" marking information,
3207 // _vo == UseMarkWord -> use mark word from object header.
3208 VerifyRootsClosure(VerifyOption vo) :
3209 _g1h(G1CollectedHeap::heap()),
3210 _vo(vo),
3211 _failures(false) { }
3212
3213 bool failures() { return _failures; }
3214
3215 template <class T> void do_oop_nv(T* p) {
3216 T heap_oop = oopDesc::load_heap_oop(p);
3217 if (!oopDesc::is_null(heap_oop)) {
3218 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
3219 if (_g1h->is_obj_dead_cond(obj, _vo)) {
3220 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
3221 "points to dead obj "PTR_FORMAT, p, (void*) obj);
3222 if (_vo == VerifyOption_G1UseMarkWord) {
3223 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
3224 }
3225 obj->print_on(gclog_or_tty);
3226 _failures = true;
3227 }
3228 }
3229 }
3230
3231 void do_oop(oop* p) { do_oop_nv(p); }
3232 void do_oop(narrowOop* p) { do_oop_nv(p); }
3233 };
3234
3235 // This is the task used for parallel heap verification.
3236
3237 class G1ParVerifyTask: public AbstractGangTask {
3238 private:
3239 G1CollectedHeap* _g1h;
3240 VerifyOption _vo;
3241 bool _failures;
3242
3243 public:
3244 // _vo == UsePrevMarking -> use "prev" marking information,
3245 // _vo == UseNextMarking -> use "next" marking information,
3246 // _vo == UseMarkWord -> use mark word from object header.
3247 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) :
3248 AbstractGangTask("Parallel verify task"),
3249 _g1h(g1h),
3250 _vo(vo),
3251 _failures(false) { }
3252
3253 bool failures() {
3254 return _failures;
3255 }
3256
3257 void work(uint worker_id) {
3258 HandleMark hm;
3259 VerifyRegionClosure blk(true, _vo);
3260 _g1h->heap_region_par_iterate_chunked(&blk, worker_id,
3261 _g1h->workers()->active_workers(),
3262 HeapRegion::ParVerifyClaimValue);
3263 if (blk.failures()) {
3264 _failures = true;
3265 }
3266 }
3267 };
3268
3269 void G1CollectedHeap::verify(bool silent) {
3270 verify(silent, VerifyOption_G1UsePrevMarking);
3271 }
3272
3273 void G1CollectedHeap::verify(bool silent,
3274 VerifyOption vo) {
3275 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
3276 if (!silent) { gclog_or_tty->print("Roots "); }
3277 VerifyRootsClosure rootsCl(vo);
3278
3279 assert(Thread::current()->is_VM_thread(),
3280 "Expected to be executed serially by the VM thread at this point");
3281
3282 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
3283 VerifyKlassClosure klassCl(this, &rootsCl);
3284
3285 // We apply the relevant closures to all the oops in the
3286 // system dictionary, the string table and the code cache.
3287 const int so = SO_AllClasses | SO_Strings | SO_CodeCache;
3288
3289 // Need cleared claim bits for the strong roots processing
3290 ClassLoaderDataGraph::clear_claimed_marks();
3291
3292 process_strong_roots(true, // activate StrongRootsScope
3293 false, // we set "is scavenging" to false,
3294 // so we don't reset the dirty cards.
3295 ScanningOption(so), // roots scanning options
3296 &rootsCl,
3297 &blobsCl,
3298 &klassCl
3299 );
3300
3301 bool failures = rootsCl.failures();
3302
3303 if (vo != VerifyOption_G1UseMarkWord) {
3304 // If we're verifying during a full GC then the region sets
3305 // will have been torn down at the start of the GC. Therefore
3306 // verifying the region sets will fail. So we only verify
3307 // the region sets when not in a full GC.
3308 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
3309 verify_region_sets();
3310 }
3311
3312 if (!silent) { gclog_or_tty->print("HeapRegions "); }
3313 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
3314 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3315 "sanity check");
3316
3317 G1ParVerifyTask task(this, vo);
3318 assert(UseDynamicNumberOfGCThreads ||
3319 workers()->active_workers() == workers()->total_workers(),
3320 "If not dynamic should be using all the workers");
3321 int n_workers = workers()->active_workers();
3322 set_par_threads(n_workers);
3323 workers()->run_task(&task);
3324 set_par_threads(0);
3325 if (task.failures()) {
3326 failures = true;
3327 }
3328
3329 // Checks that the expected amount of parallel work was done.
3330 // The implication is that n_workers is > 0.
3331 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
3332 "sanity check");
3333
3334 reset_heap_region_claim_values();
3335
3336 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3337 "sanity check");
3338 } else {
3339 VerifyRegionClosure blk(false, vo);
3340 heap_region_iterate(&blk);
3341 if (blk.failures()) {
3342 failures = true;
3343 }
3344 }
3345 if (!silent) gclog_or_tty->print("RemSet ");
3346 rem_set()->verify();
3347
3348 if (failures) {
3349 gclog_or_tty->print_cr("Heap:");
3350 // It helps to have the per-region information in the output to
3351 // help us track down what went wrong. This is why we call
3352 // print_extended_on() instead of print_on().
3353 print_extended_on(gclog_or_tty);
3354 gclog_or_tty->print_cr("");
3355 #ifndef PRODUCT
3356 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3357 concurrent_mark()->print_reachable("at-verification-failure",
3358 vo, false /* all */);
3359 }
3360 #endif
3361 gclog_or_tty->flush();
3362 }
3363 guarantee(!failures, "there should not have been any failures");
3364 } else {
3365 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) ");
3366 }
3367 }
3368
3369 class PrintRegionClosure: public HeapRegionClosure {
3370 outputStream* _st;
3371 public:
3372 PrintRegionClosure(outputStream* st) : _st(st) {}
3373 bool doHeapRegion(HeapRegion* r) {
3374 r->print_on(_st);
3375 return false;
3376 }
3377 };
3378
3379 void G1CollectedHeap::print_on(outputStream* st) const {
3380 st->print(" %-20s", "garbage-first heap");
3381 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3382 capacity()/K, used_unlocked()/K);
3383 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3384 _g1_storage.low_boundary(),
3385 _g1_storage.high(),
3386 _g1_storage.high_boundary());
3387 st->cr();
3388 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3389 uint young_regions = _young_list->length();
3390 st->print("%u young (" SIZE_FORMAT "K), ", young_regions,
3391 (size_t) young_regions * HeapRegion::GrainBytes / K);
3392 uint survivor_regions = g1_policy()->recorded_survivor_regions();
3393 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions,
3394 (size_t) survivor_regions * HeapRegion::GrainBytes / K);
3395 st->cr();
3396 MetaspaceAux::print_on(st);
3397 }
3398
3399 void G1CollectedHeap::print_extended_on(outputStream* st) const {
3400 print_on(st);
3401
3402 // Print the per-region information.
3403 st->cr();
3404 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), "
3405 "HS=humongous(starts), HC=humongous(continues), "
3406 "CS=collection set, F=free, TS=gc time stamp, "
3407 "PTAMS=previous top-at-mark-start, "
3408 "NTAMS=next top-at-mark-start)");
3409 PrintRegionClosure blk(st);
3410 heap_region_iterate(&blk);
3411 }
3412
3413 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3414 if (G1CollectedHeap::use_parallel_gc_threads()) {
3415 workers()->print_worker_threads_on(st);
3416 }
3417 _cmThread->print_on(st);
3418 st->cr();
3419 _cm->print_worker_threads_on(st);
3420 _cg1r->print_worker_threads_on(st);
3421 }
3422
3423 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3424 if (G1CollectedHeap::use_parallel_gc_threads()) {
3425 workers()->threads_do(tc);
3426 }
3427 tc->do_thread(_cmThread);
3428 _cg1r->threads_do(tc);
3429 }
3430
3431 void G1CollectedHeap::print_tracing_info() const {
3432 // We'll overload this to mean "trace GC pause statistics."
3433 if (TraceGen0Time || TraceGen1Time) {
3434 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3435 // to that.
3436 g1_policy()->print_tracing_info();
3437 }
3438 if (G1SummarizeRSetStats) {
3439 g1_rem_set()->print_summary_info();
3440 }
3441 if (G1SummarizeConcMark) {
3442 concurrent_mark()->print_summary_info();
3443 }
3444 g1_policy()->print_yg_surv_rate_info();
3445 SpecializationStats::print();
3446 }
3447
3448 #ifndef PRODUCT
3449 // Helpful for debugging RSet issues.
3450
3451 class PrintRSetsClosure : public HeapRegionClosure {
3452 private:
3453 const char* _msg;
3454 size_t _occupied_sum;
3455
3456 public:
3457 bool doHeapRegion(HeapRegion* r) {
3458 HeapRegionRemSet* hrrs = r->rem_set();
3459 size_t occupied = hrrs->occupied();
3460 _occupied_sum += occupied;
3461
3462 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3463 HR_FORMAT_PARAMS(r));
3464 if (occupied == 0) {
3465 gclog_or_tty->print_cr(" RSet is empty");
3466 } else {
3467 hrrs->print();
3468 }
3469 gclog_or_tty->print_cr("----------");
3470 return false;
3471 }
3472
3473 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3474 gclog_or_tty->cr();
3475 gclog_or_tty->print_cr("========================================");
3476 gclog_or_tty->print_cr(msg);
3477 gclog_or_tty->cr();
3478 }
3479
3480 ~PrintRSetsClosure() {
3481 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3482 gclog_or_tty->print_cr("========================================");
3483 gclog_or_tty->cr();
3484 }
3485 };
3486
3487 void G1CollectedHeap::print_cset_rsets() {
3488 PrintRSetsClosure cl("Printing CSet RSets");
3489 collection_set_iterate(&cl);
3490 }
3491
3492 void G1CollectedHeap::print_all_rsets() {
3493 PrintRSetsClosure cl("Printing All RSets");;
3494 heap_region_iterate(&cl);
3495 }
3496 #endif // PRODUCT
3497
3498 G1CollectedHeap* G1CollectedHeap::heap() {
3499 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3500 "not a garbage-first heap");
3501 return _g1h;
3502 }
3503
3504 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3505 // always_do_update_barrier = false;
3506 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3507 // Call allocation profiler
3508 AllocationProfiler::iterate_since_last_gc();
3509 // Fill TLAB's and such
3510 ensure_parsability(true);
3511 }
3512
3513 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3514 // FIXME: what is this about?
3515 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3516 // is set.
3517 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3518 "derived pointer present"));
3519 // always_do_update_barrier = true;
3520
3521 // We have just completed a GC. Update the soft reference
3522 // policy with the new heap occupancy
3523 Universe::update_heap_info_at_gc();
3524 }
3525
3526 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3527 unsigned int gc_count_before,
3528 bool* succeeded) {
3529 assert_heap_not_locked_and_not_at_safepoint();
3530 g1_policy()->record_stop_world_start();
3531 VM_G1IncCollectionPause op(gc_count_before,
3532 word_size,
3533 false, /* should_initiate_conc_mark */
3534 g1_policy()->max_pause_time_ms(),
3535 GCCause::_g1_inc_collection_pause);
3536 VMThread::execute(&op);
3537
3538 HeapWord* result = op.result();
3539 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3540 assert(result == NULL || ret_succeeded,
3541 "the result should be NULL if the VM did not succeed");
3542 *succeeded = ret_succeeded;
3543
3544 assert_heap_not_locked();
3545 return result;
3546 }
3547
3548 void
3549 G1CollectedHeap::doConcurrentMark() {
3550 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3551 if (!_cmThread->in_progress()) {
3552 _cmThread->set_started();
3553 CGC_lock->notify();
3554 }
3555 }
3556
3557 size_t G1CollectedHeap::pending_card_num() {
3558 size_t extra_cards = 0;
3559 JavaThread *curr = Threads::first();
3560 while (curr != NULL) {
3561 DirtyCardQueue& dcq = curr->dirty_card_queue();
3562 extra_cards += dcq.size();
3563 curr = curr->next();
3564 }
3565 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3566 size_t buffer_size = dcqs.buffer_size();
3567 size_t buffer_num = dcqs.completed_buffers_num();
3568
3569 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes
3570 // in bytes - not the number of 'entries'. We need to convert
3571 // into a number of cards.
3572 return (buffer_size * buffer_num + extra_cards) / oopSize;
3573 }
3574
3575 size_t G1CollectedHeap::cards_scanned() {
3576 return g1_rem_set()->cardsScanned();
3577 }
3578
3579 void
3580 G1CollectedHeap::setup_surviving_young_words() {
3581 assert(_surviving_young_words == NULL, "pre-condition");
3582 uint array_length = g1_policy()->young_cset_region_length();
3583 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC);
3584 if (_surviving_young_words == NULL) {
3585 vm_exit_out_of_memory(sizeof(size_t) * array_length,
3586 "Not enough space for young surv words summary.");
3587 }
3588 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t));
3589 #ifdef ASSERT
3590 for (uint i = 0; i < array_length; ++i) {
3591 assert( _surviving_young_words[i] == 0, "memset above" );
3592 }
3593 #endif // !ASSERT
3594 }
3595
3596 void
3597 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3598 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3599 uint array_length = g1_policy()->young_cset_region_length();
3600 for (uint i = 0; i < array_length; ++i) {
3601 _surviving_young_words[i] += surv_young_words[i];
3602 }
3603 }
3604
3605 void
3606 G1CollectedHeap::cleanup_surviving_young_words() {
3607 guarantee( _surviving_young_words != NULL, "pre-condition" );
3608 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC);
3609 _surviving_young_words = NULL;
3610 }
3611
3612 #ifdef ASSERT
3613 class VerifyCSetClosure: public HeapRegionClosure {
3614 public:
3615 bool doHeapRegion(HeapRegion* hr) {
3616 // Here we check that the CSet region's RSet is ready for parallel
3617 // iteration. The fields that we'll verify are only manipulated
3618 // when the region is part of a CSet and is collected. Afterwards,
3619 // we reset these fields when we clear the region's RSet (when the
3620 // region is freed) so they are ready when the region is
3621 // re-allocated. The only exception to this is if there's an
3622 // evacuation failure and instead of freeing the region we leave
3623 // it in the heap. In that case, we reset these fields during
3624 // evacuation failure handling.
3625 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3626
3627 // Here's a good place to add any other checks we'd like to
3628 // perform on CSet regions.
3629 return false;
3630 }
3631 };
3632 #endif // ASSERT
3633
3634 #if TASKQUEUE_STATS
3635 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3636 st->print_raw_cr("GC Task Stats");
3637 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3638 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3639 }
3640
3641 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3642 print_taskqueue_stats_hdr(st);
3643
3644 TaskQueueStats totals;
3645 const int n = workers() != NULL ? workers()->total_workers() : 1;
3646 for (int i = 0; i < n; ++i) {
3647 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3648 totals += task_queue(i)->stats;
3649 }
3650 st->print_raw("tot "); totals.print(st); st->cr();
3651
3652 DEBUG_ONLY(totals.verify());
3653 }
3654
3655 void G1CollectedHeap::reset_taskqueue_stats() {
3656 const int n = workers() != NULL ? workers()->total_workers() : 1;
3657 for (int i = 0; i < n; ++i) {
3658 task_queue(i)->stats.reset();
3659 }
3660 }
3661 #endif // TASKQUEUE_STATS
3662
3663 void G1CollectedHeap::log_gc_header() {
3664 if (!G1Log::fine()) {
3665 return;
3666 }
3667
3668 gclog_or_tty->date_stamp(PrintGCDateStamps);
3669 gclog_or_tty->stamp(PrintGCTimeStamps);
3670
3671 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause())
3672 .append(g1_policy()->gcs_are_young() ? "(young)" : "(mixed)")
3673 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : "");
3674
3675 gclog_or_tty->print("[%s", (const char*)gc_cause_str);
3676 }
3677
3678 void G1CollectedHeap::log_gc_footer(double pause_time_sec) {
3679 if (!G1Log::fine()) {
3680 return;
3681 }
3682
3683 if (G1Log::finer()) {
3684 if (evacuation_failed()) {
3685 gclog_or_tty->print(" (to-space exhausted)");
3686 }
3687 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3688 g1_policy()->phase_times()->note_gc_end();
3689 g1_policy()->phase_times()->print(pause_time_sec);
3690 g1_policy()->print_detailed_heap_transition();
3691 } else {
3692 if (evacuation_failed()) {
3693 gclog_or_tty->print("--");
3694 }
3695 g1_policy()->print_heap_transition();
3696 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3697 }
3698 gclog_or_tty->flush();
3699 }
3700
3701 bool
3702 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3703 assert_at_safepoint(true /* should_be_vm_thread */);
3704 guarantee(!is_gc_active(), "collection is not reentrant");
3705
3706 if (GC_locker::check_active_before_gc()) {
3707 return false;
3708 }
3709
3710 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3711 ResourceMark rm;
3712
3713 print_heap_before_gc();
3714
3715 HRSPhaseSetter x(HRSPhaseEvacuation);
3716 verify_region_sets_optional();
3717 verify_dirty_young_regions();
3718
3719 // This call will decide whether this pause is an initial-mark
3720 // pause. If it is, during_initial_mark_pause() will return true
3721 // for the duration of this pause.
3722 g1_policy()->decide_on_conc_mark_initiation();
3723
3724 // We do not allow initial-mark to be piggy-backed on a mixed GC.
3725 assert(!g1_policy()->during_initial_mark_pause() ||
3726 g1_policy()->gcs_are_young(), "sanity");
3727
3728 // We also do not allow mixed GCs during marking.
3729 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity");
3730
3731 // Record whether this pause is an initial mark. When the current
3732 // thread has completed its logging output and it's safe to signal
3733 // the CM thread, the flag's value in the policy has been reset.
3734 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause();
3735
3736 // Inner scope for scope based logging, timers, and stats collection
3737 {
3738 if (g1_policy()->during_initial_mark_pause()) {
3739 // We are about to start a marking cycle, so we increment the
3740 // full collection counter.
3741 increment_old_marking_cycles_started();
3742 }
3743 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
3744
3745 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3746 workers()->active_workers() : 1);
3747 double pause_start_sec = os::elapsedTime();
3748 g1_policy()->phase_times()->note_gc_start(active_workers);
3749 log_gc_header();
3750
3751 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3752 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3753
3754 // If the secondary_free_list is not empty, append it to the
3755 // free_list. No need to wait for the cleanup operation to finish;
3756 // the region allocation code will check the secondary_free_list
3757 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3758 // set, skip this step so that the region allocation code has to
3759 // get entries from the secondary_free_list.
3760 if (!G1StressConcRegionFreeing) {
3761 append_secondary_free_list_if_not_empty_with_lock();
3762 }
3763
3764 assert(check_young_list_well_formed(),
3765 "young list should be well formed");
3766
3767 // Don't dynamically change the number of GC threads this early. A value of
3768 // 0 is used to indicate serial work. When parallel work is done,
3769 // it will be set.
3770
3771 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3772 IsGCActiveMark x;
3773
3774 gc_prologue(false);
3775 increment_total_collections(false /* full gc */);
3776 increment_gc_time_stamp();
3777
3778 verify_before_gc();
3779
3780 COMPILER2_PRESENT(DerivedPointerTable::clear());
3781
3782 // Please see comment in g1CollectedHeap.hpp and
3783 // G1CollectedHeap::ref_processing_init() to see how
3784 // reference processing currently works in G1.
3785
3786 // Enable discovery in the STW reference processor
3787 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3788 true /*verify_no_refs*/);
3789
3790 {
3791 // We want to temporarily turn off discovery by the
3792 // CM ref processor, if necessary, and turn it back on
3793 // on again later if we do. Using a scoped
3794 // NoRefDiscovery object will do this.
3795 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3796
3797 // Forget the current alloc region (we might even choose it to be part
3798 // of the collection set!).
3799 release_mutator_alloc_region();
3800
3801 // We should call this after we retire the mutator alloc
3802 // region(s) so that all the ALLOC / RETIRE events are generated
3803 // before the start GC event.
3804 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3805
3806 // This timing is only used by the ergonomics to handle our pause target.
3807 // It is unclear why this should not include the full pause. We will
3808 // investigate this in CR 7178365.
3809 //
3810 // Preserving the old comment here if that helps the investigation:
3811 //
3812 // The elapsed time induced by the start time below deliberately elides
3813 // the possible verification above.
3814 double sample_start_time_sec = os::elapsedTime();
3815 size_t start_used_bytes = used();
3816
3817 #if YOUNG_LIST_VERBOSE
3818 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3819 _young_list->print();
3820 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3821 #endif // YOUNG_LIST_VERBOSE
3822
3823 g1_policy()->record_collection_pause_start(sample_start_time_sec,
3824 start_used_bytes);
3825
3826 double scan_wait_start = os::elapsedTime();
3827 // We have to wait until the CM threads finish scanning the
3828 // root regions as it's the only way to ensure that all the
3829 // objects on them have been correctly scanned before we start
3830 // moving them during the GC.
3831 bool waited = _cm->root_regions()->wait_until_scan_finished();
3832 double wait_time_ms = 0.0;
3833 if (waited) {
3834 double scan_wait_end = os::elapsedTime();
3835 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0;
3836 }
3837 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms);
3838
3839 #if YOUNG_LIST_VERBOSE
3840 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3841 _young_list->print();
3842 #endif // YOUNG_LIST_VERBOSE
3843
3844 if (g1_policy()->during_initial_mark_pause()) {
3845 concurrent_mark()->checkpointRootsInitialPre();
3846 }
3847
3848 #if YOUNG_LIST_VERBOSE
3849 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3850 _young_list->print();
3851 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3852 #endif // YOUNG_LIST_VERBOSE
3853
3854 g1_policy()->finalize_cset(target_pause_time_ms);
3855
3856 _cm->note_start_of_gc();
3857 // We should not verify the per-thread SATB buffers given that
3858 // we have not filtered them yet (we'll do so during the
3859 // GC). We also call this after finalize_cset() to
3860 // ensure that the CSet has been finalized.
3861 _cm->verify_no_cset_oops(true /* verify_stacks */,
3862 true /* verify_enqueued_buffers */,
3863 false /* verify_thread_buffers */,
3864 true /* verify_fingers */);
3865
3866 if (_hr_printer.is_active()) {
3867 HeapRegion* hr = g1_policy()->collection_set();
3868 while (hr != NULL) {
3869 G1HRPrinter::RegionType type;
3870 if (!hr->is_young()) {
3871 type = G1HRPrinter::Old;
3872 } else if (hr->is_survivor()) {
3873 type = G1HRPrinter::Survivor;
3874 } else {
3875 type = G1HRPrinter::Eden;
3876 }
3877 _hr_printer.cset(hr);
3878 hr = hr->next_in_collection_set();
3879 }
3880 }
3881
3882 #ifdef ASSERT
3883 VerifyCSetClosure cl;
3884 collection_set_iterate(&cl);
3885 #endif // ASSERT
3886
3887 setup_surviving_young_words();
3888
3889 // Initialize the GC alloc regions.
3890 init_gc_alloc_regions();
3891
3892 // Actually do the work...
3893 evacuate_collection_set();
3894
3895 // We do this to mainly verify the per-thread SATB buffers
3896 // (which have been filtered by now) since we didn't verify
3897 // them earlier. No point in re-checking the stacks / enqueued
3898 // buffers given that the CSet has not changed since last time
3899 // we checked.
3900 _cm->verify_no_cset_oops(false /* verify_stacks */,
3901 false /* verify_enqueued_buffers */,
3902 true /* verify_thread_buffers */,
3903 true /* verify_fingers */);
3904
3905 free_collection_set(g1_policy()->collection_set());
3906 g1_policy()->clear_collection_set();
3907
3908 cleanup_surviving_young_words();
3909
3910 // Start a new incremental collection set for the next pause.
3911 g1_policy()->start_incremental_cset_building();
3912
3913 // Clear the _cset_fast_test bitmap in anticipation of adding
3914 // regions to the incremental collection set for the next
3915 // evacuation pause.
3916 clear_cset_fast_test();
3917
3918 _young_list->reset_sampled_info();
3919
3920 // Don't check the whole heap at this point as the
3921 // GC alloc regions from this pause have been tagged
3922 // as survivors and moved on to the survivor list.
3923 // Survivor regions will fail the !is_young() check.
3924 assert(check_young_list_empty(false /* check_heap */),
3925 "young list should be empty");
3926
3927 #if YOUNG_LIST_VERBOSE
3928 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3929 _young_list->print();
3930 #endif // YOUNG_LIST_VERBOSE
3931
3932 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3933 _young_list->first_survivor_region(),
3934 _young_list->last_survivor_region());
3935
3936 _young_list->reset_auxilary_lists();
3937
3938 if (evacuation_failed()) {
3939 _summary_bytes_used = recalculate_used();
3940 } else {
3941 // The "used" of the the collection set have already been subtracted
3942 // when they were freed. Add in the bytes evacuated.
3943 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3944 }
3945
3946 if (g1_policy()->during_initial_mark_pause()) {
3947 // We have to do this before we notify the CM threads that
3948 // they can start working to make sure that all the
3949 // appropriate initialization is done on the CM object.
3950 concurrent_mark()->checkpointRootsInitialPost();
3951 set_marking_started();
3952 // Note that we don't actually trigger the CM thread at
3953 // this point. We do that later when we're sure that
3954 // the current thread has completed its logging output.
3955 }
3956
3957 allocate_dummy_regions();
3958
3959 #if YOUNG_LIST_VERBOSE
3960 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3961 _young_list->print();
3962 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3963 #endif // YOUNG_LIST_VERBOSE
3964
3965 init_mutator_alloc_region();
3966
3967 {
3968 size_t expand_bytes = g1_policy()->expansion_amount();
3969 if (expand_bytes > 0) {
3970 size_t bytes_before = capacity();
3971 // No need for an ergo verbose message here,
3972 // expansion_amount() does this when it returns a value > 0.
3973 if (!expand(expand_bytes)) {
3974 // We failed to expand the heap so let's verify that
3975 // committed/uncommitted amount match the backing store
3976 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3977 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3978 }
3979 }
3980 }
3981
3982 // We redo the verificaiton but now wrt to the new CSet which
3983 // has just got initialized after the previous CSet was freed.
3984 _cm->verify_no_cset_oops(true /* verify_stacks */,
3985 true /* verify_enqueued_buffers */,
3986 true /* verify_thread_buffers */,
3987 true /* verify_fingers */);
3988 _cm->note_end_of_gc();
3989
3990 // This timing is only used by the ergonomics to handle our pause target.
3991 // It is unclear why this should not include the full pause. We will
3992 // investigate this in CR 7178365.
3993 double sample_end_time_sec = os::elapsedTime();
3994 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS;
3995 g1_policy()->record_collection_pause_end(pause_time_ms);
3996
3997 MemoryService::track_memory_usage();
3998
3999 // In prepare_for_verify() below we'll need to scan the deferred
4000 // update buffers to bring the RSets up-to-date if
4001 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
4002 // the update buffers we'll probably need to scan cards on the
4003 // regions we just allocated to (i.e., the GC alloc
4004 // regions). However, during the last GC we called
4005 // set_saved_mark() on all the GC alloc regions, so card
4006 // scanning might skip the [saved_mark_word()...top()] area of
4007 // those regions (i.e., the area we allocated objects into
4008 // during the last GC). But it shouldn't. Given that
4009 // saved_mark_word() is conditional on whether the GC time stamp
4010 // on the region is current or not, by incrementing the GC time
4011 // stamp here we invalidate all the GC time stamps on all the
4012 // regions and saved_mark_word() will simply return top() for
4013 // all the regions. This is a nicer way of ensuring this rather
4014 // than iterating over the regions and fixing them. In fact, the
4015 // GC time stamp increment here also ensures that
4016 // saved_mark_word() will return top() between pauses, i.e.,
4017 // during concurrent refinement. So we don't need the
4018 // is_gc_active() check to decided which top to use when
4019 // scanning cards (see CR 7039627).
4020 increment_gc_time_stamp();
4021
4022 verify_after_gc();
4023
4024 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
4025 ref_processor_stw()->verify_no_references_recorded();
4026
4027 // CM reference discovery will be re-enabled if necessary.
4028 }
4029
4030 // We should do this after we potentially expand the heap so
4031 // that all the COMMIT events are generated before the end GC
4032 // event, and after we retire the GC alloc regions so that all
4033 // RETIRE events are generated before the end GC event.
4034 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
4035
4036 if (mark_in_progress()) {
4037 concurrent_mark()->update_g1_committed();
4038 }
4039
4040 #ifdef TRACESPINNING
4041 ParallelTaskTerminator::print_termination_counts();
4042 #endif
4043
4044 gc_epilogue(false);
4045 }
4046
4047 // Print the remainder of the GC log output.
4048 log_gc_footer(os::elapsedTime() - pause_start_sec);
4049
4050 // It is not yet to safe to tell the concurrent mark to
4051 // start as we have some optional output below. We don't want the
4052 // output from the concurrent mark thread interfering with this
4053 // logging output either.
4054
4055 _hrs.verify_optional();
4056 verify_region_sets_optional();
4057
4058 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
4059 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
4060
4061 print_heap_after_gc();
4062
4063 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
4064 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
4065 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
4066 // before any GC notifications are raised.
4067 g1mm()->update_sizes();
4068 }
4069
4070 if (G1SummarizeRSetStats &&
4071 (G1SummarizeRSetStatsPeriod > 0) &&
4072 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
4073 g1_rem_set()->print_summary_info();
4074 }
4075
4076 // It should now be safe to tell the concurrent mark thread to start
4077 // without its logging output interfering with the logging output
4078 // that came from the pause.
4079
4080 if (should_start_conc_mark) {
4081 // CAUTION: after the doConcurrentMark() call below,
4082 // the concurrent marking thread(s) could be running
4083 // concurrently with us. Make sure that anything after
4084 // this point does not assume that we are the only GC thread
4085 // running. Note: of course, the actual marking work will
4086 // not start until the safepoint itself is released in
4087 // ConcurrentGCThread::safepoint_desynchronize().
4088 doConcurrentMark();
4089 }
4090
4091 return true;
4092 }
4093
4094 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
4095 {
4096 size_t gclab_word_size;
4097 switch (purpose) {
4098 case GCAllocForSurvived:
4099 gclab_word_size = _survivor_plab_stats.desired_plab_sz();
4100 break;
4101 case GCAllocForTenured:
4102 gclab_word_size = _old_plab_stats.desired_plab_sz();
4103 break;
4104 default:
4105 assert(false, "unknown GCAllocPurpose");
4106 gclab_word_size = _old_plab_stats.desired_plab_sz();
4107 break;
4108 }
4109
4110 // Prevent humongous PLAB sizes for two reasons:
4111 // * PLABs are allocated using a similar paths as oops, but should
4112 // never be in a humongous region
4113 // * Allowing humongous PLABs needlessly churns the region free lists
4114 return MIN2(_humongous_object_threshold_in_words, gclab_word_size);
4115 }
4116
4117 void G1CollectedHeap::init_mutator_alloc_region() {
4118 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
4119 _mutator_alloc_region.init();
4120 }
4121
4122 void G1CollectedHeap::release_mutator_alloc_region() {
4123 _mutator_alloc_region.release();
4124 assert(_mutator_alloc_region.get() == NULL, "post-condition");
4125 }
4126
4127 void G1CollectedHeap::init_gc_alloc_regions() {
4128 assert_at_safepoint(true /* should_be_vm_thread */);
4129
4130 _survivor_gc_alloc_region.init();
4131 _old_gc_alloc_region.init();
4132 HeapRegion* retained_region = _retained_old_gc_alloc_region;
4133 _retained_old_gc_alloc_region = NULL;
4134
4135 // We will discard the current GC alloc region if:
4136 // a) it's in the collection set (it can happen!),
4137 // b) it's already full (no point in using it),
4138 // c) it's empty (this means that it was emptied during
4139 // a cleanup and it should be on the free list now), or
4140 // d) it's humongous (this means that it was emptied
4141 // during a cleanup and was added to the free list, but
4142 // has been subseqently used to allocate a humongous
4143 // object that may be less than the region size).
4144 if (retained_region != NULL &&
4145 !retained_region->in_collection_set() &&
4146 !(retained_region->top() == retained_region->end()) &&
4147 !retained_region->is_empty() &&
4148 !retained_region->isHumongous()) {
4149 retained_region->set_saved_mark();
4150 // The retained region was added to the old region set when it was
4151 // retired. We have to remove it now, since we don't allow regions
4152 // we allocate to in the region sets. We'll re-add it later, when
4153 // it's retired again.
4154 _old_set.remove(retained_region);
4155 bool during_im = g1_policy()->during_initial_mark_pause();
4156 retained_region->note_start_of_copying(during_im);
4157 _old_gc_alloc_region.set(retained_region);
4158 _hr_printer.reuse(retained_region);
4159 }
4160 }
4161
4162 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers) {
4163 _survivor_gc_alloc_region.release();
4164 // If we have an old GC alloc region to release, we'll save it in
4165 // _retained_old_gc_alloc_region. If we don't
4166 // _retained_old_gc_alloc_region will become NULL. This is what we
4167 // want either way so no reason to check explicitly for either
4168 // condition.
4169 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
4170
4171 if (ResizePLAB) {
4172 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4173 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4174 }
4175 }
4176
4177 void G1CollectedHeap::abandon_gc_alloc_regions() {
4178 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
4179 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
4180 _retained_old_gc_alloc_region = NULL;
4181 }
4182
4183 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
4184 _drain_in_progress = false;
4185 set_evac_failure_closure(cl);
4186 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4187 }
4188
4189 void G1CollectedHeap::finalize_for_evac_failure() {
4190 assert(_evac_failure_scan_stack != NULL &&
4191 _evac_failure_scan_stack->length() == 0,
4192 "Postcondition");
4193 assert(!_drain_in_progress, "Postcondition");
4194 delete _evac_failure_scan_stack;
4195 _evac_failure_scan_stack = NULL;
4196 }
4197
4198 void G1CollectedHeap::remove_self_forwarding_pointers() {
4199 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4200
4201 G1ParRemoveSelfForwardPtrsTask rsfp_task(this);
4202
4203 if (G1CollectedHeap::use_parallel_gc_threads()) {
4204 set_par_threads();
4205 workers()->run_task(&rsfp_task);
4206 set_par_threads(0);
4207 } else {
4208 rsfp_task.work(0);
4209 }
4210
4211 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity");
4212
4213 // Reset the claim values in the regions in the collection set.
4214 reset_cset_heap_region_claim_values();
4215
4216 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4217
4218 // Now restore saved marks, if any.
4219 while (_preserved_marks.has_data()) {
4220 G1PreserveMarkQueueEntry e = _preserved_marks.remove_first();
4221 e.obj->set_mark(e.mark);
4222 }
4223 }
4224
4225 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
4226 _evac_failure_scan_stack->push(obj);
4227 }
4228
4229 void G1CollectedHeap::drain_evac_failure_scan_stack() {
4230 assert(_evac_failure_scan_stack != NULL, "precondition");
4231
4232 while (_evac_failure_scan_stack->length() > 0) {
4233 oop obj = _evac_failure_scan_stack->pop();
4234 _evac_failure_closure->set_region(heap_region_containing(obj));
4235 obj->oop_iterate_backwards(_evac_failure_closure);
4236 }
4237 }
4238
4239 oop
4240 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4241 oop old) {
4242 assert(obj_in_cs(old),
4243 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4244 (HeapWord*) old));
4245 markOop m = old->mark();
4246 oop forward_ptr = old->forward_to_atomic(old);
4247 if (forward_ptr == NULL) {
4248 // Forward-to-self succeeded.
4249
4250 if (_evac_failure_closure != cl) {
4251 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4252 assert(!_drain_in_progress,
4253 "Should only be true while someone holds the lock.");
4254 // Set the global evac-failure closure to the current thread's.
4255 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4256 set_evac_failure_closure(cl);
4257 // Now do the common part.
4258 handle_evacuation_failure_common(old, m);
4259 // Reset to NULL.
4260 set_evac_failure_closure(NULL);
4261 } else {
4262 // The lock is already held, and this is recursive.
4263 assert(_drain_in_progress, "This should only be the recursive case.");
4264 handle_evacuation_failure_common(old, m);
4265 }
4266 return old;
4267 } else {
4268 // Forward-to-self failed. Either someone else managed to allocate
4269 // space for this object (old != forward_ptr) or they beat us in
4270 // self-forwarding it (old == forward_ptr).
4271 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4272 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4273 "should not be in the CSet",
4274 (HeapWord*) old, (HeapWord*) forward_ptr));
4275 return forward_ptr;
4276 }
4277 }
4278
4279 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4280 set_evacuation_failed(true);
4281
4282 preserve_mark_if_necessary(old, m);
4283
4284 HeapRegion* r = heap_region_containing(old);
4285 if (!r->evacuation_failed()) {
4286 r->set_evacuation_failed(true);
4287 _hr_printer.evac_failure(r);
4288 }
4289
4290 push_on_evac_failure_scan_stack(old);
4291
4292 if (!_drain_in_progress) {
4293 // prevent recursion in copy_to_survivor_space()
4294 _drain_in_progress = true;
4295 drain_evac_failure_scan_stack();
4296 _drain_in_progress = false;
4297 }
4298 }
4299
4300 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4301 assert(evacuation_failed(), "Oversaving!");
4302 // We want to call the "for_promotion_failure" version only in the
4303 // case of a promotion failure.
4304 if (m->must_be_preserved_for_promotion_failure(obj)) {
4305 _preserved_marks.append(obj, m);
4306 }
4307 }
4308
4309 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4310 size_t word_size) {
4311 if (purpose == GCAllocForSurvived) {
4312 HeapWord* result = survivor_attempt_allocation(word_size);
4313 if (result != NULL) {
4314 return result;
4315 } else {
4316 // Let's try to allocate in the old gen in case we can fit the
4317 // object there.
4318 return old_attempt_allocation(word_size);
4319 }
4320 } else {
4321 assert(purpose == GCAllocForTenured, "sanity");
4322 HeapWord* result = old_attempt_allocation(word_size);
4323 if (result != NULL) {
4324 return result;
4325 } else {
4326 // Let's try to allocate in the survivors in case we can fit the
4327 // object there.
4328 return survivor_attempt_allocation(word_size);
4329 }
4330 }
4331
4332 ShouldNotReachHere();
4333 // Trying to keep some compilers happy.
4334 return NULL;
4335 }
4336
4337 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4338 ParGCAllocBuffer(gclab_word_size), _retired(false) { }
4339
4340 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num)
4341 : _g1h(g1h),
4342 _refs(g1h->task_queue(queue_num)),
4343 _dcq(&g1h->dirty_card_queue_set()),
4344 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4345 _g1_rem(g1h->g1_rem_set()),
4346 _hash_seed(17), _queue_num(queue_num),
4347 _term_attempts(0),
4348 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4349 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4350 _age_table(false),
4351 _strong_roots_time(0), _term_time(0),
4352 _alloc_buffer_waste(0), _undo_waste(0) {
4353 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4354 // we "sacrifice" entry 0 to keep track of surviving bytes for
4355 // non-young regions (where the age is -1)
4356 // We also add a few elements at the beginning and at the end in
4357 // an attempt to eliminate cache contention
4358 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
4359 uint array_length = PADDING_ELEM_NUM +
4360 real_length +
4361 PADDING_ELEM_NUM;
4362 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
4363 if (_surviving_young_words_base == NULL)
4364 vm_exit_out_of_memory(array_length * sizeof(size_t),
4365 "Not enough space for young surv histo.");
4366 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4367 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t));
4368
4369 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4370 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4371
4372 _start = os::elapsedTime();
4373 }
4374
4375 void
4376 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4377 {
4378 st->print_raw_cr("GC Termination Stats");
4379 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4380 " ------waste (KiB)------");
4381 st->print_raw_cr("thr ms ms % ms % attempts"
4382 " total alloc undo");
4383 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4384 " ------- ------- -------");
4385 }
4386
4387 void
4388 G1ParScanThreadState::print_termination_stats(int i,
4389 outputStream* const st) const
4390 {
4391 const double elapsed_ms = elapsed_time() * 1000.0;
4392 const double s_roots_ms = strong_roots_time() * 1000.0;
4393 const double term_ms = term_time() * 1000.0;
4394 st->print_cr("%3d %9.2f %9.2f %6.2f "
4395 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4396 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4397 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4398 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4399 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4400 alloc_buffer_waste() * HeapWordSize / K,
4401 undo_waste() * HeapWordSize / K);
4402 }
4403
4404 #ifdef ASSERT
4405 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4406 assert(ref != NULL, "invariant");
4407 assert(UseCompressedOops, "sanity");
4408 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4409 oop p = oopDesc::load_decode_heap_oop(ref);
4410 assert(_g1h->is_in_g1_reserved(p),
4411 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4412 return true;
4413 }
4414
4415 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4416 assert(ref != NULL, "invariant");
4417 if (has_partial_array_mask(ref)) {
4418 // Must be in the collection set--it's already been copied.
4419 oop p = clear_partial_array_mask(ref);
4420 assert(_g1h->obj_in_cs(p),
4421 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4422 } else {
4423 oop p = oopDesc::load_decode_heap_oop(ref);
4424 assert(_g1h->is_in_g1_reserved(p),
4425 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4426 }
4427 return true;
4428 }
4429
4430 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4431 if (ref.is_narrow()) {
4432 return verify_ref((narrowOop*) ref);
4433 } else {
4434 return verify_ref((oop*) ref);
4435 }
4436 }
4437 #endif // ASSERT
4438
4439 void G1ParScanThreadState::trim_queue() {
4440 assert(_evac_cl != NULL, "not set");
4441 assert(_evac_failure_cl != NULL, "not set");
4442 assert(_partial_scan_cl != NULL, "not set");
4443
4444 StarTask ref;
4445 do {
4446 // Drain the overflow stack first, so other threads can steal.
4447 while (refs()->pop_overflow(ref)) {
4448 deal_with_reference(ref);
4449 }
4450
4451 while (refs()->pop_local(ref)) {
4452 deal_with_reference(ref);
4453 }
4454 } while (!refs()->is_empty());
4455 }
4456
4457 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1,
4458 G1ParScanThreadState* par_scan_state) :
4459 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4460 _par_scan_state(par_scan_state),
4461 _worker_id(par_scan_state->queue_num()),
4462 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4463 _mark_in_progress(_g1->mark_in_progress()) { }
4464
4465 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4466 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) {
4467 #ifdef ASSERT
4468 HeapRegion* hr = _g1->heap_region_containing(obj);
4469 assert(hr != NULL, "sanity");
4470 assert(!hr->in_collection_set(), "should not mark objects in the CSet");
4471 #endif // ASSERT
4472
4473 // We know that the object is not moving so it's safe to read its size.
4474 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id);
4475 }
4476
4477 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4478 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4479 ::mark_forwarded_object(oop from_obj, oop to_obj) {
4480 #ifdef ASSERT
4481 assert(from_obj->is_forwarded(), "from obj should be forwarded");
4482 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee");
4483 assert(from_obj != to_obj, "should not be self-forwarded");
4484
4485 HeapRegion* from_hr = _g1->heap_region_containing(from_obj);
4486 assert(from_hr != NULL, "sanity");
4487 assert(from_hr->in_collection_set(), "from obj should be in the CSet");
4488
4489 HeapRegion* to_hr = _g1->heap_region_containing(to_obj);
4490 assert(to_hr != NULL, "sanity");
4491 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet");
4492 #endif // ASSERT
4493
4494 // The object might be in the process of being copied by another
4495 // worker so we cannot trust that its to-space image is
4496 // well-formed. So we have to read its size from its from-space
4497 // image which we know should not be changing.
4498 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id);
4499 }
4500
4501 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4502 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4503 ::copy_to_survivor_space(oop old) {
4504 size_t word_sz = old->size();
4505 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4506 // +1 to make the -1 indexes valid...
4507 int young_index = from_region->young_index_in_cset()+1;
4508 assert( (from_region->is_young() && young_index > 0) ||
4509 (!from_region->is_young() && young_index == 0), "invariant" );
4510 G1CollectorPolicy* g1p = _g1->g1_policy();
4511 markOop m = old->mark();
4512 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
4513 : m->age();
4514 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4515 word_sz);
4516 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4517 #ifndef PRODUCT
4518 // Should this evacuation fail?
4519 if (_g1->evacuation_should_fail()) {
4520 if (obj_ptr != NULL) {
4521 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4522 obj_ptr = NULL;
4523 }
4524 }
4525 #endif // !PRODUCT
4526
4527 if (obj_ptr == NULL) {
4528 // This will either forward-to-self, or detect that someone else has
4529 // installed a forwarding pointer.
4530 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4531 return _g1->handle_evacuation_failure_par(cl, old);
4532 }
4533
4534 oop obj = oop(obj_ptr);
4535
4536 // We're going to allocate linearly, so might as well prefetch ahead.
4537 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4538
4539 oop forward_ptr = old->forward_to_atomic(obj);
4540 if (forward_ptr == NULL) {
4541 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4542 if (g1p->track_object_age(alloc_purpose)) {
4543 // We could simply do obj->incr_age(). However, this causes a
4544 // performance issue. obj->incr_age() will first check whether
4545 // the object has a displaced mark by checking its mark word;
4546 // getting the mark word from the new location of the object
4547 // stalls. So, given that we already have the mark word and we
4548 // are about to install it anyway, it's better to increase the
4549 // age on the mark word, when the object does not have a
4550 // displaced mark word. We're not expecting many objects to have
4551 // a displaced marked word, so that case is not optimized
4552 // further (it could be...) and we simply call obj->incr_age().
4553
4554 if (m->has_displaced_mark_helper()) {
4555 // in this case, we have to install the mark word first,
4556 // otherwise obj looks to be forwarded (the old mark word,
4557 // which contains the forward pointer, was copied)
4558 obj->set_mark(m);
4559 obj->incr_age();
4560 } else {
4561 m = m->incr_age();
4562 obj->set_mark(m);
4563 }
4564 _par_scan_state->age_table()->add(obj, word_sz);
4565 } else {
4566 obj->set_mark(m);
4567 }
4568
4569 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4570 surv_young_words[young_index] += word_sz;
4571
4572 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4573 // We keep track of the next start index in the length field of
4574 // the to-space object. The actual length can be found in the
4575 // length field of the from-space object.
4576 arrayOop(obj)->set_length(0);
4577 oop* old_p = set_partial_array_mask(old);
4578 _par_scan_state->push_on_queue(old_p);
4579 } else {
4580 // No point in using the slower heap_region_containing() method,
4581 // given that we know obj is in the heap.
4582 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4583 obj->oop_iterate_backwards(&_scanner);
4584 }
4585 } else {
4586 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4587 obj = forward_ptr;
4588 }
4589 return obj;
4590 }
4591
4592 template <class T>
4593 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) {
4594 if (_g1->heap_region_containing_raw(new_obj)->is_young()) {
4595 _scanned_klass->record_modified_oops();
4596 }
4597 }
4598
4599 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4600 template <class T>
4601 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4602 ::do_oop_work(T* p) {
4603 oop obj = oopDesc::load_decode_heap_oop(p);
4604 assert(barrier != G1BarrierRS || obj != NULL,
4605 "Precondition: G1BarrierRS implies obj is non-NULL");
4606
4607 assert(_worker_id == _par_scan_state->queue_num(), "sanity");
4608
4609 // here the null check is implicit in the cset_fast_test() test
4610 if (_g1->in_cset_fast_test(obj)) {
4611 oop forwardee;
4612 if (obj->is_forwarded()) {
4613 forwardee = obj->forwardee();
4614 } else {
4615 forwardee = copy_to_survivor_space(obj);
4616 }
4617 assert(forwardee != NULL, "forwardee should not be NULL");
4618 oopDesc::encode_store_heap_oop(p, forwardee);
4619 if (do_mark_object && forwardee != obj) {
4620 // If the object is self-forwarded we don't need to explicitly
4621 // mark it, the evacuation failure protocol will do so.
4622 mark_forwarded_object(obj, forwardee);
4623 }
4624
4625 // When scanning the RS, we only care about objs in CS.
4626 if (barrier == G1BarrierRS) {
4627 _par_scan_state->update_rs(_from, p, _worker_id);
4628 } else if (barrier == G1BarrierKlass) {
4629 do_klass_barrier(p, forwardee);
4630 }
4631 } else {
4632 // The object is not in collection set. If we're a root scanning
4633 // closure during an initial mark pause (i.e. do_mark_object will
4634 // be true) then attempt to mark the object.
4635 if (do_mark_object && _g1->is_in_g1_reserved(obj)) {
4636 mark_object(obj);
4637 }
4638 }
4639
4640 if (barrier == G1BarrierEvac && obj != NULL) {
4641 _par_scan_state->update_rs(_from, p, _worker_id);
4642 }
4643
4644 if (do_gen_barrier && obj != NULL) {
4645 par_do_barrier(p);
4646 }
4647 }
4648
4649 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4650 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4651
4652 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4653 assert(has_partial_array_mask(p), "invariant");
4654 oop from_obj = clear_partial_array_mask(p);
4655
4656 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap.");
4657 assert(from_obj->is_objArray(), "must be obj array");
4658 objArrayOop from_obj_array = objArrayOop(from_obj);
4659 // The from-space object contains the real length.
4660 int length = from_obj_array->length();
4661
4662 assert(from_obj->is_forwarded(), "must be forwarded");
4663 oop to_obj = from_obj->forwardee();
4664 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
4665 objArrayOop to_obj_array = objArrayOop(to_obj);
4666 // We keep track of the next start index in the length field of the
4667 // to-space object.
4668 int next_index = to_obj_array->length();
4669 assert(0 <= next_index && next_index < length,
4670 err_msg("invariant, next index: %d, length: %d", next_index, length));
4671
4672 int start = next_index;
4673 int end = length;
4674 int remainder = end - start;
4675 // We'll try not to push a range that's smaller than ParGCArrayScanChunk.
4676 if (remainder > 2 * ParGCArrayScanChunk) {
4677 end = start + ParGCArrayScanChunk;
4678 to_obj_array->set_length(end);
4679 // Push the remainder before we process the range in case another
4680 // worker has run out of things to do and can steal it.
4681 oop* from_obj_p = set_partial_array_mask(from_obj);
4682 _par_scan_state->push_on_queue(from_obj_p);
4683 } else {
4684 assert(length == end, "sanity");
4685 // We'll process the final range for this object. Restore the length
4686 // so that the heap remains parsable in case of evacuation failure.
4687 to_obj_array->set_length(end);
4688 }
4689 _scanner.set_region(_g1->heap_region_containing_raw(to_obj));
4690 // Process indexes [start,end). It will also process the header
4691 // along with the first chunk (i.e., the chunk with start == 0).
4692 // Note that at this point the length field of to_obj_array is not
4693 // correct given that we are using it to keep track of the next
4694 // start index. oop_iterate_range() (thankfully!) ignores the length
4695 // field and only relies on the start / end parameters. It does
4696 // however return the size of the object which will be incorrect. So
4697 // we have to ignore it even if we wanted to use it.
4698 to_obj_array->oop_iterate_range(&_scanner, start, end);
4699 }
4700
4701 class G1ParEvacuateFollowersClosure : public VoidClosure {
4702 protected:
4703 G1CollectedHeap* _g1h;
4704 G1ParScanThreadState* _par_scan_state;
4705 RefToScanQueueSet* _queues;
4706 ParallelTaskTerminator* _terminator;
4707
4708 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4709 RefToScanQueueSet* queues() { return _queues; }
4710 ParallelTaskTerminator* terminator() { return _terminator; }
4711
4712 public:
4713 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4714 G1ParScanThreadState* par_scan_state,
4715 RefToScanQueueSet* queues,
4716 ParallelTaskTerminator* terminator)
4717 : _g1h(g1h), _par_scan_state(par_scan_state),
4718 _queues(queues), _terminator(terminator) {}
4719
4720 void do_void();
4721
4722 private:
4723 inline bool offer_termination();
4724 };
4725
4726 bool G1ParEvacuateFollowersClosure::offer_termination() {
4727 G1ParScanThreadState* const pss = par_scan_state();
4728 pss->start_term_time();
4729 const bool res = terminator()->offer_termination();
4730 pss->end_term_time();
4731 return res;
4732 }
4733
4734 void G1ParEvacuateFollowersClosure::do_void() {
4735 StarTask stolen_task;
4736 G1ParScanThreadState* const pss = par_scan_state();
4737 pss->trim_queue();
4738
4739 do {
4740 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4741 assert(pss->verify_task(stolen_task), "sanity");
4742 if (stolen_task.is_narrow()) {
4743 pss->deal_with_reference((narrowOop*) stolen_task);
4744 } else {
4745 pss->deal_with_reference((oop*) stolen_task);
4746 }
4747
4748 // We've just processed a reference and we might have made
4749 // available new entries on the queues. So we have to make sure
4750 // we drain the queues as necessary.
4751 pss->trim_queue();
4752 }
4753 } while (!offer_termination());
4754
4755 pss->retire_alloc_buffers();
4756 }
4757
4758 class G1KlassScanClosure : public KlassClosure {
4759 G1ParCopyHelper* _closure;
4760 bool _process_only_dirty;
4761 int _count;
4762 public:
4763 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty)
4764 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {}
4765 void do_klass(Klass* klass) {
4766 // If the klass has not been dirtied we know that there's
4767 // no references into the young gen and we can skip it.
4768 if (!_process_only_dirty || klass->has_modified_oops()) {
4769 // Clean the klass since we're going to scavenge all the metadata.
4770 klass->clear_modified_oops();
4771
4772 // Tell the closure that this klass is the Klass to scavenge
4773 // and is the one to dirty if oops are left pointing into the young gen.
4774 _closure->set_scanned_klass(klass);
4775
4776 klass->oops_do(_closure);
4777
4778 _closure->set_scanned_klass(NULL);
4779 }
4780 _count++;
4781 }
4782 };
4783
4784 class G1ParTask : public AbstractGangTask {
4785 protected:
4786 G1CollectedHeap* _g1h;
4787 RefToScanQueueSet *_queues;
4788 ParallelTaskTerminator _terminator;
4789 uint _n_workers;
4790
4791 Mutex _stats_lock;
4792 Mutex* stats_lock() { return &_stats_lock; }
4793
4794 size_t getNCards() {
4795 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4796 / G1BlockOffsetSharedArray::N_bytes;
4797 }
4798
4799 public:
4800 G1ParTask(G1CollectedHeap* g1h,
4801 RefToScanQueueSet *task_queues)
4802 : AbstractGangTask("G1 collection"),
4803 _g1h(g1h),
4804 _queues(task_queues),
4805 _terminator(0, _queues),
4806 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true)
4807 {}
4808
4809 RefToScanQueueSet* queues() { return _queues; }
4810
4811 RefToScanQueue *work_queue(int i) {
4812 return queues()->queue(i);
4813 }
4814
4815 ParallelTaskTerminator* terminator() { return &_terminator; }
4816
4817 virtual void set_for_termination(int active_workers) {
4818 // This task calls set_n_termination() in par_non_clean_card_iterate_work()
4819 // in the young space (_par_seq_tasks) in the G1 heap
4820 // for SequentialSubTasksDone.
4821 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap
4822 // both of which need setting by set_n_termination().
4823 _g1h->SharedHeap::set_n_termination(active_workers);
4824 _g1h->set_n_termination(active_workers);
4825 terminator()->reset_for_reuse(active_workers);
4826 _n_workers = active_workers;
4827 }
4828
4829 void work(uint worker_id) {
4830 if (worker_id >= _n_workers) return; // no work needed this round
4831
4832 double start_time_ms = os::elapsedTime() * 1000.0;
4833 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms);
4834
4835 {
4836 ResourceMark rm;
4837 HandleMark hm;
4838
4839 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4840
4841 G1ParScanThreadState pss(_g1h, worker_id);
4842 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4843 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4844 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4845
4846 pss.set_evac_closure(&scan_evac_cl);
4847 pss.set_evac_failure_closure(&evac_failure_cl);
4848 pss.set_partial_scan_closure(&partial_scan_cl);
4849
4850 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4851 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp);
4852
4853 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4854 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp);
4855
4856 bool only_young = _g1h->g1_policy()->gcs_are_young();
4857 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false);
4858 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young);
4859
4860 OopClosure* scan_root_cl = &only_scan_root_cl;
4861 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s;
4862
4863 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4864 // We also need to mark copied objects.
4865 scan_root_cl = &scan_mark_root_cl;
4866 scan_klasses_cl = &scan_mark_klasses_cl_s;
4867 }
4868
4869 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4870
4871 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
4872
4873 pss.start_strong_roots();
4874 _g1h->g1_process_strong_roots(/* is scavenging */ true,
4875 SharedHeap::ScanningOption(so),
4876 scan_root_cl,
4877 &push_heap_rs_cl,
4878 scan_klasses_cl,
4879 worker_id);
4880 pss.end_strong_roots();
4881
4882 {
4883 double start = os::elapsedTime();
4884 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4885 evac.do_void();
4886 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4887 double term_ms = pss.term_time()*1000.0;
4888 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms);
4889 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts());
4890 }
4891 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4892 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4893
4894 if (ParallelGCVerbose) {
4895 MutexLocker x(stats_lock());
4896 pss.print_termination_stats(worker_id);
4897 }
4898
4899 assert(pss.refs()->is_empty(), "should be empty");
4900
4901 // Close the inner scope so that the ResourceMark and HandleMark
4902 // destructors are executed here and are included as part of the
4903 // "GC Worker Time".
4904 }
4905
4906 double end_time_ms = os::elapsedTime() * 1000.0;
4907 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms);
4908 }
4909 };
4910
4911 // *** Common G1 Evacuation Stuff
4912
4913 // Closures that support the filtering of CodeBlobs scanned during
4914 // external root scanning.
4915
4916 // Closure applied to reference fields in code blobs (specifically nmethods)
4917 // to determine whether an nmethod contains references that point into
4918 // the collection set. Used as a predicate when walking code roots so
4919 // that only nmethods that point into the collection set are added to the
4920 // 'marked' list.
4921
4922 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure {
4923
4924 class G1PointsIntoCSOopClosure : public OopClosure {
4925 G1CollectedHeap* _g1;
4926 bool _points_into_cs;
4927 public:
4928 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) :
4929 _g1(g1), _points_into_cs(false) { }
4930
4931 bool points_into_cs() const { return _points_into_cs; }
4932
4933 template <class T>
4934 void do_oop_nv(T* p) {
4935 if (!_points_into_cs) {
4936 T heap_oop = oopDesc::load_heap_oop(p);
4937 if (!oopDesc::is_null(heap_oop) &&
4938 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) {
4939 _points_into_cs = true;
4940 }
4941 }
4942 }
4943
4944 virtual void do_oop(oop* p) { do_oop_nv(p); }
4945 virtual void do_oop(narrowOop* p) { do_oop_nv(p); }
4946 };
4947
4948 G1CollectedHeap* _g1;
4949
4950 public:
4951 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) :
4952 CodeBlobToOopClosure(cl, true), _g1(g1) { }
4953
4954 virtual void do_code_blob(CodeBlob* cb) {
4955 nmethod* nm = cb->as_nmethod_or_null();
4956 if (nm != NULL && !(nm->test_oops_do_mark())) {
4957 G1PointsIntoCSOopClosure predicate_cl(_g1);
4958 nm->oops_do(&predicate_cl);
4959
4960 if (predicate_cl.points_into_cs()) {
4961 // At least one of the reference fields or the oop relocations
4962 // in the nmethod points into the collection set. We have to
4963 // 'mark' this nmethod.
4964 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob()
4965 // or MarkingCodeBlobClosure::do_code_blob() change.
4966 if (!nm->test_set_oops_do_mark()) {
4967 do_newly_marked_nmethod(nm);
4968 }
4969 }
4970 }
4971 }
4972 };
4973
4974 // This method is run in a GC worker.
4975
4976 void
4977 G1CollectedHeap::
4978 g1_process_strong_roots(bool is_scavenging,
4979 ScanningOption so,
4980 OopClosure* scan_non_heap_roots,
4981 OopsInHeapRegionClosure* scan_rs,
4982 G1KlassScanClosure* scan_klasses,
4983 int worker_i) {
4984
4985 // First scan the strong roots
4986 double ext_roots_start = os::elapsedTime();
4987 double closure_app_time_sec = 0.0;
4988
4989 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
4990
4991 // Walk the code cache w/o buffering, because StarTask cannot handle
4992 // unaligned oop locations.
4993 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots);
4994
4995 process_strong_roots(false, // no scoping; this is parallel code
4996 is_scavenging, so,
4997 &buf_scan_non_heap_roots,
4998 &eager_scan_code_roots,
4999 scan_klasses
5000 );
5001
5002 // Now the CM ref_processor roots.
5003 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
5004 // We need to treat the discovered reference lists of the
5005 // concurrent mark ref processor as roots and keep entries
5006 // (which are added by the marking threads) on them live
5007 // until they can be processed at the end of marking.
5008 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
5009 }
5010
5011 // Finish up any enqueued closure apps (attributed as object copy time).
5012 buf_scan_non_heap_roots.done();
5013
5014 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds();
5015
5016 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
5017
5018 double ext_root_time_ms =
5019 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0;
5020
5021 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
5022
5023 // During conc marking we have to filter the per-thread SATB buffers
5024 // to make sure we remove any oops into the CSet (which will show up
5025 // as implicitly live).
5026 double satb_filtering_ms = 0.0;
5027 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) {
5028 if (mark_in_progress()) {
5029 double satb_filter_start = os::elapsedTime();
5030
5031 JavaThread::satb_mark_queue_set().filter_thread_buffers();
5032
5033 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0;
5034 }
5035 }
5036 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms);
5037
5038 // Now scan the complement of the collection set.
5039 if (scan_rs != NULL) {
5040 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
5041 }
5042 _process_strong_tasks->all_tasks_completed();
5043 }
5044
5045 void
5046 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure,
5047 OopClosure* non_root_closure) {
5048 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
5049 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure);
5050 }
5051
5052 // Weak Reference Processing support
5053
5054 // An always "is_alive" closure that is used to preserve referents.
5055 // If the object is non-null then it's alive. Used in the preservation
5056 // of referent objects that are pointed to by reference objects
5057 // discovered by the CM ref processor.
5058 class G1AlwaysAliveClosure: public BoolObjectClosure {
5059 G1CollectedHeap* _g1;
5060 public:
5061 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5062 void do_object(oop p) { assert(false, "Do not call."); }
5063 bool do_object_b(oop p) {
5064 if (p != NULL) {
5065 return true;
5066 }
5067 return false;
5068 }
5069 };
5070
5071 bool G1STWIsAliveClosure::do_object_b(oop p) {
5072 // An object is reachable if it is outside the collection set,
5073 // or is inside and copied.
5074 return !_g1->obj_in_cs(p) || p->is_forwarded();
5075 }
5076
5077 // Non Copying Keep Alive closure
5078 class G1KeepAliveClosure: public OopClosure {
5079 G1CollectedHeap* _g1;
5080 public:
5081 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5082 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
5083 void do_oop( oop* p) {
5084 oop obj = *p;
5085
5086 if (_g1->obj_in_cs(obj)) {
5087 assert( obj->is_forwarded(), "invariant" );
5088 *p = obj->forwardee();
5089 }
5090 }
5091 };
5092
5093 // Copying Keep Alive closure - can be called from both
5094 // serial and parallel code as long as different worker
5095 // threads utilize different G1ParScanThreadState instances
5096 // and different queues.
5097
5098 class G1CopyingKeepAliveClosure: public OopClosure {
5099 G1CollectedHeap* _g1h;
5100 OopClosure* _copy_non_heap_obj_cl;
5101 OopsInHeapRegionClosure* _copy_metadata_obj_cl;
5102 G1ParScanThreadState* _par_scan_state;
5103
5104 public:
5105 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
5106 OopClosure* non_heap_obj_cl,
5107 OopsInHeapRegionClosure* metadata_obj_cl,
5108 G1ParScanThreadState* pss):
5109 _g1h(g1h),
5110 _copy_non_heap_obj_cl(non_heap_obj_cl),
5111 _copy_metadata_obj_cl(metadata_obj_cl),
5112 _par_scan_state(pss)
5113 {}
5114
5115 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
5116 virtual void do_oop( oop* p) { do_oop_work(p); }
5117
5118 template <class T> void do_oop_work(T* p) {
5119 oop obj = oopDesc::load_decode_heap_oop(p);
5120
5121 if (_g1h->obj_in_cs(obj)) {
5122 // If the referent object has been forwarded (either copied
5123 // to a new location or to itself in the event of an
5124 // evacuation failure) then we need to update the reference
5125 // field and, if both reference and referent are in the G1
5126 // heap, update the RSet for the referent.
5127 //
5128 // If the referent has not been forwarded then we have to keep
5129 // it alive by policy. Therefore we have copy the referent.
5130 //
5131 // If the reference field is in the G1 heap then we can push
5132 // on the PSS queue. When the queue is drained (after each
5133 // phase of reference processing) the object and it's followers
5134 // will be copied, the reference field set to point to the
5135 // new location, and the RSet updated. Otherwise we need to
5136 // use the the non-heap or metadata closures directly to copy
5137 // the refernt object and update the pointer, while avoiding
5138 // updating the RSet.
5139
5140 if (_g1h->is_in_g1_reserved(p)) {
5141 _par_scan_state->push_on_queue(p);
5142 } else {
5143 assert(!ClassLoaderDataGraph::contains((address)p),
5144 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) "
5145 PTR_FORMAT, p));
5146 _copy_non_heap_obj_cl->do_oop(p);
5147 }
5148 }
5149 }
5150 };
5151
5152 // Serial drain queue closure. Called as the 'complete_gc'
5153 // closure for each discovered list in some of the
5154 // reference processing phases.
5155
5156 class G1STWDrainQueueClosure: public VoidClosure {
5157 protected:
5158 G1CollectedHeap* _g1h;
5159 G1ParScanThreadState* _par_scan_state;
5160
5161 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
5162
5163 public:
5164 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
5165 _g1h(g1h),
5166 _par_scan_state(pss)
5167 { }
5168
5169 void do_void() {
5170 G1ParScanThreadState* const pss = par_scan_state();
5171 pss->trim_queue();
5172 }
5173 };
5174
5175 // Parallel Reference Processing closures
5176
5177 // Implementation of AbstractRefProcTaskExecutor for parallel reference
5178 // processing during G1 evacuation pauses.
5179
5180 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
5181 private:
5182 G1CollectedHeap* _g1h;
5183 RefToScanQueueSet* _queues;
5184 FlexibleWorkGang* _workers;
5185 int _active_workers;
5186
5187 public:
5188 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
5189 FlexibleWorkGang* workers,
5190 RefToScanQueueSet *task_queues,
5191 int n_workers) :
5192 _g1h(g1h),
5193 _queues(task_queues),
5194 _workers(workers),
5195 _active_workers(n_workers)
5196 {
5197 assert(n_workers > 0, "shouldn't call this otherwise");
5198 }
5199
5200 // Executes the given task using concurrent marking worker threads.
5201 virtual void execute(ProcessTask& task);
5202 virtual void execute(EnqueueTask& task);
5203 };
5204
5205 // Gang task for possibly parallel reference processing
5206
5207 class G1STWRefProcTaskProxy: public AbstractGangTask {
5208 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5209 ProcessTask& _proc_task;
5210 G1CollectedHeap* _g1h;
5211 RefToScanQueueSet *_task_queues;
5212 ParallelTaskTerminator* _terminator;
5213
5214 public:
5215 G1STWRefProcTaskProxy(ProcessTask& proc_task,
5216 G1CollectedHeap* g1h,
5217 RefToScanQueueSet *task_queues,
5218 ParallelTaskTerminator* terminator) :
5219 AbstractGangTask("Process reference objects in parallel"),
5220 _proc_task(proc_task),
5221 _g1h(g1h),
5222 _task_queues(task_queues),
5223 _terminator(terminator)
5224 {}
5225
5226 virtual void work(uint worker_id) {
5227 // The reference processing task executed by a single worker.
5228 ResourceMark rm;
5229 HandleMark hm;
5230
5231 G1STWIsAliveClosure is_alive(_g1h);
5232
5233 G1ParScanThreadState pss(_g1h, worker_id);
5234
5235 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5236 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5237 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5238
5239 pss.set_evac_closure(&scan_evac_cl);
5240 pss.set_evac_failure_closure(&evac_failure_cl);
5241 pss.set_partial_scan_closure(&partial_scan_cl);
5242
5243 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5244 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5245
5246 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5247 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5248
5249 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5250 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5251
5252 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5253 // We also need to mark copied objects.
5254 copy_non_heap_cl = ©_mark_non_heap_cl;
5255 copy_metadata_cl = ©_mark_metadata_cl;
5256 }
5257
5258 // Keep alive closure.
5259 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5260
5261 // Complete GC closure
5262 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
5263
5264 // Call the reference processing task's work routine.
5265 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue);
5266
5267 // Note we cannot assert that the refs array is empty here as not all
5268 // of the processing tasks (specifically phase2 - pp2_work) execute
5269 // the complete_gc closure (which ordinarily would drain the queue) so
5270 // the queue may not be empty.
5271 }
5272 };
5273
5274 // Driver routine for parallel reference processing.
5275 // Creates an instance of the ref processing gang
5276 // task and has the worker threads execute it.
5277 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
5278 assert(_workers != NULL, "Need parallel worker threads.");
5279
5280 ParallelTaskTerminator terminator(_active_workers, _queues);
5281 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
5282
5283 _g1h->set_par_threads(_active_workers);
5284 _workers->run_task(&proc_task_proxy);
5285 _g1h->set_par_threads(0);
5286 }
5287
5288 // Gang task for parallel reference enqueueing.
5289
5290 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
5291 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5292 EnqueueTask& _enq_task;
5293
5294 public:
5295 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
5296 AbstractGangTask("Enqueue reference objects in parallel"),
5297 _enq_task(enq_task)
5298 { }
5299
5300 virtual void work(uint worker_id) {
5301 _enq_task.work(worker_id);
5302 }
5303 };
5304
5305 // Driver routine for parallel reference enqueing.
5306 // Creates an instance of the ref enqueueing gang
5307 // task and has the worker threads execute it.
5308
5309 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5310 assert(_workers != NULL, "Need parallel worker threads.");
5311
5312 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5313
5314 _g1h->set_par_threads(_active_workers);
5315 _workers->run_task(&enq_task_proxy);
5316 _g1h->set_par_threads(0);
5317 }
5318
5319 // End of weak reference support closures
5320
5321 // Abstract task used to preserve (i.e. copy) any referent objects
5322 // that are in the collection set and are pointed to by reference
5323 // objects discovered by the CM ref processor.
5324
5325 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5326 protected:
5327 G1CollectedHeap* _g1h;
5328 RefToScanQueueSet *_queues;
5329 ParallelTaskTerminator _terminator;
5330 uint _n_workers;
5331
5332 public:
5333 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5334 AbstractGangTask("ParPreserveCMReferents"),
5335 _g1h(g1h),
5336 _queues(task_queues),
5337 _terminator(workers, _queues),
5338 _n_workers(workers)
5339 { }
5340
5341 void work(uint worker_id) {
5342 ResourceMark rm;
5343 HandleMark hm;
5344
5345 G1ParScanThreadState pss(_g1h, worker_id);
5346 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5347 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5348 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5349
5350 pss.set_evac_closure(&scan_evac_cl);
5351 pss.set_evac_failure_closure(&evac_failure_cl);
5352 pss.set_partial_scan_closure(&partial_scan_cl);
5353
5354 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5355
5356
5357 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5358 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5359
5360 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5361 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5362
5363 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5364 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5365
5366 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5367 // We also need to mark copied objects.
5368 copy_non_heap_cl = ©_mark_non_heap_cl;
5369 copy_metadata_cl = ©_mark_metadata_cl;
5370 }
5371
5372 // Is alive closure
5373 G1AlwaysAliveClosure always_alive(_g1h);
5374
5375 // Copying keep alive closure. Applied to referent objects that need
5376 // to be copied.
5377 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5378
5379 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5380
5381 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5382 uint stride = MIN2(MAX2(_n_workers, 1U), limit);
5383
5384 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5385 // So this must be true - but assert just in case someone decides to
5386 // change the worker ids.
5387 assert(0 <= worker_id && worker_id < limit, "sanity");
5388 assert(!rp->discovery_is_atomic(), "check this code");
5389
5390 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5391 for (uint idx = worker_id; idx < limit; idx += stride) {
5392 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5393
5394 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5395 while (iter.has_next()) {
5396 // Since discovery is not atomic for the CM ref processor, we
5397 // can see some null referent objects.
5398 iter.load_ptrs(DEBUG_ONLY(true));
5399 oop ref = iter.obj();
5400
5401 // This will filter nulls.
5402 if (iter.is_referent_alive()) {
5403 iter.make_referent_alive();
5404 }
5405 iter.move_to_next();
5406 }
5407 }
5408
5409 // Drain the queue - which may cause stealing
5410 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5411 drain_queue.do_void();
5412 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5413 assert(pss.refs()->is_empty(), "should be");
5414 }
5415 };
5416
5417 // Weak Reference processing during an evacuation pause (part 1).
5418 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) {
5419 double ref_proc_start = os::elapsedTime();
5420
5421 ReferenceProcessor* rp = _ref_processor_stw;
5422 assert(rp->discovery_enabled(), "should have been enabled");
5423
5424 // Any reference objects, in the collection set, that were 'discovered'
5425 // by the CM ref processor should have already been copied (either by
5426 // applying the external root copy closure to the discovered lists, or
5427 // by following an RSet entry).
5428 //
5429 // But some of the referents, that are in the collection set, that these
5430 // reference objects point to may not have been copied: the STW ref
5431 // processor would have seen that the reference object had already
5432 // been 'discovered' and would have skipped discovering the reference,
5433 // but would not have treated the reference object as a regular oop.
5434 // As a reult the copy closure would not have been applied to the
5435 // referent object.
5436 //
5437 // We need to explicitly copy these referent objects - the references
5438 // will be processed at the end of remarking.
5439 //
5440 // We also need to do this copying before we process the reference
5441 // objects discovered by the STW ref processor in case one of these
5442 // referents points to another object which is also referenced by an
5443 // object discovered by the STW ref processor.
5444
5445 assert(!G1CollectedHeap::use_parallel_gc_threads() ||
5446 no_of_gc_workers == workers()->active_workers(),
5447 "Need to reset active GC workers");
5448
5449 set_par_threads(no_of_gc_workers);
5450 G1ParPreserveCMReferentsTask keep_cm_referents(this,
5451 no_of_gc_workers,
5452 _task_queues);
5453
5454 if (G1CollectedHeap::use_parallel_gc_threads()) {
5455 workers()->run_task(&keep_cm_referents);
5456 } else {
5457 keep_cm_referents.work(0);
5458 }
5459
5460 set_par_threads(0);
5461
5462 // Closure to test whether a referent is alive.
5463 G1STWIsAliveClosure is_alive(this);
5464
5465 // Even when parallel reference processing is enabled, the processing
5466 // of JNI refs is serial and performed serially by the current thread
5467 // rather than by a worker. The following PSS will be used for processing
5468 // JNI refs.
5469
5470 // Use only a single queue for this PSS.
5471 G1ParScanThreadState pss(this, 0);
5472
5473 // We do not embed a reference processor in the copying/scanning
5474 // closures while we're actually processing the discovered
5475 // reference objects.
5476 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5477 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5478 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5479
5480 pss.set_evac_closure(&scan_evac_cl);
5481 pss.set_evac_failure_closure(&evac_failure_cl);
5482 pss.set_partial_scan_closure(&partial_scan_cl);
5483
5484 assert(pss.refs()->is_empty(), "pre-condition");
5485
5486 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5487 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL);
5488
5489 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5490 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL);
5491
5492 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5493 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5494
5495 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5496 // We also need to mark copied objects.
5497 copy_non_heap_cl = ©_mark_non_heap_cl;
5498 copy_metadata_cl = ©_mark_metadata_cl;
5499 }
5500
5501 // Keep alive closure.
5502 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss);
5503
5504 // Serial Complete GC closure
5505 G1STWDrainQueueClosure drain_queue(this, &pss);
5506
5507 // Setup the soft refs policy...
5508 rp->setup_policy(false);
5509
5510 if (!rp->processing_is_mt()) {
5511 // Serial reference processing...
5512 rp->process_discovered_references(&is_alive,
5513 &keep_alive,
5514 &drain_queue,
5515 NULL);
5516 } else {
5517 // Parallel reference processing
5518 assert(rp->num_q() == no_of_gc_workers, "sanity");
5519 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5520
5521 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5522 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5523 }
5524
5525 // We have completed copying any necessary live referent objects
5526 // (that were not copied during the actual pause) so we can
5527 // retire any active alloc buffers
5528 pss.retire_alloc_buffers();
5529 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5530
5531 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5532 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0);
5533 }
5534
5535 // Weak Reference processing during an evacuation pause (part 2).
5536 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) {
5537 double ref_enq_start = os::elapsedTime();
5538
5539 ReferenceProcessor* rp = _ref_processor_stw;
5540 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5541
5542 // Now enqueue any remaining on the discovered lists on to
5543 // the pending list.
5544 if (!rp->processing_is_mt()) {
5545 // Serial reference processing...
5546 rp->enqueue_discovered_references();
5547 } else {
5548 // Parallel reference enqueuing
5549
5550 assert(no_of_gc_workers == workers()->active_workers(),
5551 "Need to reset active workers");
5552 assert(rp->num_q() == no_of_gc_workers, "sanity");
5553 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5554
5555 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5556 rp->enqueue_discovered_references(&par_task_executor);
5557 }
5558
5559 rp->verify_no_references_recorded();
5560 assert(!rp->discovery_enabled(), "should have been disabled");
5561
5562 // FIXME
5563 // CM's reference processing also cleans up the string and symbol tables.
5564 // Should we do that here also? We could, but it is a serial operation
5565 // and could signicantly increase the pause time.
5566
5567 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5568 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0);
5569 }
5570
5571 void G1CollectedHeap::evacuate_collection_set() {
5572 _expand_heap_after_alloc_failure = true;
5573 set_evacuation_failed(false);
5574
5575 // Should G1EvacuationFailureALot be in effect for this GC?
5576 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();)
5577
5578 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5579 concurrent_g1_refine()->set_use_cache(false);
5580 concurrent_g1_refine()->clear_hot_cache_claimed_index();
5581
5582 uint n_workers;
5583 if (G1CollectedHeap::use_parallel_gc_threads()) {
5584 n_workers =
5585 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
5586 workers()->active_workers(),
5587 Threads::number_of_non_daemon_threads());
5588 assert(UseDynamicNumberOfGCThreads ||
5589 n_workers == workers()->total_workers(),
5590 "If not dynamic should be using all the workers");
5591 workers()->set_active_workers(n_workers);
5592 set_par_threads(n_workers);
5593 } else {
5594 assert(n_par_threads() == 0,
5595 "Should be the original non-parallel value");
5596 n_workers = 1;
5597 }
5598
5599 G1ParTask g1_par_task(this, _task_queues);
5600
5601 init_for_evac_failure(NULL);
5602
5603 rem_set()->prepare_for_younger_refs_iterate(true);
5604
5605 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5606 double start_par_time_sec = os::elapsedTime();
5607 double end_par_time_sec;
5608
5609 {
5610 StrongRootsScope srs(this);
5611
5612 if (G1CollectedHeap::use_parallel_gc_threads()) {
5613 // The individual threads will set their evac-failure closures.
5614 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5615 // These tasks use ShareHeap::_process_strong_tasks
5616 assert(UseDynamicNumberOfGCThreads ||
5617 workers()->active_workers() == workers()->total_workers(),
5618 "If not dynamic should be using all the workers");
5619 workers()->run_task(&g1_par_task);
5620 } else {
5621 g1_par_task.set_for_termination(n_workers);
5622 g1_par_task.work(0);
5623 }
5624 end_par_time_sec = os::elapsedTime();
5625
5626 // Closing the inner scope will execute the destructor
5627 // for the StrongRootsScope object. We record the current
5628 // elapsed time before closing the scope so that time
5629 // taken for the SRS destructor is NOT included in the
5630 // reported parallel time.
5631 }
5632
5633 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0;
5634 g1_policy()->phase_times()->record_par_time(par_time_ms);
5635
5636 double code_root_fixup_time_ms =
5637 (os::elapsedTime() - end_par_time_sec) * 1000.0;
5638 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms);
5639
5640 set_par_threads(0);
5641
5642 // Process any discovered reference objects - we have
5643 // to do this _before_ we retire the GC alloc regions
5644 // as we may have to copy some 'reachable' referent
5645 // objects (and their reachable sub-graphs) that were
5646 // not copied during the pause.
5647 process_discovered_references(n_workers);
5648
5649 // Weak root processing.
5650 // Note: when JSR 292 is enabled and code blobs can contain
5651 // non-perm oops then we will need to process the code blobs
5652 // here too.
5653 {
5654 G1STWIsAliveClosure is_alive(this);
5655 G1KeepAliveClosure keep_alive(this);
5656 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5657 }
5658
5659 release_gc_alloc_regions(n_workers);
5660 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5661
5662 concurrent_g1_refine()->clear_hot_cache();
5663 concurrent_g1_refine()->set_use_cache(true);
5664
5665 finalize_for_evac_failure();
5666
5667 if (evacuation_failed()) {
5668 remove_self_forwarding_pointers();
5669
5670 // Reset the G1EvacuationFailureALot counters and flags
5671 // Note: the values are reset only when an actual
5672 // evacuation failure occurs.
5673 NOT_PRODUCT(reset_evacuation_should_fail();)
5674 }
5675
5676 // Enqueue any remaining references remaining on the STW
5677 // reference processor's discovered lists. We need to do
5678 // this after the card table is cleaned (and verified) as
5679 // the act of enqueuing entries on to the pending list
5680 // will log these updates (and dirty their associated
5681 // cards). We need these updates logged to update any
5682 // RSets.
5683 enqueue_discovered_references(n_workers);
5684
5685 if (G1DeferredRSUpdate) {
5686 RedirtyLoggedCardTableEntryFastClosure redirty;
5687 dirty_card_queue_set().set_closure(&redirty);
5688 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5689
5690 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5691 dcq.merge_bufferlists(&dirty_card_queue_set());
5692 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5693 }
5694 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5695 }
5696
5697 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5698 size_t* pre_used,
5699 FreeRegionList* free_list,
5700 OldRegionSet* old_proxy_set,
5701 HumongousRegionSet* humongous_proxy_set,
5702 HRRSCleanupTask* hrrs_cleanup_task,
5703 bool par) {
5704 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5705 if (hr->isHumongous()) {
5706 assert(hr->startsHumongous(), "we should only see starts humongous");
5707 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5708 } else {
5709 _old_set.remove_with_proxy(hr, old_proxy_set);
5710 free_region(hr, pre_used, free_list, par);
5711 }
5712 } else {
5713 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5714 }
5715 }
5716
5717 void G1CollectedHeap::free_region(HeapRegion* hr,
5718 size_t* pre_used,
5719 FreeRegionList* free_list,
5720 bool par) {
5721 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5722 assert(!hr->is_empty(), "the region should not be empty");
5723 assert(free_list != NULL, "pre-condition");
5724
5725 *pre_used += hr->used();
5726 hr->hr_clear(par, true /* clear_space */);
5727 free_list->add_as_head(hr);
5728 }
5729
5730 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5731 size_t* pre_used,
5732 FreeRegionList* free_list,
5733 HumongousRegionSet* humongous_proxy_set,
5734 bool par) {
5735 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5736 assert(free_list != NULL, "pre-condition");
5737 assert(humongous_proxy_set != NULL, "pre-condition");
5738
5739 size_t hr_used = hr->used();
5740 size_t hr_capacity = hr->capacity();
5741 size_t hr_pre_used = 0;
5742 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5743 // We need to read this before we make the region non-humongous,
5744 // otherwise the information will be gone.
5745 uint last_index = hr->last_hc_index();
5746 hr->set_notHumongous();
5747 free_region(hr, &hr_pre_used, free_list, par);
5748
5749 uint i = hr->hrs_index() + 1;
5750 while (i < last_index) {
5751 HeapRegion* curr_hr = region_at(i);
5752 assert(curr_hr->continuesHumongous(), "invariant");
5753 curr_hr->set_notHumongous();
5754 free_region(curr_hr, &hr_pre_used, free_list, par);
5755 i += 1;
5756 }
5757 assert(hr_pre_used == hr_used,
5758 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5759 "should be the same", hr_pre_used, hr_used));
5760 *pre_used += hr_pre_used;
5761 }
5762
5763 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5764 FreeRegionList* free_list,
5765 OldRegionSet* old_proxy_set,
5766 HumongousRegionSet* humongous_proxy_set,
5767 bool par) {
5768 if (pre_used > 0) {
5769 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5770 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5771 assert(_summary_bytes_used >= pre_used,
5772 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5773 "should be >= pre_used: "SIZE_FORMAT,
5774 _summary_bytes_used, pre_used));
5775 _summary_bytes_used -= pre_used;
5776 }
5777 if (free_list != NULL && !free_list->is_empty()) {
5778 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5779 _free_list.add_as_head(free_list);
5780 }
5781 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) {
5782 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5783 _old_set.update_from_proxy(old_proxy_set);
5784 }
5785 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5786 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5787 _humongous_set.update_from_proxy(humongous_proxy_set);
5788 }
5789 }
5790
5791 class G1ParCleanupCTTask : public AbstractGangTask {
5792 CardTableModRefBS* _ct_bs;
5793 G1CollectedHeap* _g1h;
5794 HeapRegion* volatile _su_head;
5795 public:
5796 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5797 G1CollectedHeap* g1h) :
5798 AbstractGangTask("G1 Par Cleanup CT Task"),
5799 _ct_bs(ct_bs), _g1h(g1h) { }
5800
5801 void work(uint worker_id) {
5802 HeapRegion* r;
5803 while (r = _g1h->pop_dirty_cards_region()) {
5804 clear_cards(r);
5805 }
5806 }
5807
5808 void clear_cards(HeapRegion* r) {
5809 // Cards of the survivors should have already been dirtied.
5810 if (!r->is_survivor()) {
5811 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5812 }
5813 }
5814 };
5815
5816 #ifndef PRODUCT
5817 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5818 G1CollectedHeap* _g1h;
5819 CardTableModRefBS* _ct_bs;
5820 public:
5821 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5822 : _g1h(g1h), _ct_bs(ct_bs) { }
5823 virtual bool doHeapRegion(HeapRegion* r) {
5824 if (r->is_survivor()) {
5825 _g1h->verify_dirty_region(r);
5826 } else {
5827 _g1h->verify_not_dirty_region(r);
5828 }
5829 return false;
5830 }
5831 };
5832
5833 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5834 // All of the region should be clean.
5835 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5836 MemRegion mr(hr->bottom(), hr->end());
5837 ct_bs->verify_not_dirty_region(mr);
5838 }
5839
5840 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5841 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5842 // dirty allocated blocks as they allocate them. The thread that
5843 // retires each region and replaces it with a new one will do a
5844 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5845 // not dirty that area (one less thing to have to do while holding
5846 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5847 // is dirty.
5848 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5849 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5850 ct_bs->verify_dirty_region(mr);
5851 }
5852
5853 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5854 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5855 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5856 verify_dirty_region(hr);
5857 }
5858 }
5859
5860 void G1CollectedHeap::verify_dirty_young_regions() {
5861 verify_dirty_young_list(_young_list->first_region());
5862 }
5863 #endif
5864
5865 void G1CollectedHeap::cleanUpCardTable() {
5866 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5867 double start = os::elapsedTime();
5868
5869 {
5870 // Iterate over the dirty cards region list.
5871 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5872
5873 if (G1CollectedHeap::use_parallel_gc_threads()) {
5874 set_par_threads();
5875 workers()->run_task(&cleanup_task);
5876 set_par_threads(0);
5877 } else {
5878 while (_dirty_cards_region_list) {
5879 HeapRegion* r = _dirty_cards_region_list;
5880 cleanup_task.clear_cards(r);
5881 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5882 if (_dirty_cards_region_list == r) {
5883 // The last region.
5884 _dirty_cards_region_list = NULL;
5885 }
5886 r->set_next_dirty_cards_region(NULL);
5887 }
5888 }
5889 #ifndef PRODUCT
5890 if (G1VerifyCTCleanup || VerifyAfterGC) {
5891 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5892 heap_region_iterate(&cleanup_verifier);
5893 }
5894 #endif
5895 }
5896
5897 double elapsed = os::elapsedTime() - start;
5898 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0);
5899 }
5900
5901 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5902 size_t pre_used = 0;
5903 FreeRegionList local_free_list("Local List for CSet Freeing");
5904
5905 double young_time_ms = 0.0;
5906 double non_young_time_ms = 0.0;
5907
5908 // Since the collection set is a superset of the the young list,
5909 // all we need to do to clear the young list is clear its
5910 // head and length, and unlink any young regions in the code below
5911 _young_list->clear();
5912
5913 G1CollectorPolicy* policy = g1_policy();
5914
5915 double start_sec = os::elapsedTime();
5916 bool non_young = true;
5917
5918 HeapRegion* cur = cs_head;
5919 int age_bound = -1;
5920 size_t rs_lengths = 0;
5921
5922 while (cur != NULL) {
5923 assert(!is_on_master_free_list(cur), "sanity");
5924 if (non_young) {
5925 if (cur->is_young()) {
5926 double end_sec = os::elapsedTime();
5927 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5928 non_young_time_ms += elapsed_ms;
5929
5930 start_sec = os::elapsedTime();
5931 non_young = false;
5932 }
5933 } else {
5934 if (!cur->is_young()) {
5935 double end_sec = os::elapsedTime();
5936 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5937 young_time_ms += elapsed_ms;
5938
5939 start_sec = os::elapsedTime();
5940 non_young = true;
5941 }
5942 }
5943
5944 rs_lengths += cur->rem_set()->occupied();
5945
5946 HeapRegion* next = cur->next_in_collection_set();
5947 assert(cur->in_collection_set(), "bad CS");
5948 cur->set_next_in_collection_set(NULL);
5949 cur->set_in_collection_set(false);
5950
5951 if (cur->is_young()) {
5952 int index = cur->young_index_in_cset();
5953 assert(index != -1, "invariant");
5954 assert((uint) index < policy->young_cset_region_length(), "invariant");
5955 size_t words_survived = _surviving_young_words[index];
5956 cur->record_surv_words_in_group(words_survived);
5957
5958 // At this point the we have 'popped' cur from the collection set
5959 // (linked via next_in_collection_set()) but it is still in the
5960 // young list (linked via next_young_region()). Clear the
5961 // _next_young_region field.
5962 cur->set_next_young_region(NULL);
5963 } else {
5964 int index = cur->young_index_in_cset();
5965 assert(index == -1, "invariant");
5966 }
5967
5968 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
5969 (!cur->is_young() && cur->young_index_in_cset() == -1),
5970 "invariant" );
5971
5972 if (!cur->evacuation_failed()) {
5973 MemRegion used_mr = cur->used_region();
5974
5975 // And the region is empty.
5976 assert(!used_mr.is_empty(), "Should not have empty regions in a CS.");
5977 free_region(cur, &pre_used, &local_free_list, false /* par */);
5978 } else {
5979 cur->uninstall_surv_rate_group();
5980 if (cur->is_young()) {
5981 cur->set_young_index_in_cset(-1);
5982 }
5983 cur->set_not_young();
5984 cur->set_evacuation_failed(false);
5985 // The region is now considered to be old.
5986 _old_set.add(cur);
5987 }
5988 cur = next;
5989 }
5990
5991 policy->record_max_rs_lengths(rs_lengths);
5992 policy->cset_regions_freed();
5993
5994 double end_sec = os::elapsedTime();
5995 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5996
5997 if (non_young) {
5998 non_young_time_ms += elapsed_ms;
5999 } else {
6000 young_time_ms += elapsed_ms;
6001 }
6002
6003 update_sets_after_freeing_regions(pre_used, &local_free_list,
6004 NULL /* old_proxy_set */,
6005 NULL /* humongous_proxy_set */,
6006 false /* par */);
6007 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms);
6008 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms);
6009 }
6010
6011 // This routine is similar to the above but does not record
6012 // any policy statistics or update free lists; we are abandoning
6013 // the current incremental collection set in preparation of a
6014 // full collection. After the full GC we will start to build up
6015 // the incremental collection set again.
6016 // This is only called when we're doing a full collection
6017 // and is immediately followed by the tearing down of the young list.
6018
6019 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
6020 HeapRegion* cur = cs_head;
6021
6022 while (cur != NULL) {
6023 HeapRegion* next = cur->next_in_collection_set();
6024 assert(cur->in_collection_set(), "bad CS");
6025 cur->set_next_in_collection_set(NULL);
6026 cur->set_in_collection_set(false);
6027 cur->set_young_index_in_cset(-1);
6028 cur = next;
6029 }
6030 }
6031
6032 void G1CollectedHeap::set_free_regions_coming() {
6033 if (G1ConcRegionFreeingVerbose) {
6034 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6035 "setting free regions coming");
6036 }
6037
6038 assert(!free_regions_coming(), "pre-condition");
6039 _free_regions_coming = true;
6040 }
6041
6042 void G1CollectedHeap::reset_free_regions_coming() {
6043 assert(free_regions_coming(), "pre-condition");
6044
6045 {
6046 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6047 _free_regions_coming = false;
6048 SecondaryFreeList_lock->notify_all();
6049 }
6050
6051 if (G1ConcRegionFreeingVerbose) {
6052 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6053 "reset free regions coming");
6054 }
6055 }
6056
6057 void G1CollectedHeap::wait_while_free_regions_coming() {
6058 // Most of the time we won't have to wait, so let's do a quick test
6059 // first before we take the lock.
6060 if (!free_regions_coming()) {
6061 return;
6062 }
6063
6064 if (G1ConcRegionFreeingVerbose) {
6065 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6066 "waiting for free regions");
6067 }
6068
6069 {
6070 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6071 while (free_regions_coming()) {
6072 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
6073 }
6074 }
6075
6076 if (G1ConcRegionFreeingVerbose) {
6077 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6078 "done waiting for free regions");
6079 }
6080 }
6081
6082 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
6083 assert(heap_lock_held_for_gc(),
6084 "the heap lock should already be held by or for this thread");
6085 _young_list->push_region(hr);
6086 }
6087
6088 class NoYoungRegionsClosure: public HeapRegionClosure {
6089 private:
6090 bool _success;
6091 public:
6092 NoYoungRegionsClosure() : _success(true) { }
6093 bool doHeapRegion(HeapRegion* r) {
6094 if (r->is_young()) {
6095 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
6096 r->bottom(), r->end());
6097 _success = false;
6098 }
6099 return false;
6100 }
6101 bool success() { return _success; }
6102 };
6103
6104 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
6105 bool ret = _young_list->check_list_empty(check_sample);
6106
6107 if (check_heap) {
6108 NoYoungRegionsClosure closure;
6109 heap_region_iterate(&closure);
6110 ret = ret && closure.success();
6111 }
6112
6113 return ret;
6114 }
6115
6116 class TearDownRegionSetsClosure : public HeapRegionClosure {
6117 private:
6118 OldRegionSet *_old_set;
6119
6120 public:
6121 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { }
6122
6123 bool doHeapRegion(HeapRegion* r) {
6124 if (r->is_empty()) {
6125 // We ignore empty regions, we'll empty the free list afterwards
6126 } else if (r->is_young()) {
6127 // We ignore young regions, we'll empty the young list afterwards
6128 } else if (r->isHumongous()) {
6129 // We ignore humongous regions, we're not tearing down the
6130 // humongous region set
6131 } else {
6132 // The rest should be old
6133 _old_set->remove(r);
6134 }
6135 return false;
6136 }
6137
6138 ~TearDownRegionSetsClosure() {
6139 assert(_old_set->is_empty(), "post-condition");
6140 }
6141 };
6142
6143 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) {
6144 assert_at_safepoint(true /* should_be_vm_thread */);
6145
6146 if (!free_list_only) {
6147 TearDownRegionSetsClosure cl(&_old_set);
6148 heap_region_iterate(&cl);
6149
6150 // Need to do this after the heap iteration to be able to
6151 // recognize the young regions and ignore them during the iteration.
6152 _young_list->empty_list();
6153 }
6154 _free_list.remove_all();
6155 }
6156
6157 class RebuildRegionSetsClosure : public HeapRegionClosure {
6158 private:
6159 bool _free_list_only;
6160 OldRegionSet* _old_set;
6161 FreeRegionList* _free_list;
6162 size_t _total_used;
6163
6164 public:
6165 RebuildRegionSetsClosure(bool free_list_only,
6166 OldRegionSet* old_set, FreeRegionList* free_list) :
6167 _free_list_only(free_list_only),
6168 _old_set(old_set), _free_list(free_list), _total_used(0) {
6169 assert(_free_list->is_empty(), "pre-condition");
6170 if (!free_list_only) {
6171 assert(_old_set->is_empty(), "pre-condition");
6172 }
6173 }
6174
6175 bool doHeapRegion(HeapRegion* r) {
6176 if (r->continuesHumongous()) {
6177 return false;
6178 }
6179
6180 if (r->is_empty()) {
6181 // Add free regions to the free list
6182 _free_list->add_as_tail(r);
6183 } else if (!_free_list_only) {
6184 assert(!r->is_young(), "we should not come across young regions");
6185
6186 if (r->isHumongous()) {
6187 // We ignore humongous regions, we left the humongous set unchanged
6188 } else {
6189 // The rest should be old, add them to the old set
6190 _old_set->add(r);
6191 }
6192 _total_used += r->used();
6193 }
6194
6195 return false;
6196 }
6197
6198 size_t total_used() {
6199 return _total_used;
6200 }
6201 };
6202
6203 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) {
6204 assert_at_safepoint(true /* should_be_vm_thread */);
6205
6206 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list);
6207 heap_region_iterate(&cl);
6208
6209 if (!free_list_only) {
6210 _summary_bytes_used = cl.total_used();
6211 }
6212 assert(_summary_bytes_used == recalculate_used(),
6213 err_msg("inconsistent _summary_bytes_used, "
6214 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT,
6215 _summary_bytes_used, recalculate_used()));
6216 }
6217
6218 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
6219 _refine_cte_cl->set_concurrent(concurrent);
6220 }
6221
6222 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
6223 HeapRegion* hr = heap_region_containing(p);
6224 if (hr == NULL) {
6225 return false;
6226 } else {
6227 return hr->is_in(p);
6228 }
6229 }
6230
6231 // Methods for the mutator alloc region
6232
6233 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
6234 bool force) {
6235 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6236 assert(!force || g1_policy()->can_expand_young_list(),
6237 "if force is true we should be able to expand the young list");
6238 bool young_list_full = g1_policy()->is_young_list_full();
6239 if (force || !young_list_full) {
6240 HeapRegion* new_alloc_region = new_region(word_size,
6241 false /* do_expand */);
6242 if (new_alloc_region != NULL) {
6243 set_region_short_lived_locked(new_alloc_region);
6244 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
6245 return new_alloc_region;
6246 }
6247 }
6248 return NULL;
6249 }
6250
6251 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
6252 size_t allocated_bytes) {
6253 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6254 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
6255
6256 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
6257 _summary_bytes_used += allocated_bytes;
6258 _hr_printer.retire(alloc_region);
6259 // We update the eden sizes here, when the region is retired,
6260 // instead of when it's allocated, since this is the point that its
6261 // used space has been recored in _summary_bytes_used.
6262 g1mm()->update_eden_size();
6263 }
6264
6265 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
6266 bool force) {
6267 return _g1h->new_mutator_alloc_region(word_size, force);
6268 }
6269
6270 void G1CollectedHeap::set_par_threads() {
6271 // Don't change the number of workers. Use the value previously set
6272 // in the workgroup.
6273 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise");
6274 uint n_workers = workers()->active_workers();
6275 assert(UseDynamicNumberOfGCThreads ||
6276 n_workers == workers()->total_workers(),
6277 "Otherwise should be using the total number of workers");
6278 if (n_workers == 0) {
6279 assert(false, "Should have been set in prior evacuation pause.");
6280 n_workers = ParallelGCThreads;
6281 workers()->set_active_workers(n_workers);
6282 }
6283 set_par_threads(n_workers);
6284 }
6285
6286 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
6287 size_t allocated_bytes) {
6288 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
6289 }
6290
6291 // Methods for the GC alloc regions
6292
6293 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
6294 uint count,
6295 GCAllocPurpose ap) {
6296 assert(FreeList_lock->owned_by_self(), "pre-condition");
6297
6298 if (count < g1_policy()->max_regions(ap)) {
6299 HeapRegion* new_alloc_region = new_region(word_size,
6300 true /* do_expand */);
6301 if (new_alloc_region != NULL) {
6302 // We really only need to do this for old regions given that we
6303 // should never scan survivors. But it doesn't hurt to do it
6304 // for survivors too.
6305 new_alloc_region->set_saved_mark();
6306 if (ap == GCAllocForSurvived) {
6307 new_alloc_region->set_survivor();
6308 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
6309 } else {
6310 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
6311 }
6312 bool during_im = g1_policy()->during_initial_mark_pause();
6313 new_alloc_region->note_start_of_copying(during_im);
6314 return new_alloc_region;
6315 } else {
6316 g1_policy()->note_alloc_region_limit_reached(ap);
6317 }
6318 }
6319 return NULL;
6320 }
6321
6322 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
6323 size_t allocated_bytes,
6324 GCAllocPurpose ap) {
6325 bool during_im = g1_policy()->during_initial_mark_pause();
6326 alloc_region->note_end_of_copying(during_im);
6327 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
6328 if (ap == GCAllocForSurvived) {
6329 young_list()->add_survivor_region(alloc_region);
6330 } else {
6331 _old_set.add(alloc_region);
6332 }
6333 _hr_printer.retire(alloc_region);
6334 }
6335
6336 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
6337 bool force) {
6338 assert(!force, "not supported for GC alloc regions");
6339 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
6340 }
6341
6342 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
6343 size_t allocated_bytes) {
6344 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6345 GCAllocForSurvived);
6346 }
6347
6348 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
6349 bool force) {
6350 assert(!force, "not supported for GC alloc regions");
6351 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
6352 }
6353
6354 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
6355 size_t allocated_bytes) {
6356 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6357 GCAllocForTenured);
6358 }
6359 // Heap region set verification
6360
6361 class VerifyRegionListsClosure : public HeapRegionClosure {
6362 private:
6363 FreeRegionList* _free_list;
6364 OldRegionSet* _old_set;
6365 HumongousRegionSet* _humongous_set;
6366 uint _region_count;
6367
6368 public:
6369 VerifyRegionListsClosure(OldRegionSet* old_set,
6370 HumongousRegionSet* humongous_set,
6371 FreeRegionList* free_list) :
6372 _old_set(old_set), _humongous_set(humongous_set),
6373 _free_list(free_list), _region_count(0) { }
6374
6375 uint region_count() { return _region_count; }
6376
6377 bool doHeapRegion(HeapRegion* hr) {
6378 _region_count += 1;
6379
6380 if (hr->continuesHumongous()) {
6381 return false;
6382 }
6383
6384 if (hr->is_young()) {
6385 // TODO
6386 } else if (hr->startsHumongous()) {
6387 _humongous_set->verify_next_region(hr);
6388 } else if (hr->is_empty()) {
6389 _free_list->verify_next_region(hr);
6390 } else {
6391 _old_set->verify_next_region(hr);
6392 }
6393 return false;
6394 }
6395 };
6396
6397 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index,
6398 HeapWord* bottom) {
6399 HeapWord* end = bottom + HeapRegion::GrainWords;
6400 MemRegion mr(bottom, end);
6401 assert(_g1_reserved.contains(mr), "invariant");
6402 // This might return NULL if the allocation fails
6403 return new HeapRegion(hrs_index, _bot_shared, mr);
6404 }
6405
6406 void G1CollectedHeap::verify_region_sets() {
6407 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6408
6409 // First, check the explicit lists.
6410 _free_list.verify();
6411 {
6412 // Given that a concurrent operation might be adding regions to
6413 // the secondary free list we have to take the lock before
6414 // verifying it.
6415 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6416 _secondary_free_list.verify();
6417 }
6418 _old_set.verify();
6419 _humongous_set.verify();
6420
6421 // If a concurrent region freeing operation is in progress it will
6422 // be difficult to correctly attributed any free regions we come
6423 // across to the correct free list given that they might belong to
6424 // one of several (free_list, secondary_free_list, any local lists,
6425 // etc.). So, if that's the case we will skip the rest of the
6426 // verification operation. Alternatively, waiting for the concurrent
6427 // operation to complete will have a non-trivial effect on the GC's
6428 // operation (no concurrent operation will last longer than the
6429 // interval between two calls to verification) and it might hide
6430 // any issues that we would like to catch during testing.
6431 if (free_regions_coming()) {
6432 return;
6433 }
6434
6435 // Make sure we append the secondary_free_list on the free_list so
6436 // that all free regions we will come across can be safely
6437 // attributed to the free_list.
6438 append_secondary_free_list_if_not_empty_with_lock();
6439
6440 // Finally, make sure that the region accounting in the lists is
6441 // consistent with what we see in the heap.
6442 _old_set.verify_start();
6443 _humongous_set.verify_start();
6444 _free_list.verify_start();
6445
6446 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list);
6447 heap_region_iterate(&cl);
6448
6449 _old_set.verify_end();
6450 _humongous_set.verify_end();
6451 _free_list.verify_end();
6452 }