1 /*
2 * Copyright (c) 2001, 2013, 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/g1RemSet.inline.hpp"
41 #include "gc_implementation/g1/heapRegion.inline.hpp"
42 #include "gc_implementation/g1/heapRegionRemSet.hpp"
43 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
44 #include "gc_implementation/g1/vm_operations_g1.hpp"
45 #include "gc_implementation/shared/isGCActiveMark.hpp"
46 #include "memory/gcLocker.inline.hpp"
47 #include "memory/genOopClosures.inline.hpp"
48 #include "memory/generationSpec.hpp"
49 #include "memory/referenceProcessor.hpp"
50 #include "oops/oop.inline.hpp"
51 #include "oops/oop.pcgc.inline.hpp"
52 #include "runtime/aprofiler.hpp"
53 #include "runtime/vmThread.hpp"
54
55 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0;
56
57 // turn it on so that the contents of the young list (scan-only /
58 // to-be-collected) are printed at "strategic" points before / during
59 // / after the collection --- this is useful for debugging
60 #define YOUNG_LIST_VERBOSE 0
61 // CURRENT STATUS
62 // This file is under construction. Search for "FIXME".
63
64 // INVARIANTS/NOTES
65 //
66 // All allocation activity covered by the G1CollectedHeap interface is
67 // serialized by acquiring the HeapLock. This happens in mem_allocate
68 // and allocate_new_tlab, which are the "entry" points to the
69 // allocation code from the rest of the JVM. (Note that this does not
70 // apply to TLAB allocation, which is not part of this interface: it
71 // is done by clients of this interface.)
72
73 // Notes on implementation of parallelism in different tasks.
74 //
75 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism.
76 // The number of GC workers is passed to heap_region_par_iterate_chunked().
77 // It does use run_task() which sets _n_workers in the task.
78 // G1ParTask executes g1_process_strong_roots() ->
79 // SharedHeap::process_strong_roots() which calls eventuall to
80 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses
81 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also
82 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap).
83 //
84
85 // Local to this file.
86
87 class RefineCardTableEntryClosure: public CardTableEntryClosure {
88 SuspendibleThreadSet* _sts;
89 G1RemSet* _g1rs;
90 ConcurrentG1Refine* _cg1r;
91 bool _concurrent;
92 public:
93 RefineCardTableEntryClosure(SuspendibleThreadSet* sts,
94 G1RemSet* g1rs,
95 ConcurrentG1Refine* cg1r) :
96 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true)
97 {}
98 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
99 bool oops_into_cset = _g1rs->refine_card(card_ptr, worker_i, false);
100 // This path is executed by the concurrent refine or mutator threads,
101 // concurrently, and so we do not care if card_ptr contains references
102 // that point into the collection set.
103 assert(!oops_into_cset, "should be");
104
105 if (_concurrent && _sts->should_yield()) {
106 // Caller will actually yield.
107 return false;
108 }
109 // Otherwise, we finished successfully; return true.
110 return true;
111 }
112 void set_concurrent(bool b) { _concurrent = b; }
113 };
114
115
116 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure {
117 int _calls;
118 G1CollectedHeap* _g1h;
119 CardTableModRefBS* _ctbs;
120 int _histo[256];
121 public:
122 ClearLoggedCardTableEntryClosure() :
123 _calls(0)
124 {
125 _g1h = G1CollectedHeap::heap();
126 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
127 for (int i = 0; i < 256; i++) _histo[i] = 0;
128 }
129 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
130 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
131 _calls++;
132 unsigned char* ujb = (unsigned char*)card_ptr;
133 int ind = (int)(*ujb);
134 _histo[ind]++;
135 *card_ptr = -1;
136 }
137 return true;
138 }
139 int calls() { return _calls; }
140 void print_histo() {
141 gclog_or_tty->print_cr("Card table value histogram:");
142 for (int i = 0; i < 256; i++) {
143 if (_histo[i] != 0) {
144 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]);
145 }
146 }
147 }
148 };
149
150 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure {
151 int _calls;
152 G1CollectedHeap* _g1h;
153 CardTableModRefBS* _ctbs;
154 public:
155 RedirtyLoggedCardTableEntryClosure() :
156 _calls(0)
157 {
158 _g1h = G1CollectedHeap::heap();
159 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
160 }
161 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
162 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
163 _calls++;
164 *card_ptr = 0;
165 }
166 return true;
167 }
168 int calls() { return _calls; }
169 };
170
171 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure {
172 public:
173 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
174 *card_ptr = CardTableModRefBS::dirty_card_val();
175 return true;
176 }
177 };
178
179 YoungList::YoungList(G1CollectedHeap* g1h) :
180 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0),
181 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) {
182 guarantee(check_list_empty(false), "just making sure...");
183 }
184
185 void YoungList::push_region(HeapRegion *hr) {
186 assert(!hr->is_young(), "should not already be young");
187 assert(hr->get_next_young_region() == NULL, "cause it should!");
188
189 hr->set_next_young_region(_head);
190 _head = hr;
191
192 _g1h->g1_policy()->set_region_eden(hr, (int) _length);
193 ++_length;
194 }
195
196 void YoungList::add_survivor_region(HeapRegion* hr) {
197 assert(hr->is_survivor(), "should be flagged as survivor region");
198 assert(hr->get_next_young_region() == NULL, "cause it should!");
199
200 hr->set_next_young_region(_survivor_head);
201 if (_survivor_head == NULL) {
202 _survivor_tail = hr;
203 }
204 _survivor_head = hr;
205 ++_survivor_length;
206 }
207
208 void YoungList::empty_list(HeapRegion* list) {
209 while (list != NULL) {
210 HeapRegion* next = list->get_next_young_region();
211 list->set_next_young_region(NULL);
212 list->uninstall_surv_rate_group();
213 list->set_not_young();
214 list = next;
215 }
216 }
217
218 void YoungList::empty_list() {
219 assert(check_list_well_formed(), "young list should be well formed");
220
221 empty_list(_head);
222 _head = NULL;
223 _length = 0;
224
225 empty_list(_survivor_head);
226 _survivor_head = NULL;
227 _survivor_tail = NULL;
228 _survivor_length = 0;
229
230 _last_sampled_rs_lengths = 0;
231
232 assert(check_list_empty(false), "just making sure...");
233 }
234
235 bool YoungList::check_list_well_formed() {
236 bool ret = true;
237
238 uint length = 0;
239 HeapRegion* curr = _head;
240 HeapRegion* last = NULL;
241 while (curr != NULL) {
242 if (!curr->is_young()) {
243 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" "
244 "incorrectly tagged (y: %d, surv: %d)",
245 curr->bottom(), curr->end(),
246 curr->is_young(), curr->is_survivor());
247 ret = false;
248 }
249 ++length;
250 last = curr;
251 curr = curr->get_next_young_region();
252 }
253 ret = ret && (length == _length);
254
255 if (!ret) {
256 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!");
257 gclog_or_tty->print_cr("### list has %u entries, _length is %u",
258 length, _length);
259 }
260
261 return ret;
262 }
263
264 bool YoungList::check_list_empty(bool check_sample) {
265 bool ret = true;
266
267 if (_length != 0) {
268 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u",
269 _length);
270 ret = false;
271 }
272 if (check_sample && _last_sampled_rs_lengths != 0) {
273 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths");
274 ret = false;
275 }
276 if (_head != NULL) {
277 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head");
278 ret = false;
279 }
280 if (!ret) {
281 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty");
282 }
283
284 return ret;
285 }
286
287 void
288 YoungList::rs_length_sampling_init() {
289 _sampled_rs_lengths = 0;
290 _curr = _head;
291 }
292
293 bool
294 YoungList::rs_length_sampling_more() {
295 return _curr != NULL;
296 }
297
298 void
299 YoungList::rs_length_sampling_next() {
300 assert( _curr != NULL, "invariant" );
301 size_t rs_length = _curr->rem_set()->occupied();
302
303 _sampled_rs_lengths += rs_length;
304
305 // The current region may not yet have been added to the
306 // incremental collection set (it gets added when it is
307 // retired as the current allocation region).
308 if (_curr->in_collection_set()) {
309 // Update the collection set policy information for this region
310 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length);
311 }
312
313 _curr = _curr->get_next_young_region();
314 if (_curr == NULL) {
315 _last_sampled_rs_lengths = _sampled_rs_lengths;
316 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths);
317 }
318 }
319
320 void
321 YoungList::reset_auxilary_lists() {
322 guarantee( is_empty(), "young list should be empty" );
323 assert(check_list_well_formed(), "young list should be well formed");
324
325 // Add survivor regions to SurvRateGroup.
326 _g1h->g1_policy()->note_start_adding_survivor_regions();
327 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */);
328
329 int young_index_in_cset = 0;
330 for (HeapRegion* curr = _survivor_head;
331 curr != NULL;
332 curr = curr->get_next_young_region()) {
333 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset);
334
335 // The region is a non-empty survivor so let's add it to
336 // the incremental collection set for the next evacuation
337 // pause.
338 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr);
339 young_index_in_cset += 1;
340 }
341 assert((uint) young_index_in_cset == _survivor_length, "post-condition");
342 _g1h->g1_policy()->note_stop_adding_survivor_regions();
343
344 _head = _survivor_head;
345 _length = _survivor_length;
346 if (_survivor_head != NULL) {
347 assert(_survivor_tail != NULL, "cause it shouldn't be");
348 assert(_survivor_length > 0, "invariant");
349 _survivor_tail->set_next_young_region(NULL);
350 }
351
352 // Don't clear the survivor list handles until the start of
353 // the next evacuation pause - we need it in order to re-tag
354 // the survivor regions from this evacuation pause as 'young'
355 // at the start of the next.
356
357 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */);
358
359 assert(check_list_well_formed(), "young list should be well formed");
360 }
361
362 void YoungList::print() {
363 HeapRegion* lists[] = {_head, _survivor_head};
364 const char* names[] = {"YOUNG", "SURVIVOR"};
365
366 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) {
367 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]);
368 HeapRegion *curr = lists[list];
369 if (curr == NULL)
370 gclog_or_tty->print_cr(" empty");
371 while (curr != NULL) {
372 gclog_or_tty->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d",
373 HR_FORMAT_PARAMS(curr),
374 curr->prev_top_at_mark_start(),
375 curr->next_top_at_mark_start(),
376 curr->age_in_surv_rate_group_cond());
377 curr = curr->get_next_young_region();
378 }
379 }
380
381 gclog_or_tty->print_cr("");
382 }
383
384 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr)
385 {
386 // Claim the right to put the region on the dirty cards region list
387 // by installing a self pointer.
388 HeapRegion* next = hr->get_next_dirty_cards_region();
389 if (next == NULL) {
390 HeapRegion* res = (HeapRegion*)
391 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(),
392 NULL);
393 if (res == NULL) {
394 HeapRegion* head;
395 do {
396 // Put the region to the dirty cards region list.
397 head = _dirty_cards_region_list;
398 next = (HeapRegion*)
399 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head);
400 if (next == head) {
401 assert(hr->get_next_dirty_cards_region() == hr,
402 "hr->get_next_dirty_cards_region() != hr");
403 if (next == NULL) {
404 // The last region in the list points to itself.
405 hr->set_next_dirty_cards_region(hr);
406 } else {
407 hr->set_next_dirty_cards_region(next);
408 }
409 }
410 } while (next != head);
411 }
412 }
413 }
414
415 HeapRegion* G1CollectedHeap::pop_dirty_cards_region()
416 {
417 HeapRegion* head;
418 HeapRegion* hr;
419 do {
420 head = _dirty_cards_region_list;
421 if (head == NULL) {
422 return NULL;
423 }
424 HeapRegion* new_head = head->get_next_dirty_cards_region();
425 if (head == new_head) {
426 // The last region.
427 new_head = NULL;
428 }
429 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list,
430 head);
431 } while (hr != head);
432 assert(hr != NULL, "invariant");
433 hr->set_next_dirty_cards_region(NULL);
434 return hr;
435 }
436
437 void G1CollectedHeap::stop_conc_gc_threads() {
438 _cg1r->stop();
439 _cmThread->stop();
440 }
441
442 #ifdef ASSERT
443 // A region is added to the collection set as it is retired
444 // so an address p can point to a region which will be in the
445 // collection set but has not yet been retired. This method
446 // therefore is only accurate during a GC pause after all
447 // regions have been retired. It is used for debugging
448 // to check if an nmethod has references to objects that can
449 // be move during a partial collection. Though it can be
450 // inaccurate, it is sufficient for G1 because the conservative
451 // implementation of is_scavengable() for G1 will indicate that
452 // all nmethods must be scanned during a partial collection.
453 bool G1CollectedHeap::is_in_partial_collection(const void* p) {
454 HeapRegion* hr = heap_region_containing(p);
455 return hr != NULL && hr->in_collection_set();
456 }
457 #endif
458
459 // Returns true if the reference points to an object that
460 // can move in an incremental collecction.
461 bool G1CollectedHeap::is_scavengable(const void* p) {
462 G1CollectedHeap* g1h = G1CollectedHeap::heap();
463 G1CollectorPolicy* g1p = g1h->g1_policy();
464 HeapRegion* hr = heap_region_containing(p);
465 if (hr == NULL) {
466 // null
467 assert(p == NULL, err_msg("Not NULL " PTR_FORMAT ,p));
468 return false;
469 } else {
470 return !hr->isHumongous();
471 }
472 }
473
474 void G1CollectedHeap::check_ct_logs_at_safepoint() {
475 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
476 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
477
478 // Count the dirty cards at the start.
479 CountNonCleanMemRegionClosure count1(this);
480 ct_bs->mod_card_iterate(&count1);
481 int orig_count = count1.n();
482
483 // First clear the logged cards.
484 ClearLoggedCardTableEntryClosure clear;
485 dcqs.set_closure(&clear);
486 dcqs.apply_closure_to_all_completed_buffers();
487 dcqs.iterate_closure_all_threads(false);
488 clear.print_histo();
489
490 // Now ensure that there's no dirty cards.
491 CountNonCleanMemRegionClosure count2(this);
492 ct_bs->mod_card_iterate(&count2);
493 if (count2.n() != 0) {
494 gclog_or_tty->print_cr("Card table has %d entries; %d originally",
495 count2.n(), orig_count);
496 }
497 guarantee(count2.n() == 0, "Card table should be clean.");
498
499 RedirtyLoggedCardTableEntryClosure redirty;
500 JavaThread::dirty_card_queue_set().set_closure(&redirty);
501 dcqs.apply_closure_to_all_completed_buffers();
502 dcqs.iterate_closure_all_threads(false);
503 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.",
504 clear.calls(), orig_count);
505 guarantee(redirty.calls() == clear.calls(),
506 "Or else mechanism is broken.");
507
508 CountNonCleanMemRegionClosure count3(this);
509 ct_bs->mod_card_iterate(&count3);
510 if (count3.n() != orig_count) {
511 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.",
512 orig_count, count3.n());
513 guarantee(count3.n() >= orig_count, "Should have restored them all.");
514 }
515
516 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
517 }
518
519 // Private class members.
520
521 G1CollectedHeap* G1CollectedHeap::_g1h;
522
523 // Private methods.
524
525 HeapRegion*
526 G1CollectedHeap::new_region_try_secondary_free_list() {
527 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
528 while (!_secondary_free_list.is_empty() || free_regions_coming()) {
529 if (!_secondary_free_list.is_empty()) {
530 if (G1ConcRegionFreeingVerbose) {
531 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
532 "secondary_free_list has %u entries",
533 _secondary_free_list.length());
534 }
535 // It looks as if there are free regions available on the
536 // secondary_free_list. Let's move them to the free_list and try
537 // again to allocate from it.
538 append_secondary_free_list();
539
540 assert(!_free_list.is_empty(), "if the secondary_free_list was not "
541 "empty we should have moved at least one entry to the free_list");
542 HeapRegion* res = _free_list.remove_head();
543 if (G1ConcRegionFreeingVerbose) {
544 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
545 "allocated "HR_FORMAT" from secondary_free_list",
546 HR_FORMAT_PARAMS(res));
547 }
548 return res;
549 }
550
551 // Wait here until we get notifed either when (a) there are no
552 // more free regions coming or (b) some regions have been moved on
553 // the secondary_free_list.
554 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
555 }
556
557 if (G1ConcRegionFreeingVerbose) {
558 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
559 "could not allocate from secondary_free_list");
560 }
561 return NULL;
562 }
563
564 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) {
565 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords,
566 "the only time we use this to allocate a humongous region is "
567 "when we are allocating a single humongous region");
568
569 HeapRegion* res;
570 if (G1StressConcRegionFreeing) {
571 if (!_secondary_free_list.is_empty()) {
572 if (G1ConcRegionFreeingVerbose) {
573 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
574 "forced to look at the secondary_free_list");
575 }
576 res = new_region_try_secondary_free_list();
577 if (res != NULL) {
578 return res;
579 }
580 }
581 }
582 res = _free_list.remove_head_or_null();
583 if (res == NULL) {
584 if (G1ConcRegionFreeingVerbose) {
585 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
586 "res == NULL, trying the secondary_free_list");
587 }
588 res = new_region_try_secondary_free_list();
589 }
590 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) {
591 // Currently, only attempts to allocate GC alloc regions set
592 // do_expand to true. So, we should only reach here during a
593 // safepoint. If this assumption changes we might have to
594 // reconsider the use of _expand_heap_after_alloc_failure.
595 assert(SafepointSynchronize::is_at_safepoint(), "invariant");
596
597 ergo_verbose1(ErgoHeapSizing,
598 "attempt heap expansion",
599 ergo_format_reason("region allocation request failed")
600 ergo_format_byte("allocation request"),
601 word_size * HeapWordSize);
602 if (expand(word_size * HeapWordSize)) {
603 // Given that expand() succeeded in expanding the heap, and we
604 // always expand the heap by an amount aligned to the heap
605 // region size, the free list should in theory not be empty. So
606 // it would probably be OK to use remove_head(). But the extra
607 // check for NULL is unlikely to be a performance issue here (we
608 // just expanded the heap!) so let's just be conservative and
609 // use remove_head_or_null().
610 res = _free_list.remove_head_or_null();
611 } else {
612 _expand_heap_after_alloc_failure = false;
613 }
614 }
615 return res;
616 }
617
618 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions,
619 size_t word_size) {
620 assert(isHumongous(word_size), "word_size should be humongous");
621 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
622
623 uint first = G1_NULL_HRS_INDEX;
624 if (num_regions == 1) {
625 // Only one region to allocate, no need to go through the slower
626 // path. The caller will attempt the expasion if this fails, so
627 // let's not try to expand here too.
628 HeapRegion* hr = new_region(word_size, false /* do_expand */);
629 if (hr != NULL) {
630 first = hr->hrs_index();
631 } else {
632 first = G1_NULL_HRS_INDEX;
633 }
634 } else {
635 // We can't allocate humongous regions while cleanupComplete() is
636 // running, since some of the regions we find to be empty might not
637 // yet be added to the free list and it is not straightforward to
638 // know which list they are on so that we can remove them. Note
639 // that we only need to do this if we need to allocate more than
640 // one region to satisfy the current humongous allocation
641 // request. If we are only allocating one region we use the common
642 // region allocation code (see above).
643 wait_while_free_regions_coming();
644 append_secondary_free_list_if_not_empty_with_lock();
645
646 if (free_regions() >= num_regions) {
647 first = _hrs.find_contiguous(num_regions);
648 if (first != G1_NULL_HRS_INDEX) {
649 for (uint i = first; i < first + num_regions; ++i) {
650 HeapRegion* hr = region_at(i);
651 assert(hr->is_empty(), "sanity");
652 assert(is_on_master_free_list(hr), "sanity");
653 hr->set_pending_removal(true);
654 }
655 _free_list.remove_all_pending(num_regions);
656 }
657 }
658 }
659 return first;
660 }
661
662 HeapWord*
663 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first,
664 uint num_regions,
665 size_t word_size) {
666 assert(first != G1_NULL_HRS_INDEX, "pre-condition");
667 assert(isHumongous(word_size), "word_size should be humongous");
668 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
669
670 // Index of last region in the series + 1.
671 uint last = first + num_regions;
672
673 // We need to initialize the region(s) we just discovered. This is
674 // a bit tricky given that it can happen concurrently with
675 // refinement threads refining cards on these regions and
676 // potentially wanting to refine the BOT as they are scanning
677 // those cards (this can happen shortly after a cleanup; see CR
678 // 6991377). So we have to set up the region(s) carefully and in
679 // a specific order.
680
681 // The word size sum of all the regions we will allocate.
682 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords;
683 assert(word_size <= word_size_sum, "sanity");
684
685 // This will be the "starts humongous" region.
686 HeapRegion* first_hr = region_at(first);
687 // The header of the new object will be placed at the bottom of
688 // the first region.
689 HeapWord* new_obj = first_hr->bottom();
690 // This will be the new end of the first region in the series that
691 // should also match the end of the last region in the seriers.
692 HeapWord* new_end = new_obj + word_size_sum;
693 // This will be the new top of the first region that will reflect
694 // this allocation.
695 HeapWord* new_top = new_obj + word_size;
696
697 // First, we need to zero the header of the space that we will be
698 // allocating. When we update top further down, some refinement
699 // threads might try to scan the region. By zeroing the header we
700 // ensure that any thread that will try to scan the region will
701 // come across the zero klass word and bail out.
702 //
703 // NOTE: It would not have been correct to have used
704 // CollectedHeap::fill_with_object() and make the space look like
705 // an int array. The thread that is doing the allocation will
706 // later update the object header to a potentially different array
707 // type and, for a very short period of time, the klass and length
708 // fields will be inconsistent. This could cause a refinement
709 // thread to calculate the object size incorrectly.
710 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0);
711
712 // We will set up the first region as "starts humongous". This
713 // will also update the BOT covering all the regions to reflect
714 // that there is a single object that starts at the bottom of the
715 // first region.
716 first_hr->set_startsHumongous(new_top, new_end);
717
718 // Then, if there are any, we will set up the "continues
719 // humongous" regions.
720 HeapRegion* hr = NULL;
721 for (uint i = first + 1; i < last; ++i) {
722 hr = region_at(i);
723 hr->set_continuesHumongous(first_hr);
724 }
725 // If we have "continues humongous" regions (hr != NULL), then the
726 // end of the last one should match new_end.
727 assert(hr == NULL || hr->end() == new_end, "sanity");
728
729 // Up to this point no concurrent thread would have been able to
730 // do any scanning on any region in this series. All the top
731 // fields still point to bottom, so the intersection between
732 // [bottom,top] and [card_start,card_end] will be empty. Before we
733 // update the top fields, we'll do a storestore to make sure that
734 // no thread sees the update to top before the zeroing of the
735 // object header and the BOT initialization.
736 OrderAccess::storestore();
737
738 // Now that the BOT and the object header have been initialized,
739 // we can update top of the "starts humongous" region.
740 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(),
741 "new_top should be in this region");
742 first_hr->set_top(new_top);
743 if (_hr_printer.is_active()) {
744 HeapWord* bottom = first_hr->bottom();
745 HeapWord* end = first_hr->orig_end();
746 if ((first + 1) == last) {
747 // the series has a single humongous region
748 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top);
749 } else {
750 // the series has more than one humongous regions
751 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end);
752 }
753 }
754
755 // Now, we will update the top fields of the "continues humongous"
756 // regions. The reason we need to do this is that, otherwise,
757 // these regions would look empty and this will confuse parts of
758 // G1. For example, the code that looks for a consecutive number
759 // of empty regions will consider them empty and try to
760 // re-allocate them. We can extend is_empty() to also include
761 // !continuesHumongous(), but it is easier to just update the top
762 // fields here. The way we set top for all regions (i.e., top ==
763 // end for all regions but the last one, top == new_top for the
764 // last one) is actually used when we will free up the humongous
765 // region in free_humongous_region().
766 hr = NULL;
767 for (uint i = first + 1; i < last; ++i) {
768 hr = region_at(i);
769 if ((i + 1) == last) {
770 // last continues humongous region
771 assert(hr->bottom() < new_top && new_top <= hr->end(),
772 "new_top should fall on this region");
773 hr->set_top(new_top);
774 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top);
775 } else {
776 // not last one
777 assert(new_top > hr->end(), "new_top should be above this region");
778 hr->set_top(hr->end());
779 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end());
780 }
781 }
782 // If we have continues humongous regions (hr != NULL), then the
783 // end of the last one should match new_end and its top should
784 // match new_top.
785 assert(hr == NULL ||
786 (hr->end() == new_end && hr->top() == new_top), "sanity");
787
788 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
789 _summary_bytes_used += first_hr->used();
790 _humongous_set.add(first_hr);
791
792 return new_obj;
793 }
794
795 // If could fit into free regions w/o expansion, try.
796 // Otherwise, if can expand, do so.
797 // Otherwise, if using ex regions might help, try with ex given back.
798 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) {
799 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
800
801 verify_region_sets_optional();
802
803 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords);
804 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords);
805 uint x_num = expansion_regions();
806 uint fs = _hrs.free_suffix();
807 uint first = humongous_obj_allocate_find_first(num_regions, word_size);
808 if (first == G1_NULL_HRS_INDEX) {
809 // The only thing we can do now is attempt expansion.
810 if (fs + x_num >= num_regions) {
811 // If the number of regions we're trying to allocate for this
812 // object is at most the number of regions in the free suffix,
813 // then the call to humongous_obj_allocate_find_first() above
814 // should have succeeded and we wouldn't be here.
815 //
816 // We should only be trying to expand when the free suffix is
817 // not sufficient for the object _and_ we have some expansion
818 // room available.
819 assert(num_regions > fs, "earlier allocation should have succeeded");
820
821 ergo_verbose1(ErgoHeapSizing,
822 "attempt heap expansion",
823 ergo_format_reason("humongous allocation request failed")
824 ergo_format_byte("allocation request"),
825 word_size * HeapWordSize);
826 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) {
827 // Even though the heap was expanded, it might not have
828 // reached the desired size. So, we cannot assume that the
829 // allocation will succeed.
830 first = humongous_obj_allocate_find_first(num_regions, word_size);
831 }
832 }
833 }
834
835 HeapWord* result = NULL;
836 if (first != G1_NULL_HRS_INDEX) {
837 result =
838 humongous_obj_allocate_initialize_regions(first, num_regions, word_size);
839 assert(result != NULL, "it should always return a valid result");
840
841 // A successful humongous object allocation changes the used space
842 // information of the old generation so we need to recalculate the
843 // sizes and update the jstat counters here.
844 g1mm()->update_sizes();
845 }
846
847 verify_region_sets_optional();
848
849 return result;
850 }
851
852 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) {
853 assert_heap_not_locked_and_not_at_safepoint();
854 assert(!isHumongous(word_size), "we do not allow humongous TLABs");
855
856 unsigned int dummy_gc_count_before;
857 int dummy_gclocker_retry_count = 0;
858 return attempt_allocation(word_size, &dummy_gc_count_before, &dummy_gclocker_retry_count);
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, gclocker_retry_count = 0; /* 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, &gclocker_retry_count);
873 } else {
874 result = attempt_allocation_humongous(word_size, &gc_count_before, &gclocker_retry_count);
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 if (gclocker_retry_count > GCLockerRetryAllocationCount) {
899 return NULL;
900 }
901 assert(op.result() == NULL,
902 "the result should be NULL if the VM op did not succeed");
903 }
904
905 // Give a warning if we seem to be looping forever.
906 if ((QueuedAllocationWarningCount > 0) &&
907 (try_count % QueuedAllocationWarningCount == 0)) {
908 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
909 }
910 }
911
912 ShouldNotReachHere();
913 return NULL;
914 }
915
916 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
917 unsigned int *gc_count_before_ret,
918 int* gclocker_retry_count_ret) {
919 // Make sure you read the note in attempt_allocation_humongous().
920
921 assert_heap_not_locked_and_not_at_safepoint();
922 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
923 "be called for humongous allocation requests");
924
925 // We should only get here after the first-level allocation attempt
926 // (attempt_allocation()) failed to allocate.
927
928 // We will loop until a) we manage to successfully perform the
929 // allocation or b) we successfully schedule a collection which
930 // fails to perform the allocation. b) is the only case when we'll
931 // return NULL.
932 HeapWord* result = NULL;
933 for (int try_count = 1; /* we'll return */; try_count += 1) {
934 bool should_try_gc;
935 unsigned int gc_count_before;
936
937 {
938 MutexLockerEx x(Heap_lock);
939
940 result = _mutator_alloc_region.attempt_allocation_locked(word_size,
941 false /* bot_updates */);
942 if (result != NULL) {
943 return result;
944 }
945
946 // If we reach here, attempt_allocation_locked() above failed to
947 // allocate a new region. So the mutator alloc region should be NULL.
948 assert(_mutator_alloc_region.get() == NULL, "only way to get here");
949
950 if (GC_locker::is_active_and_needs_gc()) {
951 if (g1_policy()->can_expand_young_list()) {
952 // No need for an ergo verbose message here,
953 // can_expand_young_list() does this when it returns true.
954 result = _mutator_alloc_region.attempt_allocation_force(word_size,
955 false /* bot_updates */);
956 if (result != NULL) {
957 return result;
958 }
959 }
960 should_try_gc = false;
961 } else {
962 // The GCLocker may not be active but the GCLocker initiated
963 // GC may not yet have been performed (GCLocker::needs_gc()
964 // returns true). In this case we do not try this GC and
965 // wait until the GCLocker initiated GC is performed, and
966 // then retry the allocation.
967 if (GC_locker::needs_gc()) {
968 should_try_gc = false;
969 } else {
970 // Read the GC count while still holding the Heap_lock.
971 gc_count_before = total_collections();
972 should_try_gc = true;
973 }
974 }
975 }
976
977 if (should_try_gc) {
978 bool succeeded;
979 result = do_collection_pause(word_size, gc_count_before, &succeeded);
980 if (result != NULL) {
981 assert(succeeded, "only way to get back a non-NULL result");
982 return result;
983 }
984
985 if (succeeded) {
986 // If we get here we successfully scheduled a collection which
987 // failed to allocate. No point in trying to allocate
988 // further. We'll just return NULL.
989 MutexLockerEx x(Heap_lock);
990 *gc_count_before_ret = total_collections();
991 return NULL;
992 }
993 } else {
994 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) {
995 MutexLockerEx x(Heap_lock);
996 *gc_count_before_ret = total_collections();
997 return NULL;
998 }
999 // The GCLocker is either active or the GCLocker initiated
1000 // GC has not yet been performed. Stall until it is and
1001 // then retry the allocation.
1002 GC_locker::stall_until_clear();
1003 (*gclocker_retry_count_ret) += 1;
1004 }
1005
1006 // We can reach here if we were unsuccessul in scheduling a
1007 // collection (because another thread beat us to it) or if we were
1008 // stalled due to the GC locker. In either can we should retry the
1009 // allocation attempt in case another thread successfully
1010 // performed a collection and reclaimed enough space. We do the
1011 // first attempt (without holding the Heap_lock) here and the
1012 // follow-on attempt will be at the start of the next loop
1013 // iteration (after taking the Heap_lock).
1014 result = _mutator_alloc_region.attempt_allocation(word_size,
1015 false /* bot_updates */);
1016 if (result != NULL) {
1017 return result;
1018 }
1019
1020 // Give a warning if we seem to be looping forever.
1021 if ((QueuedAllocationWarningCount > 0) &&
1022 (try_count % QueuedAllocationWarningCount == 0)) {
1023 warning("G1CollectedHeap::attempt_allocation_slow() "
1024 "retries %d times", try_count);
1025 }
1026 }
1027
1028 ShouldNotReachHere();
1029 return NULL;
1030 }
1031
1032 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
1033 unsigned int * gc_count_before_ret,
1034 int* gclocker_retry_count_ret) {
1035 // The structure of this method has a lot of similarities to
1036 // attempt_allocation_slow(). The reason these two were not merged
1037 // into a single one is that such a method would require several "if
1038 // allocation is not humongous do this, otherwise do that"
1039 // conditional paths which would obscure its flow. In fact, an early
1040 // version of this code did use a unified method which was harder to
1041 // follow and, as a result, it had subtle bugs that were hard to
1042 // track down. So keeping these two methods separate allows each to
1043 // be more readable. It will be good to keep these two in sync as
1044 // much as possible.
1045
1046 assert_heap_not_locked_and_not_at_safepoint();
1047 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1048 "should only be called for humongous allocations");
1049
1050 // Humongous objects can exhaust the heap quickly, so we should check if we
1051 // need to start a marking cycle at each humongous object allocation. We do
1052 // the check before we do the actual allocation. The reason for doing it
1053 // before the allocation is that we avoid having to keep track of the newly
1054 // allocated memory while we do a GC.
1055 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation",
1056 word_size)) {
1057 collect(GCCause::_g1_humongous_allocation);
1058 }
1059
1060 // We will loop until a) we manage to successfully perform the
1061 // allocation or b) we successfully schedule a collection which
1062 // fails to perform the allocation. b) is the only case when we'll
1063 // return NULL.
1064 HeapWord* result = NULL;
1065 for (int try_count = 1; /* we'll return */; try_count += 1) {
1066 bool should_try_gc;
1067 unsigned int gc_count_before;
1068
1069 {
1070 MutexLockerEx x(Heap_lock);
1071
1072 // Given that humongous objects are not allocated in young
1073 // regions, we'll first try to do the allocation without doing a
1074 // collection hoping that there's enough space in the heap.
1075 result = humongous_obj_allocate(word_size);
1076 if (result != NULL) {
1077 return result;
1078 }
1079
1080 if (GC_locker::is_active_and_needs_gc()) {
1081 should_try_gc = false;
1082 } else {
1083 // The GCLocker may not be active but the GCLocker initiated
1084 // GC may not yet have been performed (GCLocker::needs_gc()
1085 // returns true). In this case we do not try this GC and
1086 // wait until the GCLocker initiated GC is performed, and
1087 // then retry the allocation.
1088 if (GC_locker::needs_gc()) {
1089 should_try_gc = false;
1090 } else {
1091 // Read the GC count while still holding the Heap_lock.
1092 gc_count_before = total_collections();
1093 should_try_gc = true;
1094 }
1095 }
1096 }
1097
1098 if (should_try_gc) {
1099 // If we failed to allocate the humongous object, we should try to
1100 // do a collection pause (if we're allowed) in case it reclaims
1101 // enough space for the allocation to succeed after the pause.
1102
1103 bool succeeded;
1104 result = do_collection_pause(word_size, gc_count_before, &succeeded);
1105 if (result != NULL) {
1106 assert(succeeded, "only way to get back a non-NULL result");
1107 return result;
1108 }
1109
1110 if (succeeded) {
1111 // If we get here we successfully scheduled a collection which
1112 // failed to allocate. No point in trying to allocate
1113 // further. We'll just return NULL.
1114 MutexLockerEx x(Heap_lock);
1115 *gc_count_before_ret = total_collections();
1116 return NULL;
1117 }
1118 } else {
1119 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) {
1120 MutexLockerEx x(Heap_lock);
1121 *gc_count_before_ret = total_collections();
1122 return NULL;
1123 }
1124 // The GCLocker is either active or the GCLocker initiated
1125 // GC has not yet been performed. Stall until it is and
1126 // then retry the allocation.
1127 GC_locker::stall_until_clear();
1128 (*gclocker_retry_count_ret) += 1;
1129 }
1130
1131 // We can reach here if we were unsuccessul in scheduling a
1132 // collection (because another thread beat us to it) or if we were
1133 // stalled due to the GC locker. In either can we should retry the
1134 // allocation attempt in case another thread successfully
1135 // performed a collection and reclaimed enough space. Give a
1136 // warning if we seem to be looping forever.
1137
1138 if ((QueuedAllocationWarningCount > 0) &&
1139 (try_count % QueuedAllocationWarningCount == 0)) {
1140 warning("G1CollectedHeap::attempt_allocation_humongous() "
1141 "retries %d times", try_count);
1142 }
1143 }
1144
1145 ShouldNotReachHere();
1146 return NULL;
1147 }
1148
1149 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1150 bool expect_null_mutator_alloc_region) {
1151 assert_at_safepoint(true /* should_be_vm_thread */);
1152 assert(_mutator_alloc_region.get() == NULL ||
1153 !expect_null_mutator_alloc_region,
1154 "the current alloc region was unexpectedly found to be non-NULL");
1155
1156 if (!isHumongous(word_size)) {
1157 return _mutator_alloc_region.attempt_allocation_locked(word_size,
1158 false /* bot_updates */);
1159 } else {
1160 HeapWord* result = humongous_obj_allocate(word_size);
1161 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) {
1162 g1_policy()->set_initiate_conc_mark_if_possible();
1163 }
1164 return result;
1165 }
1166
1167 ShouldNotReachHere();
1168 }
1169
1170 class PostMCRemSetClearClosure: public HeapRegionClosure {
1171 G1CollectedHeap* _g1h;
1172 ModRefBarrierSet* _mr_bs;
1173 public:
1174 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) :
1175 _g1h(g1h), _mr_bs(mr_bs) { }
1176 bool doHeapRegion(HeapRegion* r) {
1177 if (r->continuesHumongous()) {
1178 return false;
1179 }
1180 _g1h->reset_gc_time_stamps(r);
1181 HeapRegionRemSet* hrrs = r->rem_set();
1182 if (hrrs != NULL) hrrs->clear();
1183 // You might think here that we could clear just the cards
1184 // corresponding to the used region. But no: if we leave a dirty card
1185 // in a region we might allocate into, then it would prevent that card
1186 // from being enqueued, and cause it to be missed.
1187 // Re: the performance cost: we shouldn't be doing full GC anyway!
1188 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1189 return false;
1190 }
1191 };
1192
1193 void G1CollectedHeap::clear_rsets_post_compaction() {
1194 PostMCRemSetClearClosure rs_clear(this, mr_bs());
1195 heap_region_iterate(&rs_clear);
1196 }
1197
1198 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1199 G1CollectedHeap* _g1h;
1200 UpdateRSOopClosure _cl;
1201 int _worker_i;
1202 public:
1203 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1204 _cl(g1->g1_rem_set(), worker_i),
1205 _worker_i(worker_i),
1206 _g1h(g1)
1207 { }
1208
1209 bool doHeapRegion(HeapRegion* r) {
1210 if (!r->continuesHumongous()) {
1211 _cl.set_from(r);
1212 r->oop_iterate(&_cl);
1213 }
1214 return false;
1215 }
1216 };
1217
1218 class ParRebuildRSTask: public AbstractGangTask {
1219 G1CollectedHeap* _g1;
1220 public:
1221 ParRebuildRSTask(G1CollectedHeap* g1)
1222 : AbstractGangTask("ParRebuildRSTask"),
1223 _g1(g1)
1224 { }
1225
1226 void work(uint worker_id) {
1227 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id);
1228 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id,
1229 _g1->workers()->active_workers(),
1230 HeapRegion::RebuildRSClaimValue);
1231 }
1232 };
1233
1234 class PostCompactionPrinterClosure: public HeapRegionClosure {
1235 private:
1236 G1HRPrinter* _hr_printer;
1237 public:
1238 bool doHeapRegion(HeapRegion* hr) {
1239 assert(!hr->is_young(), "not expecting to find young regions");
1240 // We only generate output for non-empty regions.
1241 if (!hr->is_empty()) {
1242 if (!hr->isHumongous()) {
1243 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1244 } else if (hr->startsHumongous()) {
1245 if (hr->region_num() == 1) {
1246 // single humongous region
1247 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1248 } else {
1249 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1250 }
1251 } else {
1252 assert(hr->continuesHumongous(), "only way to get here");
1253 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1254 }
1255 }
1256 return false;
1257 }
1258
1259 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1260 : _hr_printer(hr_printer) { }
1261 };
1262
1263 void G1CollectedHeap::print_hrs_post_compaction() {
1264 PostCompactionPrinterClosure cl(hr_printer());
1265 heap_region_iterate(&cl);
1266 }
1267
1268 double G1CollectedHeap::verify(bool guard, const char* msg) {
1269 double verify_time_ms = 0.0;
1270
1271 if (guard && total_collections() >= VerifyGCStartAt) {
1272 double verify_start = os::elapsedTime();
1273 HandleMark hm; // Discard invalid handles created during verification
1274 prepare_for_verify();
1275 Universe::verify(VerifyOption_G1UsePrevMarking, msg);
1276 verify_time_ms = (os::elapsedTime() - verify_start) * 1000;
1277 }
1278
1279 return verify_time_ms;
1280 }
1281
1282 void G1CollectedHeap::verify_before_gc() {
1283 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:");
1284 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms);
1285 }
1286
1287 void G1CollectedHeap::verify_after_gc() {
1288 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:");
1289 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms);
1290 }
1291
1292 bool G1CollectedHeap::do_collection(bool explicit_gc,
1293 bool clear_all_soft_refs,
1294 size_t word_size) {
1295 assert_at_safepoint(true /* should_be_vm_thread */);
1296
1297 if (GC_locker::check_active_before_gc()) {
1298 return false;
1299 }
1300
1301 SvcGCMarker sgcm(SvcGCMarker::FULL);
1302 ResourceMark rm;
1303
1304 print_heap_before_gc();
1305
1306 size_t metadata_prev_used = MetaspaceAux::allocated_used_bytes();
1307
1308 HRSPhaseSetter x(HRSPhaseFullGC);
1309 verify_region_sets_optional();
1310
1311 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1312 collector_policy()->should_clear_all_soft_refs();
1313
1314 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1315
1316 {
1317 IsGCActiveMark x;
1318
1319 // Timing
1320 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant");
1321 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps);
1322 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
1323
1324 {
1325 TraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, gclog_or_tty);
1326 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1327 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1328
1329 double start = os::elapsedTime();
1330 g1_policy()->record_full_collection_start();
1331
1332 // Note: When we have a more flexible GC logging framework that
1333 // allows us to add optional attributes to a GC log record we
1334 // could consider timing and reporting how long we wait in the
1335 // following two methods.
1336 wait_while_free_regions_coming();
1337 // If we start the compaction before the CM threads finish
1338 // scanning the root regions we might trip them over as we'll
1339 // be moving objects / updating references. So let's wait until
1340 // they are done. By telling them to abort, they should complete
1341 // early.
1342 _cm->root_regions()->abort();
1343 _cm->root_regions()->wait_until_scan_finished();
1344 append_secondary_free_list_if_not_empty_with_lock();
1345
1346 gc_prologue(true);
1347 increment_total_collections(true /* full gc */);
1348 increment_old_marking_cycles_started();
1349
1350 assert(used() == recalculate_used(), "Should be equal");
1351
1352 verify_before_gc();
1353
1354 pre_full_gc_dump();
1355
1356 COMPILER2_PRESENT(DerivedPointerTable::clear());
1357
1358 // Disable discovery and empty the discovered lists
1359 // for the CM ref processor.
1360 ref_processor_cm()->disable_discovery();
1361 ref_processor_cm()->abandon_partial_discovery();
1362 ref_processor_cm()->verify_no_references_recorded();
1363
1364 // Abandon current iterations of concurrent marking and concurrent
1365 // refinement, if any are in progress. We have to do this before
1366 // wait_until_scan_finished() below.
1367 concurrent_mark()->abort();
1368
1369 // Make sure we'll choose a new allocation region afterwards.
1370 release_mutator_alloc_region();
1371 abandon_gc_alloc_regions();
1372 g1_rem_set()->cleanupHRRS();
1373
1374 // We should call this after we retire any currently active alloc
1375 // regions so that all the ALLOC / RETIRE events are generated
1376 // before the start GC event.
1377 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1378
1379 // We may have added regions to the current incremental collection
1380 // set between the last GC or pause and now. We need to clear the
1381 // incremental collection set and then start rebuilding it afresh
1382 // after this full GC.
1383 abandon_collection_set(g1_policy()->inc_cset_head());
1384 g1_policy()->clear_incremental_cset();
1385 g1_policy()->stop_incremental_cset_building();
1386
1387 tear_down_region_sets(false /* free_list_only */);
1388 g1_policy()->set_gcs_are_young(true);
1389
1390 // See the comments in g1CollectedHeap.hpp and
1391 // G1CollectedHeap::ref_processing_init() about
1392 // how reference processing currently works in G1.
1393
1394 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1395 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1396
1397 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1398 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1399
1400 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1401 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1402
1403 // Do collection work
1404 {
1405 HandleMark hm; // Discard invalid handles created during gc
1406 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1407 }
1408
1409 assert(free_regions() == 0, "we should not have added any free regions");
1410 rebuild_region_sets(false /* free_list_only */);
1411
1412 // Enqueue any discovered reference objects that have
1413 // not been removed from the discovered lists.
1414 ref_processor_stw()->enqueue_discovered_references();
1415
1416 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1417
1418 MemoryService::track_memory_usage();
1419
1420 verify_after_gc();
1421
1422 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1423 ref_processor_stw()->verify_no_references_recorded();
1424
1425 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1426 ClassLoaderDataGraph::purge();
1427 MetaspaceAux::verify_metrics();
1428
1429 // Note: since we've just done a full GC, concurrent
1430 // marking is no longer active. Therefore we need not
1431 // re-enable reference discovery for the CM ref processor.
1432 // That will be done at the start of the next marking cycle.
1433 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1434 ref_processor_cm()->verify_no_references_recorded();
1435
1436 reset_gc_time_stamp();
1437 // Since everything potentially moved, we will clear all remembered
1438 // sets, and clear all cards. Later we will rebuild remebered
1439 // sets. We will also reset the GC time stamps of the regions.
1440 clear_rsets_post_compaction();
1441 check_gc_time_stamps();
1442
1443 // Resize the heap if necessary.
1444 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1445
1446 if (_hr_printer.is_active()) {
1447 // We should do this after we potentially resize the heap so
1448 // that all the COMMIT / UNCOMMIT events are generated before
1449 // the end GC event.
1450
1451 print_hrs_post_compaction();
1452 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1453 }
1454
1455 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache();
1456 if (hot_card_cache->use_cache()) {
1457 hot_card_cache->reset_card_counts();
1458 hot_card_cache->reset_hot_cache();
1459 }
1460
1461 // Rebuild remembered sets of all regions.
1462 if (G1CollectedHeap::use_parallel_gc_threads()) {
1463 uint n_workers =
1464 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
1465 workers()->active_workers(),
1466 Threads::number_of_non_daemon_threads());
1467 assert(UseDynamicNumberOfGCThreads ||
1468 n_workers == workers()->total_workers(),
1469 "If not dynamic should be using all the workers");
1470 workers()->set_active_workers(n_workers);
1471 // Set parallel threads in the heap (_n_par_threads) only
1472 // before a parallel phase and always reset it to 0 after
1473 // the phase so that the number of parallel threads does
1474 // no get carried forward to a serial phase where there
1475 // may be code that is "possibly_parallel".
1476 set_par_threads(n_workers);
1477
1478 ParRebuildRSTask rebuild_rs_task(this);
1479 assert(check_heap_region_claim_values(
1480 HeapRegion::InitialClaimValue), "sanity check");
1481 assert(UseDynamicNumberOfGCThreads ||
1482 workers()->active_workers() == workers()->total_workers(),
1483 "Unless dynamic should use total workers");
1484 // Use the most recent number of active workers
1485 assert(workers()->active_workers() > 0,
1486 "Active workers not properly set");
1487 set_par_threads(workers()->active_workers());
1488 workers()->run_task(&rebuild_rs_task);
1489 set_par_threads(0);
1490 assert(check_heap_region_claim_values(
1491 HeapRegion::RebuildRSClaimValue), "sanity check");
1492 reset_heap_region_claim_values();
1493 } else {
1494 RebuildRSOutOfRegionClosure rebuild_rs(this);
1495 heap_region_iterate(&rebuild_rs);
1496 }
1497
1498 if (true) { // FIXME
1499 MetaspaceGC::compute_new_size();
1500 }
1501
1502 #ifdef TRACESPINNING
1503 ParallelTaskTerminator::print_termination_counts();
1504 #endif
1505
1506 // Discard all rset updates
1507 JavaThread::dirty_card_queue_set().abandon_logs();
1508 assert(!G1DeferredRSUpdate
1509 || (G1DeferredRSUpdate &&
1510 (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
1511
1512 _young_list->reset_sampled_info();
1513 // At this point there should be no regions in the
1514 // entire heap tagged as young.
1515 assert(check_young_list_empty(true /* check_heap */),
1516 "young list should be empty at this point");
1517
1518 // Update the number of full collections that have been completed.
1519 increment_old_marking_cycles_completed(false /* concurrent */);
1520
1521 _hrs.verify_optional();
1522 verify_region_sets_optional();
1523
1524 // Start a new incremental collection set for the next pause
1525 assert(g1_policy()->collection_set() == NULL, "must be");
1526 g1_policy()->start_incremental_cset_building();
1527
1528 // Clear the _cset_fast_test bitmap in anticipation of adding
1529 // regions to the incremental collection set for the next
1530 // evacuation pause.
1531 clear_cset_fast_test();
1532
1533 init_mutator_alloc_region();
1534
1535 double end = os::elapsedTime();
1536 g1_policy()->record_full_collection_end();
1537
1538 if (G1Log::fine()) {
1539 g1_policy()->print_heap_transition();
1540 }
1541
1542 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
1543 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
1544 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
1545 // before any GC notifications are raised.
1546 g1mm()->update_sizes();
1547
1548 gc_epilogue(true);
1549 }
1550
1551 if (G1Log::finer()) {
1552 g1_policy()->print_detailed_heap_transition();
1553 }
1554
1555 print_heap_after_gc();
1556
1557 post_full_gc_dump();
1558 }
1559
1560 return true;
1561 }
1562
1563 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1564 // do_collection() will return whether it succeeded in performing
1565 // the GC. Currently, there is no facility on the
1566 // do_full_collection() API to notify the caller than the collection
1567 // did not succeed (e.g., because it was locked out by the GC
1568 // locker). So, right now, we'll ignore the return value.
1569 bool dummy = do_collection(true, /* explicit_gc */
1570 clear_all_soft_refs,
1571 0 /* word_size */);
1572 }
1573
1574 // This code is mostly copied from TenuredGeneration.
1575 void
1576 G1CollectedHeap::
1577 resize_if_necessary_after_full_collection(size_t word_size) {
1578 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check");
1579
1580 // Include the current allocation, if any, and bytes that will be
1581 // pre-allocated to support collections, as "used".
1582 const size_t used_after_gc = used();
1583 const size_t capacity_after_gc = capacity();
1584 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1585
1586 // This is enforced in arguments.cpp.
1587 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1588 "otherwise the code below doesn't make sense");
1589
1590 // We don't have floating point command-line arguments
1591 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1592 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1593 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1594 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1595
1596 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1597 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1598
1599 // We have to be careful here as these two calculations can overflow
1600 // 32-bit size_t's.
1601 double used_after_gc_d = (double) used_after_gc;
1602 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1603 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1604
1605 // Let's make sure that they are both under the max heap size, which
1606 // by default will make them fit into a size_t.
1607 double desired_capacity_upper_bound = (double) max_heap_size;
1608 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1609 desired_capacity_upper_bound);
1610 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1611 desired_capacity_upper_bound);
1612
1613 // We can now safely turn them into size_t's.
1614 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1615 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1616
1617 // This assert only makes sense here, before we adjust them
1618 // with respect to the min and max heap size.
1619 assert(minimum_desired_capacity <= maximum_desired_capacity,
1620 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1621 "maximum_desired_capacity = "SIZE_FORMAT,
1622 minimum_desired_capacity, maximum_desired_capacity));
1623
1624 // Should not be greater than the heap max size. No need to adjust
1625 // it with respect to the heap min size as it's a lower bound (i.e.,
1626 // we'll try to make the capacity larger than it, not smaller).
1627 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1628 // Should not be less than the heap min size. No need to adjust it
1629 // with respect to the heap max size as it's an upper bound (i.e.,
1630 // we'll try to make the capacity smaller than it, not greater).
1631 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1632
1633 if (capacity_after_gc < minimum_desired_capacity) {
1634 // Don't expand unless it's significant
1635 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1636 ergo_verbose4(ErgoHeapSizing,
1637 "attempt heap expansion",
1638 ergo_format_reason("capacity lower than "
1639 "min desired capacity after Full GC")
1640 ergo_format_byte("capacity")
1641 ergo_format_byte("occupancy")
1642 ergo_format_byte_perc("min desired capacity"),
1643 capacity_after_gc, used_after_gc,
1644 minimum_desired_capacity, (double) MinHeapFreeRatio);
1645 expand(expand_bytes);
1646
1647 // No expansion, now see if we want to shrink
1648 } else if (capacity_after_gc > maximum_desired_capacity) {
1649 // Capacity too large, compute shrinking size
1650 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1651 ergo_verbose4(ErgoHeapSizing,
1652 "attempt heap shrinking",
1653 ergo_format_reason("capacity higher than "
1654 "max desired capacity after Full GC")
1655 ergo_format_byte("capacity")
1656 ergo_format_byte("occupancy")
1657 ergo_format_byte_perc("max desired capacity"),
1658 capacity_after_gc, used_after_gc,
1659 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1660 shrink(shrink_bytes);
1661 }
1662 }
1663
1664
1665 HeapWord*
1666 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1667 bool* succeeded) {
1668 assert_at_safepoint(true /* should_be_vm_thread */);
1669
1670 *succeeded = true;
1671 // Let's attempt the allocation first.
1672 HeapWord* result =
1673 attempt_allocation_at_safepoint(word_size,
1674 false /* expect_null_mutator_alloc_region */);
1675 if (result != NULL) {
1676 assert(*succeeded, "sanity");
1677 return result;
1678 }
1679
1680 // In a G1 heap, we're supposed to keep allocation from failing by
1681 // incremental pauses. Therefore, at least for now, we'll favor
1682 // expansion over collection. (This might change in the future if we can
1683 // do something smarter than full collection to satisfy a failed alloc.)
1684 result = expand_and_allocate(word_size);
1685 if (result != NULL) {
1686 assert(*succeeded, "sanity");
1687 return result;
1688 }
1689
1690 // Expansion didn't work, we'll try to do a Full GC.
1691 bool gc_succeeded = do_collection(false, /* explicit_gc */
1692 false, /* clear_all_soft_refs */
1693 word_size);
1694 if (!gc_succeeded) {
1695 *succeeded = false;
1696 return NULL;
1697 }
1698
1699 // Retry the allocation
1700 result = attempt_allocation_at_safepoint(word_size,
1701 true /* expect_null_mutator_alloc_region */);
1702 if (result != NULL) {
1703 assert(*succeeded, "sanity");
1704 return result;
1705 }
1706
1707 // Then, try a Full GC that will collect all soft references.
1708 gc_succeeded = do_collection(false, /* explicit_gc */
1709 true, /* clear_all_soft_refs */
1710 word_size);
1711 if (!gc_succeeded) {
1712 *succeeded = false;
1713 return NULL;
1714 }
1715
1716 // Retry the allocation once more
1717 result = attempt_allocation_at_safepoint(word_size,
1718 true /* expect_null_mutator_alloc_region */);
1719 if (result != NULL) {
1720 assert(*succeeded, "sanity");
1721 return result;
1722 }
1723
1724 assert(!collector_policy()->should_clear_all_soft_refs(),
1725 "Flag should have been handled and cleared prior to this point");
1726
1727 // What else? We might try synchronous finalization later. If the total
1728 // space available is large enough for the allocation, then a more
1729 // complete compaction phase than we've tried so far might be
1730 // appropriate.
1731 assert(*succeeded, "sanity");
1732 return NULL;
1733 }
1734
1735 // Attempting to expand the heap sufficiently
1736 // to support an allocation of the given "word_size". If
1737 // successful, perform the allocation and return the address of the
1738 // allocated block, or else "NULL".
1739
1740 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) {
1741 assert_at_safepoint(true /* should_be_vm_thread */);
1742
1743 verify_region_sets_optional();
1744
1745 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1746 ergo_verbose1(ErgoHeapSizing,
1747 "attempt heap expansion",
1748 ergo_format_reason("allocation request failed")
1749 ergo_format_byte("allocation request"),
1750 word_size * HeapWordSize);
1751 if (expand(expand_bytes)) {
1752 _hrs.verify_optional();
1753 verify_region_sets_optional();
1754 return attempt_allocation_at_safepoint(word_size,
1755 false /* expect_null_mutator_alloc_region */);
1756 }
1757 return NULL;
1758 }
1759
1760 void G1CollectedHeap::update_committed_space(HeapWord* old_end,
1761 HeapWord* new_end) {
1762 assert(old_end != new_end, "don't call this otherwise");
1763 assert((HeapWord*) _g1_storage.high() == new_end, "invariant");
1764
1765 // Update the committed mem region.
1766 _g1_committed.set_end(new_end);
1767 // Tell the card table about the update.
1768 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed);
1769 // Tell the BOT about the update.
1770 _bot_shared->resize(_g1_committed.word_size());
1771 // Tell the hot card cache about the update
1772 _cg1r->hot_card_cache()->resize_card_counts(capacity());
1773 }
1774
1775 bool G1CollectedHeap::expand(size_t expand_bytes) {
1776 size_t old_mem_size = _g1_storage.committed_size();
1777 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1778 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1779 HeapRegion::GrainBytes);
1780 ergo_verbose2(ErgoHeapSizing,
1781 "expand the heap",
1782 ergo_format_byte("requested expansion amount")
1783 ergo_format_byte("attempted expansion amount"),
1784 expand_bytes, aligned_expand_bytes);
1785
1786 // First commit the memory.
1787 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1788 bool successful = _g1_storage.expand_by(aligned_expand_bytes);
1789 if (successful) {
1790 // Then propagate this update to the necessary data structures.
1791 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1792 update_committed_space(old_end, new_end);
1793
1794 FreeRegionList expansion_list("Local Expansion List");
1795 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list);
1796 assert(mr.start() == old_end, "post-condition");
1797 // mr might be a smaller region than what was requested if
1798 // expand_by() was unable to allocate the HeapRegion instances
1799 assert(mr.end() <= new_end, "post-condition");
1800
1801 size_t actual_expand_bytes = mr.byte_size();
1802 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1803 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(),
1804 "post-condition");
1805 if (actual_expand_bytes < aligned_expand_bytes) {
1806 // We could not expand _hrs to the desired size. In this case we
1807 // need to shrink the committed space accordingly.
1808 assert(mr.end() < new_end, "invariant");
1809
1810 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes;
1811 // First uncommit the memory.
1812 _g1_storage.shrink_by(diff_bytes);
1813 // Then propagate this update to the necessary data structures.
1814 update_committed_space(new_end, mr.end());
1815 }
1816 _free_list.add_as_tail(&expansion_list);
1817
1818 if (_hr_printer.is_active()) {
1819 HeapWord* curr = mr.start();
1820 while (curr < mr.end()) {
1821 HeapWord* curr_end = curr + HeapRegion::GrainWords;
1822 _hr_printer.commit(curr, curr_end);
1823 curr = curr_end;
1824 }
1825 assert(curr == mr.end(), "post-condition");
1826 }
1827 g1_policy()->record_new_heap_size(n_regions());
1828 } else {
1829 ergo_verbose0(ErgoHeapSizing,
1830 "did not expand the heap",
1831 ergo_format_reason("heap expansion operation failed"));
1832 // The expansion of the virtual storage space was unsuccessful.
1833 // Let's see if it was because we ran out of swap.
1834 if (G1ExitOnExpansionFailure &&
1835 _g1_storage.uncommitted_size() >= aligned_expand_bytes) {
1836 // We had head room...
1837 vm_exit_out_of_memory(aligned_expand_bytes, OOM_MMAP_ERROR, "G1 heap expansion");
1838 }
1839 }
1840 return successful;
1841 }
1842
1843 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1844 size_t old_mem_size = _g1_storage.committed_size();
1845 size_t aligned_shrink_bytes =
1846 ReservedSpace::page_align_size_down(shrink_bytes);
1847 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1848 HeapRegion::GrainBytes);
1849 uint num_regions_to_remove = (uint)(shrink_bytes / HeapRegion::GrainBytes);
1850
1851 uint num_regions_removed = _hrs.shrink_by(num_regions_to_remove);
1852 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1853 size_t shrunk_bytes = num_regions_removed * HeapRegion::GrainBytes;
1854
1855 ergo_verbose3(ErgoHeapSizing,
1856 "shrink the heap",
1857 ergo_format_byte("requested shrinking amount")
1858 ergo_format_byte("aligned shrinking amount")
1859 ergo_format_byte("attempted shrinking amount"),
1860 shrink_bytes, aligned_shrink_bytes, shrunk_bytes);
1861 if (num_regions_removed > 0) {
1862 _g1_storage.shrink_by(shrunk_bytes);
1863 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1864
1865 if (_hr_printer.is_active()) {
1866 HeapWord* curr = old_end;
1867 while (curr > new_end) {
1868 HeapWord* curr_end = curr;
1869 curr -= HeapRegion::GrainWords;
1870 _hr_printer.uncommit(curr, curr_end);
1871 }
1872 }
1873
1874 _expansion_regions += num_regions_removed;
1875 update_committed_space(old_end, new_end);
1876 HeapRegionRemSet::shrink_heap(n_regions());
1877 g1_policy()->record_new_heap_size(n_regions());
1878 } else {
1879 ergo_verbose0(ErgoHeapSizing,
1880 "did not shrink the heap",
1881 ergo_format_reason("heap shrinking operation failed"));
1882 }
1883 }
1884
1885 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1886 verify_region_sets_optional();
1887
1888 // We should only reach here at the end of a Full GC which means we
1889 // should not not be holding to any GC alloc regions. The method
1890 // below will make sure of that and do any remaining clean up.
1891 abandon_gc_alloc_regions();
1892
1893 // Instead of tearing down / rebuilding the free lists here, we
1894 // could instead use the remove_all_pending() method on free_list to
1895 // remove only the ones that we need to remove.
1896 tear_down_region_sets(true /* free_list_only */);
1897 shrink_helper(shrink_bytes);
1898 rebuild_region_sets(true /* free_list_only */);
1899
1900 _hrs.verify_optional();
1901 verify_region_sets_optional();
1902 }
1903
1904 // Public methods.
1905
1906 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1907 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1908 #endif // _MSC_VER
1909
1910
1911 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1912 SharedHeap(policy_),
1913 _g1_policy(policy_),
1914 _dirty_card_queue_set(false),
1915 _into_cset_dirty_card_queue_set(false),
1916 _is_alive_closure_cm(this),
1917 _is_alive_closure_stw(this),
1918 _ref_processor_cm(NULL),
1919 _ref_processor_stw(NULL),
1920 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1921 _bot_shared(NULL),
1922 _evac_failure_scan_stack(NULL) ,
1923 _mark_in_progress(false),
1924 _cg1r(NULL), _summary_bytes_used(0),
1925 _g1mm(NULL),
1926 _refine_cte_cl(NULL),
1927 _full_collection(false),
1928 _free_list("Master Free List"),
1929 _secondary_free_list("Secondary Free List"),
1930 _old_set("Old Set"),
1931 _humongous_set("Master Humongous Set"),
1932 _free_regions_coming(false),
1933 _young_list(new YoungList(this)),
1934 _gc_time_stamp(0),
1935 _retained_old_gc_alloc_region(NULL),
1936 _survivor_plab_stats(YoungPLABSize, PLABWeight),
1937 _old_plab_stats(OldPLABSize, PLABWeight),
1938 _expand_heap_after_alloc_failure(true),
1939 _surviving_young_words(NULL),
1940 _old_marking_cycles_started(0),
1941 _old_marking_cycles_completed(0),
1942 _in_cset_fast_test(NULL),
1943 _in_cset_fast_test_base(NULL),
1944 _dirty_cards_region_list(NULL),
1945 _worker_cset_start_region(NULL),
1946 _worker_cset_start_region_time_stamp(NULL) {
1947 _g1h = this; // To catch bugs.
1948 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1949 vm_exit_during_initialization("Failed necessary allocation.");
1950 }
1951
1952 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1953
1954 int n_queues = MAX2((int)ParallelGCThreads, 1);
1955 _task_queues = new RefToScanQueueSet(n_queues);
1956
1957 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1958 assert(n_rem_sets > 0, "Invariant.");
1959
1960 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC);
1961 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC);
1962
1963 for (int i = 0; i < n_queues; i++) {
1964 RefToScanQueue* q = new RefToScanQueue();
1965 q->initialize();
1966 _task_queues->register_queue(i, q);
1967 }
1968
1969 clear_cset_start_regions();
1970
1971 // Initialize the G1EvacuationFailureALot counters and flags.
1972 NOT_PRODUCT(reset_evacuation_should_fail();)
1973
1974 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1975 }
1976
1977 jint G1CollectedHeap::initialize() {
1978 CollectedHeap::pre_initialize();
1979 os::enable_vtime();
1980
1981 G1Log::init();
1982
1983 // Necessary to satisfy locking discipline assertions.
1984
1985 MutexLocker x(Heap_lock);
1986
1987 // We have to initialize the printer before committing the heap, as
1988 // it will be used then.
1989 _hr_printer.set_active(G1PrintHeapRegions);
1990
1991 // While there are no constraints in the GC code that HeapWordSize
1992 // be any particular value, there are multiple other areas in the
1993 // system which believe this to be true (e.g. oop->object_size in some
1994 // cases incorrectly returns the size in wordSize units rather than
1995 // HeapWordSize).
1996 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1997
1998 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1999 size_t max_byte_size = collector_policy()->max_heap_byte_size();
2000
2001 // Ensure that the sizes are properly aligned.
2002 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
2003 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
2004
2005 _cg1r = new ConcurrentG1Refine(this);
2006
2007 // Reserve the maximum.
2008
2009 // When compressed oops are enabled, the preferred heap base
2010 // is calculated by subtracting the requested size from the
2011 // 32Gb boundary and using the result as the base address for
2012 // heap reservation. If the requested size is not aligned to
2013 // HeapRegion::GrainBytes (i.e. the alignment that is passed
2014 // into the ReservedHeapSpace constructor) then the actual
2015 // base of the reserved heap may end up differing from the
2016 // address that was requested (i.e. the preferred heap base).
2017 // If this happens then we could end up using a non-optimal
2018 // compressed oops mode.
2019
2020 // Since max_byte_size is aligned to the size of a heap region (checked
2021 // above).
2022 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
2023
2024 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size,
2025 HeapRegion::GrainBytes);
2026
2027 // It is important to do this in a way such that concurrent readers can't
2028 // temporarily think somethings in the heap. (I've actually seen this
2029 // happen in asserts: DLD.)
2030 _reserved.set_word_size(0);
2031 _reserved.set_start((HeapWord*)heap_rs.base());
2032 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
2033
2034 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes);
2035
2036 // Create the gen rem set (and barrier set) for the entire reserved region.
2037 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
2038 set_barrier_set(rem_set()->bs());
2039 if (barrier_set()->is_a(BarrierSet::ModRef)) {
2040 _mr_bs = (ModRefBarrierSet*)_barrier_set;
2041 } else {
2042 vm_exit_during_initialization("G1 requires a mod ref bs.");
2043 return JNI_ENOMEM;
2044 }
2045
2046 // Also create a G1 rem set.
2047 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
2048 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
2049 } else {
2050 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
2051 return JNI_ENOMEM;
2052 }
2053
2054 // Carve out the G1 part of the heap.
2055
2056 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
2057 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
2058 g1_rs.size()/HeapWordSize);
2059
2060 _g1_storage.initialize(g1_rs, 0);
2061 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
2062 _hrs.initialize((HeapWord*) _g1_reserved.start(),
2063 (HeapWord*) _g1_reserved.end(),
2064 _expansion_regions);
2065
2066 // Do later initialization work for concurrent refinement.
2067 _cg1r->init();
2068
2069 // 6843694 - ensure that the maximum region index can fit
2070 // in the remembered set structures.
2071 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
2072 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
2073
2074 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
2075 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
2076 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
2077 "too many cards per region");
2078
2079 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
2080
2081 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
2082 heap_word_size(init_byte_size));
2083
2084 _g1h = this;
2085
2086 _in_cset_fast_test_length = max_regions();
2087 _in_cset_fast_test_base =
2088 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC);
2089
2090 // We're biasing _in_cset_fast_test to avoid subtracting the
2091 // beginning of the heap every time we want to index; basically
2092 // it's the same with what we do with the card table.
2093 _in_cset_fast_test = _in_cset_fast_test_base -
2094 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
2095
2096 // Clear the _cset_fast_test bitmap in anticipation of adding
2097 // regions to the incremental collection set for the first
2098 // evacuation pause.
2099 clear_cset_fast_test();
2100
2101 // Create the ConcurrentMark data structure and thread.
2102 // (Must do this late, so that "max_regions" is defined.)
2103 _cm = new ConcurrentMark(this, heap_rs);
2104 if (_cm == NULL || !_cm->completed_initialization()) {
2105 vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark");
2106 return JNI_ENOMEM;
2107 }
2108 _cmThread = _cm->cmThread();
2109
2110 // Initialize the from_card cache structure of HeapRegionRemSet.
2111 HeapRegionRemSet::init_heap(max_regions());
2112
2113 // Now expand into the initial heap size.
2114 if (!expand(init_byte_size)) {
2115 vm_shutdown_during_initialization("Failed to allocate initial heap.");
2116 return JNI_ENOMEM;
2117 }
2118
2119 // Perform any initialization actions delegated to the policy.
2120 g1_policy()->init();
2121
2122 _refine_cte_cl =
2123 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2124 g1_rem_set(),
2125 concurrent_g1_refine());
2126 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2127
2128 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2129 SATB_Q_FL_lock,
2130 G1SATBProcessCompletedThreshold,
2131 Shared_SATB_Q_lock);
2132
2133 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2134 DirtyCardQ_FL_lock,
2135 concurrent_g1_refine()->yellow_zone(),
2136 concurrent_g1_refine()->red_zone(),
2137 Shared_DirtyCardQ_lock);
2138
2139 if (G1DeferredRSUpdate) {
2140 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2141 DirtyCardQ_FL_lock,
2142 -1, // never trigger processing
2143 -1, // no limit on length
2144 Shared_DirtyCardQ_lock,
2145 &JavaThread::dirty_card_queue_set());
2146 }
2147
2148 // Initialize the card queue set used to hold cards containing
2149 // references into the collection set.
2150 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2151 DirtyCardQ_FL_lock,
2152 -1, // never trigger processing
2153 -1, // no limit on length
2154 Shared_DirtyCardQ_lock,
2155 &JavaThread::dirty_card_queue_set());
2156
2157 // In case we're keeping closure specialization stats, initialize those
2158 // counts and that mechanism.
2159 SpecializationStats::clear();
2160
2161 // Here we allocate the dummy full region that is required by the
2162 // G1AllocRegion class. If we don't pass an address in the reserved
2163 // space here, lots of asserts fire.
2164
2165 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2166 _g1_reserved.start());
2167 // We'll re-use the same region whether the alloc region will
2168 // require BOT updates or not and, if it doesn't, then a non-young
2169 // region will complain that it cannot support allocations without
2170 // BOT updates. So we'll tag the dummy region as young to avoid that.
2171 dummy_region->set_young();
2172 // Make sure it's full.
2173 dummy_region->set_top(dummy_region->end());
2174 G1AllocRegion::setup(this, dummy_region);
2175
2176 init_mutator_alloc_region();
2177
2178 // Do create of the monitoring and management support so that
2179 // values in the heap have been properly initialized.
2180 _g1mm = new G1MonitoringSupport(this);
2181
2182 return JNI_OK;
2183 }
2184
2185 void G1CollectedHeap::ref_processing_init() {
2186 // Reference processing in G1 currently works as follows:
2187 //
2188 // * There are two reference processor instances. One is
2189 // used to record and process discovered references
2190 // during concurrent marking; the other is used to
2191 // record and process references during STW pauses
2192 // (both full and incremental).
2193 // * Both ref processors need to 'span' the entire heap as
2194 // the regions in the collection set may be dotted around.
2195 //
2196 // * For the concurrent marking ref processor:
2197 // * Reference discovery is enabled at initial marking.
2198 // * Reference discovery is disabled and the discovered
2199 // references processed etc during remarking.
2200 // * Reference discovery is MT (see below).
2201 // * Reference discovery requires a barrier (see below).
2202 // * Reference processing may or may not be MT
2203 // (depending on the value of ParallelRefProcEnabled
2204 // and ParallelGCThreads).
2205 // * A full GC disables reference discovery by the CM
2206 // ref processor and abandons any entries on it's
2207 // discovered lists.
2208 //
2209 // * For the STW processor:
2210 // * Non MT discovery is enabled at the start of a full GC.
2211 // * Processing and enqueueing during a full GC is non-MT.
2212 // * During a full GC, references are processed after marking.
2213 //
2214 // * Discovery (may or may not be MT) is enabled at the start
2215 // of an incremental evacuation pause.
2216 // * References are processed near the end of a STW evacuation pause.
2217 // * For both types of GC:
2218 // * Discovery is atomic - i.e. not concurrent.
2219 // * Reference discovery will not need a barrier.
2220
2221 SharedHeap::ref_processing_init();
2222 MemRegion mr = reserved_region();
2223
2224 // Concurrent Mark ref processor
2225 _ref_processor_cm =
2226 new ReferenceProcessor(mr, // span
2227 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2228 // mt processing
2229 (int) ParallelGCThreads,
2230 // degree of mt processing
2231 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2232 // mt discovery
2233 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2234 // degree of mt discovery
2235 false,
2236 // Reference discovery is not atomic
2237 &_is_alive_closure_cm,
2238 // is alive closure
2239 // (for efficiency/performance)
2240 true);
2241 // Setting next fields of discovered
2242 // lists requires a barrier.
2243
2244 // STW ref processor
2245 _ref_processor_stw =
2246 new ReferenceProcessor(mr, // span
2247 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2248 // mt processing
2249 MAX2((int)ParallelGCThreads, 1),
2250 // degree of mt processing
2251 (ParallelGCThreads > 1),
2252 // mt discovery
2253 MAX2((int)ParallelGCThreads, 1),
2254 // degree of mt discovery
2255 true,
2256 // Reference discovery is atomic
2257 &_is_alive_closure_stw,
2258 // is alive closure
2259 // (for efficiency/performance)
2260 false);
2261 // Setting next fields of discovered
2262 // lists requires a barrier.
2263 }
2264
2265 size_t G1CollectedHeap::capacity() const {
2266 return _g1_committed.byte_size();
2267 }
2268
2269 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) {
2270 assert(!hr->continuesHumongous(), "pre-condition");
2271 hr->reset_gc_time_stamp();
2272 if (hr->startsHumongous()) {
2273 uint first_index = hr->hrs_index() + 1;
2274 uint last_index = hr->last_hc_index();
2275 for (uint i = first_index; i < last_index; i += 1) {
2276 HeapRegion* chr = region_at(i);
2277 assert(chr->continuesHumongous(), "sanity");
2278 chr->reset_gc_time_stamp();
2279 }
2280 }
2281 }
2282
2283 #ifndef PRODUCT
2284 class CheckGCTimeStampsHRClosure : public HeapRegionClosure {
2285 private:
2286 unsigned _gc_time_stamp;
2287 bool _failures;
2288
2289 public:
2290 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) :
2291 _gc_time_stamp(gc_time_stamp), _failures(false) { }
2292
2293 virtual bool doHeapRegion(HeapRegion* hr) {
2294 unsigned region_gc_time_stamp = hr->get_gc_time_stamp();
2295 if (_gc_time_stamp != region_gc_time_stamp) {
2296 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, "
2297 "expected %d", HR_FORMAT_PARAMS(hr),
2298 region_gc_time_stamp, _gc_time_stamp);
2299 _failures = true;
2300 }
2301 return false;
2302 }
2303
2304 bool failures() { return _failures; }
2305 };
2306
2307 void G1CollectedHeap::check_gc_time_stamps() {
2308 CheckGCTimeStampsHRClosure cl(_gc_time_stamp);
2309 heap_region_iterate(&cl);
2310 guarantee(!cl.failures(), "all GC time stamps should have been reset");
2311 }
2312 #endif // PRODUCT
2313
2314 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2315 DirtyCardQueue* into_cset_dcq,
2316 bool concurrent,
2317 int worker_i) {
2318 // Clean cards in the hot card cache
2319 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache();
2320 hot_card_cache->drain(worker_i, g1_rem_set(), into_cset_dcq);
2321
2322 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2323 int n_completed_buffers = 0;
2324 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2325 n_completed_buffers++;
2326 }
2327 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers);
2328 dcqs.clear_n_completed_buffers();
2329 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2330 }
2331
2332
2333 // Computes the sum of the storage used by the various regions.
2334
2335 size_t G1CollectedHeap::used() const {
2336 assert(Heap_lock->owner() != NULL,
2337 "Should be owned on this thread's behalf.");
2338 size_t result = _summary_bytes_used;
2339 // Read only once in case it is set to NULL concurrently
2340 HeapRegion* hr = _mutator_alloc_region.get();
2341 if (hr != NULL)
2342 result += hr->used();
2343 return result;
2344 }
2345
2346 size_t G1CollectedHeap::used_unlocked() const {
2347 size_t result = _summary_bytes_used;
2348 return result;
2349 }
2350
2351 class SumUsedClosure: public HeapRegionClosure {
2352 size_t _used;
2353 public:
2354 SumUsedClosure() : _used(0) {}
2355 bool doHeapRegion(HeapRegion* r) {
2356 if (!r->continuesHumongous()) {
2357 _used += r->used();
2358 }
2359 return false;
2360 }
2361 size_t result() { return _used; }
2362 };
2363
2364 size_t G1CollectedHeap::recalculate_used() const {
2365 SumUsedClosure blk;
2366 heap_region_iterate(&blk);
2367 return blk.result();
2368 }
2369
2370 size_t G1CollectedHeap::unsafe_max_alloc() {
2371 if (free_regions() > 0) return HeapRegion::GrainBytes;
2372 // otherwise, is there space in the current allocation region?
2373
2374 // We need to store the current allocation region in a local variable
2375 // here. The problem is that this method doesn't take any locks and
2376 // there may be other threads which overwrite the current allocation
2377 // region field. attempt_allocation(), for example, sets it to NULL
2378 // and this can happen *after* the NULL check here but before the call
2379 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2380 // to be a problem in the optimized build, since the two loads of the
2381 // current allocation region field are optimized away.
2382 HeapRegion* hr = _mutator_alloc_region.get();
2383 if (hr == NULL) {
2384 return 0;
2385 }
2386 return hr->free();
2387 }
2388
2389 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2390 switch (cause) {
2391 case GCCause::_gc_locker: return GCLockerInvokesConcurrent;
2392 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent;
2393 case GCCause::_g1_humongous_allocation: return true;
2394 default: return false;
2395 }
2396 }
2397
2398 #ifndef PRODUCT
2399 void G1CollectedHeap::allocate_dummy_regions() {
2400 // Let's fill up most of the region
2401 size_t word_size = HeapRegion::GrainWords - 1024;
2402 // And as a result the region we'll allocate will be humongous.
2403 guarantee(isHumongous(word_size), "sanity");
2404
2405 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2406 // Let's use the existing mechanism for the allocation
2407 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2408 if (dummy_obj != NULL) {
2409 MemRegion mr(dummy_obj, word_size);
2410 CollectedHeap::fill_with_object(mr);
2411 } else {
2412 // If we can't allocate once, we probably cannot allocate
2413 // again. Let's get out of the loop.
2414 break;
2415 }
2416 }
2417 }
2418 #endif // !PRODUCT
2419
2420 void G1CollectedHeap::increment_old_marking_cycles_started() {
2421 assert(_old_marking_cycles_started == _old_marking_cycles_completed ||
2422 _old_marking_cycles_started == _old_marking_cycles_completed + 1,
2423 err_msg("Wrong marking cycle count (started: %d, completed: %d)",
2424 _old_marking_cycles_started, _old_marking_cycles_completed));
2425
2426 _old_marking_cycles_started++;
2427 }
2428
2429 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) {
2430 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2431
2432 // We assume that if concurrent == true, then the caller is a
2433 // concurrent thread that was joined the Suspendible Thread
2434 // Set. If there's ever a cheap way to check this, we should add an
2435 // assert here.
2436
2437 // Given that this method is called at the end of a Full GC or of a
2438 // concurrent cycle, and those can be nested (i.e., a Full GC can
2439 // interrupt a concurrent cycle), the number of full collections
2440 // completed should be either one (in the case where there was no
2441 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2442 // behind the number of full collections started.
2443
2444 // This is the case for the inner caller, i.e. a Full GC.
2445 assert(concurrent ||
2446 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) ||
2447 (_old_marking_cycles_started == _old_marking_cycles_completed + 2),
2448 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u "
2449 "is inconsistent with _old_marking_cycles_completed = %u",
2450 _old_marking_cycles_started, _old_marking_cycles_completed));
2451
2452 // This is the case for the outer caller, i.e. the concurrent cycle.
2453 assert(!concurrent ||
2454 (_old_marking_cycles_started == _old_marking_cycles_completed + 1),
2455 err_msg("for outer caller (concurrent cycle): "
2456 "_old_marking_cycles_started = %u "
2457 "is inconsistent with _old_marking_cycles_completed = %u",
2458 _old_marking_cycles_started, _old_marking_cycles_completed));
2459
2460 _old_marking_cycles_completed += 1;
2461
2462 // We need to clear the "in_progress" flag in the CM thread before
2463 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2464 // is set) so that if a waiter requests another System.gc() it doesn't
2465 // incorrectly see that a marking cyle is still in progress.
2466 if (concurrent) {
2467 _cmThread->clear_in_progress();
2468 }
2469
2470 // This notify_all() will ensure that a thread that called
2471 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2472 // and it's waiting for a full GC to finish will be woken up. It is
2473 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2474 FullGCCount_lock->notify_all();
2475 }
2476
2477 void G1CollectedHeap::collect(GCCause::Cause cause) {
2478 assert_heap_not_locked();
2479
2480 unsigned int gc_count_before;
2481 unsigned int old_marking_count_before;
2482 bool retry_gc;
2483
2484 do {
2485 retry_gc = false;
2486
2487 {
2488 MutexLocker ml(Heap_lock);
2489
2490 // Read the GC count while holding the Heap_lock
2491 gc_count_before = total_collections();
2492 old_marking_count_before = _old_marking_cycles_started;
2493 }
2494
2495 if (should_do_concurrent_full_gc(cause)) {
2496 // Schedule an initial-mark evacuation pause that will start a
2497 // concurrent cycle. We're setting word_size to 0 which means that
2498 // we are not requesting a post-GC allocation.
2499 VM_G1IncCollectionPause op(gc_count_before,
2500 0, /* word_size */
2501 true, /* should_initiate_conc_mark */
2502 g1_policy()->max_pause_time_ms(),
2503 cause);
2504
2505 VMThread::execute(&op);
2506 if (!op.pause_succeeded()) {
2507 if (old_marking_count_before == _old_marking_cycles_started) {
2508 retry_gc = op.should_retry_gc();
2509 } else {
2510 // A Full GC happened while we were trying to schedule the
2511 // initial-mark GC. No point in starting a new cycle given
2512 // that the whole heap was collected anyway.
2513 }
2514
2515 if (retry_gc) {
2516 if (GC_locker::is_active_and_needs_gc()) {
2517 GC_locker::stall_until_clear();
2518 }
2519 }
2520 }
2521 } else {
2522 if (cause == GCCause::_gc_locker
2523 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2524
2525 // Schedule a standard evacuation pause. We're setting word_size
2526 // to 0 which means that we are not requesting a post-GC allocation.
2527 VM_G1IncCollectionPause op(gc_count_before,
2528 0, /* word_size */
2529 false, /* should_initiate_conc_mark */
2530 g1_policy()->max_pause_time_ms(),
2531 cause);
2532 VMThread::execute(&op);
2533 } else {
2534 // Schedule a Full GC.
2535 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause);
2536 VMThread::execute(&op);
2537 }
2538 }
2539 } while (retry_gc);
2540 }
2541
2542 bool G1CollectedHeap::is_in(const void* p) const {
2543 if (_g1_committed.contains(p)) {
2544 // Given that we know that p is in the committed space,
2545 // heap_region_containing_raw() should successfully
2546 // return the containing region.
2547 HeapRegion* hr = heap_region_containing_raw(p);
2548 return hr->is_in(p);
2549 } else {
2550 return false;
2551 }
2552 }
2553
2554 // Iteration functions.
2555
2556 // Iterates an OopClosure over all ref-containing fields of objects
2557 // within a HeapRegion.
2558
2559 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2560 MemRegion _mr;
2561 ExtendedOopClosure* _cl;
2562 public:
2563 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl)
2564 : _mr(mr), _cl(cl) {}
2565 bool doHeapRegion(HeapRegion* r) {
2566 if (!r->continuesHumongous()) {
2567 r->oop_iterate(_cl);
2568 }
2569 return false;
2570 }
2571 };
2572
2573 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) {
2574 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2575 heap_region_iterate(&blk);
2576 }
2577
2578 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
2579 IterateOopClosureRegionClosure blk(mr, cl);
2580 heap_region_iterate(&blk);
2581 }
2582
2583 // Iterates an ObjectClosure over all objects within a HeapRegion.
2584
2585 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2586 ObjectClosure* _cl;
2587 public:
2588 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2589 bool doHeapRegion(HeapRegion* r) {
2590 if (! r->continuesHumongous()) {
2591 r->object_iterate(_cl);
2592 }
2593 return false;
2594 }
2595 };
2596
2597 void G1CollectedHeap::object_iterate(ObjectClosure* cl) {
2598 IterateObjectClosureRegionClosure blk(cl);
2599 heap_region_iterate(&blk);
2600 }
2601
2602 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2603 // FIXME: is this right?
2604 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2605 }
2606
2607 // Calls a SpaceClosure on a HeapRegion.
2608
2609 class SpaceClosureRegionClosure: public HeapRegionClosure {
2610 SpaceClosure* _cl;
2611 public:
2612 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2613 bool doHeapRegion(HeapRegion* r) {
2614 _cl->do_space(r);
2615 return false;
2616 }
2617 };
2618
2619 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2620 SpaceClosureRegionClosure blk(cl);
2621 heap_region_iterate(&blk);
2622 }
2623
2624 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2625 _hrs.iterate(cl);
2626 }
2627
2628 void
2629 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2630 uint worker_id,
2631 uint no_of_par_workers,
2632 jint claim_value) {
2633 const uint regions = n_regions();
2634 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
2635 no_of_par_workers :
2636 1);
2637 assert(UseDynamicNumberOfGCThreads ||
2638 no_of_par_workers == workers()->total_workers(),
2639 "Non dynamic should use fixed number of workers");
2640 // try to spread out the starting points of the workers
2641 const HeapRegion* start_hr =
2642 start_region_for_worker(worker_id, no_of_par_workers);
2643 const uint start_index = start_hr->hrs_index();
2644
2645 // each worker will actually look at all regions
2646 for (uint count = 0; count < regions; ++count) {
2647 const uint index = (start_index + count) % regions;
2648 assert(0 <= index && index < regions, "sanity");
2649 HeapRegion* r = region_at(index);
2650 // we'll ignore "continues humongous" regions (we'll process them
2651 // when we come across their corresponding "start humongous"
2652 // region) and regions already claimed
2653 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2654 continue;
2655 }
2656 // OK, try to claim it
2657 if (r->claimHeapRegion(claim_value)) {
2658 // success!
2659 assert(!r->continuesHumongous(), "sanity");
2660 if (r->startsHumongous()) {
2661 // If the region is "starts humongous" we'll iterate over its
2662 // "continues humongous" first; in fact we'll do them
2663 // first. The order is important. In on case, calling the
2664 // closure on the "starts humongous" region might de-allocate
2665 // and clear all its "continues humongous" regions and, as a
2666 // result, we might end up processing them twice. So, we'll do
2667 // them first (notice: most closures will ignore them anyway) and
2668 // then we'll do the "starts humongous" region.
2669 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) {
2670 HeapRegion* chr = region_at(ch_index);
2671
2672 // if the region has already been claimed or it's not
2673 // "continues humongous" we're done
2674 if (chr->claim_value() == claim_value ||
2675 !chr->continuesHumongous()) {
2676 break;
2677 }
2678
2679 // Noone should have claimed it directly. We can given
2680 // that we claimed its "starts humongous" region.
2681 assert(chr->claim_value() != claim_value, "sanity");
2682 assert(chr->humongous_start_region() == r, "sanity");
2683
2684 if (chr->claimHeapRegion(claim_value)) {
2685 // we should always be able to claim it; noone else should
2686 // be trying to claim this region
2687
2688 bool res2 = cl->doHeapRegion(chr);
2689 assert(!res2, "Should not abort");
2690
2691 // Right now, this holds (i.e., no closure that actually
2692 // does something with "continues humongous" regions
2693 // clears them). We might have to weaken it in the future,
2694 // but let's leave these two asserts here for extra safety.
2695 assert(chr->continuesHumongous(), "should still be the case");
2696 assert(chr->humongous_start_region() == r, "sanity");
2697 } else {
2698 guarantee(false, "we should not reach here");
2699 }
2700 }
2701 }
2702
2703 assert(!r->continuesHumongous(), "sanity");
2704 bool res = cl->doHeapRegion(r);
2705 assert(!res, "Should not abort");
2706 }
2707 }
2708 }
2709
2710 class ResetClaimValuesClosure: public HeapRegionClosure {
2711 public:
2712 bool doHeapRegion(HeapRegion* r) {
2713 r->set_claim_value(HeapRegion::InitialClaimValue);
2714 return false;
2715 }
2716 };
2717
2718 void G1CollectedHeap::reset_heap_region_claim_values() {
2719 ResetClaimValuesClosure blk;
2720 heap_region_iterate(&blk);
2721 }
2722
2723 void G1CollectedHeap::reset_cset_heap_region_claim_values() {
2724 ResetClaimValuesClosure blk;
2725 collection_set_iterate(&blk);
2726 }
2727
2728 #ifdef ASSERT
2729 // This checks whether all regions in the heap have the correct claim
2730 // value. I also piggy-backed on this a check to ensure that the
2731 // humongous_start_region() information on "continues humongous"
2732 // regions is correct.
2733
2734 class CheckClaimValuesClosure : public HeapRegionClosure {
2735 private:
2736 jint _claim_value;
2737 uint _failures;
2738 HeapRegion* _sh_region;
2739
2740 public:
2741 CheckClaimValuesClosure(jint claim_value) :
2742 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2743 bool doHeapRegion(HeapRegion* r) {
2744 if (r->claim_value() != _claim_value) {
2745 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2746 "claim value = %d, should be %d",
2747 HR_FORMAT_PARAMS(r),
2748 r->claim_value(), _claim_value);
2749 ++_failures;
2750 }
2751 if (!r->isHumongous()) {
2752 _sh_region = NULL;
2753 } else if (r->startsHumongous()) {
2754 _sh_region = r;
2755 } else if (r->continuesHumongous()) {
2756 if (r->humongous_start_region() != _sh_region) {
2757 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2758 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2759 HR_FORMAT_PARAMS(r),
2760 r->humongous_start_region(),
2761 _sh_region);
2762 ++_failures;
2763 }
2764 }
2765 return false;
2766 }
2767 uint failures() { return _failures; }
2768 };
2769
2770 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2771 CheckClaimValuesClosure cl(claim_value);
2772 heap_region_iterate(&cl);
2773 return cl.failures() == 0;
2774 }
2775
2776 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure {
2777 private:
2778 jint _claim_value;
2779 uint _failures;
2780
2781 public:
2782 CheckClaimValuesInCSetHRClosure(jint claim_value) :
2783 _claim_value(claim_value), _failures(0) { }
2784
2785 uint failures() { return _failures; }
2786
2787 bool doHeapRegion(HeapRegion* hr) {
2788 assert(hr->in_collection_set(), "how?");
2789 assert(!hr->isHumongous(), "H-region in CSet");
2790 if (hr->claim_value() != _claim_value) {
2791 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", "
2792 "claim value = %d, should be %d",
2793 HR_FORMAT_PARAMS(hr),
2794 hr->claim_value(), _claim_value);
2795 _failures += 1;
2796 }
2797 return false;
2798 }
2799 };
2800
2801 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) {
2802 CheckClaimValuesInCSetHRClosure cl(claim_value);
2803 collection_set_iterate(&cl);
2804 return cl.failures() == 0;
2805 }
2806 #endif // ASSERT
2807
2808 // Clear the cached CSet starting regions and (more importantly)
2809 // the time stamps. Called when we reset the GC time stamp.
2810 void G1CollectedHeap::clear_cset_start_regions() {
2811 assert(_worker_cset_start_region != NULL, "sanity");
2812 assert(_worker_cset_start_region_time_stamp != NULL, "sanity");
2813
2814 int n_queues = MAX2((int)ParallelGCThreads, 1);
2815 for (int i = 0; i < n_queues; i++) {
2816 _worker_cset_start_region[i] = NULL;
2817 _worker_cset_start_region_time_stamp[i] = 0;
2818 }
2819 }
2820
2821 // Given the id of a worker, obtain or calculate a suitable
2822 // starting region for iterating over the current collection set.
2823 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) {
2824 assert(get_gc_time_stamp() > 0, "should have been updated by now");
2825
2826 HeapRegion* result = NULL;
2827 unsigned gc_time_stamp = get_gc_time_stamp();
2828
2829 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) {
2830 // Cached starting region for current worker was set
2831 // during the current pause - so it's valid.
2832 // Note: the cached starting heap region may be NULL
2833 // (when the collection set is empty).
2834 result = _worker_cset_start_region[worker_i];
2835 assert(result == NULL || result->in_collection_set(), "sanity");
2836 return result;
2837 }
2838
2839 // The cached entry was not valid so let's calculate
2840 // a suitable starting heap region for this worker.
2841
2842 // We want the parallel threads to start their collection
2843 // set iteration at different collection set regions to
2844 // avoid contention.
2845 // If we have:
2846 // n collection set regions
2847 // p threads
2848 // Then thread t will start at region floor ((t * n) / p)
2849
2850 result = g1_policy()->collection_set();
2851 if (G1CollectedHeap::use_parallel_gc_threads()) {
2852 uint cs_size = g1_policy()->cset_region_length();
2853 uint active_workers = workers()->active_workers();
2854 assert(UseDynamicNumberOfGCThreads ||
2855 active_workers == workers()->total_workers(),
2856 "Unless dynamic should use total workers");
2857
2858 uint end_ind = (cs_size * worker_i) / active_workers;
2859 uint start_ind = 0;
2860
2861 if (worker_i > 0 &&
2862 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) {
2863 // Previous workers starting region is valid
2864 // so let's iterate from there
2865 start_ind = (cs_size * (worker_i - 1)) / active_workers;
2866 result = _worker_cset_start_region[worker_i - 1];
2867 }
2868
2869 for (uint i = start_ind; i < end_ind; i++) {
2870 result = result->next_in_collection_set();
2871 }
2872 }
2873
2874 // Note: the calculated starting heap region may be NULL
2875 // (when the collection set is empty).
2876 assert(result == NULL || result->in_collection_set(), "sanity");
2877 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp,
2878 "should be updated only once per pause");
2879 _worker_cset_start_region[worker_i] = result;
2880 OrderAccess::storestore();
2881 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp;
2882 return result;
2883 }
2884
2885 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i,
2886 uint no_of_par_workers) {
2887 uint worker_num =
2888 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U;
2889 assert(UseDynamicNumberOfGCThreads ||
2890 no_of_par_workers == workers()->total_workers(),
2891 "Non dynamic should use fixed number of workers");
2892 const uint start_index = n_regions() * worker_i / worker_num;
2893 return region_at(start_index);
2894 }
2895
2896 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2897 HeapRegion* r = g1_policy()->collection_set();
2898 while (r != NULL) {
2899 HeapRegion* next = r->next_in_collection_set();
2900 if (cl->doHeapRegion(r)) {
2901 cl->incomplete();
2902 return;
2903 }
2904 r = next;
2905 }
2906 }
2907
2908 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2909 HeapRegionClosure *cl) {
2910 if (r == NULL) {
2911 // The CSet is empty so there's nothing to do.
2912 return;
2913 }
2914
2915 assert(r->in_collection_set(),
2916 "Start region must be a member of the collection set.");
2917 HeapRegion* cur = r;
2918 while (cur != NULL) {
2919 HeapRegion* next = cur->next_in_collection_set();
2920 if (cl->doHeapRegion(cur) && false) {
2921 cl->incomplete();
2922 return;
2923 }
2924 cur = next;
2925 }
2926 cur = g1_policy()->collection_set();
2927 while (cur != r) {
2928 HeapRegion* next = cur->next_in_collection_set();
2929 if (cl->doHeapRegion(cur) && false) {
2930 cl->incomplete();
2931 return;
2932 }
2933 cur = next;
2934 }
2935 }
2936
2937 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2938 return n_regions() > 0 ? region_at(0) : NULL;
2939 }
2940
2941
2942 Space* G1CollectedHeap::space_containing(const void* addr) const {
2943 Space* res = heap_region_containing(addr);
2944 return res;
2945 }
2946
2947 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2948 Space* sp = space_containing(addr);
2949 if (sp != NULL) {
2950 return sp->block_start(addr);
2951 }
2952 return NULL;
2953 }
2954
2955 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2956 Space* sp = space_containing(addr);
2957 assert(sp != NULL, "block_size of address outside of heap");
2958 return sp->block_size(addr);
2959 }
2960
2961 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2962 Space* sp = space_containing(addr);
2963 return sp->block_is_obj(addr);
2964 }
2965
2966 bool G1CollectedHeap::supports_tlab_allocation() const {
2967 return true;
2968 }
2969
2970 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2971 return HeapRegion::GrainBytes;
2972 }
2973
2974 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2975 // Return the remaining space in the cur alloc region, but not less than
2976 // the min TLAB size.
2977
2978 // Also, this value can be at most the humongous object threshold,
2979 // since we can't allow tlabs to grow big enough to accomodate
2980 // humongous objects.
2981
2982 HeapRegion* hr = _mutator_alloc_region.get();
2983 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2984 if (hr == NULL) {
2985 return max_tlab_size;
2986 } else {
2987 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2988 }
2989 }
2990
2991 size_t G1CollectedHeap::max_capacity() const {
2992 return _g1_reserved.byte_size();
2993 }
2994
2995 jlong G1CollectedHeap::millis_since_last_gc() {
2996 // assert(false, "NYI");
2997 return 0;
2998 }
2999
3000 void G1CollectedHeap::prepare_for_verify() {
3001 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
3002 ensure_parsability(false);
3003 }
3004 g1_rem_set()->prepare_for_verify();
3005 }
3006
3007 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr,
3008 VerifyOption vo) {
3009 switch (vo) {
3010 case VerifyOption_G1UsePrevMarking:
3011 return hr->obj_allocated_since_prev_marking(obj);
3012 case VerifyOption_G1UseNextMarking:
3013 return hr->obj_allocated_since_next_marking(obj);
3014 case VerifyOption_G1UseMarkWord:
3015 return false;
3016 default:
3017 ShouldNotReachHere();
3018 }
3019 return false; // keep some compilers happy
3020 }
3021
3022 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) {
3023 switch (vo) {
3024 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start();
3025 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start();
3026 case VerifyOption_G1UseMarkWord: return NULL;
3027 default: ShouldNotReachHere();
3028 }
3029 return NULL; // keep some compilers happy
3030 }
3031
3032 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) {
3033 switch (vo) {
3034 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj);
3035 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj);
3036 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked();
3037 default: ShouldNotReachHere();
3038 }
3039 return false; // keep some compilers happy
3040 }
3041
3042 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) {
3043 switch (vo) {
3044 case VerifyOption_G1UsePrevMarking: return "PTAMS";
3045 case VerifyOption_G1UseNextMarking: return "NTAMS";
3046 case VerifyOption_G1UseMarkWord: return "NONE";
3047 default: ShouldNotReachHere();
3048 }
3049 return NULL; // keep some compilers happy
3050 }
3051
3052 class VerifyLivenessOopClosure: public OopClosure {
3053 G1CollectedHeap* _g1h;
3054 VerifyOption _vo;
3055 public:
3056 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
3057 _g1h(g1h), _vo(vo)
3058 { }
3059 void do_oop(narrowOop *p) { do_oop_work(p); }
3060 void do_oop( oop *p) { do_oop_work(p); }
3061
3062 template <class T> void do_oop_work(T *p) {
3063 oop obj = oopDesc::load_decode_heap_oop(p);
3064 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
3065 "Dead object referenced by a not dead object");
3066 }
3067 };
3068
3069 class VerifyObjsInRegionClosure: public ObjectClosure {
3070 private:
3071 G1CollectedHeap* _g1h;
3072 size_t _live_bytes;
3073 HeapRegion *_hr;
3074 VerifyOption _vo;
3075 public:
3076 // _vo == UsePrevMarking -> use "prev" marking information,
3077 // _vo == UseNextMarking -> use "next" marking information,
3078 // _vo == UseMarkWord -> use mark word from object header.
3079 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
3080 : _live_bytes(0), _hr(hr), _vo(vo) {
3081 _g1h = G1CollectedHeap::heap();
3082 }
3083 void do_object(oop o) {
3084 VerifyLivenessOopClosure isLive(_g1h, _vo);
3085 assert(o != NULL, "Huh?");
3086 if (!_g1h->is_obj_dead_cond(o, _vo)) {
3087 // If the object is alive according to the mark word,
3088 // then verify that the marking information agrees.
3089 // Note we can't verify the contra-positive of the
3090 // above: if the object is dead (according to the mark
3091 // word), it may not be marked, or may have been marked
3092 // but has since became dead, or may have been allocated
3093 // since the last marking.
3094 if (_vo == VerifyOption_G1UseMarkWord) {
3095 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
3096 }
3097
3098 o->oop_iterate_no_header(&isLive);
3099 if (!_hr->obj_allocated_since_prev_marking(o)) {
3100 size_t obj_size = o->size(); // Make sure we don't overflow
3101 _live_bytes += (obj_size * HeapWordSize);
3102 }
3103 }
3104 }
3105 size_t live_bytes() { return _live_bytes; }
3106 };
3107
3108 class PrintObjsInRegionClosure : public ObjectClosure {
3109 HeapRegion *_hr;
3110 G1CollectedHeap *_g1;
3111 public:
3112 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
3113 _g1 = G1CollectedHeap::heap();
3114 };
3115
3116 void do_object(oop o) {
3117 if (o != NULL) {
3118 HeapWord *start = (HeapWord *) o;
3119 size_t word_sz = o->size();
3120 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
3121 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
3122 (void*) o, word_sz,
3123 _g1->isMarkedPrev(o),
3124 _g1->isMarkedNext(o),
3125 _hr->obj_allocated_since_prev_marking(o));
3126 HeapWord *end = start + word_sz;
3127 HeapWord *cur;
3128 int *val;
3129 for (cur = start; cur < end; cur++) {
3130 val = (int *) cur;
3131 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
3132 }
3133 }
3134 }
3135 };
3136
3137 class VerifyRegionClosure: public HeapRegionClosure {
3138 private:
3139 bool _par;
3140 VerifyOption _vo;
3141 bool _failures;
3142 public:
3143 // _vo == UsePrevMarking -> use "prev" marking information,
3144 // _vo == UseNextMarking -> use "next" marking information,
3145 // _vo == UseMarkWord -> use mark word from object header.
3146 VerifyRegionClosure(bool par, VerifyOption vo)
3147 : _par(par),
3148 _vo(vo),
3149 _failures(false) {}
3150
3151 bool failures() {
3152 return _failures;
3153 }
3154
3155 bool doHeapRegion(HeapRegion* r) {
3156 if (!r->continuesHumongous()) {
3157 bool failures = false;
3158 r->verify(_vo, &failures);
3159 if (failures) {
3160 _failures = true;
3161 } else {
3162 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
3163 r->object_iterate(¬_dead_yet_cl);
3164 if (_vo != VerifyOption_G1UseNextMarking) {
3165 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
3166 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
3167 "max_live_bytes "SIZE_FORMAT" "
3168 "< calculated "SIZE_FORMAT,
3169 r->bottom(), r->end(),
3170 r->max_live_bytes(),
3171 not_dead_yet_cl.live_bytes());
3172 _failures = true;
3173 }
3174 } else {
3175 // When vo == UseNextMarking we cannot currently do a sanity
3176 // check on the live bytes as the calculation has not been
3177 // finalized yet.
3178 }
3179 }
3180 }
3181 return false; // stop the region iteration if we hit a failure
3182 }
3183 };
3184
3185 class YoungRefCounterClosure : public OopClosure {
3186 G1CollectedHeap* _g1h;
3187 int _count;
3188 public:
3189 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {}
3190 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } }
3191 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3192
3193 int count() { return _count; }
3194 void reset_count() { _count = 0; };
3195 };
3196
3197 class VerifyKlassClosure: public KlassClosure {
3198 YoungRefCounterClosure _young_ref_counter_closure;
3199 OopClosure *_oop_closure;
3200 public:
3201 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {}
3202 void do_klass(Klass* k) {
3203 k->oops_do(_oop_closure);
3204
3205 _young_ref_counter_closure.reset_count();
3206 k->oops_do(&_young_ref_counter_closure);
3207 if (_young_ref_counter_closure.count() > 0) {
3208 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k));
3209 }
3210 }
3211 };
3212
3213 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can
3214 // pass it as the perm_blk to SharedHeap::process_strong_roots.
3215 // When process_strong_roots stop calling perm_blk->younger_refs_iterate
3216 // we can change this closure to extend the simpler OopClosure.
3217 class VerifyRootsClosure: public OopsInGenClosure {
3218 private:
3219 G1CollectedHeap* _g1h;
3220 VerifyOption _vo;
3221 bool _failures;
3222 public:
3223 // _vo == UsePrevMarking -> use "prev" marking information,
3224 // _vo == UseNextMarking -> use "next" marking information,
3225 // _vo == UseMarkWord -> use mark word from object header.
3226 VerifyRootsClosure(VerifyOption vo) :
3227 _g1h(G1CollectedHeap::heap()),
3228 _vo(vo),
3229 _failures(false) { }
3230
3231 bool failures() { return _failures; }
3232
3233 template <class T> void do_oop_nv(T* p) {
3234 T heap_oop = oopDesc::load_heap_oop(p);
3235 if (!oopDesc::is_null(heap_oop)) {
3236 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
3237 if (_g1h->is_obj_dead_cond(obj, _vo)) {
3238 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
3239 "points to dead obj "PTR_FORMAT, p, (void*) obj);
3240 if (_vo == VerifyOption_G1UseMarkWord) {
3241 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
3242 }
3243 obj->print_on(gclog_or_tty);
3244 _failures = true;
3245 }
3246 }
3247 }
3248
3249 void do_oop(oop* p) { do_oop_nv(p); }
3250 void do_oop(narrowOop* p) { do_oop_nv(p); }
3251 };
3252
3253 // This is the task used for parallel heap verification.
3254
3255 class G1ParVerifyTask: public AbstractGangTask {
3256 private:
3257 G1CollectedHeap* _g1h;
3258 VerifyOption _vo;
3259 bool _failures;
3260
3261 public:
3262 // _vo == UsePrevMarking -> use "prev" marking information,
3263 // _vo == UseNextMarking -> use "next" marking information,
3264 // _vo == UseMarkWord -> use mark word from object header.
3265 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) :
3266 AbstractGangTask("Parallel verify task"),
3267 _g1h(g1h),
3268 _vo(vo),
3269 _failures(false) { }
3270
3271 bool failures() {
3272 return _failures;
3273 }
3274
3275 void work(uint worker_id) {
3276 HandleMark hm;
3277 VerifyRegionClosure blk(true, _vo);
3278 _g1h->heap_region_par_iterate_chunked(&blk, worker_id,
3279 _g1h->workers()->active_workers(),
3280 HeapRegion::ParVerifyClaimValue);
3281 if (blk.failures()) {
3282 _failures = true;
3283 }
3284 }
3285 };
3286
3287 void G1CollectedHeap::verify(bool silent) {
3288 verify(silent, VerifyOption_G1UsePrevMarking);
3289 }
3290
3291 void G1CollectedHeap::verify(bool silent,
3292 VerifyOption vo) {
3293 if (SafepointSynchronize::is_at_safepoint()) {
3294 if (!silent) { gclog_or_tty->print("Roots "); }
3295 VerifyRootsClosure rootsCl(vo);
3296
3297 assert(Thread::current()->is_VM_thread(),
3298 "Expected to be executed serially by the VM thread at this point");
3299
3300 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
3301 VerifyKlassClosure klassCl(this, &rootsCl);
3302
3303 // We apply the relevant closures to all the oops in the
3304 // system dictionary, the string table and the code cache.
3305 const int so = SO_AllClasses | SO_Strings | SO_CodeCache;
3306
3307 // Need cleared claim bits for the strong roots processing
3308 ClassLoaderDataGraph::clear_claimed_marks();
3309
3310 process_strong_roots(true, // activate StrongRootsScope
3311 false, // we set "is scavenging" to false,
3312 // so we don't reset the dirty cards.
3313 ScanningOption(so), // roots scanning options
3314 &rootsCl,
3315 &blobsCl,
3316 &klassCl
3317 );
3318
3319 bool failures = rootsCl.failures();
3320
3321 if (vo != VerifyOption_G1UseMarkWord) {
3322 // If we're verifying during a full GC then the region sets
3323 // will have been torn down at the start of the GC. Therefore
3324 // verifying the region sets will fail. So we only verify
3325 // the region sets when not in a full GC.
3326 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
3327 verify_region_sets();
3328 }
3329
3330 if (!silent) { gclog_or_tty->print("HeapRegions "); }
3331 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
3332 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3333 "sanity check");
3334
3335 G1ParVerifyTask task(this, vo);
3336 assert(UseDynamicNumberOfGCThreads ||
3337 workers()->active_workers() == workers()->total_workers(),
3338 "If not dynamic should be using all the workers");
3339 int n_workers = workers()->active_workers();
3340 set_par_threads(n_workers);
3341 workers()->run_task(&task);
3342 set_par_threads(0);
3343 if (task.failures()) {
3344 failures = true;
3345 }
3346
3347 // Checks that the expected amount of parallel work was done.
3348 // The implication is that n_workers is > 0.
3349 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
3350 "sanity check");
3351
3352 reset_heap_region_claim_values();
3353
3354 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3355 "sanity check");
3356 } else {
3357 VerifyRegionClosure blk(false, vo);
3358 heap_region_iterate(&blk);
3359 if (blk.failures()) {
3360 failures = true;
3361 }
3362 }
3363 if (!silent) gclog_or_tty->print("RemSet ");
3364 rem_set()->verify();
3365
3366 if (failures) {
3367 gclog_or_tty->print_cr("Heap:");
3368 // It helps to have the per-region information in the output to
3369 // help us track down what went wrong. This is why we call
3370 // print_extended_on() instead of print_on().
3371 print_extended_on(gclog_or_tty);
3372 gclog_or_tty->print_cr("");
3373 #ifndef PRODUCT
3374 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3375 concurrent_mark()->print_reachable("at-verification-failure",
3376 vo, false /* all */);
3377 }
3378 #endif
3379 gclog_or_tty->flush();
3380 }
3381 guarantee(!failures, "there should not have been any failures");
3382 } else {
3383 if (!silent)
3384 gclog_or_tty->print("(SKIPPING roots, heapRegionSets, heapRegions, remset) ");
3385 }
3386 }
3387
3388 class PrintRegionClosure: public HeapRegionClosure {
3389 outputStream* _st;
3390 public:
3391 PrintRegionClosure(outputStream* st) : _st(st) {}
3392 bool doHeapRegion(HeapRegion* r) {
3393 r->print_on(_st);
3394 return false;
3395 }
3396 };
3397
3398 void G1CollectedHeap::print_on(outputStream* st) const {
3399 st->print(" %-20s", "garbage-first heap");
3400 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3401 capacity()/K, used_unlocked()/K);
3402 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3403 _g1_storage.low_boundary(),
3404 _g1_storage.high(),
3405 _g1_storage.high_boundary());
3406 st->cr();
3407 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3408 uint young_regions = _young_list->length();
3409 st->print("%u young (" SIZE_FORMAT "K), ", young_regions,
3410 (size_t) young_regions * HeapRegion::GrainBytes / K);
3411 uint survivor_regions = g1_policy()->recorded_survivor_regions();
3412 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions,
3413 (size_t) survivor_regions * HeapRegion::GrainBytes / K);
3414 st->cr();
3415 MetaspaceAux::print_on(st);
3416 }
3417
3418 void G1CollectedHeap::print_extended_on(outputStream* st) const {
3419 print_on(st);
3420
3421 // Print the per-region information.
3422 st->cr();
3423 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), "
3424 "HS=humongous(starts), HC=humongous(continues), "
3425 "CS=collection set, F=free, TS=gc time stamp, "
3426 "PTAMS=previous top-at-mark-start, "
3427 "NTAMS=next top-at-mark-start)");
3428 PrintRegionClosure blk(st);
3429 heap_region_iterate(&blk);
3430 }
3431
3432 void G1CollectedHeap::print_on_error(outputStream* st) const {
3433 this->CollectedHeap::print_on_error(st);
3434
3435 if (_cm != NULL) {
3436 st->cr();
3437 _cm->print_on_error(st);
3438 }
3439 }
3440
3441 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3442 if (G1CollectedHeap::use_parallel_gc_threads()) {
3443 workers()->print_worker_threads_on(st);
3444 }
3445 _cmThread->print_on(st);
3446 st->cr();
3447 _cm->print_worker_threads_on(st);
3448 _cg1r->print_worker_threads_on(st);
3449 }
3450
3451 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3452 if (G1CollectedHeap::use_parallel_gc_threads()) {
3453 workers()->threads_do(tc);
3454 }
3455 tc->do_thread(_cmThread);
3456 _cg1r->threads_do(tc);
3457 }
3458
3459 void G1CollectedHeap::print_tracing_info() const {
3460 // We'll overload this to mean "trace GC pause statistics."
3461 if (TraceGen0Time || TraceGen1Time) {
3462 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3463 // to that.
3464 g1_policy()->print_tracing_info();
3465 }
3466 if (G1SummarizeRSetStats) {
3467 g1_rem_set()->print_summary_info();
3468 }
3469 if (G1SummarizeConcMark) {
3470 concurrent_mark()->print_summary_info();
3471 }
3472 g1_policy()->print_yg_surv_rate_info();
3473 SpecializationStats::print();
3474 }
3475
3476 #ifndef PRODUCT
3477 // Helpful for debugging RSet issues.
3478
3479 class PrintRSetsClosure : public HeapRegionClosure {
3480 private:
3481 const char* _msg;
3482 size_t _occupied_sum;
3483
3484 public:
3485 bool doHeapRegion(HeapRegion* r) {
3486 HeapRegionRemSet* hrrs = r->rem_set();
3487 size_t occupied = hrrs->occupied();
3488 _occupied_sum += occupied;
3489
3490 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3491 HR_FORMAT_PARAMS(r));
3492 if (occupied == 0) {
3493 gclog_or_tty->print_cr(" RSet is empty");
3494 } else {
3495 hrrs->print();
3496 }
3497 gclog_or_tty->print_cr("----------");
3498 return false;
3499 }
3500
3501 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3502 gclog_or_tty->cr();
3503 gclog_or_tty->print_cr("========================================");
3504 gclog_or_tty->print_cr(msg);
3505 gclog_or_tty->cr();
3506 }
3507
3508 ~PrintRSetsClosure() {
3509 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3510 gclog_or_tty->print_cr("========================================");
3511 gclog_or_tty->cr();
3512 }
3513 };
3514
3515 void G1CollectedHeap::print_cset_rsets() {
3516 PrintRSetsClosure cl("Printing CSet RSets");
3517 collection_set_iterate(&cl);
3518 }
3519
3520 void G1CollectedHeap::print_all_rsets() {
3521 PrintRSetsClosure cl("Printing All RSets");;
3522 heap_region_iterate(&cl);
3523 }
3524 #endif // PRODUCT
3525
3526 G1CollectedHeap* G1CollectedHeap::heap() {
3527 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3528 "not a garbage-first heap");
3529 return _g1h;
3530 }
3531
3532 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3533 // always_do_update_barrier = false;
3534 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3535 // Call allocation profiler
3536 AllocationProfiler::iterate_since_last_gc();
3537 // Fill TLAB's and such
3538 ensure_parsability(true);
3539 }
3540
3541 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3542 // FIXME: what is this about?
3543 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3544 // is set.
3545 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3546 "derived pointer present"));
3547 // always_do_update_barrier = true;
3548
3549 // We have just completed a GC. Update the soft reference
3550 // policy with the new heap occupancy
3551 Universe::update_heap_info_at_gc();
3552 }
3553
3554 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3555 unsigned int gc_count_before,
3556 bool* succeeded) {
3557 assert_heap_not_locked_and_not_at_safepoint();
3558 g1_policy()->record_stop_world_start();
3559 VM_G1IncCollectionPause op(gc_count_before,
3560 word_size,
3561 false, /* should_initiate_conc_mark */
3562 g1_policy()->max_pause_time_ms(),
3563 GCCause::_g1_inc_collection_pause);
3564 VMThread::execute(&op);
3565
3566 HeapWord* result = op.result();
3567 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3568 assert(result == NULL || ret_succeeded,
3569 "the result should be NULL if the VM did not succeed");
3570 *succeeded = ret_succeeded;
3571
3572 assert_heap_not_locked();
3573 return result;
3574 }
3575
3576 void
3577 G1CollectedHeap::doConcurrentMark() {
3578 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3579 if (!_cmThread->in_progress()) {
3580 _cmThread->set_started();
3581 CGC_lock->notify();
3582 }
3583 }
3584
3585 size_t G1CollectedHeap::pending_card_num() {
3586 size_t extra_cards = 0;
3587 JavaThread *curr = Threads::first();
3588 while (curr != NULL) {
3589 DirtyCardQueue& dcq = curr->dirty_card_queue();
3590 extra_cards += dcq.size();
3591 curr = curr->next();
3592 }
3593 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3594 size_t buffer_size = dcqs.buffer_size();
3595 size_t buffer_num = dcqs.completed_buffers_num();
3596
3597 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes
3598 // in bytes - not the number of 'entries'. We need to convert
3599 // into a number of cards.
3600 return (buffer_size * buffer_num + extra_cards) / oopSize;
3601 }
3602
3603 size_t G1CollectedHeap::cards_scanned() {
3604 return g1_rem_set()->cardsScanned();
3605 }
3606
3607 void
3608 G1CollectedHeap::setup_surviving_young_words() {
3609 assert(_surviving_young_words == NULL, "pre-condition");
3610 uint array_length = g1_policy()->young_cset_region_length();
3611 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC);
3612 if (_surviving_young_words == NULL) {
3613 vm_exit_out_of_memory(sizeof(size_t) * array_length, OOM_MALLOC_ERROR,
3614 "Not enough space for young surv words summary.");
3615 }
3616 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t));
3617 #ifdef ASSERT
3618 for (uint i = 0; i < array_length; ++i) {
3619 assert( _surviving_young_words[i] == 0, "memset above" );
3620 }
3621 #endif // !ASSERT
3622 }
3623
3624 void
3625 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3626 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3627 uint array_length = g1_policy()->young_cset_region_length();
3628 for (uint i = 0; i < array_length; ++i) {
3629 _surviving_young_words[i] += surv_young_words[i];
3630 }
3631 }
3632
3633 void
3634 G1CollectedHeap::cleanup_surviving_young_words() {
3635 guarantee( _surviving_young_words != NULL, "pre-condition" );
3636 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC);
3637 _surviving_young_words = NULL;
3638 }
3639
3640 #ifdef ASSERT
3641 class VerifyCSetClosure: public HeapRegionClosure {
3642 public:
3643 bool doHeapRegion(HeapRegion* hr) {
3644 // Here we check that the CSet region's RSet is ready for parallel
3645 // iteration. The fields that we'll verify are only manipulated
3646 // when the region is part of a CSet and is collected. Afterwards,
3647 // we reset these fields when we clear the region's RSet (when the
3648 // region is freed) so they are ready when the region is
3649 // re-allocated. The only exception to this is if there's an
3650 // evacuation failure and instead of freeing the region we leave
3651 // it in the heap. In that case, we reset these fields during
3652 // evacuation failure handling.
3653 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3654
3655 // Here's a good place to add any other checks we'd like to
3656 // perform on CSet regions.
3657 return false;
3658 }
3659 };
3660 #endif // ASSERT
3661
3662 #if TASKQUEUE_STATS
3663 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3664 st->print_raw_cr("GC Task Stats");
3665 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3666 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3667 }
3668
3669 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3670 print_taskqueue_stats_hdr(st);
3671
3672 TaskQueueStats totals;
3673 const int n = workers() != NULL ? workers()->total_workers() : 1;
3674 for (int i = 0; i < n; ++i) {
3675 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3676 totals += task_queue(i)->stats;
3677 }
3678 st->print_raw("tot "); totals.print(st); st->cr();
3679
3680 DEBUG_ONLY(totals.verify());
3681 }
3682
3683 void G1CollectedHeap::reset_taskqueue_stats() {
3684 const int n = workers() != NULL ? workers()->total_workers() : 1;
3685 for (int i = 0; i < n; ++i) {
3686 task_queue(i)->stats.reset();
3687 }
3688 }
3689 #endif // TASKQUEUE_STATS
3690
3691 void G1CollectedHeap::log_gc_header() {
3692 if (!G1Log::fine()) {
3693 return;
3694 }
3695
3696 gclog_or_tty->date_stamp(PrintGCDateStamps);
3697 gclog_or_tty->stamp(PrintGCTimeStamps);
3698
3699 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause())
3700 .append(g1_policy()->gcs_are_young() ? "(young)" : "(mixed)")
3701 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : "");
3702
3703 gclog_or_tty->print("[%s", (const char*)gc_cause_str);
3704 }
3705
3706 void G1CollectedHeap::log_gc_footer(double pause_time_sec) {
3707 if (!G1Log::fine()) {
3708 return;
3709 }
3710
3711 if (G1Log::finer()) {
3712 if (evacuation_failed()) {
3713 gclog_or_tty->print(" (to-space exhausted)");
3714 }
3715 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3716 g1_policy()->phase_times()->note_gc_end();
3717 g1_policy()->phase_times()->print(pause_time_sec);
3718 g1_policy()->print_detailed_heap_transition();
3719 } else {
3720 if (evacuation_failed()) {
3721 gclog_or_tty->print("--");
3722 }
3723 g1_policy()->print_heap_transition();
3724 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3725 }
3726 gclog_or_tty->flush();
3727 }
3728
3729 bool
3730 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3731 assert_at_safepoint(true /* should_be_vm_thread */);
3732 guarantee(!is_gc_active(), "collection is not reentrant");
3733
3734 if (GC_locker::check_active_before_gc()) {
3735 return false;
3736 }
3737
3738 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3739 ResourceMark rm;
3740
3741 print_heap_before_gc();
3742
3743 HRSPhaseSetter x(HRSPhaseEvacuation);
3744 verify_region_sets_optional();
3745 verify_dirty_young_regions();
3746
3747 // This call will decide whether this pause is an initial-mark
3748 // pause. If it is, during_initial_mark_pause() will return true
3749 // for the duration of this pause.
3750 g1_policy()->decide_on_conc_mark_initiation();
3751
3752 // We do not allow initial-mark to be piggy-backed on a mixed GC.
3753 assert(!g1_policy()->during_initial_mark_pause() ||
3754 g1_policy()->gcs_are_young(), "sanity");
3755
3756 // We also do not allow mixed GCs during marking.
3757 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity");
3758
3759 // Record whether this pause is an initial mark. When the current
3760 // thread has completed its logging output and it's safe to signal
3761 // the CM thread, the flag's value in the policy has been reset.
3762 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause();
3763
3764 // Inner scope for scope based logging, timers, and stats collection
3765 {
3766 if (g1_policy()->during_initial_mark_pause()) {
3767 // We are about to start a marking cycle, so we increment the
3768 // full collection counter.
3769 increment_old_marking_cycles_started();
3770 }
3771 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
3772
3773 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3774 workers()->active_workers() : 1);
3775 double pause_start_sec = os::elapsedTime();
3776 g1_policy()->phase_times()->note_gc_start(active_workers);
3777 log_gc_header();
3778
3779 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3780 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3781
3782 // If the secondary_free_list is not empty, append it to the
3783 // free_list. No need to wait for the cleanup operation to finish;
3784 // the region allocation code will check the secondary_free_list
3785 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3786 // set, skip this step so that the region allocation code has to
3787 // get entries from the secondary_free_list.
3788 if (!G1StressConcRegionFreeing) {
3789 append_secondary_free_list_if_not_empty_with_lock();
3790 }
3791
3792 assert(check_young_list_well_formed(),
3793 "young list should be well formed");
3794
3795 // Don't dynamically change the number of GC threads this early. A value of
3796 // 0 is used to indicate serial work. When parallel work is done,
3797 // it will be set.
3798
3799 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3800 IsGCActiveMark x;
3801
3802 gc_prologue(false);
3803 increment_total_collections(false /* full gc */);
3804 increment_gc_time_stamp();
3805
3806 verify_before_gc();
3807
3808 COMPILER2_PRESENT(DerivedPointerTable::clear());
3809
3810 // Please see comment in g1CollectedHeap.hpp and
3811 // G1CollectedHeap::ref_processing_init() to see how
3812 // reference processing currently works in G1.
3813
3814 // Enable discovery in the STW reference processor
3815 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3816 true /*verify_no_refs*/);
3817
3818 {
3819 // We want to temporarily turn off discovery by the
3820 // CM ref processor, if necessary, and turn it back on
3821 // on again later if we do. Using a scoped
3822 // NoRefDiscovery object will do this.
3823 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3824
3825 // Forget the current alloc region (we might even choose it to be part
3826 // of the collection set!).
3827 release_mutator_alloc_region();
3828
3829 // We should call this after we retire the mutator alloc
3830 // region(s) so that all the ALLOC / RETIRE events are generated
3831 // before the start GC event.
3832 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3833
3834 // This timing is only used by the ergonomics to handle our pause target.
3835 // It is unclear why this should not include the full pause. We will
3836 // investigate this in CR 7178365.
3837 //
3838 // Preserving the old comment here if that helps the investigation:
3839 //
3840 // The elapsed time induced by the start time below deliberately elides
3841 // the possible verification above.
3842 double sample_start_time_sec = os::elapsedTime();
3843
3844 #if YOUNG_LIST_VERBOSE
3845 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3846 _young_list->print();
3847 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3848 #endif // YOUNG_LIST_VERBOSE
3849
3850 g1_policy()->record_collection_pause_start(sample_start_time_sec);
3851
3852 double scan_wait_start = os::elapsedTime();
3853 // We have to wait until the CM threads finish scanning the
3854 // root regions as it's the only way to ensure that all the
3855 // objects on them have been correctly scanned before we start
3856 // moving them during the GC.
3857 bool waited = _cm->root_regions()->wait_until_scan_finished();
3858 double wait_time_ms = 0.0;
3859 if (waited) {
3860 double scan_wait_end = os::elapsedTime();
3861 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0;
3862 }
3863 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms);
3864
3865 #if YOUNG_LIST_VERBOSE
3866 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3867 _young_list->print();
3868 #endif // YOUNG_LIST_VERBOSE
3869
3870 if (g1_policy()->during_initial_mark_pause()) {
3871 concurrent_mark()->checkpointRootsInitialPre();
3872 }
3873
3874 #if YOUNG_LIST_VERBOSE
3875 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3876 _young_list->print();
3877 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3878 #endif // YOUNG_LIST_VERBOSE
3879
3880 g1_policy()->finalize_cset(target_pause_time_ms);
3881
3882 _cm->note_start_of_gc();
3883 // We should not verify the per-thread SATB buffers given that
3884 // we have not filtered them yet (we'll do so during the
3885 // GC). We also call this after finalize_cset() to
3886 // ensure that the CSet has been finalized.
3887 _cm->verify_no_cset_oops(true /* verify_stacks */,
3888 true /* verify_enqueued_buffers */,
3889 false /* verify_thread_buffers */,
3890 true /* verify_fingers */);
3891
3892 if (_hr_printer.is_active()) {
3893 HeapRegion* hr = g1_policy()->collection_set();
3894 while (hr != NULL) {
3895 G1HRPrinter::RegionType type;
3896 if (!hr->is_young()) {
3897 type = G1HRPrinter::Old;
3898 } else if (hr->is_survivor()) {
3899 type = G1HRPrinter::Survivor;
3900 } else {
3901 type = G1HRPrinter::Eden;
3902 }
3903 _hr_printer.cset(hr);
3904 hr = hr->next_in_collection_set();
3905 }
3906 }
3907
3908 #ifdef ASSERT
3909 VerifyCSetClosure cl;
3910 collection_set_iterate(&cl);
3911 #endif // ASSERT
3912
3913 setup_surviving_young_words();
3914
3915 // Initialize the GC alloc regions.
3916 init_gc_alloc_regions();
3917
3918 // Actually do the work...
3919 evacuate_collection_set();
3920
3921 // We do this to mainly verify the per-thread SATB buffers
3922 // (which have been filtered by now) since we didn't verify
3923 // them earlier. No point in re-checking the stacks / enqueued
3924 // buffers given that the CSet has not changed since last time
3925 // we checked.
3926 _cm->verify_no_cset_oops(false /* verify_stacks */,
3927 false /* verify_enqueued_buffers */,
3928 true /* verify_thread_buffers */,
3929 true /* verify_fingers */);
3930
3931 free_collection_set(g1_policy()->collection_set());
3932 g1_policy()->clear_collection_set();
3933
3934 cleanup_surviving_young_words();
3935
3936 // Start a new incremental collection set for the next pause.
3937 g1_policy()->start_incremental_cset_building();
3938
3939 // Clear the _cset_fast_test bitmap in anticipation of adding
3940 // regions to the incremental collection set for the next
3941 // evacuation pause.
3942 clear_cset_fast_test();
3943
3944 _young_list->reset_sampled_info();
3945
3946 // Don't check the whole heap at this point as the
3947 // GC alloc regions from this pause have been tagged
3948 // as survivors and moved on to the survivor list.
3949 // Survivor regions will fail the !is_young() check.
3950 assert(check_young_list_empty(false /* check_heap */),
3951 "young list should be empty");
3952
3953 #if YOUNG_LIST_VERBOSE
3954 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3955 _young_list->print();
3956 #endif // YOUNG_LIST_VERBOSE
3957
3958 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3959 _young_list->first_survivor_region(),
3960 _young_list->last_survivor_region());
3961
3962 _young_list->reset_auxilary_lists();
3963
3964 if (evacuation_failed()) {
3965 _summary_bytes_used = recalculate_used();
3966 } else {
3967 // The "used" of the the collection set have already been subtracted
3968 // when they were freed. Add in the bytes evacuated.
3969 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3970 }
3971
3972 if (g1_policy()->during_initial_mark_pause()) {
3973 // We have to do this before we notify the CM threads that
3974 // they can start working to make sure that all the
3975 // appropriate initialization is done on the CM object.
3976 concurrent_mark()->checkpointRootsInitialPost();
3977 set_marking_started();
3978 // Note that we don't actually trigger the CM thread at
3979 // this point. We do that later when we're sure that
3980 // the current thread has completed its logging output.
3981 }
3982
3983 allocate_dummy_regions();
3984
3985 #if YOUNG_LIST_VERBOSE
3986 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3987 _young_list->print();
3988 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3989 #endif // YOUNG_LIST_VERBOSE
3990
3991 init_mutator_alloc_region();
3992
3993 {
3994 size_t expand_bytes = g1_policy()->expansion_amount();
3995 if (expand_bytes > 0) {
3996 size_t bytes_before = capacity();
3997 // No need for an ergo verbose message here,
3998 // expansion_amount() does this when it returns a value > 0.
3999 if (!expand(expand_bytes)) {
4000 // We failed to expand the heap so let's verify that
4001 // committed/uncommitted amount match the backing store
4002 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
4003 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
4004 }
4005 }
4006 }
4007
4008 // We redo the verificaiton but now wrt to the new CSet which
4009 // has just got initialized after the previous CSet was freed.
4010 _cm->verify_no_cset_oops(true /* verify_stacks */,
4011 true /* verify_enqueued_buffers */,
4012 true /* verify_thread_buffers */,
4013 true /* verify_fingers */);
4014 _cm->note_end_of_gc();
4015
4016 // This timing is only used by the ergonomics to handle our pause target.
4017 // It is unclear why this should not include the full pause. We will
4018 // investigate this in CR 7178365.
4019 double sample_end_time_sec = os::elapsedTime();
4020 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS;
4021 g1_policy()->record_collection_pause_end(pause_time_ms);
4022
4023 MemoryService::track_memory_usage();
4024
4025 // In prepare_for_verify() below we'll need to scan the deferred
4026 // update buffers to bring the RSets up-to-date if
4027 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
4028 // the update buffers we'll probably need to scan cards on the
4029 // regions we just allocated to (i.e., the GC alloc
4030 // regions). However, during the last GC we called
4031 // set_saved_mark() on all the GC alloc regions, so card
4032 // scanning might skip the [saved_mark_word()...top()] area of
4033 // those regions (i.e., the area we allocated objects into
4034 // during the last GC). But it shouldn't. Given that
4035 // saved_mark_word() is conditional on whether the GC time stamp
4036 // on the region is current or not, by incrementing the GC time
4037 // stamp here we invalidate all the GC time stamps on all the
4038 // regions and saved_mark_word() will simply return top() for
4039 // all the regions. This is a nicer way of ensuring this rather
4040 // than iterating over the regions and fixing them. In fact, the
4041 // GC time stamp increment here also ensures that
4042 // saved_mark_word() will return top() between pauses, i.e.,
4043 // during concurrent refinement. So we don't need the
4044 // is_gc_active() check to decided which top to use when
4045 // scanning cards (see CR 7039627).
4046 increment_gc_time_stamp();
4047
4048 verify_after_gc();
4049
4050 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
4051 ref_processor_stw()->verify_no_references_recorded();
4052
4053 // CM reference discovery will be re-enabled if necessary.
4054 }
4055
4056 // We should do this after we potentially expand the heap so
4057 // that all the COMMIT events are generated before the end GC
4058 // event, and after we retire the GC alloc regions so that all
4059 // RETIRE events are generated before the end GC event.
4060 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
4061
4062 if (mark_in_progress()) {
4063 concurrent_mark()->update_g1_committed();
4064 }
4065
4066 #ifdef TRACESPINNING
4067 ParallelTaskTerminator::print_termination_counts();
4068 #endif
4069
4070 gc_epilogue(false);
4071 }
4072
4073 // Print the remainder of the GC log output.
4074 log_gc_footer(os::elapsedTime() - pause_start_sec);
4075
4076 // It is not yet to safe to tell the concurrent mark to
4077 // start as we have some optional output below. We don't want the
4078 // output from the concurrent mark thread interfering with this
4079 // logging output either.
4080
4081 _hrs.verify_optional();
4082 verify_region_sets_optional();
4083
4084 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
4085 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
4086
4087 print_heap_after_gc();
4088
4089 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
4090 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
4091 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
4092 // before any GC notifications are raised.
4093 g1mm()->update_sizes();
4094 }
4095
4096 if (G1SummarizeRSetStats &&
4097 (G1SummarizeRSetStatsPeriod > 0) &&
4098 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
4099 g1_rem_set()->print_summary_info();
4100 }
4101
4102 // It should now be safe to tell the concurrent mark thread to start
4103 // without its logging output interfering with the logging output
4104 // that came from the pause.
4105
4106 if (should_start_conc_mark) {
4107 // CAUTION: after the doConcurrentMark() call below,
4108 // the concurrent marking thread(s) could be running
4109 // concurrently with us. Make sure that anything after
4110 // this point does not assume that we are the only GC thread
4111 // running. Note: of course, the actual marking work will
4112 // not start until the safepoint itself is released in
4113 // ConcurrentGCThread::safepoint_desynchronize().
4114 doConcurrentMark();
4115 }
4116
4117 return true;
4118 }
4119
4120 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
4121 {
4122 size_t gclab_word_size;
4123 switch (purpose) {
4124 case GCAllocForSurvived:
4125 gclab_word_size = _survivor_plab_stats.desired_plab_sz();
4126 break;
4127 case GCAllocForTenured:
4128 gclab_word_size = _old_plab_stats.desired_plab_sz();
4129 break;
4130 default:
4131 assert(false, "unknown GCAllocPurpose");
4132 gclab_word_size = _old_plab_stats.desired_plab_sz();
4133 break;
4134 }
4135
4136 // Prevent humongous PLAB sizes for two reasons:
4137 // * PLABs are allocated using a similar paths as oops, but should
4138 // never be in a humongous region
4139 // * Allowing humongous PLABs needlessly churns the region free lists
4140 return MIN2(_humongous_object_threshold_in_words, gclab_word_size);
4141 }
4142
4143 void G1CollectedHeap::init_mutator_alloc_region() {
4144 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
4145 _mutator_alloc_region.init();
4146 }
4147
4148 void G1CollectedHeap::release_mutator_alloc_region() {
4149 _mutator_alloc_region.release();
4150 assert(_mutator_alloc_region.get() == NULL, "post-condition");
4151 }
4152
4153 void G1CollectedHeap::init_gc_alloc_regions() {
4154 assert_at_safepoint(true /* should_be_vm_thread */);
4155
4156 _survivor_gc_alloc_region.init();
4157 _old_gc_alloc_region.init();
4158 HeapRegion* retained_region = _retained_old_gc_alloc_region;
4159 _retained_old_gc_alloc_region = NULL;
4160
4161 // We will discard the current GC alloc region if:
4162 // a) it's in the collection set (it can happen!),
4163 // b) it's already full (no point in using it),
4164 // c) it's empty (this means that it was emptied during
4165 // a cleanup and it should be on the free list now), or
4166 // d) it's humongous (this means that it was emptied
4167 // during a cleanup and was added to the free list, but
4168 // has been subseqently used to allocate a humongous
4169 // object that may be less than the region size).
4170 if (retained_region != NULL &&
4171 !retained_region->in_collection_set() &&
4172 !(retained_region->top() == retained_region->end()) &&
4173 !retained_region->is_empty() &&
4174 !retained_region->isHumongous()) {
4175 retained_region->set_saved_mark();
4176 // The retained region was added to the old region set when it was
4177 // retired. We have to remove it now, since we don't allow regions
4178 // we allocate to in the region sets. We'll re-add it later, when
4179 // it's retired again.
4180 _old_set.remove(retained_region);
4181 bool during_im = g1_policy()->during_initial_mark_pause();
4182 retained_region->note_start_of_copying(during_im);
4183 _old_gc_alloc_region.set(retained_region);
4184 _hr_printer.reuse(retained_region);
4185 }
4186 }
4187
4188 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers) {
4189 _survivor_gc_alloc_region.release();
4190 // If we have an old GC alloc region to release, we'll save it in
4191 // _retained_old_gc_alloc_region. If we don't
4192 // _retained_old_gc_alloc_region will become NULL. This is what we
4193 // want either way so no reason to check explicitly for either
4194 // condition.
4195 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
4196
4197 if (ResizePLAB) {
4198 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4199 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4200 }
4201 }
4202
4203 void G1CollectedHeap::abandon_gc_alloc_regions() {
4204 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
4205 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
4206 _retained_old_gc_alloc_region = NULL;
4207 }
4208
4209 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
4210 _drain_in_progress = false;
4211 set_evac_failure_closure(cl);
4212 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4213 }
4214
4215 void G1CollectedHeap::finalize_for_evac_failure() {
4216 assert(_evac_failure_scan_stack != NULL &&
4217 _evac_failure_scan_stack->length() == 0,
4218 "Postcondition");
4219 assert(!_drain_in_progress, "Postcondition");
4220 delete _evac_failure_scan_stack;
4221 _evac_failure_scan_stack = NULL;
4222 }
4223
4224 void G1CollectedHeap::remove_self_forwarding_pointers() {
4225 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4226
4227 G1ParRemoveSelfForwardPtrsTask rsfp_task(this);
4228
4229 if (G1CollectedHeap::use_parallel_gc_threads()) {
4230 set_par_threads();
4231 workers()->run_task(&rsfp_task);
4232 set_par_threads(0);
4233 } else {
4234 rsfp_task.work(0);
4235 }
4236
4237 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity");
4238
4239 // Reset the claim values in the regions in the collection set.
4240 reset_cset_heap_region_claim_values();
4241
4242 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4243
4244 // Now restore saved marks, if any.
4245 assert(_objs_with_preserved_marks.size() ==
4246 _preserved_marks_of_objs.size(), "Both or none.");
4247 while (!_objs_with_preserved_marks.is_empty()) {
4248 oop obj = _objs_with_preserved_marks.pop();
4249 markOop m = _preserved_marks_of_objs.pop();
4250 obj->set_mark(m);
4251 }
4252 _objs_with_preserved_marks.clear(true);
4253 _preserved_marks_of_objs.clear(true);
4254 }
4255
4256 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
4257 _evac_failure_scan_stack->push(obj);
4258 }
4259
4260 void G1CollectedHeap::drain_evac_failure_scan_stack() {
4261 assert(_evac_failure_scan_stack != NULL, "precondition");
4262
4263 while (_evac_failure_scan_stack->length() > 0) {
4264 oop obj = _evac_failure_scan_stack->pop();
4265 _evac_failure_closure->set_region(heap_region_containing(obj));
4266 obj->oop_iterate_backwards(_evac_failure_closure);
4267 }
4268 }
4269
4270 oop
4271 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4272 oop old) {
4273 assert(obj_in_cs(old),
4274 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4275 (HeapWord*) old));
4276 markOop m = old->mark();
4277 oop forward_ptr = old->forward_to_atomic(old);
4278 if (forward_ptr == NULL) {
4279 // Forward-to-self succeeded.
4280
4281 if (_evac_failure_closure != cl) {
4282 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4283 assert(!_drain_in_progress,
4284 "Should only be true while someone holds the lock.");
4285 // Set the global evac-failure closure to the current thread's.
4286 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4287 set_evac_failure_closure(cl);
4288 // Now do the common part.
4289 handle_evacuation_failure_common(old, m);
4290 // Reset to NULL.
4291 set_evac_failure_closure(NULL);
4292 } else {
4293 // The lock is already held, and this is recursive.
4294 assert(_drain_in_progress, "This should only be the recursive case.");
4295 handle_evacuation_failure_common(old, m);
4296 }
4297 return old;
4298 } else {
4299 // Forward-to-self failed. Either someone else managed to allocate
4300 // space for this object (old != forward_ptr) or they beat us in
4301 // self-forwarding it (old == forward_ptr).
4302 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4303 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4304 "should not be in the CSet",
4305 (HeapWord*) old, (HeapWord*) forward_ptr));
4306 return forward_ptr;
4307 }
4308 }
4309
4310 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4311 set_evacuation_failed(true);
4312
4313 preserve_mark_if_necessary(old, m);
4314
4315 HeapRegion* r = heap_region_containing(old);
4316 if (!r->evacuation_failed()) {
4317 r->set_evacuation_failed(true);
4318 _hr_printer.evac_failure(r);
4319 }
4320
4321 push_on_evac_failure_scan_stack(old);
4322
4323 if (!_drain_in_progress) {
4324 // prevent recursion in copy_to_survivor_space()
4325 _drain_in_progress = true;
4326 drain_evac_failure_scan_stack();
4327 _drain_in_progress = false;
4328 }
4329 }
4330
4331 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4332 assert(evacuation_failed(), "Oversaving!");
4333 // We want to call the "for_promotion_failure" version only in the
4334 // case of a promotion failure.
4335 if (m->must_be_preserved_for_promotion_failure(obj)) {
4336 _objs_with_preserved_marks.push(obj);
4337 _preserved_marks_of_objs.push(m);
4338 }
4339 }
4340
4341 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4342 size_t word_size) {
4343 if (purpose == GCAllocForSurvived) {
4344 HeapWord* result = survivor_attempt_allocation(word_size);
4345 if (result != NULL) {
4346 return result;
4347 } else {
4348 // Let's try to allocate in the old gen in case we can fit the
4349 // object there.
4350 return old_attempt_allocation(word_size);
4351 }
4352 } else {
4353 assert(purpose == GCAllocForTenured, "sanity");
4354 HeapWord* result = old_attempt_allocation(word_size);
4355 if (result != NULL) {
4356 return result;
4357 } else {
4358 // Let's try to allocate in the survivors in case we can fit the
4359 // object there.
4360 return survivor_attempt_allocation(word_size);
4361 }
4362 }
4363
4364 ShouldNotReachHere();
4365 // Trying to keep some compilers happy.
4366 return NULL;
4367 }
4368
4369 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4370 ParGCAllocBuffer(gclab_word_size), _retired(false) { }
4371
4372 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num)
4373 : _g1h(g1h),
4374 _refs(g1h->task_queue(queue_num)),
4375 _dcq(&g1h->dirty_card_queue_set()),
4376 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4377 _g1_rem(g1h->g1_rem_set()),
4378 _hash_seed(17), _queue_num(queue_num),
4379 _term_attempts(0),
4380 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4381 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4382 _age_table(false),
4383 _strong_roots_time(0), _term_time(0),
4384 _alloc_buffer_waste(0), _undo_waste(0) {
4385 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4386 // we "sacrifice" entry 0 to keep track of surviving bytes for
4387 // non-young regions (where the age is -1)
4388 // We also add a few elements at the beginning and at the end in
4389 // an attempt to eliminate cache contention
4390 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
4391 uint array_length = PADDING_ELEM_NUM +
4392 real_length +
4393 PADDING_ELEM_NUM;
4394 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
4395 if (_surviving_young_words_base == NULL)
4396 vm_exit_out_of_memory(array_length * sizeof(size_t), OOM_MALLOC_ERROR,
4397 "Not enough space for young surv histo.");
4398 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4399 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t));
4400
4401 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4402 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4403
4404 _start = os::elapsedTime();
4405 }
4406
4407 void
4408 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4409 {
4410 st->print_raw_cr("GC Termination Stats");
4411 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4412 " ------waste (KiB)------");
4413 st->print_raw_cr("thr ms ms % ms % attempts"
4414 " total alloc undo");
4415 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4416 " ------- ------- -------");
4417 }
4418
4419 void
4420 G1ParScanThreadState::print_termination_stats(int i,
4421 outputStream* const st) const
4422 {
4423 const double elapsed_ms = elapsed_time() * 1000.0;
4424 const double s_roots_ms = strong_roots_time() * 1000.0;
4425 const double term_ms = term_time() * 1000.0;
4426 st->print_cr("%3d %9.2f %9.2f %6.2f "
4427 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4428 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4429 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4430 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4431 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4432 alloc_buffer_waste() * HeapWordSize / K,
4433 undo_waste() * HeapWordSize / K);
4434 }
4435
4436 #ifdef ASSERT
4437 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4438 assert(ref != NULL, "invariant");
4439 assert(UseCompressedOops, "sanity");
4440 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4441 oop p = oopDesc::load_decode_heap_oop(ref);
4442 assert(_g1h->is_in_g1_reserved(p),
4443 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4444 return true;
4445 }
4446
4447 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4448 assert(ref != NULL, "invariant");
4449 if (has_partial_array_mask(ref)) {
4450 // Must be in the collection set--it's already been copied.
4451 oop p = clear_partial_array_mask(ref);
4452 assert(_g1h->obj_in_cs(p),
4453 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4454 } else {
4455 oop p = oopDesc::load_decode_heap_oop(ref);
4456 assert(_g1h->is_in_g1_reserved(p),
4457 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4458 }
4459 return true;
4460 }
4461
4462 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4463 if (ref.is_narrow()) {
4464 return verify_ref((narrowOop*) ref);
4465 } else {
4466 return verify_ref((oop*) ref);
4467 }
4468 }
4469 #endif // ASSERT
4470
4471 void G1ParScanThreadState::trim_queue() {
4472 assert(_evac_cl != NULL, "not set");
4473 assert(_evac_failure_cl != NULL, "not set");
4474 assert(_partial_scan_cl != NULL, "not set");
4475
4476 StarTask ref;
4477 do {
4478 // Drain the overflow stack first, so other threads can steal.
4479 while (refs()->pop_overflow(ref)) {
4480 deal_with_reference(ref);
4481 }
4482
4483 while (refs()->pop_local(ref)) {
4484 deal_with_reference(ref);
4485 }
4486 } while (!refs()->is_empty());
4487 }
4488
4489 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1,
4490 G1ParScanThreadState* par_scan_state) :
4491 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4492 _par_scan_state(par_scan_state),
4493 _worker_id(par_scan_state->queue_num()),
4494 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4495 _mark_in_progress(_g1->mark_in_progress()) { }
4496
4497 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4498 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) {
4499 #ifdef ASSERT
4500 HeapRegion* hr = _g1->heap_region_containing(obj);
4501 assert(hr != NULL, "sanity");
4502 assert(!hr->in_collection_set(), "should not mark objects in the CSet");
4503 #endif // ASSERT
4504
4505 // We know that the object is not moving so it's safe to read its size.
4506 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id);
4507 }
4508
4509 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4510 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4511 ::mark_forwarded_object(oop from_obj, oop to_obj) {
4512 #ifdef ASSERT
4513 assert(from_obj->is_forwarded(), "from obj should be forwarded");
4514 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee");
4515 assert(from_obj != to_obj, "should not be self-forwarded");
4516
4517 HeapRegion* from_hr = _g1->heap_region_containing(from_obj);
4518 assert(from_hr != NULL, "sanity");
4519 assert(from_hr->in_collection_set(), "from obj should be in the CSet");
4520
4521 HeapRegion* to_hr = _g1->heap_region_containing(to_obj);
4522 assert(to_hr != NULL, "sanity");
4523 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet");
4524 #endif // ASSERT
4525
4526 // The object might be in the process of being copied by another
4527 // worker so we cannot trust that its to-space image is
4528 // well-formed. So we have to read its size from its from-space
4529 // image which we know should not be changing.
4530 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id);
4531 }
4532
4533 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4534 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4535 ::copy_to_survivor_space(oop old) {
4536 size_t word_sz = old->size();
4537 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4538 // +1 to make the -1 indexes valid...
4539 int young_index = from_region->young_index_in_cset()+1;
4540 assert( (from_region->is_young() && young_index > 0) ||
4541 (!from_region->is_young() && young_index == 0), "invariant" );
4542 G1CollectorPolicy* g1p = _g1->g1_policy();
4543 markOop m = old->mark();
4544 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
4545 : m->age();
4546 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4547 word_sz);
4548 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4549 #ifndef PRODUCT
4550 // Should this evacuation fail?
4551 if (_g1->evacuation_should_fail()) {
4552 if (obj_ptr != NULL) {
4553 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4554 obj_ptr = NULL;
4555 }
4556 }
4557 #endif // !PRODUCT
4558
4559 if (obj_ptr == NULL) {
4560 // This will either forward-to-self, or detect that someone else has
4561 // installed a forwarding pointer.
4562 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4563 return _g1->handle_evacuation_failure_par(cl, old);
4564 }
4565
4566 oop obj = oop(obj_ptr);
4567
4568 // We're going to allocate linearly, so might as well prefetch ahead.
4569 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4570
4571 oop forward_ptr = old->forward_to_atomic(obj);
4572 if (forward_ptr == NULL) {
4573 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4574 if (g1p->track_object_age(alloc_purpose)) {
4575 // We could simply do obj->incr_age(). However, this causes a
4576 // performance issue. obj->incr_age() will first check whether
4577 // the object has a displaced mark by checking its mark word;
4578 // getting the mark word from the new location of the object
4579 // stalls. So, given that we already have the mark word and we
4580 // are about to install it anyway, it's better to increase the
4581 // age on the mark word, when the object does not have a
4582 // displaced mark word. We're not expecting many objects to have
4583 // a displaced marked word, so that case is not optimized
4584 // further (it could be...) and we simply call obj->incr_age().
4585
4586 if (m->has_displaced_mark_helper()) {
4587 // in this case, we have to install the mark word first,
4588 // otherwise obj looks to be forwarded (the old mark word,
4589 // which contains the forward pointer, was copied)
4590 obj->set_mark(m);
4591 obj->incr_age();
4592 } else {
4593 m = m->incr_age();
4594 obj->set_mark(m);
4595 }
4596 _par_scan_state->age_table()->add(obj, word_sz);
4597 } else {
4598 obj->set_mark(m);
4599 }
4600
4601 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4602 surv_young_words[young_index] += word_sz;
4603
4604 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4605 // We keep track of the next start index in the length field of
4606 // the to-space object. The actual length can be found in the
4607 // length field of the from-space object.
4608 arrayOop(obj)->set_length(0);
4609 oop* old_p = set_partial_array_mask(old);
4610 _par_scan_state->push_on_queue(old_p);
4611 } else {
4612 // No point in using the slower heap_region_containing() method,
4613 // given that we know obj is in the heap.
4614 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4615 obj->oop_iterate_backwards(&_scanner);
4616 }
4617 } else {
4618 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4619 obj = forward_ptr;
4620 }
4621 return obj;
4622 }
4623
4624 template <class T>
4625 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) {
4626 if (_g1->heap_region_containing_raw(new_obj)->is_young()) {
4627 _scanned_klass->record_modified_oops();
4628 }
4629 }
4630
4631 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4632 template <class T>
4633 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4634 ::do_oop_work(T* p) {
4635 oop obj = oopDesc::load_decode_heap_oop(p);
4636 assert(barrier != G1BarrierRS || obj != NULL,
4637 "Precondition: G1BarrierRS implies obj is non-NULL");
4638
4639 assert(_worker_id == _par_scan_state->queue_num(), "sanity");
4640
4641 // here the null check is implicit in the cset_fast_test() test
4642 if (_g1->in_cset_fast_test(obj)) {
4643 oop forwardee;
4644 if (obj->is_forwarded()) {
4645 forwardee = obj->forwardee();
4646 } else {
4647 forwardee = copy_to_survivor_space(obj);
4648 }
4649 assert(forwardee != NULL, "forwardee should not be NULL");
4650 oopDesc::encode_store_heap_oop(p, forwardee);
4651 if (do_mark_object && forwardee != obj) {
4652 // If the object is self-forwarded we don't need to explicitly
4653 // mark it, the evacuation failure protocol will do so.
4654 mark_forwarded_object(obj, forwardee);
4655 }
4656
4657 // When scanning the RS, we only care about objs in CS.
4658 if (barrier == G1BarrierRS) {
4659 _par_scan_state->update_rs(_from, p, _worker_id);
4660 } else if (barrier == G1BarrierKlass) {
4661 do_klass_barrier(p, forwardee);
4662 }
4663 } else {
4664 // The object is not in collection set. If we're a root scanning
4665 // closure during an initial mark pause (i.e. do_mark_object will
4666 // be true) then attempt to mark the object.
4667 if (do_mark_object && _g1->is_in_g1_reserved(obj)) {
4668 mark_object(obj);
4669 }
4670 }
4671
4672 if (barrier == G1BarrierEvac && obj != NULL) {
4673 _par_scan_state->update_rs(_from, p, _worker_id);
4674 }
4675
4676 if (do_gen_barrier && obj != NULL) {
4677 par_do_barrier(p);
4678 }
4679 }
4680
4681 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4682 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4683
4684 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4685 assert(has_partial_array_mask(p), "invariant");
4686 oop from_obj = clear_partial_array_mask(p);
4687
4688 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap.");
4689 assert(from_obj->is_objArray(), "must be obj array");
4690 objArrayOop from_obj_array = objArrayOop(from_obj);
4691 // The from-space object contains the real length.
4692 int length = from_obj_array->length();
4693
4694 assert(from_obj->is_forwarded(), "must be forwarded");
4695 oop to_obj = from_obj->forwardee();
4696 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
4697 objArrayOop to_obj_array = objArrayOop(to_obj);
4698 // We keep track of the next start index in the length field of the
4699 // to-space object.
4700 int next_index = to_obj_array->length();
4701 assert(0 <= next_index && next_index < length,
4702 err_msg("invariant, next index: %d, length: %d", next_index, length));
4703
4704 int start = next_index;
4705 int end = length;
4706 int remainder = end - start;
4707 // We'll try not to push a range that's smaller than ParGCArrayScanChunk.
4708 if (remainder > 2 * ParGCArrayScanChunk) {
4709 end = start + ParGCArrayScanChunk;
4710 to_obj_array->set_length(end);
4711 // Push the remainder before we process the range in case another
4712 // worker has run out of things to do and can steal it.
4713 oop* from_obj_p = set_partial_array_mask(from_obj);
4714 _par_scan_state->push_on_queue(from_obj_p);
4715 } else {
4716 assert(length == end, "sanity");
4717 // We'll process the final range for this object. Restore the length
4718 // so that the heap remains parsable in case of evacuation failure.
4719 to_obj_array->set_length(end);
4720 }
4721 _scanner.set_region(_g1->heap_region_containing_raw(to_obj));
4722 // Process indexes [start,end). It will also process the header
4723 // along with the first chunk (i.e., the chunk with start == 0).
4724 // Note that at this point the length field of to_obj_array is not
4725 // correct given that we are using it to keep track of the next
4726 // start index. oop_iterate_range() (thankfully!) ignores the length
4727 // field and only relies on the start / end parameters. It does
4728 // however return the size of the object which will be incorrect. So
4729 // we have to ignore it even if we wanted to use it.
4730 to_obj_array->oop_iterate_range(&_scanner, start, end);
4731 }
4732
4733 class G1ParEvacuateFollowersClosure : public VoidClosure {
4734 protected:
4735 G1CollectedHeap* _g1h;
4736 G1ParScanThreadState* _par_scan_state;
4737 RefToScanQueueSet* _queues;
4738 ParallelTaskTerminator* _terminator;
4739
4740 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4741 RefToScanQueueSet* queues() { return _queues; }
4742 ParallelTaskTerminator* terminator() { return _terminator; }
4743
4744 public:
4745 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4746 G1ParScanThreadState* par_scan_state,
4747 RefToScanQueueSet* queues,
4748 ParallelTaskTerminator* terminator)
4749 : _g1h(g1h), _par_scan_state(par_scan_state),
4750 _queues(queues), _terminator(terminator) {}
4751
4752 void do_void();
4753
4754 private:
4755 inline bool offer_termination();
4756 };
4757
4758 bool G1ParEvacuateFollowersClosure::offer_termination() {
4759 G1ParScanThreadState* const pss = par_scan_state();
4760 pss->start_term_time();
4761 const bool res = terminator()->offer_termination();
4762 pss->end_term_time();
4763 return res;
4764 }
4765
4766 void G1ParEvacuateFollowersClosure::do_void() {
4767 StarTask stolen_task;
4768 G1ParScanThreadState* const pss = par_scan_state();
4769 pss->trim_queue();
4770
4771 do {
4772 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4773 assert(pss->verify_task(stolen_task), "sanity");
4774 if (stolen_task.is_narrow()) {
4775 pss->deal_with_reference((narrowOop*) stolen_task);
4776 } else {
4777 pss->deal_with_reference((oop*) stolen_task);
4778 }
4779
4780 // We've just processed a reference and we might have made
4781 // available new entries on the queues. So we have to make sure
4782 // we drain the queues as necessary.
4783 pss->trim_queue();
4784 }
4785 } while (!offer_termination());
4786
4787 pss->retire_alloc_buffers();
4788 }
4789
4790 class G1KlassScanClosure : public KlassClosure {
4791 G1ParCopyHelper* _closure;
4792 bool _process_only_dirty;
4793 int _count;
4794 public:
4795 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty)
4796 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {}
4797 void do_klass(Klass* klass) {
4798 // If the klass has not been dirtied we know that there's
4799 // no references into the young gen and we can skip it.
4800 if (!_process_only_dirty || klass->has_modified_oops()) {
4801 // Clean the klass since we're going to scavenge all the metadata.
4802 klass->clear_modified_oops();
4803
4804 // Tell the closure that this klass is the Klass to scavenge
4805 // and is the one to dirty if oops are left pointing into the young gen.
4806 _closure->set_scanned_klass(klass);
4807
4808 klass->oops_do(_closure);
4809
4810 _closure->set_scanned_klass(NULL);
4811 }
4812 _count++;
4813 }
4814 };
4815
4816 class G1ParTask : public AbstractGangTask {
4817 protected:
4818 G1CollectedHeap* _g1h;
4819 RefToScanQueueSet *_queues;
4820 ParallelTaskTerminator _terminator;
4821 uint _n_workers;
4822
4823 Mutex _stats_lock;
4824 Mutex* stats_lock() { return &_stats_lock; }
4825
4826 size_t getNCards() {
4827 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4828 / G1BlockOffsetSharedArray::N_bytes;
4829 }
4830
4831 public:
4832 G1ParTask(G1CollectedHeap* g1h,
4833 RefToScanQueueSet *task_queues)
4834 : AbstractGangTask("G1 collection"),
4835 _g1h(g1h),
4836 _queues(task_queues),
4837 _terminator(0, _queues),
4838 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true)
4839 {}
4840
4841 RefToScanQueueSet* queues() { return _queues; }
4842
4843 RefToScanQueue *work_queue(int i) {
4844 return queues()->queue(i);
4845 }
4846
4847 ParallelTaskTerminator* terminator() { return &_terminator; }
4848
4849 virtual void set_for_termination(int active_workers) {
4850 // This task calls set_n_termination() in par_non_clean_card_iterate_work()
4851 // in the young space (_par_seq_tasks) in the G1 heap
4852 // for SequentialSubTasksDone.
4853 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap
4854 // both of which need setting by set_n_termination().
4855 _g1h->SharedHeap::set_n_termination(active_workers);
4856 _g1h->set_n_termination(active_workers);
4857 terminator()->reset_for_reuse(active_workers);
4858 _n_workers = active_workers;
4859 }
4860
4861 void work(uint worker_id) {
4862 if (worker_id >= _n_workers) return; // no work needed this round
4863
4864 double start_time_ms = os::elapsedTime() * 1000.0;
4865 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms);
4866
4867 {
4868 ResourceMark rm;
4869 HandleMark hm;
4870
4871 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4872
4873 G1ParScanThreadState pss(_g1h, worker_id);
4874 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4875 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4876 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4877
4878 pss.set_evac_closure(&scan_evac_cl);
4879 pss.set_evac_failure_closure(&evac_failure_cl);
4880 pss.set_partial_scan_closure(&partial_scan_cl);
4881
4882 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4883 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp);
4884
4885 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4886 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp);
4887
4888 bool only_young = _g1h->g1_policy()->gcs_are_young();
4889 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false);
4890 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young);
4891
4892 OopClosure* scan_root_cl = &only_scan_root_cl;
4893 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s;
4894
4895 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4896 // We also need to mark copied objects.
4897 scan_root_cl = &scan_mark_root_cl;
4898 scan_klasses_cl = &scan_mark_klasses_cl_s;
4899 }
4900
4901 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4902
4903 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
4904
4905 pss.start_strong_roots();
4906 _g1h->g1_process_strong_roots(/* is scavenging */ true,
4907 SharedHeap::ScanningOption(so),
4908 scan_root_cl,
4909 &push_heap_rs_cl,
4910 scan_klasses_cl,
4911 worker_id);
4912 pss.end_strong_roots();
4913
4914 {
4915 double start = os::elapsedTime();
4916 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4917 evac.do_void();
4918 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4919 double term_ms = pss.term_time()*1000.0;
4920 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms);
4921 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts());
4922 }
4923 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4924 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4925
4926 if (ParallelGCVerbose) {
4927 MutexLocker x(stats_lock());
4928 pss.print_termination_stats(worker_id);
4929 }
4930
4931 assert(pss.refs()->is_empty(), "should be empty");
4932
4933 // Close the inner scope so that the ResourceMark and HandleMark
4934 // destructors are executed here and are included as part of the
4935 // "GC Worker Time".
4936 }
4937
4938 double end_time_ms = os::elapsedTime() * 1000.0;
4939 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms);
4940 }
4941 };
4942
4943 // *** Common G1 Evacuation Stuff
4944
4945 // Closures that support the filtering of CodeBlobs scanned during
4946 // external root scanning.
4947
4948 // Closure applied to reference fields in code blobs (specifically nmethods)
4949 // to determine whether an nmethod contains references that point into
4950 // the collection set. Used as a predicate when walking code roots so
4951 // that only nmethods that point into the collection set are added to the
4952 // 'marked' list.
4953
4954 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure {
4955
4956 class G1PointsIntoCSOopClosure : public OopClosure {
4957 G1CollectedHeap* _g1;
4958 bool _points_into_cs;
4959 public:
4960 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) :
4961 _g1(g1), _points_into_cs(false) { }
4962
4963 bool points_into_cs() const { return _points_into_cs; }
4964
4965 template <class T>
4966 void do_oop_nv(T* p) {
4967 if (!_points_into_cs) {
4968 T heap_oop = oopDesc::load_heap_oop(p);
4969 if (!oopDesc::is_null(heap_oop) &&
4970 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) {
4971 _points_into_cs = true;
4972 }
4973 }
4974 }
4975
4976 virtual void do_oop(oop* p) { do_oop_nv(p); }
4977 virtual void do_oop(narrowOop* p) { do_oop_nv(p); }
4978 };
4979
4980 G1CollectedHeap* _g1;
4981
4982 public:
4983 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) :
4984 CodeBlobToOopClosure(cl, true), _g1(g1) { }
4985
4986 virtual void do_code_blob(CodeBlob* cb) {
4987 nmethod* nm = cb->as_nmethod_or_null();
4988 if (nm != NULL && !(nm->test_oops_do_mark())) {
4989 G1PointsIntoCSOopClosure predicate_cl(_g1);
4990 nm->oops_do(&predicate_cl);
4991
4992 if (predicate_cl.points_into_cs()) {
4993 // At least one of the reference fields or the oop relocations
4994 // in the nmethod points into the collection set. We have to
4995 // 'mark' this nmethod.
4996 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob()
4997 // or MarkingCodeBlobClosure::do_code_blob() change.
4998 if (!nm->test_set_oops_do_mark()) {
4999 do_newly_marked_nmethod(nm);
5000 }
5001 }
5002 }
5003 }
5004 };
5005
5006 // This method is run in a GC worker.
5007
5008 void
5009 G1CollectedHeap::
5010 g1_process_strong_roots(bool is_scavenging,
5011 ScanningOption so,
5012 OopClosure* scan_non_heap_roots,
5013 OopsInHeapRegionClosure* scan_rs,
5014 G1KlassScanClosure* scan_klasses,
5015 int worker_i) {
5016
5017 // First scan the strong roots
5018 double ext_roots_start = os::elapsedTime();
5019 double closure_app_time_sec = 0.0;
5020
5021 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
5022
5023 // Walk the code cache w/o buffering, because StarTask cannot handle
5024 // unaligned oop locations.
5025 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots);
5026
5027 process_strong_roots(false, // no scoping; this is parallel code
5028 is_scavenging, so,
5029 &buf_scan_non_heap_roots,
5030 &eager_scan_code_roots,
5031 scan_klasses
5032 );
5033
5034 // Now the CM ref_processor roots.
5035 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
5036 // We need to treat the discovered reference lists of the
5037 // concurrent mark ref processor as roots and keep entries
5038 // (which are added by the marking threads) on them live
5039 // until they can be processed at the end of marking.
5040 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
5041 }
5042
5043 // Finish up any enqueued closure apps (attributed as object copy time).
5044 buf_scan_non_heap_roots.done();
5045
5046 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds();
5047
5048 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
5049
5050 double ext_root_time_ms =
5051 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0;
5052
5053 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
5054
5055 // During conc marking we have to filter the per-thread SATB buffers
5056 // to make sure we remove any oops into the CSet (which will show up
5057 // as implicitly live).
5058 double satb_filtering_ms = 0.0;
5059 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) {
5060 if (mark_in_progress()) {
5061 double satb_filter_start = os::elapsedTime();
5062
5063 JavaThread::satb_mark_queue_set().filter_thread_buffers();
5064
5065 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0;
5066 }
5067 }
5068 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms);
5069
5070 // Now scan the complement of the collection set.
5071 if (scan_rs != NULL) {
5072 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
5073 }
5074 _process_strong_tasks->all_tasks_completed();
5075 }
5076
5077 void
5078 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure) {
5079 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
5080 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs);
5081 }
5082
5083 // Weak Reference Processing support
5084
5085 // An always "is_alive" closure that is used to preserve referents.
5086 // If the object is non-null then it's alive. Used in the preservation
5087 // of referent objects that are pointed to by reference objects
5088 // discovered by the CM ref processor.
5089 class G1AlwaysAliveClosure: public BoolObjectClosure {
5090 G1CollectedHeap* _g1;
5091 public:
5092 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5093 void do_object(oop p) { assert(false, "Do not call."); }
5094 bool do_object_b(oop p) {
5095 if (p != NULL) {
5096 return true;
5097 }
5098 return false;
5099 }
5100 };
5101
5102 bool G1STWIsAliveClosure::do_object_b(oop p) {
5103 // An object is reachable if it is outside the collection set,
5104 // or is inside and copied.
5105 return !_g1->obj_in_cs(p) || p->is_forwarded();
5106 }
5107
5108 // Non Copying Keep Alive closure
5109 class G1KeepAliveClosure: public OopClosure {
5110 G1CollectedHeap* _g1;
5111 public:
5112 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5113 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
5114 void do_oop( oop* p) {
5115 oop obj = *p;
5116
5117 if (_g1->obj_in_cs(obj)) {
5118 assert( obj->is_forwarded(), "invariant" );
5119 *p = obj->forwardee();
5120 }
5121 }
5122 };
5123
5124 // Copying Keep Alive closure - can be called from both
5125 // serial and parallel code as long as different worker
5126 // threads utilize different G1ParScanThreadState instances
5127 // and different queues.
5128
5129 class G1CopyingKeepAliveClosure: public OopClosure {
5130 G1CollectedHeap* _g1h;
5131 OopClosure* _copy_non_heap_obj_cl;
5132 OopsInHeapRegionClosure* _copy_metadata_obj_cl;
5133 G1ParScanThreadState* _par_scan_state;
5134
5135 public:
5136 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
5137 OopClosure* non_heap_obj_cl,
5138 OopsInHeapRegionClosure* metadata_obj_cl,
5139 G1ParScanThreadState* pss):
5140 _g1h(g1h),
5141 _copy_non_heap_obj_cl(non_heap_obj_cl),
5142 _copy_metadata_obj_cl(metadata_obj_cl),
5143 _par_scan_state(pss)
5144 {}
5145
5146 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
5147 virtual void do_oop( oop* p) { do_oop_work(p); }
5148
5149 template <class T> void do_oop_work(T* p) {
5150 oop obj = oopDesc::load_decode_heap_oop(p);
5151
5152 if (_g1h->obj_in_cs(obj)) {
5153 // If the referent object has been forwarded (either copied
5154 // to a new location or to itself in the event of an
5155 // evacuation failure) then we need to update the reference
5156 // field and, if both reference and referent are in the G1
5157 // heap, update the RSet for the referent.
5158 //
5159 // If the referent has not been forwarded then we have to keep
5160 // it alive by policy. Therefore we have copy the referent.
5161 //
5162 // If the reference field is in the G1 heap then we can push
5163 // on the PSS queue. When the queue is drained (after each
5164 // phase of reference processing) the object and it's followers
5165 // will be copied, the reference field set to point to the
5166 // new location, and the RSet updated. Otherwise we need to
5167 // use the the non-heap or metadata closures directly to copy
5168 // the refernt object and update the pointer, while avoiding
5169 // updating the RSet.
5170
5171 if (_g1h->is_in_g1_reserved(p)) {
5172 _par_scan_state->push_on_queue(p);
5173 } else {
5174 assert(!ClassLoaderDataGraph::contains((address)p),
5175 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) "
5176 PTR_FORMAT, p));
5177 _copy_non_heap_obj_cl->do_oop(p);
5178 }
5179 }
5180 }
5181 };
5182
5183 // Serial drain queue closure. Called as the 'complete_gc'
5184 // closure for each discovered list in some of the
5185 // reference processing phases.
5186
5187 class G1STWDrainQueueClosure: public VoidClosure {
5188 protected:
5189 G1CollectedHeap* _g1h;
5190 G1ParScanThreadState* _par_scan_state;
5191
5192 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
5193
5194 public:
5195 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
5196 _g1h(g1h),
5197 _par_scan_state(pss)
5198 { }
5199
5200 void do_void() {
5201 G1ParScanThreadState* const pss = par_scan_state();
5202 pss->trim_queue();
5203 }
5204 };
5205
5206 // Parallel Reference Processing closures
5207
5208 // Implementation of AbstractRefProcTaskExecutor for parallel reference
5209 // processing during G1 evacuation pauses.
5210
5211 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
5212 private:
5213 G1CollectedHeap* _g1h;
5214 RefToScanQueueSet* _queues;
5215 FlexibleWorkGang* _workers;
5216 int _active_workers;
5217
5218 public:
5219 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
5220 FlexibleWorkGang* workers,
5221 RefToScanQueueSet *task_queues,
5222 int n_workers) :
5223 _g1h(g1h),
5224 _queues(task_queues),
5225 _workers(workers),
5226 _active_workers(n_workers)
5227 {
5228 assert(n_workers > 0, "shouldn't call this otherwise");
5229 }
5230
5231 // Executes the given task using concurrent marking worker threads.
5232 virtual void execute(ProcessTask& task);
5233 virtual void execute(EnqueueTask& task);
5234 };
5235
5236 // Gang task for possibly parallel reference processing
5237
5238 class G1STWRefProcTaskProxy: public AbstractGangTask {
5239 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5240 ProcessTask& _proc_task;
5241 G1CollectedHeap* _g1h;
5242 RefToScanQueueSet *_task_queues;
5243 ParallelTaskTerminator* _terminator;
5244
5245 public:
5246 G1STWRefProcTaskProxy(ProcessTask& proc_task,
5247 G1CollectedHeap* g1h,
5248 RefToScanQueueSet *task_queues,
5249 ParallelTaskTerminator* terminator) :
5250 AbstractGangTask("Process reference objects in parallel"),
5251 _proc_task(proc_task),
5252 _g1h(g1h),
5253 _task_queues(task_queues),
5254 _terminator(terminator)
5255 {}
5256
5257 virtual void work(uint worker_id) {
5258 // The reference processing task executed by a single worker.
5259 ResourceMark rm;
5260 HandleMark hm;
5261
5262 G1STWIsAliveClosure is_alive(_g1h);
5263
5264 G1ParScanThreadState pss(_g1h, worker_id);
5265
5266 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5267 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5268 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5269
5270 pss.set_evac_closure(&scan_evac_cl);
5271 pss.set_evac_failure_closure(&evac_failure_cl);
5272 pss.set_partial_scan_closure(&partial_scan_cl);
5273
5274 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5275 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5276
5277 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5278 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5279
5280 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5281 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5282
5283 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5284 // We also need to mark copied objects.
5285 copy_non_heap_cl = ©_mark_non_heap_cl;
5286 copy_metadata_cl = ©_mark_metadata_cl;
5287 }
5288
5289 // Keep alive closure.
5290 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5291
5292 // Complete GC closure
5293 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
5294
5295 // Call the reference processing task's work routine.
5296 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue);
5297
5298 // Note we cannot assert that the refs array is empty here as not all
5299 // of the processing tasks (specifically phase2 - pp2_work) execute
5300 // the complete_gc closure (which ordinarily would drain the queue) so
5301 // the queue may not be empty.
5302 }
5303 };
5304
5305 // Driver routine for parallel reference processing.
5306 // Creates an instance of the ref processing gang
5307 // task and has the worker threads execute it.
5308 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
5309 assert(_workers != NULL, "Need parallel worker threads.");
5310
5311 ParallelTaskTerminator terminator(_active_workers, _queues);
5312 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
5313
5314 _g1h->set_par_threads(_active_workers);
5315 _workers->run_task(&proc_task_proxy);
5316 _g1h->set_par_threads(0);
5317 }
5318
5319 // Gang task for parallel reference enqueueing.
5320
5321 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
5322 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5323 EnqueueTask& _enq_task;
5324
5325 public:
5326 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
5327 AbstractGangTask("Enqueue reference objects in parallel"),
5328 _enq_task(enq_task)
5329 { }
5330
5331 virtual void work(uint worker_id) {
5332 _enq_task.work(worker_id);
5333 }
5334 };
5335
5336 // Driver routine for parallel reference enqueing.
5337 // Creates an instance of the ref enqueueing gang
5338 // task and has the worker threads execute it.
5339
5340 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5341 assert(_workers != NULL, "Need parallel worker threads.");
5342
5343 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5344
5345 _g1h->set_par_threads(_active_workers);
5346 _workers->run_task(&enq_task_proxy);
5347 _g1h->set_par_threads(0);
5348 }
5349
5350 // End of weak reference support closures
5351
5352 // Abstract task used to preserve (i.e. copy) any referent objects
5353 // that are in the collection set and are pointed to by reference
5354 // objects discovered by the CM ref processor.
5355
5356 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5357 protected:
5358 G1CollectedHeap* _g1h;
5359 RefToScanQueueSet *_queues;
5360 ParallelTaskTerminator _terminator;
5361 uint _n_workers;
5362
5363 public:
5364 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5365 AbstractGangTask("ParPreserveCMReferents"),
5366 _g1h(g1h),
5367 _queues(task_queues),
5368 _terminator(workers, _queues),
5369 _n_workers(workers)
5370 { }
5371
5372 void work(uint worker_id) {
5373 ResourceMark rm;
5374 HandleMark hm;
5375
5376 G1ParScanThreadState pss(_g1h, worker_id);
5377 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5378 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5379 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5380
5381 pss.set_evac_closure(&scan_evac_cl);
5382 pss.set_evac_failure_closure(&evac_failure_cl);
5383 pss.set_partial_scan_closure(&partial_scan_cl);
5384
5385 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5386
5387
5388 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5389 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5390
5391 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5392 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5393
5394 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5395 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5396
5397 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5398 // We also need to mark copied objects.
5399 copy_non_heap_cl = ©_mark_non_heap_cl;
5400 copy_metadata_cl = ©_mark_metadata_cl;
5401 }
5402
5403 // Is alive closure
5404 G1AlwaysAliveClosure always_alive(_g1h);
5405
5406 // Copying keep alive closure. Applied to referent objects that need
5407 // to be copied.
5408 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5409
5410 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5411
5412 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5413 uint stride = MIN2(MAX2(_n_workers, 1U), limit);
5414
5415 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5416 // So this must be true - but assert just in case someone decides to
5417 // change the worker ids.
5418 assert(0 <= worker_id && worker_id < limit, "sanity");
5419 assert(!rp->discovery_is_atomic(), "check this code");
5420
5421 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5422 for (uint idx = worker_id; idx < limit; idx += stride) {
5423 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5424
5425 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5426 while (iter.has_next()) {
5427 // Since discovery is not atomic for the CM ref processor, we
5428 // can see some null referent objects.
5429 iter.load_ptrs(DEBUG_ONLY(true));
5430 oop ref = iter.obj();
5431
5432 // This will filter nulls.
5433 if (iter.is_referent_alive()) {
5434 iter.make_referent_alive();
5435 }
5436 iter.move_to_next();
5437 }
5438 }
5439
5440 // Drain the queue - which may cause stealing
5441 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5442 drain_queue.do_void();
5443 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5444 assert(pss.refs()->is_empty(), "should be");
5445 }
5446 };
5447
5448 // Weak Reference processing during an evacuation pause (part 1).
5449 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) {
5450 double ref_proc_start = os::elapsedTime();
5451
5452 ReferenceProcessor* rp = _ref_processor_stw;
5453 assert(rp->discovery_enabled(), "should have been enabled");
5454
5455 // Any reference objects, in the collection set, that were 'discovered'
5456 // by the CM ref processor should have already been copied (either by
5457 // applying the external root copy closure to the discovered lists, or
5458 // by following an RSet entry).
5459 //
5460 // But some of the referents, that are in the collection set, that these
5461 // reference objects point to may not have been copied: the STW ref
5462 // processor would have seen that the reference object had already
5463 // been 'discovered' and would have skipped discovering the reference,
5464 // but would not have treated the reference object as a regular oop.
5465 // As a reult the copy closure would not have been applied to the
5466 // referent object.
5467 //
5468 // We need to explicitly copy these referent objects - the references
5469 // will be processed at the end of remarking.
5470 //
5471 // We also need to do this copying before we process the reference
5472 // objects discovered by the STW ref processor in case one of these
5473 // referents points to another object which is also referenced by an
5474 // object discovered by the STW ref processor.
5475
5476 assert(!G1CollectedHeap::use_parallel_gc_threads() ||
5477 no_of_gc_workers == workers()->active_workers(),
5478 "Need to reset active GC workers");
5479
5480 set_par_threads(no_of_gc_workers);
5481 G1ParPreserveCMReferentsTask keep_cm_referents(this,
5482 no_of_gc_workers,
5483 _task_queues);
5484
5485 if (G1CollectedHeap::use_parallel_gc_threads()) {
5486 workers()->run_task(&keep_cm_referents);
5487 } else {
5488 keep_cm_referents.work(0);
5489 }
5490
5491 set_par_threads(0);
5492
5493 // Closure to test whether a referent is alive.
5494 G1STWIsAliveClosure is_alive(this);
5495
5496 // Even when parallel reference processing is enabled, the processing
5497 // of JNI refs is serial and performed serially by the current thread
5498 // rather than by a worker. The following PSS will be used for processing
5499 // JNI refs.
5500
5501 // Use only a single queue for this PSS.
5502 G1ParScanThreadState pss(this, 0);
5503
5504 // We do not embed a reference processor in the copying/scanning
5505 // closures while we're actually processing the discovered
5506 // reference objects.
5507 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5508 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5509 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5510
5511 pss.set_evac_closure(&scan_evac_cl);
5512 pss.set_evac_failure_closure(&evac_failure_cl);
5513 pss.set_partial_scan_closure(&partial_scan_cl);
5514
5515 assert(pss.refs()->is_empty(), "pre-condition");
5516
5517 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5518 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL);
5519
5520 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5521 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL);
5522
5523 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5524 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5525
5526 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5527 // We also need to mark copied objects.
5528 copy_non_heap_cl = ©_mark_non_heap_cl;
5529 copy_metadata_cl = ©_mark_metadata_cl;
5530 }
5531
5532 // Keep alive closure.
5533 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss);
5534
5535 // Serial Complete GC closure
5536 G1STWDrainQueueClosure drain_queue(this, &pss);
5537
5538 // Setup the soft refs policy...
5539 rp->setup_policy(false);
5540
5541 if (!rp->processing_is_mt()) {
5542 // Serial reference processing...
5543 rp->process_discovered_references(&is_alive,
5544 &keep_alive,
5545 &drain_queue,
5546 NULL);
5547 } else {
5548 // Parallel reference processing
5549 assert(rp->num_q() == no_of_gc_workers, "sanity");
5550 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5551
5552 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5553 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5554 }
5555
5556 // We have completed copying any necessary live referent objects
5557 // (that were not copied during the actual pause) so we can
5558 // retire any active alloc buffers
5559 pss.retire_alloc_buffers();
5560 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5561
5562 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5563 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0);
5564 }
5565
5566 // Weak Reference processing during an evacuation pause (part 2).
5567 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) {
5568 double ref_enq_start = os::elapsedTime();
5569
5570 ReferenceProcessor* rp = _ref_processor_stw;
5571 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5572
5573 // Now enqueue any remaining on the discovered lists on to
5574 // the pending list.
5575 if (!rp->processing_is_mt()) {
5576 // Serial reference processing...
5577 rp->enqueue_discovered_references();
5578 } else {
5579 // Parallel reference enqueuing
5580
5581 assert(no_of_gc_workers == workers()->active_workers(),
5582 "Need to reset active workers");
5583 assert(rp->num_q() == no_of_gc_workers, "sanity");
5584 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5585
5586 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5587 rp->enqueue_discovered_references(&par_task_executor);
5588 }
5589
5590 rp->verify_no_references_recorded();
5591 assert(!rp->discovery_enabled(), "should have been disabled");
5592
5593 // FIXME
5594 // CM's reference processing also cleans up the string and symbol tables.
5595 // Should we do that here also? We could, but it is a serial operation
5596 // and could signicantly increase the pause time.
5597
5598 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5599 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0);
5600 }
5601
5602 void G1CollectedHeap::evacuate_collection_set() {
5603 _expand_heap_after_alloc_failure = true;
5604 set_evacuation_failed(false);
5605
5606 // Should G1EvacuationFailureALot be in effect for this GC?
5607 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();)
5608
5609 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5610
5611 // Disable the hot card cache.
5612 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache();
5613 hot_card_cache->reset_hot_cache_claimed_index();
5614 hot_card_cache->set_use_cache(false);
5615
5616 uint n_workers;
5617 if (G1CollectedHeap::use_parallel_gc_threads()) {
5618 n_workers =
5619 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
5620 workers()->active_workers(),
5621 Threads::number_of_non_daemon_threads());
5622 assert(UseDynamicNumberOfGCThreads ||
5623 n_workers == workers()->total_workers(),
5624 "If not dynamic should be using all the workers");
5625 workers()->set_active_workers(n_workers);
5626 set_par_threads(n_workers);
5627 } else {
5628 assert(n_par_threads() == 0,
5629 "Should be the original non-parallel value");
5630 n_workers = 1;
5631 }
5632
5633 G1ParTask g1_par_task(this, _task_queues);
5634
5635 init_for_evac_failure(NULL);
5636
5637 rem_set()->prepare_for_younger_refs_iterate(true);
5638
5639 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5640 double start_par_time_sec = os::elapsedTime();
5641 double end_par_time_sec;
5642
5643 {
5644 StrongRootsScope srs(this);
5645
5646 if (G1CollectedHeap::use_parallel_gc_threads()) {
5647 // The individual threads will set their evac-failure closures.
5648 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5649 // These tasks use ShareHeap::_process_strong_tasks
5650 assert(UseDynamicNumberOfGCThreads ||
5651 workers()->active_workers() == workers()->total_workers(),
5652 "If not dynamic should be using all the workers");
5653 workers()->run_task(&g1_par_task);
5654 } else {
5655 g1_par_task.set_for_termination(n_workers);
5656 g1_par_task.work(0);
5657 }
5658 end_par_time_sec = os::elapsedTime();
5659
5660 // Closing the inner scope will execute the destructor
5661 // for the StrongRootsScope object. We record the current
5662 // elapsed time before closing the scope so that time
5663 // taken for the SRS destructor is NOT included in the
5664 // reported parallel time.
5665 }
5666
5667 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0;
5668 g1_policy()->phase_times()->record_par_time(par_time_ms);
5669
5670 double code_root_fixup_time_ms =
5671 (os::elapsedTime() - end_par_time_sec) * 1000.0;
5672 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms);
5673
5674 set_par_threads(0);
5675
5676 // Process any discovered reference objects - we have
5677 // to do this _before_ we retire the GC alloc regions
5678 // as we may have to copy some 'reachable' referent
5679 // objects (and their reachable sub-graphs) that were
5680 // not copied during the pause.
5681 process_discovered_references(n_workers);
5682
5683 // Weak root processing.
5684 // Note: when JSR 292 is enabled and code blobs can contain
5685 // non-perm oops then we will need to process the code blobs
5686 // here too.
5687 {
5688 G1STWIsAliveClosure is_alive(this);
5689 G1KeepAliveClosure keep_alive(this);
5690 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5691 }
5692
5693 release_gc_alloc_regions(n_workers);
5694 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5695
5696 // Reset and re-enable the hot card cache.
5697 // Note the counts for the cards in the regions in the
5698 // collection set are reset when the collection set is freed.
5699 hot_card_cache->reset_hot_cache();
5700 hot_card_cache->set_use_cache(true);
5701
5702 finalize_for_evac_failure();
5703
5704 if (evacuation_failed()) {
5705 remove_self_forwarding_pointers();
5706
5707 // Reset the G1EvacuationFailureALot counters and flags
5708 // Note: the values are reset only when an actual
5709 // evacuation failure occurs.
5710 NOT_PRODUCT(reset_evacuation_should_fail();)
5711 }
5712
5713 // Enqueue any remaining references remaining on the STW
5714 // reference processor's discovered lists. We need to do
5715 // this after the card table is cleaned (and verified) as
5716 // the act of enqueuing entries on to the pending list
5717 // will log these updates (and dirty their associated
5718 // cards). We need these updates logged to update any
5719 // RSets.
5720 enqueue_discovered_references(n_workers);
5721
5722 if (G1DeferredRSUpdate) {
5723 RedirtyLoggedCardTableEntryFastClosure redirty;
5724 dirty_card_queue_set().set_closure(&redirty);
5725 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5726
5727 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5728 dcq.merge_bufferlists(&dirty_card_queue_set());
5729 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5730 }
5731 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5732 }
5733
5734 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5735 size_t* pre_used,
5736 FreeRegionList* free_list,
5737 OldRegionSet* old_proxy_set,
5738 HumongousRegionSet* humongous_proxy_set,
5739 HRRSCleanupTask* hrrs_cleanup_task,
5740 bool par) {
5741 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5742 if (hr->isHumongous()) {
5743 assert(hr->startsHumongous(), "we should only see starts humongous");
5744 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5745 } else {
5746 _old_set.remove_with_proxy(hr, old_proxy_set);
5747 free_region(hr, pre_used, free_list, par);
5748 }
5749 } else {
5750 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5751 }
5752 }
5753
5754 void G1CollectedHeap::free_region(HeapRegion* hr,
5755 size_t* pre_used,
5756 FreeRegionList* free_list,
5757 bool par) {
5758 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5759 assert(!hr->is_empty(), "the region should not be empty");
5760 assert(free_list != NULL, "pre-condition");
5761
5762 // Clear the card counts for this region.
5763 // Note: we only need to do this if the region is not young
5764 // (since we don't refine cards in young regions).
5765 if (!hr->is_young()) {
5766 _cg1r->hot_card_cache()->reset_card_counts(hr);
5767 }
5768 *pre_used += hr->used();
5769 hr->hr_clear(par, true /* clear_space */);
5770 free_list->add_as_head(hr);
5771 }
5772
5773 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5774 size_t* pre_used,
5775 FreeRegionList* free_list,
5776 HumongousRegionSet* humongous_proxy_set,
5777 bool par) {
5778 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5779 assert(free_list != NULL, "pre-condition");
5780 assert(humongous_proxy_set != NULL, "pre-condition");
5781
5782 size_t hr_used = hr->used();
5783 size_t hr_capacity = hr->capacity();
5784 size_t hr_pre_used = 0;
5785 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5786 // We need to read this before we make the region non-humongous,
5787 // otherwise the information will be gone.
5788 uint last_index = hr->last_hc_index();
5789 hr->set_notHumongous();
5790 free_region(hr, &hr_pre_used, free_list, par);
5791
5792 uint i = hr->hrs_index() + 1;
5793 while (i < last_index) {
5794 HeapRegion* curr_hr = region_at(i);
5795 assert(curr_hr->continuesHumongous(), "invariant");
5796 curr_hr->set_notHumongous();
5797 free_region(curr_hr, &hr_pre_used, free_list, par);
5798 i += 1;
5799 }
5800 assert(hr_pre_used == hr_used,
5801 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5802 "should be the same", hr_pre_used, hr_used));
5803 *pre_used += hr_pre_used;
5804 }
5805
5806 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5807 FreeRegionList* free_list,
5808 OldRegionSet* old_proxy_set,
5809 HumongousRegionSet* humongous_proxy_set,
5810 bool par) {
5811 if (pre_used > 0) {
5812 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5813 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5814 assert(_summary_bytes_used >= pre_used,
5815 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5816 "should be >= pre_used: "SIZE_FORMAT,
5817 _summary_bytes_used, pre_used));
5818 _summary_bytes_used -= pre_used;
5819 }
5820 if (free_list != NULL && !free_list->is_empty()) {
5821 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5822 _free_list.add_as_head(free_list);
5823 }
5824 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) {
5825 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5826 _old_set.update_from_proxy(old_proxy_set);
5827 }
5828 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5829 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5830 _humongous_set.update_from_proxy(humongous_proxy_set);
5831 }
5832 }
5833
5834 class G1ParCleanupCTTask : public AbstractGangTask {
5835 CardTableModRefBS* _ct_bs;
5836 G1CollectedHeap* _g1h;
5837 HeapRegion* volatile _su_head;
5838 public:
5839 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5840 G1CollectedHeap* g1h) :
5841 AbstractGangTask("G1 Par Cleanup CT Task"),
5842 _ct_bs(ct_bs), _g1h(g1h) { }
5843
5844 void work(uint worker_id) {
5845 HeapRegion* r;
5846 while (r = _g1h->pop_dirty_cards_region()) {
5847 clear_cards(r);
5848 }
5849 }
5850
5851 void clear_cards(HeapRegion* r) {
5852 // Cards of the survivors should have already been dirtied.
5853 if (!r->is_survivor()) {
5854 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5855 }
5856 }
5857 };
5858
5859 #ifndef PRODUCT
5860 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5861 G1CollectedHeap* _g1h;
5862 CardTableModRefBS* _ct_bs;
5863 public:
5864 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5865 : _g1h(g1h), _ct_bs(ct_bs) { }
5866 virtual bool doHeapRegion(HeapRegion* r) {
5867 if (r->is_survivor()) {
5868 _g1h->verify_dirty_region(r);
5869 } else {
5870 _g1h->verify_not_dirty_region(r);
5871 }
5872 return false;
5873 }
5874 };
5875
5876 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5877 // All of the region should be clean.
5878 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5879 MemRegion mr(hr->bottom(), hr->end());
5880 ct_bs->verify_not_dirty_region(mr);
5881 }
5882
5883 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5884 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5885 // dirty allocated blocks as they allocate them. The thread that
5886 // retires each region and replaces it with a new one will do a
5887 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5888 // not dirty that area (one less thing to have to do while holding
5889 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5890 // is dirty.
5891 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5892 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5893 ct_bs->verify_dirty_region(mr);
5894 }
5895
5896 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5897 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5898 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5899 verify_dirty_region(hr);
5900 }
5901 }
5902
5903 void G1CollectedHeap::verify_dirty_young_regions() {
5904 verify_dirty_young_list(_young_list->first_region());
5905 }
5906 #endif
5907
5908 void G1CollectedHeap::cleanUpCardTable() {
5909 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5910 double start = os::elapsedTime();
5911
5912 {
5913 // Iterate over the dirty cards region list.
5914 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5915
5916 if (G1CollectedHeap::use_parallel_gc_threads()) {
5917 set_par_threads();
5918 workers()->run_task(&cleanup_task);
5919 set_par_threads(0);
5920 } else {
5921 while (_dirty_cards_region_list) {
5922 HeapRegion* r = _dirty_cards_region_list;
5923 cleanup_task.clear_cards(r);
5924 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5925 if (_dirty_cards_region_list == r) {
5926 // The last region.
5927 _dirty_cards_region_list = NULL;
5928 }
5929 r->set_next_dirty_cards_region(NULL);
5930 }
5931 }
5932 #ifndef PRODUCT
5933 if (G1VerifyCTCleanup || VerifyAfterGC) {
5934 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5935 heap_region_iterate(&cleanup_verifier);
5936 }
5937 #endif
5938 }
5939
5940 double elapsed = os::elapsedTime() - start;
5941 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0);
5942 }
5943
5944 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5945 size_t pre_used = 0;
5946 FreeRegionList local_free_list("Local List for CSet Freeing");
5947
5948 double young_time_ms = 0.0;
5949 double non_young_time_ms = 0.0;
5950
5951 // Since the collection set is a superset of the the young list,
5952 // all we need to do to clear the young list is clear its
5953 // head and length, and unlink any young regions in the code below
5954 _young_list->clear();
5955
5956 G1CollectorPolicy* policy = g1_policy();
5957
5958 double start_sec = os::elapsedTime();
5959 bool non_young = true;
5960
5961 HeapRegion* cur = cs_head;
5962 int age_bound = -1;
5963 size_t rs_lengths = 0;
5964
5965 while (cur != NULL) {
5966 assert(!is_on_master_free_list(cur), "sanity");
5967 if (non_young) {
5968 if (cur->is_young()) {
5969 double end_sec = os::elapsedTime();
5970 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5971 non_young_time_ms += elapsed_ms;
5972
5973 start_sec = os::elapsedTime();
5974 non_young = false;
5975 }
5976 } else {
5977 if (!cur->is_young()) {
5978 double end_sec = os::elapsedTime();
5979 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5980 young_time_ms += elapsed_ms;
5981
5982 start_sec = os::elapsedTime();
5983 non_young = true;
5984 }
5985 }
5986
5987 rs_lengths += cur->rem_set()->occupied();
5988
5989 HeapRegion* next = cur->next_in_collection_set();
5990 assert(cur->in_collection_set(), "bad CS");
5991 cur->set_next_in_collection_set(NULL);
5992 cur->set_in_collection_set(false);
5993
5994 if (cur->is_young()) {
5995 int index = cur->young_index_in_cset();
5996 assert(index != -1, "invariant");
5997 assert((uint) index < policy->young_cset_region_length(), "invariant");
5998 size_t words_survived = _surviving_young_words[index];
5999 cur->record_surv_words_in_group(words_survived);
6000
6001 // At this point the we have 'popped' cur from the collection set
6002 // (linked via next_in_collection_set()) but it is still in the
6003 // young list (linked via next_young_region()). Clear the
6004 // _next_young_region field.
6005 cur->set_next_young_region(NULL);
6006 } else {
6007 int index = cur->young_index_in_cset();
6008 assert(index == -1, "invariant");
6009 }
6010
6011 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
6012 (!cur->is_young() && cur->young_index_in_cset() == -1),
6013 "invariant" );
6014
6015 if (!cur->evacuation_failed()) {
6016 MemRegion used_mr = cur->used_region();
6017
6018 // And the region is empty.
6019 assert(!used_mr.is_empty(), "Should not have empty regions in a CS.");
6020 free_region(cur, &pre_used, &local_free_list, false /* par */);
6021 } else {
6022 cur->uninstall_surv_rate_group();
6023 if (cur->is_young()) {
6024 cur->set_young_index_in_cset(-1);
6025 }
6026 cur->set_not_young();
6027 cur->set_evacuation_failed(false);
6028 // The region is now considered to be old.
6029 _old_set.add(cur);
6030 }
6031 cur = next;
6032 }
6033
6034 policy->record_max_rs_lengths(rs_lengths);
6035 policy->cset_regions_freed();
6036
6037 double end_sec = os::elapsedTime();
6038 double elapsed_ms = (end_sec - start_sec) * 1000.0;
6039
6040 if (non_young) {
6041 non_young_time_ms += elapsed_ms;
6042 } else {
6043 young_time_ms += elapsed_ms;
6044 }
6045
6046 update_sets_after_freeing_regions(pre_used, &local_free_list,
6047 NULL /* old_proxy_set */,
6048 NULL /* humongous_proxy_set */,
6049 false /* par */);
6050 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms);
6051 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms);
6052 }
6053
6054 // This routine is similar to the above but does not record
6055 // any policy statistics or update free lists; we are abandoning
6056 // the current incremental collection set in preparation of a
6057 // full collection. After the full GC we will start to build up
6058 // the incremental collection set again.
6059 // This is only called when we're doing a full collection
6060 // and is immediately followed by the tearing down of the young list.
6061
6062 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
6063 HeapRegion* cur = cs_head;
6064
6065 while (cur != NULL) {
6066 HeapRegion* next = cur->next_in_collection_set();
6067 assert(cur->in_collection_set(), "bad CS");
6068 cur->set_next_in_collection_set(NULL);
6069 cur->set_in_collection_set(false);
6070 cur->set_young_index_in_cset(-1);
6071 cur = next;
6072 }
6073 }
6074
6075 void G1CollectedHeap::set_free_regions_coming() {
6076 if (G1ConcRegionFreeingVerbose) {
6077 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6078 "setting free regions coming");
6079 }
6080
6081 assert(!free_regions_coming(), "pre-condition");
6082 _free_regions_coming = true;
6083 }
6084
6085 void G1CollectedHeap::reset_free_regions_coming() {
6086 assert(free_regions_coming(), "pre-condition");
6087
6088 {
6089 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6090 _free_regions_coming = false;
6091 SecondaryFreeList_lock->notify_all();
6092 }
6093
6094 if (G1ConcRegionFreeingVerbose) {
6095 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6096 "reset free regions coming");
6097 }
6098 }
6099
6100 void G1CollectedHeap::wait_while_free_regions_coming() {
6101 // Most of the time we won't have to wait, so let's do a quick test
6102 // first before we take the lock.
6103 if (!free_regions_coming()) {
6104 return;
6105 }
6106
6107 if (G1ConcRegionFreeingVerbose) {
6108 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6109 "waiting for free regions");
6110 }
6111
6112 {
6113 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6114 while (free_regions_coming()) {
6115 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
6116 }
6117 }
6118
6119 if (G1ConcRegionFreeingVerbose) {
6120 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6121 "done waiting for free regions");
6122 }
6123 }
6124
6125 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
6126 assert(heap_lock_held_for_gc(),
6127 "the heap lock should already be held by or for this thread");
6128 _young_list->push_region(hr);
6129 }
6130
6131 class NoYoungRegionsClosure: public HeapRegionClosure {
6132 private:
6133 bool _success;
6134 public:
6135 NoYoungRegionsClosure() : _success(true) { }
6136 bool doHeapRegion(HeapRegion* r) {
6137 if (r->is_young()) {
6138 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
6139 r->bottom(), r->end());
6140 _success = false;
6141 }
6142 return false;
6143 }
6144 bool success() { return _success; }
6145 };
6146
6147 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
6148 bool ret = _young_list->check_list_empty(check_sample);
6149
6150 if (check_heap) {
6151 NoYoungRegionsClosure closure;
6152 heap_region_iterate(&closure);
6153 ret = ret && closure.success();
6154 }
6155
6156 return ret;
6157 }
6158
6159 class TearDownRegionSetsClosure : public HeapRegionClosure {
6160 private:
6161 OldRegionSet *_old_set;
6162
6163 public:
6164 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { }
6165
6166 bool doHeapRegion(HeapRegion* r) {
6167 if (r->is_empty()) {
6168 // We ignore empty regions, we'll empty the free list afterwards
6169 } else if (r->is_young()) {
6170 // We ignore young regions, we'll empty the young list afterwards
6171 } else if (r->isHumongous()) {
6172 // We ignore humongous regions, we're not tearing down the
6173 // humongous region set
6174 } else {
6175 // The rest should be old
6176 _old_set->remove(r);
6177 }
6178 return false;
6179 }
6180
6181 ~TearDownRegionSetsClosure() {
6182 assert(_old_set->is_empty(), "post-condition");
6183 }
6184 };
6185
6186 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) {
6187 assert_at_safepoint(true /* should_be_vm_thread */);
6188
6189 if (!free_list_only) {
6190 TearDownRegionSetsClosure cl(&_old_set);
6191 heap_region_iterate(&cl);
6192
6193 // Need to do this after the heap iteration to be able to
6194 // recognize the young regions and ignore them during the iteration.
6195 _young_list->empty_list();
6196 }
6197 _free_list.remove_all();
6198 }
6199
6200 class RebuildRegionSetsClosure : public HeapRegionClosure {
6201 private:
6202 bool _free_list_only;
6203 OldRegionSet* _old_set;
6204 FreeRegionList* _free_list;
6205 size_t _total_used;
6206
6207 public:
6208 RebuildRegionSetsClosure(bool free_list_only,
6209 OldRegionSet* old_set, FreeRegionList* free_list) :
6210 _free_list_only(free_list_only),
6211 _old_set(old_set), _free_list(free_list), _total_used(0) {
6212 assert(_free_list->is_empty(), "pre-condition");
6213 if (!free_list_only) {
6214 assert(_old_set->is_empty(), "pre-condition");
6215 }
6216 }
6217
6218 bool doHeapRegion(HeapRegion* r) {
6219 if (r->continuesHumongous()) {
6220 return false;
6221 }
6222
6223 if (r->is_empty()) {
6224 // Add free regions to the free list
6225 _free_list->add_as_tail(r);
6226 } else if (!_free_list_only) {
6227 assert(!r->is_young(), "we should not come across young regions");
6228
6229 if (r->isHumongous()) {
6230 // We ignore humongous regions, we left the humongous set unchanged
6231 } else {
6232 // The rest should be old, add them to the old set
6233 _old_set->add(r);
6234 }
6235 _total_used += r->used();
6236 }
6237
6238 return false;
6239 }
6240
6241 size_t total_used() {
6242 return _total_used;
6243 }
6244 };
6245
6246 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) {
6247 assert_at_safepoint(true /* should_be_vm_thread */);
6248
6249 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list);
6250 heap_region_iterate(&cl);
6251
6252 if (!free_list_only) {
6253 _summary_bytes_used = cl.total_used();
6254 }
6255 assert(_summary_bytes_used == recalculate_used(),
6256 err_msg("inconsistent _summary_bytes_used, "
6257 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT,
6258 _summary_bytes_used, recalculate_used()));
6259 }
6260
6261 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
6262 _refine_cte_cl->set_concurrent(concurrent);
6263 }
6264
6265 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
6266 HeapRegion* hr = heap_region_containing(p);
6267 if (hr == NULL) {
6268 return false;
6269 } else {
6270 return hr->is_in(p);
6271 }
6272 }
6273
6274 // Methods for the mutator alloc region
6275
6276 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
6277 bool force) {
6278 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6279 assert(!force || g1_policy()->can_expand_young_list(),
6280 "if force is true we should be able to expand the young list");
6281 bool young_list_full = g1_policy()->is_young_list_full();
6282 if (force || !young_list_full) {
6283 HeapRegion* new_alloc_region = new_region(word_size,
6284 false /* do_expand */);
6285 if (new_alloc_region != NULL) {
6286 set_region_short_lived_locked(new_alloc_region);
6287 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
6288 return new_alloc_region;
6289 }
6290 }
6291 return NULL;
6292 }
6293
6294 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
6295 size_t allocated_bytes) {
6296 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6297 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
6298
6299 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
6300 _summary_bytes_used += allocated_bytes;
6301 _hr_printer.retire(alloc_region);
6302 // We update the eden sizes here, when the region is retired,
6303 // instead of when it's allocated, since this is the point that its
6304 // used space has been recored in _summary_bytes_used.
6305 g1mm()->update_eden_size();
6306 }
6307
6308 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
6309 bool force) {
6310 return _g1h->new_mutator_alloc_region(word_size, force);
6311 }
6312
6313 void G1CollectedHeap::set_par_threads() {
6314 // Don't change the number of workers. Use the value previously set
6315 // in the workgroup.
6316 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise");
6317 uint n_workers = workers()->active_workers();
6318 assert(UseDynamicNumberOfGCThreads ||
6319 n_workers == workers()->total_workers(),
6320 "Otherwise should be using the total number of workers");
6321 if (n_workers == 0) {
6322 assert(false, "Should have been set in prior evacuation pause.");
6323 n_workers = ParallelGCThreads;
6324 workers()->set_active_workers(n_workers);
6325 }
6326 set_par_threads(n_workers);
6327 }
6328
6329 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
6330 size_t allocated_bytes) {
6331 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
6332 }
6333
6334 // Methods for the GC alloc regions
6335
6336 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
6337 uint count,
6338 GCAllocPurpose ap) {
6339 assert(FreeList_lock->owned_by_self(), "pre-condition");
6340
6341 if (count < g1_policy()->max_regions(ap)) {
6342 HeapRegion* new_alloc_region = new_region(word_size,
6343 true /* do_expand */);
6344 if (new_alloc_region != NULL) {
6345 // We really only need to do this for old regions given that we
6346 // should never scan survivors. But it doesn't hurt to do it
6347 // for survivors too.
6348 new_alloc_region->set_saved_mark();
6349 if (ap == GCAllocForSurvived) {
6350 new_alloc_region->set_survivor();
6351 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
6352 } else {
6353 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
6354 }
6355 bool during_im = g1_policy()->during_initial_mark_pause();
6356 new_alloc_region->note_start_of_copying(during_im);
6357 return new_alloc_region;
6358 } else {
6359 g1_policy()->note_alloc_region_limit_reached(ap);
6360 }
6361 }
6362 return NULL;
6363 }
6364
6365 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
6366 size_t allocated_bytes,
6367 GCAllocPurpose ap) {
6368 bool during_im = g1_policy()->during_initial_mark_pause();
6369 alloc_region->note_end_of_copying(during_im);
6370 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
6371 if (ap == GCAllocForSurvived) {
6372 young_list()->add_survivor_region(alloc_region);
6373 } else {
6374 _old_set.add(alloc_region);
6375 }
6376 _hr_printer.retire(alloc_region);
6377 }
6378
6379 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
6380 bool force) {
6381 assert(!force, "not supported for GC alloc regions");
6382 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
6383 }
6384
6385 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
6386 size_t allocated_bytes) {
6387 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6388 GCAllocForSurvived);
6389 }
6390
6391 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
6392 bool force) {
6393 assert(!force, "not supported for GC alloc regions");
6394 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
6395 }
6396
6397 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
6398 size_t allocated_bytes) {
6399 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6400 GCAllocForTenured);
6401 }
6402 // Heap region set verification
6403
6404 class VerifyRegionListsClosure : public HeapRegionClosure {
6405 private:
6406 FreeRegionList* _free_list;
6407 OldRegionSet* _old_set;
6408 HumongousRegionSet* _humongous_set;
6409 uint _region_count;
6410
6411 public:
6412 VerifyRegionListsClosure(OldRegionSet* old_set,
6413 HumongousRegionSet* humongous_set,
6414 FreeRegionList* free_list) :
6415 _old_set(old_set), _humongous_set(humongous_set),
6416 _free_list(free_list), _region_count(0) { }
6417
6418 uint region_count() { return _region_count; }
6419
6420 bool doHeapRegion(HeapRegion* hr) {
6421 _region_count += 1;
6422
6423 if (hr->continuesHumongous()) {
6424 return false;
6425 }
6426
6427 if (hr->is_young()) {
6428 // TODO
6429 } else if (hr->startsHumongous()) {
6430 _humongous_set->verify_next_region(hr);
6431 } else if (hr->is_empty()) {
6432 _free_list->verify_next_region(hr);
6433 } else {
6434 _old_set->verify_next_region(hr);
6435 }
6436 return false;
6437 }
6438 };
6439
6440 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index,
6441 HeapWord* bottom) {
6442 HeapWord* end = bottom + HeapRegion::GrainWords;
6443 MemRegion mr(bottom, end);
6444 assert(_g1_reserved.contains(mr), "invariant");
6445 // This might return NULL if the allocation fails
6446 return new HeapRegion(hrs_index, _bot_shared, mr);
6447 }
6448
6449 void G1CollectedHeap::verify_region_sets() {
6450 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6451
6452 // First, check the explicit lists.
6453 _free_list.verify();
6454 {
6455 // Given that a concurrent operation might be adding regions to
6456 // the secondary free list we have to take the lock before
6457 // verifying it.
6458 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6459 _secondary_free_list.verify();
6460 }
6461 _old_set.verify();
6462 _humongous_set.verify();
6463
6464 // If a concurrent region freeing operation is in progress it will
6465 // be difficult to correctly attributed any free regions we come
6466 // across to the correct free list given that they might belong to
6467 // one of several (free_list, secondary_free_list, any local lists,
6468 // etc.). So, if that's the case we will skip the rest of the
6469 // verification operation. Alternatively, waiting for the concurrent
6470 // operation to complete will have a non-trivial effect on the GC's
6471 // operation (no concurrent operation will last longer than the
6472 // interval between two calls to verification) and it might hide
6473 // any issues that we would like to catch during testing.
6474 if (free_regions_coming()) {
6475 return;
6476 }
6477
6478 // Make sure we append the secondary_free_list on the free_list so
6479 // that all free regions we will come across can be safely
6480 // attributed to the free_list.
6481 append_secondary_free_list_if_not_empty_with_lock();
6482
6483 // Finally, make sure that the region accounting in the lists is
6484 // consistent with what we see in the heap.
6485 _old_set.verify_start();
6486 _humongous_set.verify_start();
6487 _free_list.verify_start();
6488
6489 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list);
6490 heap_region_iterate(&cl);
6491
6492 _old_set.verify_end();
6493 _humongous_set.verify_end();
6494 _free_list.verify_end();
6495 }