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