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 gclog_or_tty->print(msg);
1275 prepare_for_verify();
1276 Universe::verify(false /* silent */, VerifyOption_G1UsePrevMarking);
1277 verify_time_ms = (os::elapsedTime() - verify_start) * 1000;
1278 }
1279
1280 return verify_time_ms;
1281 }
1282
1283 void G1CollectedHeap::verify_before_gc() {
1284 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:");
1285 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms);
1286 }
1287
1288 void G1CollectedHeap::verify_after_gc() {
1289 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:");
1290 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms);
1291 }
1292
1293 bool G1CollectedHeap::do_collection(bool explicit_gc,
1294 bool clear_all_soft_refs,
1295 size_t word_size) {
1296 assert_at_safepoint(true /* should_be_vm_thread */);
1297
1298 if (GC_locker::check_active_before_gc()) {
1299 return false;
1300 }
1301
1302 SvcGCMarker sgcm(SvcGCMarker::FULL);
1303 ResourceMark rm;
1304
1305 print_heap_before_gc();
1306
1307 size_t metadata_prev_used = MetaspaceAux::used_in_bytes();
1308
1309 HRSPhaseSetter x(HRSPhaseFullGC);
1310 verify_region_sets_optional();
1311
1312 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1313 collector_policy()->should_clear_all_soft_refs();
1314
1315 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1316
1317 {
1318 IsGCActiveMark x;
1319
1320 // Timing
1321 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant");
1322 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps);
1323 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
1324
1325 {
1326 TraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, gclog_or_tty);
1327 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1328 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1329
1330 double start = os::elapsedTime();
1331 g1_policy()->record_full_collection_start();
1332
1333 // Note: When we have a more flexible GC logging framework that
1334 // allows us to add optional attributes to a GC log record we
1335 // could consider timing and reporting how long we wait in the
1336 // following two methods.
1337 wait_while_free_regions_coming();
1338 // If we start the compaction before the CM threads finish
1339 // scanning the root regions we might trip them over as we'll
1340 // be moving objects / updating references. So let's wait until
1341 // they are done. By telling them to abort, they should complete
1342 // early.
1343 _cm->root_regions()->abort();
1344 _cm->root_regions()->wait_until_scan_finished();
1345 append_secondary_free_list_if_not_empty_with_lock();
1346
1347 gc_prologue(true);
1348 increment_total_collections(true /* full gc */);
1349 increment_old_marking_cycles_started();
1350
1351 assert(used() == recalculate_used(), "Should be equal");
1352
1353 verify_before_gc();
1354
1355 pre_full_gc_dump();
1356
1357 COMPILER2_PRESENT(DerivedPointerTable::clear());
1358
1359 // Disable discovery and empty the discovered lists
1360 // for the CM ref processor.
1361 ref_processor_cm()->disable_discovery();
1362 ref_processor_cm()->abandon_partial_discovery();
1363 ref_processor_cm()->verify_no_references_recorded();
1364
1365 // Abandon current iterations of concurrent marking and concurrent
1366 // refinement, if any are in progress. We have to do this before
1367 // wait_until_scan_finished() below.
1368 concurrent_mark()->abort();
1369
1370 // Make sure we'll choose a new allocation region afterwards.
1371 release_mutator_alloc_region();
1372 abandon_gc_alloc_regions();
1373 g1_rem_set()->cleanupHRRS();
1374
1375 // We should call this after we retire any currently active alloc
1376 // regions so that all the ALLOC / RETIRE events are generated
1377 // before the start GC event.
1378 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1379
1380 // We may have added regions to the current incremental collection
1381 // set between the last GC or pause and now. We need to clear the
1382 // incremental collection set and then start rebuilding it afresh
1383 // after this full GC.
1384 abandon_collection_set(g1_policy()->inc_cset_head());
1385 g1_policy()->clear_incremental_cset();
1386 g1_policy()->stop_incremental_cset_building();
1387
1388 tear_down_region_sets(false /* free_list_only */);
1389 g1_policy()->set_gcs_are_young(true);
1390
1391 // See the comments in g1CollectedHeap.hpp and
1392 // G1CollectedHeap::ref_processing_init() about
1393 // how reference processing currently works in G1.
1394
1395 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1396 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1397
1398 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1399 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1400
1401 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1402 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1403
1404 // Do collection work
1405 {
1406 HandleMark hm; // Discard invalid handles created during gc
1407 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1408 }
1409
1410 assert(free_regions() == 0, "we should not have added any free regions");
1411 rebuild_region_sets(false /* free_list_only */);
1412
1413 // Enqueue any discovered reference objects that have
1414 // not been removed from the discovered lists.
1415 ref_processor_stw()->enqueue_discovered_references();
1416
1417 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1418
1419 MemoryService::track_memory_usage();
1420
1421 verify_after_gc();
1422
1423 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1424 ref_processor_stw()->verify_no_references_recorded();
1425
1426 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1427 ClassLoaderDataGraph::purge();
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, "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_deleted = 0;
1847 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted);
1848 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1849 assert(mr.end() == old_end, "post-condition");
1850
1851 ergo_verbose3(ErgoHeapSizing,
1852 "shrink the heap",
1853 ergo_format_byte("requested shrinking amount")
1854 ergo_format_byte("aligned shrinking amount")
1855 ergo_format_byte("attempted shrinking amount"),
1856 shrink_bytes, aligned_shrink_bytes, mr.byte_size());
1857 if (mr.byte_size() > 0) {
1858 if (_hr_printer.is_active()) {
1859 HeapWord* curr = mr.end();
1860 while (curr > mr.start()) {
1861 HeapWord* curr_end = curr;
1862 curr -= HeapRegion::GrainWords;
1863 _hr_printer.uncommit(curr, curr_end);
1864 }
1865 assert(curr == mr.start(), "post-condition");
1866 }
1867
1868 _g1_storage.shrink_by(mr.byte_size());
1869 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1870 assert(mr.start() == new_end, "post-condition");
1871
1872 _expansion_regions += num_regions_deleted;
1873 update_committed_space(old_end, new_end);
1874 HeapRegionRemSet::shrink_heap(n_regions());
1875 g1_policy()->record_new_heap_size(n_regions());
1876 } else {
1877 ergo_verbose0(ErgoHeapSizing,
1878 "did not shrink the heap",
1879 ergo_format_reason("heap shrinking operation failed"));
1880 }
1881 }
1882
1883 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1884 verify_region_sets_optional();
1885
1886 // We should only reach here at the end of a Full GC which means we
1887 // should not not be holding to any GC alloc regions. The method
1888 // below will make sure of that and do any remaining clean up.
1889 abandon_gc_alloc_regions();
1890
1891 // Instead of tearing down / rebuilding the free lists here, we
1892 // could instead use the remove_all_pending() method on free_list to
1893 // remove only the ones that we need to remove.
1894 tear_down_region_sets(true /* free_list_only */);
1895 shrink_helper(shrink_bytes);
1896 rebuild_region_sets(true /* free_list_only */);
1897
1898 _hrs.verify_optional();
1899 verify_region_sets_optional();
1900 }
1901
1902 // Public methods.
1903
1904 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1905 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1906 #endif // _MSC_VER
1907
1908
1909 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1910 SharedHeap(policy_),
1911 _g1_policy(policy_),
1912 _dirty_card_queue_set(false),
1913 _into_cset_dirty_card_queue_set(false),
1914 _is_alive_closure_cm(this),
1915 _is_alive_closure_stw(this),
1916 _ref_processor_cm(NULL),
1917 _ref_processor_stw(NULL),
1918 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1919 _bot_shared(NULL),
1920 _evac_failure_scan_stack(NULL) ,
1921 _mark_in_progress(false),
1922 _cg1r(NULL), _summary_bytes_used(0),
1923 _g1mm(NULL),
1924 _refine_cte_cl(NULL),
1925 _full_collection(false),
1926 _free_list("Master Free List"),
1927 _secondary_free_list("Secondary Free List"),
1928 _old_set("Old Set"),
1929 _humongous_set("Master Humongous Set"),
1930 _free_regions_coming(false),
1931 _young_list(new YoungList(this)),
1932 _gc_time_stamp(0),
1933 _retained_old_gc_alloc_region(NULL),
1934 _survivor_plab_stats(YoungPLABSize, PLABWeight),
1935 _old_plab_stats(OldPLABSize, PLABWeight),
1936 _expand_heap_after_alloc_failure(true),
1937 _surviving_young_words(NULL),
1938 _old_marking_cycles_started(0),
1939 _old_marking_cycles_completed(0),
1940 _in_cset_fast_test(NULL),
1941 _in_cset_fast_test_base(NULL),
1942 _dirty_cards_region_list(NULL),
1943 _worker_cset_start_region(NULL),
1944 _worker_cset_start_region_time_stamp(NULL) {
1945 _g1h = this; // To catch bugs.
1946 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1947 vm_exit_during_initialization("Failed necessary allocation.");
1948 }
1949
1950 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1951
1952 int n_queues = MAX2((int)ParallelGCThreads, 1);
1953 _task_queues = new RefToScanQueueSet(n_queues);
1954
1955 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1956 assert(n_rem_sets > 0, "Invariant.");
1957
1958 HeapRegionRemSetIterator** iter_arr =
1959 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues, mtGC);
1960 for (int i = 0; i < n_queues; i++) {
1961 iter_arr[i] = new HeapRegionRemSetIterator();
1962 }
1963 _rem_set_iterator = iter_arr;
1964
1965 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC);
1966 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC);
1967
1968 for (int i = 0; i < n_queues; i++) {
1969 RefToScanQueue* q = new RefToScanQueue();
1970 q->initialize();
1971 _task_queues->register_queue(i, q);
1972 }
1973
1974 clear_cset_start_regions();
1975
1976 // Initialize the G1EvacuationFailureALot counters and flags.
1977 NOT_PRODUCT(reset_evacuation_should_fail();)
1978
1979 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1980 }
1981
1982 jint G1CollectedHeap::initialize() {
1983 CollectedHeap::pre_initialize();
1984 os::enable_vtime();
1985
1986 G1Log::init();
1987
1988 // Necessary to satisfy locking discipline assertions.
1989
1990 MutexLocker x(Heap_lock);
1991
1992 // We have to initialize the printer before committing the heap, as
1993 // it will be used then.
1994 _hr_printer.set_active(G1PrintHeapRegions);
1995
1996 // While there are no constraints in the GC code that HeapWordSize
1997 // be any particular value, there are multiple other areas in the
1998 // system which believe this to be true (e.g. oop->object_size in some
1999 // cases incorrectly returns the size in wordSize units rather than
2000 // HeapWordSize).
2001 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
2002
2003 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
2004 size_t max_byte_size = collector_policy()->max_heap_byte_size();
2005
2006 // Ensure that the sizes are properly aligned.
2007 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
2008 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
2009
2010 _cg1r = new ConcurrentG1Refine();
2011
2012 // Reserve the maximum.
2013
2014 // When compressed oops are enabled, the preferred heap base
2015 // is calculated by subtracting the requested size from the
2016 // 32Gb boundary and using the result as the base address for
2017 // heap reservation. If the requested size is not aligned to
2018 // HeapRegion::GrainBytes (i.e. the alignment that is passed
2019 // into the ReservedHeapSpace constructor) then the actual
2020 // base of the reserved heap may end up differing from the
2021 // address that was requested (i.e. the preferred heap base).
2022 // If this happens then we could end up using a non-optimal
2023 // compressed oops mode.
2024
2025 // Since max_byte_size is aligned to the size of a heap region (checked
2026 // above).
2027 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
2028
2029 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size,
2030 HeapRegion::GrainBytes);
2031
2032 // It is important to do this in a way such that concurrent readers can't
2033 // temporarily think somethings in the heap. (I've actually seen this
2034 // happen in asserts: DLD.)
2035 _reserved.set_word_size(0);
2036 _reserved.set_start((HeapWord*)heap_rs.base());
2037 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
2038
2039 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes);
2040
2041 // Create the gen rem set (and barrier set) for the entire reserved region.
2042 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
2043 set_barrier_set(rem_set()->bs());
2044 if (barrier_set()->is_a(BarrierSet::ModRef)) {
2045 _mr_bs = (ModRefBarrierSet*)_barrier_set;
2046 } else {
2047 vm_exit_during_initialization("G1 requires a mod ref bs.");
2048 return JNI_ENOMEM;
2049 }
2050
2051 // Also create a G1 rem set.
2052 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
2053 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
2054 } else {
2055 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
2056 return JNI_ENOMEM;
2057 }
2058
2059 // Carve out the G1 part of the heap.
2060
2061 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
2062 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
2063 g1_rs.size()/HeapWordSize);
2064
2065 _g1_storage.initialize(g1_rs, 0);
2066 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
2067 _hrs.initialize((HeapWord*) _g1_reserved.start(),
2068 (HeapWord*) _g1_reserved.end(),
2069 _expansion_regions);
2070
2071 // 6843694 - ensure that the maximum region index can fit
2072 // in the remembered set structures.
2073 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
2074 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
2075
2076 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
2077 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
2078 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
2079 "too many cards per region");
2080
2081 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
2082
2083 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
2084 heap_word_size(init_byte_size));
2085
2086 _g1h = this;
2087
2088 _in_cset_fast_test_length = max_regions();
2089 _in_cset_fast_test_base =
2090 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC);
2091
2092 // We're biasing _in_cset_fast_test to avoid subtracting the
2093 // beginning of the heap every time we want to index; basically
2094 // it's the same with what we do with the card table.
2095 _in_cset_fast_test = _in_cset_fast_test_base -
2096 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
2097
2098 // Clear the _cset_fast_test bitmap in anticipation of adding
2099 // regions to the incremental collection set for the first
2100 // evacuation pause.
2101 clear_cset_fast_test();
2102
2103 // Create the ConcurrentMark data structure and thread.
2104 // (Must do this late, so that "max_regions" is defined.)
2105 _cm = new ConcurrentMark(this, heap_rs);
2106 if (_cm == NULL || !_cm->completed_initialization()) {
2107 vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark");
2108 return JNI_ENOMEM;
2109 }
2110 _cmThread = _cm->cmThread();
2111
2112 // Initialize the from_card cache structure of HeapRegionRemSet.
2113 HeapRegionRemSet::init_heap(max_regions());
2114
2115 // Now expand into the initial heap size.
2116 if (!expand(init_byte_size)) {
2117 vm_shutdown_during_initialization("Failed to allocate initial heap.");
2118 return JNI_ENOMEM;
2119 }
2120
2121 // Perform any initialization actions delegated to the policy.
2122 g1_policy()->init();
2123
2124 _refine_cte_cl =
2125 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2126 g1_rem_set(),
2127 concurrent_g1_refine());
2128 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2129
2130 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2131 SATB_Q_FL_lock,
2132 G1SATBProcessCompletedThreshold,
2133 Shared_SATB_Q_lock);
2134
2135 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2136 DirtyCardQ_FL_lock,
2137 concurrent_g1_refine()->yellow_zone(),
2138 concurrent_g1_refine()->red_zone(),
2139 Shared_DirtyCardQ_lock);
2140
2141 if (G1DeferredRSUpdate) {
2142 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2143 DirtyCardQ_FL_lock,
2144 -1, // never trigger processing
2145 -1, // no limit on length
2146 Shared_DirtyCardQ_lock,
2147 &JavaThread::dirty_card_queue_set());
2148 }
2149
2150 // Initialize the card queue set used to hold cards containing
2151 // references into the collection set.
2152 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2153 DirtyCardQ_FL_lock,
2154 -1, // never trigger processing
2155 -1, // no limit on length
2156 Shared_DirtyCardQ_lock,
2157 &JavaThread::dirty_card_queue_set());
2158
2159 // In case we're keeping closure specialization stats, initialize those
2160 // counts and that mechanism.
2161 SpecializationStats::clear();
2162
2163 // Do later initialization work for concurrent refinement.
2164 _cg1r->init();
2165
2166 // Here we allocate the dummy full region that is required by the
2167 // G1AllocRegion class. If we don't pass an address in the reserved
2168 // space here, lots of asserts fire.
2169
2170 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2171 _g1_reserved.start());
2172 // We'll re-use the same region whether the alloc region will
2173 // require BOT updates or not and, if it doesn't, then a non-young
2174 // region will complain that it cannot support allocations without
2175 // BOT updates. So we'll tag the dummy region as young to avoid that.
2176 dummy_region->set_young();
2177 // Make sure it's full.
2178 dummy_region->set_top(dummy_region->end());
2179 G1AllocRegion::setup(this, dummy_region);
2180
2181 init_mutator_alloc_region();
2182
2183 // Do create of the monitoring and management support so that
2184 // values in the heap have been properly initialized.
2185 _g1mm = new G1MonitoringSupport(this);
2186
2187 return JNI_OK;
2188 }
2189
2190 void G1CollectedHeap::ref_processing_init() {
2191 // Reference processing in G1 currently works as follows:
2192 //
2193 // * There are two reference processor instances. One is
2194 // used to record and process discovered references
2195 // during concurrent marking; the other is used to
2196 // record and process references during STW pauses
2197 // (both full and incremental).
2198 // * Both ref processors need to 'span' the entire heap as
2199 // the regions in the collection set may be dotted around.
2200 //
2201 // * For the concurrent marking ref processor:
2202 // * Reference discovery is enabled at initial marking.
2203 // * Reference discovery is disabled and the discovered
2204 // references processed etc during remarking.
2205 // * Reference discovery is MT (see below).
2206 // * Reference discovery requires a barrier (see below).
2207 // * Reference processing may or may not be MT
2208 // (depending on the value of ParallelRefProcEnabled
2209 // and ParallelGCThreads).
2210 // * A full GC disables reference discovery by the CM
2211 // ref processor and abandons any entries on it's
2212 // discovered lists.
2213 //
2214 // * For the STW processor:
2215 // * Non MT discovery is enabled at the start of a full GC.
2216 // * Processing and enqueueing during a full GC is non-MT.
2217 // * During a full GC, references are processed after marking.
2218 //
2219 // * Discovery (may or may not be MT) is enabled at the start
2220 // of an incremental evacuation pause.
2221 // * References are processed near the end of a STW evacuation pause.
2222 // * For both types of GC:
2223 // * Discovery is atomic - i.e. not concurrent.
2224 // * Reference discovery will not need a barrier.
2225
2226 SharedHeap::ref_processing_init();
2227 MemRegion mr = reserved_region();
2228
2229 // Concurrent Mark ref processor
2230 _ref_processor_cm =
2231 new ReferenceProcessor(mr, // span
2232 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2233 // mt processing
2234 (int) ParallelGCThreads,
2235 // degree of mt processing
2236 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2237 // mt discovery
2238 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2239 // degree of mt discovery
2240 false,
2241 // Reference discovery is not atomic
2242 &_is_alive_closure_cm,
2243 // is alive closure
2244 // (for efficiency/performance)
2245 true);
2246 // Setting next fields of discovered
2247 // lists requires a barrier.
2248
2249 // STW ref processor
2250 _ref_processor_stw =
2251 new ReferenceProcessor(mr, // span
2252 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2253 // mt processing
2254 MAX2((int)ParallelGCThreads, 1),
2255 // degree of mt processing
2256 (ParallelGCThreads > 1),
2257 // mt discovery
2258 MAX2((int)ParallelGCThreads, 1),
2259 // degree of mt discovery
2260 true,
2261 // Reference discovery is atomic
2262 &_is_alive_closure_stw,
2263 // is alive closure
2264 // (for efficiency/performance)
2265 false);
2266 // Setting next fields of discovered
2267 // lists requires a barrier.
2268 }
2269
2270 size_t G1CollectedHeap::capacity() const {
2271 return _g1_committed.byte_size();
2272 }
2273
2274 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) {
2275 assert(!hr->continuesHumongous(), "pre-condition");
2276 hr->reset_gc_time_stamp();
2277 if (hr->startsHumongous()) {
2278 uint first_index = hr->hrs_index() + 1;
2279 uint last_index = hr->last_hc_index();
2280 for (uint i = first_index; i < last_index; i += 1) {
2281 HeapRegion* chr = region_at(i);
2282 assert(chr->continuesHumongous(), "sanity");
2283 chr->reset_gc_time_stamp();
2284 }
2285 }
2286 }
2287
2288 #ifndef PRODUCT
2289 class CheckGCTimeStampsHRClosure : public HeapRegionClosure {
2290 private:
2291 unsigned _gc_time_stamp;
2292 bool _failures;
2293
2294 public:
2295 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) :
2296 _gc_time_stamp(gc_time_stamp), _failures(false) { }
2297
2298 virtual bool doHeapRegion(HeapRegion* hr) {
2299 unsigned region_gc_time_stamp = hr->get_gc_time_stamp();
2300 if (_gc_time_stamp != region_gc_time_stamp) {
2301 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, "
2302 "expected %d", HR_FORMAT_PARAMS(hr),
2303 region_gc_time_stamp, _gc_time_stamp);
2304 _failures = true;
2305 }
2306 return false;
2307 }
2308
2309 bool failures() { return _failures; }
2310 };
2311
2312 void G1CollectedHeap::check_gc_time_stamps() {
2313 CheckGCTimeStampsHRClosure cl(_gc_time_stamp);
2314 heap_region_iterate(&cl);
2315 guarantee(!cl.failures(), "all GC time stamps should have been reset");
2316 }
2317 #endif // PRODUCT
2318
2319 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2320 DirtyCardQueue* into_cset_dcq,
2321 bool concurrent,
2322 int worker_i) {
2323 // Clean cards in the hot card cache
2324 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq);
2325
2326 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2327 int n_completed_buffers = 0;
2328 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2329 n_completed_buffers++;
2330 }
2331 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers);
2332 dcqs.clear_n_completed_buffers();
2333 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2334 }
2335
2336
2337 // Computes the sum of the storage used by the various regions.
2338
2339 size_t G1CollectedHeap::used() const {
2340 assert(Heap_lock->owner() != NULL,
2341 "Should be owned on this thread's behalf.");
2342 size_t result = _summary_bytes_used;
2343 // Read only once in case it is set to NULL concurrently
2344 HeapRegion* hr = _mutator_alloc_region.get();
2345 if (hr != NULL)
2346 result += hr->used();
2347 return result;
2348 }
2349
2350 size_t G1CollectedHeap::used_unlocked() const {
2351 size_t result = _summary_bytes_used;
2352 return result;
2353 }
2354
2355 class SumUsedClosure: public HeapRegionClosure {
2356 size_t _used;
2357 public:
2358 SumUsedClosure() : _used(0) {}
2359 bool doHeapRegion(HeapRegion* r) {
2360 if (!r->continuesHumongous()) {
2361 _used += r->used();
2362 }
2363 return false;
2364 }
2365 size_t result() { return _used; }
2366 };
2367
2368 size_t G1CollectedHeap::recalculate_used() const {
2369 SumUsedClosure blk;
2370 heap_region_iterate(&blk);
2371 return blk.result();
2372 }
2373
2374 size_t G1CollectedHeap::unsafe_max_alloc() {
2375 if (free_regions() > 0) return HeapRegion::GrainBytes;
2376 // otherwise, is there space in the current allocation region?
2377
2378 // We need to store the current allocation region in a local variable
2379 // here. The problem is that this method doesn't take any locks and
2380 // there may be other threads which overwrite the current allocation
2381 // region field. attempt_allocation(), for example, sets it to NULL
2382 // and this can happen *after* the NULL check here but before the call
2383 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2384 // to be a problem in the optimized build, since the two loads of the
2385 // current allocation region field are optimized away.
2386 HeapRegion* hr = _mutator_alloc_region.get();
2387 if (hr == NULL) {
2388 return 0;
2389 }
2390 return hr->free();
2391 }
2392
2393 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2394 switch (cause) {
2395 case GCCause::_gc_locker: return GCLockerInvokesConcurrent;
2396 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent;
2397 case GCCause::_g1_humongous_allocation: return true;
2398 default: return false;
2399 }
2400 }
2401
2402 #ifndef PRODUCT
2403 void G1CollectedHeap::allocate_dummy_regions() {
2404 // Let's fill up most of the region
2405 size_t word_size = HeapRegion::GrainWords - 1024;
2406 // And as a result the region we'll allocate will be humongous.
2407 guarantee(isHumongous(word_size), "sanity");
2408
2409 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2410 // Let's use the existing mechanism for the allocation
2411 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2412 if (dummy_obj != NULL) {
2413 MemRegion mr(dummy_obj, word_size);
2414 CollectedHeap::fill_with_object(mr);
2415 } else {
2416 // If we can't allocate once, we probably cannot allocate
2417 // again. Let's get out of the loop.
2418 break;
2419 }
2420 }
2421 }
2422 #endif // !PRODUCT
2423
2424 void G1CollectedHeap::increment_old_marking_cycles_started() {
2425 assert(_old_marking_cycles_started == _old_marking_cycles_completed ||
2426 _old_marking_cycles_started == _old_marking_cycles_completed + 1,
2427 err_msg("Wrong marking cycle count (started: %d, completed: %d)",
2428 _old_marking_cycles_started, _old_marking_cycles_completed));
2429
2430 _old_marking_cycles_started++;
2431 }
2432
2433 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) {
2434 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2435
2436 // We assume that if concurrent == true, then the caller is a
2437 // concurrent thread that was joined the Suspendible Thread
2438 // Set. If there's ever a cheap way to check this, we should add an
2439 // assert here.
2440
2441 // Given that this method is called at the end of a Full GC or of a
2442 // concurrent cycle, and those can be nested (i.e., a Full GC can
2443 // interrupt a concurrent cycle), the number of full collections
2444 // completed should be either one (in the case where there was no
2445 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2446 // behind the number of full collections started.
2447
2448 // This is the case for the inner caller, i.e. a Full GC.
2449 assert(concurrent ||
2450 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) ||
2451 (_old_marking_cycles_started == _old_marking_cycles_completed + 2),
2452 err_msg("for inner caller (Full GC): _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 // This is the case for the outer caller, i.e. the concurrent cycle.
2457 assert(!concurrent ||
2458 (_old_marking_cycles_started == _old_marking_cycles_completed + 1),
2459 err_msg("for outer caller (concurrent cycle): "
2460 "_old_marking_cycles_started = %u "
2461 "is inconsistent with _old_marking_cycles_completed = %u",
2462 _old_marking_cycles_started, _old_marking_cycles_completed));
2463
2464 _old_marking_cycles_completed += 1;
2465
2466 // We need to clear the "in_progress" flag in the CM thread before
2467 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2468 // is set) so that if a waiter requests another System.gc() it doesn't
2469 // incorrectly see that a marking cyle is still in progress.
2470 if (concurrent) {
2471 _cmThread->clear_in_progress();
2472 }
2473
2474 // This notify_all() will ensure that a thread that called
2475 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2476 // and it's waiting for a full GC to finish will be woken up. It is
2477 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2478 FullGCCount_lock->notify_all();
2479 }
2480
2481 void G1CollectedHeap::collect(GCCause::Cause cause) {
2482 assert_heap_not_locked();
2483
2484 unsigned int gc_count_before;
2485 unsigned int old_marking_count_before;
2486 bool retry_gc;
2487
2488 do {
2489 retry_gc = false;
2490
2491 {
2492 MutexLocker ml(Heap_lock);
2493
2494 // Read the GC count while holding the Heap_lock
2495 gc_count_before = total_collections();
2496 old_marking_count_before = _old_marking_cycles_started;
2497 }
2498
2499 if (should_do_concurrent_full_gc(cause)) {
2500 // Schedule an initial-mark evacuation pause that will start a
2501 // concurrent cycle. We're setting word_size to 0 which means that
2502 // we are not requesting a post-GC allocation.
2503 VM_G1IncCollectionPause op(gc_count_before,
2504 0, /* word_size */
2505 true, /* should_initiate_conc_mark */
2506 g1_policy()->max_pause_time_ms(),
2507 cause);
2508
2509 VMThread::execute(&op);
2510 if (!op.pause_succeeded()) {
2511 if (old_marking_count_before == _old_marking_cycles_started) {
2512 retry_gc = op.should_retry_gc();
2513 } else {
2514 // A Full GC happened while we were trying to schedule the
2515 // initial-mark GC. No point in starting a new cycle given
2516 // that the whole heap was collected anyway.
2517 }
2518
2519 if (retry_gc) {
2520 if (GC_locker::is_active_and_needs_gc()) {
2521 GC_locker::stall_until_clear();
2522 }
2523 }
2524 }
2525 } else {
2526 if (cause == GCCause::_gc_locker
2527 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2528
2529 // Schedule a standard evacuation pause. We're setting word_size
2530 // to 0 which means that we are not requesting a post-GC allocation.
2531 VM_G1IncCollectionPause op(gc_count_before,
2532 0, /* word_size */
2533 false, /* should_initiate_conc_mark */
2534 g1_policy()->max_pause_time_ms(),
2535 cause);
2536 VMThread::execute(&op);
2537 } else {
2538 // Schedule a Full GC.
2539 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause);
2540 VMThread::execute(&op);
2541 }
2542 }
2543 } while (retry_gc);
2544 }
2545
2546 bool G1CollectedHeap::is_in(const void* p) const {
2547 if (_g1_committed.contains(p)) {
2548 // Given that we know that p is in the committed space,
2549 // heap_region_containing_raw() should successfully
2550 // return the containing region.
2551 HeapRegion* hr = heap_region_containing_raw(p);
2552 return hr->is_in(p);
2553 } else {
2554 return false;
2555 }
2556 }
2557
2558 // Iteration functions.
2559
2560 // Iterates an OopClosure over all ref-containing fields of objects
2561 // within a HeapRegion.
2562
2563 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2564 MemRegion _mr;
2565 ExtendedOopClosure* _cl;
2566 public:
2567 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl)
2568 : _mr(mr), _cl(cl) {}
2569 bool doHeapRegion(HeapRegion* r) {
2570 if (!r->continuesHumongous()) {
2571 r->oop_iterate(_cl);
2572 }
2573 return false;
2574 }
2575 };
2576
2577 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) {
2578 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2579 heap_region_iterate(&blk);
2580 }
2581
2582 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
2583 IterateOopClosureRegionClosure blk(mr, cl);
2584 heap_region_iterate(&blk);
2585 }
2586
2587 // Iterates an ObjectClosure over all objects within a HeapRegion.
2588
2589 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2590 ObjectClosure* _cl;
2591 public:
2592 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2593 bool doHeapRegion(HeapRegion* r) {
2594 if (! r->continuesHumongous()) {
2595 r->object_iterate(_cl);
2596 }
2597 return false;
2598 }
2599 };
2600
2601 void G1CollectedHeap::object_iterate(ObjectClosure* cl) {
2602 IterateObjectClosureRegionClosure blk(cl);
2603 heap_region_iterate(&blk);
2604 }
2605
2606 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2607 // FIXME: is this right?
2608 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2609 }
2610
2611 // Calls a SpaceClosure on a HeapRegion.
2612
2613 class SpaceClosureRegionClosure: public HeapRegionClosure {
2614 SpaceClosure* _cl;
2615 public:
2616 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2617 bool doHeapRegion(HeapRegion* r) {
2618 _cl->do_space(r);
2619 return false;
2620 }
2621 };
2622
2623 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2624 SpaceClosureRegionClosure blk(cl);
2625 heap_region_iterate(&blk);
2626 }
2627
2628 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2629 _hrs.iterate(cl);
2630 }
2631
2632 void
2633 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2634 uint worker_id,
2635 uint no_of_par_workers,
2636 jint claim_value) {
2637 const uint regions = n_regions();
2638 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
2639 no_of_par_workers :
2640 1);
2641 assert(UseDynamicNumberOfGCThreads ||
2642 no_of_par_workers == workers()->total_workers(),
2643 "Non dynamic should use fixed number of workers");
2644 // try to spread out the starting points of the workers
2645 const HeapRegion* start_hr =
2646 start_region_for_worker(worker_id, no_of_par_workers);
2647 const uint start_index = start_hr->hrs_index();
2648
2649 // each worker will actually look at all regions
2650 for (uint count = 0; count < regions; ++count) {
2651 const uint index = (start_index + count) % regions;
2652 assert(0 <= index && index < regions, "sanity");
2653 HeapRegion* r = region_at(index);
2654 // we'll ignore "continues humongous" regions (we'll process them
2655 // when we come across their corresponding "start humongous"
2656 // region) and regions already claimed
2657 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2658 continue;
2659 }
2660 // OK, try to claim it
2661 if (r->claimHeapRegion(claim_value)) {
2662 // success!
2663 assert(!r->continuesHumongous(), "sanity");
2664 if (r->startsHumongous()) {
2665 // If the region is "starts humongous" we'll iterate over its
2666 // "continues humongous" first; in fact we'll do them
2667 // first. The order is important. In on case, calling the
2668 // closure on the "starts humongous" region might de-allocate
2669 // and clear all its "continues humongous" regions and, as a
2670 // result, we might end up processing them twice. So, we'll do
2671 // them first (notice: most closures will ignore them anyway) and
2672 // then we'll do the "starts humongous" region.
2673 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) {
2674 HeapRegion* chr = region_at(ch_index);
2675
2676 // if the region has already been claimed or it's not
2677 // "continues humongous" we're done
2678 if (chr->claim_value() == claim_value ||
2679 !chr->continuesHumongous()) {
2680 break;
2681 }
2682
2683 // Noone should have claimed it directly. We can given
2684 // that we claimed its "starts humongous" region.
2685 assert(chr->claim_value() != claim_value, "sanity");
2686 assert(chr->humongous_start_region() == r, "sanity");
2687
2688 if (chr->claimHeapRegion(claim_value)) {
2689 // we should always be able to claim it; noone else should
2690 // be trying to claim this region
2691
2692 bool res2 = cl->doHeapRegion(chr);
2693 assert(!res2, "Should not abort");
2694
2695 // Right now, this holds (i.e., no closure that actually
2696 // does something with "continues humongous" regions
2697 // clears them). We might have to weaken it in the future,
2698 // but let's leave these two asserts here for extra safety.
2699 assert(chr->continuesHumongous(), "should still be the case");
2700 assert(chr->humongous_start_region() == r, "sanity");
2701 } else {
2702 guarantee(false, "we should not reach here");
2703 }
2704 }
2705 }
2706
2707 assert(!r->continuesHumongous(), "sanity");
2708 bool res = cl->doHeapRegion(r);
2709 assert(!res, "Should not abort");
2710 }
2711 }
2712 }
2713
2714 class ResetClaimValuesClosure: public HeapRegionClosure {
2715 public:
2716 bool doHeapRegion(HeapRegion* r) {
2717 r->set_claim_value(HeapRegion::InitialClaimValue);
2718 return false;
2719 }
2720 };
2721
2722 void G1CollectedHeap::reset_heap_region_claim_values() {
2723 ResetClaimValuesClosure blk;
2724 heap_region_iterate(&blk);
2725 }
2726
2727 void G1CollectedHeap::reset_cset_heap_region_claim_values() {
2728 ResetClaimValuesClosure blk;
2729 collection_set_iterate(&blk);
2730 }
2731
2732 #ifdef ASSERT
2733 // This checks whether all regions in the heap have the correct claim
2734 // value. I also piggy-backed on this a check to ensure that the
2735 // humongous_start_region() information on "continues humongous"
2736 // regions is correct.
2737
2738 class CheckClaimValuesClosure : public HeapRegionClosure {
2739 private:
2740 jint _claim_value;
2741 uint _failures;
2742 HeapRegion* _sh_region;
2743
2744 public:
2745 CheckClaimValuesClosure(jint claim_value) :
2746 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2747 bool doHeapRegion(HeapRegion* r) {
2748 if (r->claim_value() != _claim_value) {
2749 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2750 "claim value = %d, should be %d",
2751 HR_FORMAT_PARAMS(r),
2752 r->claim_value(), _claim_value);
2753 ++_failures;
2754 }
2755 if (!r->isHumongous()) {
2756 _sh_region = NULL;
2757 } else if (r->startsHumongous()) {
2758 _sh_region = r;
2759 } else if (r->continuesHumongous()) {
2760 if (r->humongous_start_region() != _sh_region) {
2761 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2762 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2763 HR_FORMAT_PARAMS(r),
2764 r->humongous_start_region(),
2765 _sh_region);
2766 ++_failures;
2767 }
2768 }
2769 return false;
2770 }
2771 uint failures() { return _failures; }
2772 };
2773
2774 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2775 CheckClaimValuesClosure cl(claim_value);
2776 heap_region_iterate(&cl);
2777 return cl.failures() == 0;
2778 }
2779
2780 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure {
2781 private:
2782 jint _claim_value;
2783 uint _failures;
2784
2785 public:
2786 CheckClaimValuesInCSetHRClosure(jint claim_value) :
2787 _claim_value(claim_value), _failures(0) { }
2788
2789 uint failures() { return _failures; }
2790
2791 bool doHeapRegion(HeapRegion* hr) {
2792 assert(hr->in_collection_set(), "how?");
2793 assert(!hr->isHumongous(), "H-region in CSet");
2794 if (hr->claim_value() != _claim_value) {
2795 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", "
2796 "claim value = %d, should be %d",
2797 HR_FORMAT_PARAMS(hr),
2798 hr->claim_value(), _claim_value);
2799 _failures += 1;
2800 }
2801 return false;
2802 }
2803 };
2804
2805 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) {
2806 CheckClaimValuesInCSetHRClosure cl(claim_value);
2807 collection_set_iterate(&cl);
2808 return cl.failures() == 0;
2809 }
2810 #endif // ASSERT
2811
2812 // Clear the cached CSet starting regions and (more importantly)
2813 // the time stamps. Called when we reset the GC time stamp.
2814 void G1CollectedHeap::clear_cset_start_regions() {
2815 assert(_worker_cset_start_region != NULL, "sanity");
2816 assert(_worker_cset_start_region_time_stamp != NULL, "sanity");
2817
2818 int n_queues = MAX2((int)ParallelGCThreads, 1);
2819 for (int i = 0; i < n_queues; i++) {
2820 _worker_cset_start_region[i] = NULL;
2821 _worker_cset_start_region_time_stamp[i] = 0;
2822 }
2823 }
2824
2825 // Given the id of a worker, obtain or calculate a suitable
2826 // starting region for iterating over the current collection set.
2827 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) {
2828 assert(get_gc_time_stamp() > 0, "should have been updated by now");
2829
2830 HeapRegion* result = NULL;
2831 unsigned gc_time_stamp = get_gc_time_stamp();
2832
2833 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) {
2834 // Cached starting region for current worker was set
2835 // during the current pause - so it's valid.
2836 // Note: the cached starting heap region may be NULL
2837 // (when the collection set is empty).
2838 result = _worker_cset_start_region[worker_i];
2839 assert(result == NULL || result->in_collection_set(), "sanity");
2840 return result;
2841 }
2842
2843 // The cached entry was not valid so let's calculate
2844 // a suitable starting heap region for this worker.
2845
2846 // We want the parallel threads to start their collection
2847 // set iteration at different collection set regions to
2848 // avoid contention.
2849 // If we have:
2850 // n collection set regions
2851 // p threads
2852 // Then thread t will start at region floor ((t * n) / p)
2853
2854 result = g1_policy()->collection_set();
2855 if (G1CollectedHeap::use_parallel_gc_threads()) {
2856 uint cs_size = g1_policy()->cset_region_length();
2857 uint active_workers = workers()->active_workers();
2858 assert(UseDynamicNumberOfGCThreads ||
2859 active_workers == workers()->total_workers(),
2860 "Unless dynamic should use total workers");
2861
2862 uint end_ind = (cs_size * worker_i) / active_workers;
2863 uint start_ind = 0;
2864
2865 if (worker_i > 0 &&
2866 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) {
2867 // Previous workers starting region is valid
2868 // so let's iterate from there
2869 start_ind = (cs_size * (worker_i - 1)) / active_workers;
2870 result = _worker_cset_start_region[worker_i - 1];
2871 }
2872
2873 for (uint i = start_ind; i < end_ind; i++) {
2874 result = result->next_in_collection_set();
2875 }
2876 }
2877
2878 // Note: the calculated starting heap region may be NULL
2879 // (when the collection set is empty).
2880 assert(result == NULL || result->in_collection_set(), "sanity");
2881 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp,
2882 "should be updated only once per pause");
2883 _worker_cset_start_region[worker_i] = result;
2884 OrderAccess::storestore();
2885 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp;
2886 return result;
2887 }
2888
2889 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i,
2890 uint no_of_par_workers) {
2891 uint worker_num =
2892 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U;
2893 assert(UseDynamicNumberOfGCThreads ||
2894 no_of_par_workers == workers()->total_workers(),
2895 "Non dynamic should use fixed number of workers");
2896 const uint start_index = n_regions() * worker_i / worker_num;
2897 return region_at(start_index);
2898 }
2899
2900 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2901 HeapRegion* r = g1_policy()->collection_set();
2902 while (r != NULL) {
2903 HeapRegion* next = r->next_in_collection_set();
2904 if (cl->doHeapRegion(r)) {
2905 cl->incomplete();
2906 return;
2907 }
2908 r = next;
2909 }
2910 }
2911
2912 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2913 HeapRegionClosure *cl) {
2914 if (r == NULL) {
2915 // The CSet is empty so there's nothing to do.
2916 return;
2917 }
2918
2919 assert(r->in_collection_set(),
2920 "Start region must be a member of the collection set.");
2921 HeapRegion* cur = r;
2922 while (cur != NULL) {
2923 HeapRegion* next = cur->next_in_collection_set();
2924 if (cl->doHeapRegion(cur) && false) {
2925 cl->incomplete();
2926 return;
2927 }
2928 cur = next;
2929 }
2930 cur = g1_policy()->collection_set();
2931 while (cur != r) {
2932 HeapRegion* next = cur->next_in_collection_set();
2933 if (cl->doHeapRegion(cur) && false) {
2934 cl->incomplete();
2935 return;
2936 }
2937 cur = next;
2938 }
2939 }
2940
2941 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2942 return n_regions() > 0 ? region_at(0) : NULL;
2943 }
2944
2945
2946 Space* G1CollectedHeap::space_containing(const void* addr) const {
2947 Space* res = heap_region_containing(addr);
2948 return res;
2949 }
2950
2951 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2952 Space* sp = space_containing(addr);
2953 if (sp != NULL) {
2954 return sp->block_start(addr);
2955 }
2956 return NULL;
2957 }
2958
2959 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2960 Space* sp = space_containing(addr);
2961 assert(sp != NULL, "block_size of address outside of heap");
2962 return sp->block_size(addr);
2963 }
2964
2965 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2966 Space* sp = space_containing(addr);
2967 return sp->block_is_obj(addr);
2968 }
2969
2970 bool G1CollectedHeap::supports_tlab_allocation() const {
2971 return true;
2972 }
2973
2974 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2975 return HeapRegion::GrainBytes;
2976 }
2977
2978 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2979 // Return the remaining space in the cur alloc region, but not less than
2980 // the min TLAB size.
2981
2982 // Also, this value can be at most the humongous object threshold,
2983 // since we can't allow tlabs to grow big enough to accomodate
2984 // humongous objects.
2985
2986 HeapRegion* hr = _mutator_alloc_region.get();
2987 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2988 if (hr == NULL) {
2989 return max_tlab_size;
2990 } else {
2991 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2992 }
2993 }
2994
2995 size_t G1CollectedHeap::max_capacity() const {
2996 return _g1_reserved.byte_size();
2997 }
2998
2999 jlong G1CollectedHeap::millis_since_last_gc() {
3000 // assert(false, "NYI");
3001 return 0;
3002 }
3003
3004 void G1CollectedHeap::prepare_for_verify() {
3005 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
3006 ensure_parsability(false);
3007 }
3008 g1_rem_set()->prepare_for_verify();
3009 }
3010
3011 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr,
3012 VerifyOption vo) {
3013 switch (vo) {
3014 case VerifyOption_G1UsePrevMarking:
3015 return hr->obj_allocated_since_prev_marking(obj);
3016 case VerifyOption_G1UseNextMarking:
3017 return hr->obj_allocated_since_next_marking(obj);
3018 case VerifyOption_G1UseMarkWord:
3019 return false;
3020 default:
3021 ShouldNotReachHere();
3022 }
3023 return false; // keep some compilers happy
3024 }
3025
3026 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) {
3027 switch (vo) {
3028 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start();
3029 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start();
3030 case VerifyOption_G1UseMarkWord: return NULL;
3031 default: ShouldNotReachHere();
3032 }
3033 return NULL; // keep some compilers happy
3034 }
3035
3036 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) {
3037 switch (vo) {
3038 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj);
3039 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj);
3040 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked();
3041 default: ShouldNotReachHere();
3042 }
3043 return false; // keep some compilers happy
3044 }
3045
3046 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) {
3047 switch (vo) {
3048 case VerifyOption_G1UsePrevMarking: return "PTAMS";
3049 case VerifyOption_G1UseNextMarking: return "NTAMS";
3050 case VerifyOption_G1UseMarkWord: return "NONE";
3051 default: ShouldNotReachHere();
3052 }
3053 return NULL; // keep some compilers happy
3054 }
3055
3056 class VerifyLivenessOopClosure: public OopClosure {
3057 G1CollectedHeap* _g1h;
3058 VerifyOption _vo;
3059 public:
3060 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
3061 _g1h(g1h), _vo(vo)
3062 { }
3063 void do_oop(narrowOop *p) { do_oop_work(p); }
3064 void do_oop( oop *p) { do_oop_work(p); }
3065
3066 template <class T> void do_oop_work(T *p) {
3067 oop obj = oopDesc::load_decode_heap_oop(p);
3068 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
3069 "Dead object referenced by a not dead object");
3070 }
3071 };
3072
3073 class VerifyObjsInRegionClosure: public ObjectClosure {
3074 private:
3075 G1CollectedHeap* _g1h;
3076 size_t _live_bytes;
3077 HeapRegion *_hr;
3078 VerifyOption _vo;
3079 public:
3080 // _vo == UsePrevMarking -> use "prev" marking information,
3081 // _vo == UseNextMarking -> use "next" marking information,
3082 // _vo == UseMarkWord -> use mark word from object header.
3083 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
3084 : _live_bytes(0), _hr(hr), _vo(vo) {
3085 _g1h = G1CollectedHeap::heap();
3086 }
3087 void do_object(oop o) {
3088 VerifyLivenessOopClosure isLive(_g1h, _vo);
3089 assert(o != NULL, "Huh?");
3090 if (!_g1h->is_obj_dead_cond(o, _vo)) {
3091 // If the object is alive according to the mark word,
3092 // then verify that the marking information agrees.
3093 // Note we can't verify the contra-positive of the
3094 // above: if the object is dead (according to the mark
3095 // word), it may not be marked, or may have been marked
3096 // but has since became dead, or may have been allocated
3097 // since the last marking.
3098 if (_vo == VerifyOption_G1UseMarkWord) {
3099 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
3100 }
3101
3102 o->oop_iterate_no_header(&isLive);
3103 if (!_hr->obj_allocated_since_prev_marking(o)) {
3104 size_t obj_size = o->size(); // Make sure we don't overflow
3105 _live_bytes += (obj_size * HeapWordSize);
3106 }
3107 }
3108 }
3109 size_t live_bytes() { return _live_bytes; }
3110 };
3111
3112 class PrintObjsInRegionClosure : public ObjectClosure {
3113 HeapRegion *_hr;
3114 G1CollectedHeap *_g1;
3115 public:
3116 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
3117 _g1 = G1CollectedHeap::heap();
3118 };
3119
3120 void do_object(oop o) {
3121 if (o != NULL) {
3122 HeapWord *start = (HeapWord *) o;
3123 size_t word_sz = o->size();
3124 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
3125 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
3126 (void*) o, word_sz,
3127 _g1->isMarkedPrev(o),
3128 _g1->isMarkedNext(o),
3129 _hr->obj_allocated_since_prev_marking(o));
3130 HeapWord *end = start + word_sz;
3131 HeapWord *cur;
3132 int *val;
3133 for (cur = start; cur < end; cur++) {
3134 val = (int *) cur;
3135 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
3136 }
3137 }
3138 }
3139 };
3140
3141 class VerifyRegionClosure: public HeapRegionClosure {
3142 private:
3143 bool _par;
3144 VerifyOption _vo;
3145 bool _failures;
3146 public:
3147 // _vo == UsePrevMarking -> use "prev" marking information,
3148 // _vo == UseNextMarking -> use "next" marking information,
3149 // _vo == UseMarkWord -> use mark word from object header.
3150 VerifyRegionClosure(bool par, VerifyOption vo)
3151 : _par(par),
3152 _vo(vo),
3153 _failures(false) {}
3154
3155 bool failures() {
3156 return _failures;
3157 }
3158
3159 bool doHeapRegion(HeapRegion* r) {
3160 if (!r->continuesHumongous()) {
3161 bool failures = false;
3162 r->verify(_vo, &failures);
3163 if (failures) {
3164 _failures = true;
3165 } else {
3166 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
3167 r->object_iterate(¬_dead_yet_cl);
3168 if (_vo != VerifyOption_G1UseNextMarking) {
3169 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
3170 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
3171 "max_live_bytes "SIZE_FORMAT" "
3172 "< calculated "SIZE_FORMAT,
3173 r->bottom(), r->end(),
3174 r->max_live_bytes(),
3175 not_dead_yet_cl.live_bytes());
3176 _failures = true;
3177 }
3178 } else {
3179 // When vo == UseNextMarking we cannot currently do a sanity
3180 // check on the live bytes as the calculation has not been
3181 // finalized yet.
3182 }
3183 }
3184 }
3185 return false; // stop the region iteration if we hit a failure
3186 }
3187 };
3188
3189 class YoungRefCounterClosure : public OopClosure {
3190 G1CollectedHeap* _g1h;
3191 int _count;
3192 public:
3193 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {}
3194 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } }
3195 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3196
3197 int count() { return _count; }
3198 void reset_count() { _count = 0; };
3199 };
3200
3201 class VerifyKlassClosure: public KlassClosure {
3202 YoungRefCounterClosure _young_ref_counter_closure;
3203 OopClosure *_oop_closure;
3204 public:
3205 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {}
3206 void do_klass(Klass* k) {
3207 k->oops_do(_oop_closure);
3208
3209 _young_ref_counter_closure.reset_count();
3210 k->oops_do(&_young_ref_counter_closure);
3211 if (_young_ref_counter_closure.count() > 0) {
3212 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k));
3213 }
3214 }
3215 };
3216
3217 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can
3218 // pass it as the perm_blk to SharedHeap::process_strong_roots.
3219 // When process_strong_roots stop calling perm_blk->younger_refs_iterate
3220 // we can change this closure to extend the simpler OopClosure.
3221 class VerifyRootsClosure: public OopsInGenClosure {
3222 private:
3223 G1CollectedHeap* _g1h;
3224 VerifyOption _vo;
3225 bool _failures;
3226 public:
3227 // _vo == UsePrevMarking -> use "prev" marking information,
3228 // _vo == UseNextMarking -> use "next" marking information,
3229 // _vo == UseMarkWord -> use mark word from object header.
3230 VerifyRootsClosure(VerifyOption vo) :
3231 _g1h(G1CollectedHeap::heap()),
3232 _vo(vo),
3233 _failures(false) { }
3234
3235 bool failures() { return _failures; }
3236
3237 template <class T> void do_oop_nv(T* p) {
3238 T heap_oop = oopDesc::load_heap_oop(p);
3239 if (!oopDesc::is_null(heap_oop)) {
3240 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
3241 if (_g1h->is_obj_dead_cond(obj, _vo)) {
3242 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
3243 "points to dead obj "PTR_FORMAT, p, (void*) obj);
3244 if (_vo == VerifyOption_G1UseMarkWord) {
3245 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
3246 }
3247 obj->print_on(gclog_or_tty);
3248 _failures = true;
3249 }
3250 }
3251 }
3252
3253 void do_oop(oop* p) { do_oop_nv(p); }
3254 void do_oop(narrowOop* p) { do_oop_nv(p); }
3255 };
3256
3257 // This is the task used for parallel heap verification.
3258
3259 class G1ParVerifyTask: public AbstractGangTask {
3260 private:
3261 G1CollectedHeap* _g1h;
3262 VerifyOption _vo;
3263 bool _failures;
3264
3265 public:
3266 // _vo == UsePrevMarking -> use "prev" marking information,
3267 // _vo == UseNextMarking -> use "next" marking information,
3268 // _vo == UseMarkWord -> use mark word from object header.
3269 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) :
3270 AbstractGangTask("Parallel verify task"),
3271 _g1h(g1h),
3272 _vo(vo),
3273 _failures(false) { }
3274
3275 bool failures() {
3276 return _failures;
3277 }
3278
3279 void work(uint worker_id) {
3280 HandleMark hm;
3281 VerifyRegionClosure blk(true, _vo);
3282 _g1h->heap_region_par_iterate_chunked(&blk, worker_id,
3283 _g1h->workers()->active_workers(),
3284 HeapRegion::ParVerifyClaimValue);
3285 if (blk.failures()) {
3286 _failures = true;
3287 }
3288 }
3289 };
3290
3291 void G1CollectedHeap::verify(bool silent) {
3292 verify(silent, VerifyOption_G1UsePrevMarking);
3293 }
3294
3295 void G1CollectedHeap::verify(bool silent,
3296 VerifyOption vo) {
3297 if (SafepointSynchronize::is_at_safepoint()) {
3298 if (!silent) { gclog_or_tty->print("Roots "); }
3299 VerifyRootsClosure rootsCl(vo);
3300
3301 assert(Thread::current()->is_VM_thread(),
3302 "Expected to be executed serially by the VM thread at this point");
3303
3304 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
3305 VerifyKlassClosure klassCl(this, &rootsCl);
3306
3307 // We apply the relevant closures to all the oops in the
3308 // system dictionary, the string table and the code cache.
3309 const int so = SO_AllClasses | SO_Strings | SO_CodeCache;
3310
3311 // Need cleared claim bits for the strong roots processing
3312 ClassLoaderDataGraph::clear_claimed_marks();
3313
3314 process_strong_roots(true, // activate StrongRootsScope
3315 false, // we set "is scavenging" to false,
3316 // so we don't reset the dirty cards.
3317 ScanningOption(so), // roots scanning options
3318 &rootsCl,
3319 &blobsCl,
3320 &klassCl
3321 );
3322
3323 bool failures = rootsCl.failures();
3324
3325 if (vo != VerifyOption_G1UseMarkWord) {
3326 // If we're verifying during a full GC then the region sets
3327 // will have been torn down at the start of the GC. Therefore
3328 // verifying the region sets will fail. So we only verify
3329 // the region sets when not in a full GC.
3330 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
3331 verify_region_sets();
3332 }
3333
3334 if (!silent) { gclog_or_tty->print("HeapRegions "); }
3335 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
3336 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3337 "sanity check");
3338
3339 G1ParVerifyTask task(this, vo);
3340 assert(UseDynamicNumberOfGCThreads ||
3341 workers()->active_workers() == workers()->total_workers(),
3342 "If not dynamic should be using all the workers");
3343 int n_workers = workers()->active_workers();
3344 set_par_threads(n_workers);
3345 workers()->run_task(&task);
3346 set_par_threads(0);
3347 if (task.failures()) {
3348 failures = true;
3349 }
3350
3351 // Checks that the expected amount of parallel work was done.
3352 // The implication is that n_workers is > 0.
3353 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
3354 "sanity check");
3355
3356 reset_heap_region_claim_values();
3357
3358 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3359 "sanity check");
3360 } else {
3361 VerifyRegionClosure blk(false, vo);
3362 heap_region_iterate(&blk);
3363 if (blk.failures()) {
3364 failures = true;
3365 }
3366 }
3367 if (!silent) gclog_or_tty->print("RemSet ");
3368 rem_set()->verify();
3369
3370 if (failures) {
3371 gclog_or_tty->print_cr("Heap:");
3372 // It helps to have the per-region information in the output to
3373 // help us track down what went wrong. This is why we call
3374 // print_extended_on() instead of print_on().
3375 print_extended_on(gclog_or_tty);
3376 gclog_or_tty->print_cr("");
3377 #ifndef PRODUCT
3378 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3379 concurrent_mark()->print_reachable("at-verification-failure",
3380 vo, false /* all */);
3381 }
3382 #endif
3383 gclog_or_tty->flush();
3384 }
3385 guarantee(!failures, "there should not have been any failures");
3386 } else {
3387 if (!silent)
3388 gclog_or_tty->print("(SKIPPING roots, heapRegionSets, heapRegions, remset) ");
3389 }
3390 }
3391
3392 class PrintRegionClosure: public HeapRegionClosure {
3393 outputStream* _st;
3394 public:
3395 PrintRegionClosure(outputStream* st) : _st(st) {}
3396 bool doHeapRegion(HeapRegion* r) {
3397 r->print_on(_st);
3398 return false;
3399 }
3400 };
3401
3402 void G1CollectedHeap::print_on(outputStream* st) const {
3403 st->print(" %-20s", "garbage-first heap");
3404 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3405 capacity()/K, used_unlocked()/K);
3406 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3407 _g1_storage.low_boundary(),
3408 _g1_storage.high(),
3409 _g1_storage.high_boundary());
3410 st->cr();
3411 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3412 uint young_regions = _young_list->length();
3413 st->print("%u young (" SIZE_FORMAT "K), ", young_regions,
3414 (size_t) young_regions * HeapRegion::GrainBytes / K);
3415 uint survivor_regions = g1_policy()->recorded_survivor_regions();
3416 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions,
3417 (size_t) survivor_regions * HeapRegion::GrainBytes / K);
3418 st->cr();
3419 MetaspaceAux::print_on(st);
3420 }
3421
3422 void G1CollectedHeap::print_extended_on(outputStream* st) const {
3423 print_on(st);
3424
3425 // Print the per-region information.
3426 st->cr();
3427 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), "
3428 "HS=humongous(starts), HC=humongous(continues), "
3429 "CS=collection set, F=free, TS=gc time stamp, "
3430 "PTAMS=previous top-at-mark-start, "
3431 "NTAMS=next top-at-mark-start)");
3432 PrintRegionClosure blk(st);
3433 heap_region_iterate(&blk);
3434 }
3435
3436 void G1CollectedHeap::print_on_error(outputStream* st) const {
3437 this->CollectedHeap::print_on_error(st);
3438
3439 if (_cm != NULL) {
3440 st->cr();
3441 _cm->print_on_error(st);
3442 }
3443 }
3444
3445 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3446 if (G1CollectedHeap::use_parallel_gc_threads()) {
3447 workers()->print_worker_threads_on(st);
3448 }
3449 _cmThread->print_on(st);
3450 st->cr();
3451 _cm->print_worker_threads_on(st);
3452 _cg1r->print_worker_threads_on(st);
3453 }
3454
3455 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3456 if (G1CollectedHeap::use_parallel_gc_threads()) {
3457 workers()->threads_do(tc);
3458 }
3459 tc->do_thread(_cmThread);
3460 _cg1r->threads_do(tc);
3461 }
3462
3463 void G1CollectedHeap::print_tracing_info() const {
3464 // We'll overload this to mean "trace GC pause statistics."
3465 if (TraceGen0Time || TraceGen1Time) {
3466 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3467 // to that.
3468 g1_policy()->print_tracing_info();
3469 }
3470 if (G1SummarizeRSetStats) {
3471 g1_rem_set()->print_summary_info();
3472 }
3473 if (G1SummarizeConcMark) {
3474 concurrent_mark()->print_summary_info();
3475 }
3476 g1_policy()->print_yg_surv_rate_info();
3477 SpecializationStats::print();
3478 }
3479
3480 #ifndef PRODUCT
3481 // Helpful for debugging RSet issues.
3482
3483 class PrintRSetsClosure : public HeapRegionClosure {
3484 private:
3485 const char* _msg;
3486 size_t _occupied_sum;
3487
3488 public:
3489 bool doHeapRegion(HeapRegion* r) {
3490 HeapRegionRemSet* hrrs = r->rem_set();
3491 size_t occupied = hrrs->occupied();
3492 _occupied_sum += occupied;
3493
3494 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3495 HR_FORMAT_PARAMS(r));
3496 if (occupied == 0) {
3497 gclog_or_tty->print_cr(" RSet is empty");
3498 } else {
3499 hrrs->print();
3500 }
3501 gclog_or_tty->print_cr("----------");
3502 return false;
3503 }
3504
3505 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3506 gclog_or_tty->cr();
3507 gclog_or_tty->print_cr("========================================");
3508 gclog_or_tty->print_cr(msg);
3509 gclog_or_tty->cr();
3510 }
3511
3512 ~PrintRSetsClosure() {
3513 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3514 gclog_or_tty->print_cr("========================================");
3515 gclog_or_tty->cr();
3516 }
3517 };
3518
3519 void G1CollectedHeap::print_cset_rsets() {
3520 PrintRSetsClosure cl("Printing CSet RSets");
3521 collection_set_iterate(&cl);
3522 }
3523
3524 void G1CollectedHeap::print_all_rsets() {
3525 PrintRSetsClosure cl("Printing All RSets");;
3526 heap_region_iterate(&cl);
3527 }
3528 #endif // PRODUCT
3529
3530 G1CollectedHeap* G1CollectedHeap::heap() {
3531 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3532 "not a garbage-first heap");
3533 return _g1h;
3534 }
3535
3536 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3537 // always_do_update_barrier = false;
3538 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3539 // Call allocation profiler
3540 AllocationProfiler::iterate_since_last_gc();
3541 // Fill TLAB's and such
3542 ensure_parsability(true);
3543 }
3544
3545 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3546 // FIXME: what is this about?
3547 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3548 // is set.
3549 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3550 "derived pointer present"));
3551 // always_do_update_barrier = true;
3552
3553 // We have just completed a GC. Update the soft reference
3554 // policy with the new heap occupancy
3555 Universe::update_heap_info_at_gc();
3556 }
3557
3558 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3559 unsigned int gc_count_before,
3560 bool* succeeded) {
3561 assert_heap_not_locked_and_not_at_safepoint();
3562 g1_policy()->record_stop_world_start();
3563 VM_G1IncCollectionPause op(gc_count_before,
3564 word_size,
3565 false, /* should_initiate_conc_mark */
3566 g1_policy()->max_pause_time_ms(),
3567 GCCause::_g1_inc_collection_pause);
3568 VMThread::execute(&op);
3569
3570 HeapWord* result = op.result();
3571 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3572 assert(result == NULL || ret_succeeded,
3573 "the result should be NULL if the VM did not succeed");
3574 *succeeded = ret_succeeded;
3575
3576 assert_heap_not_locked();
3577 return result;
3578 }
3579
3580 void
3581 G1CollectedHeap::doConcurrentMark() {
3582 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3583 if (!_cmThread->in_progress()) {
3584 _cmThread->set_started();
3585 CGC_lock->notify();
3586 }
3587 }
3588
3589 size_t G1CollectedHeap::pending_card_num() {
3590 size_t extra_cards = 0;
3591 JavaThread *curr = Threads::first();
3592 while (curr != NULL) {
3593 DirtyCardQueue& dcq = curr->dirty_card_queue();
3594 extra_cards += dcq.size();
3595 curr = curr->next();
3596 }
3597 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3598 size_t buffer_size = dcqs.buffer_size();
3599 size_t buffer_num = dcqs.completed_buffers_num();
3600
3601 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes
3602 // in bytes - not the number of 'entries'. We need to convert
3603 // into a number of cards.
3604 return (buffer_size * buffer_num + extra_cards) / oopSize;
3605 }
3606
3607 size_t G1CollectedHeap::cards_scanned() {
3608 return g1_rem_set()->cardsScanned();
3609 }
3610
3611 void
3612 G1CollectedHeap::setup_surviving_young_words() {
3613 assert(_surviving_young_words == NULL, "pre-condition");
3614 uint array_length = g1_policy()->young_cset_region_length();
3615 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC);
3616 if (_surviving_young_words == NULL) {
3617 vm_exit_out_of_memory(sizeof(size_t) * array_length,
3618 "Not enough space for young surv words summary.");
3619 }
3620 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t));
3621 #ifdef ASSERT
3622 for (uint i = 0; i < array_length; ++i) {
3623 assert( _surviving_young_words[i] == 0, "memset above" );
3624 }
3625 #endif // !ASSERT
3626 }
3627
3628 void
3629 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3630 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3631 uint array_length = g1_policy()->young_cset_region_length();
3632 for (uint i = 0; i < array_length; ++i) {
3633 _surviving_young_words[i] += surv_young_words[i];
3634 }
3635 }
3636
3637 void
3638 G1CollectedHeap::cleanup_surviving_young_words() {
3639 guarantee( _surviving_young_words != NULL, "pre-condition" );
3640 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC);
3641 _surviving_young_words = NULL;
3642 }
3643
3644 #ifdef ASSERT
3645 class VerifyCSetClosure: public HeapRegionClosure {
3646 public:
3647 bool doHeapRegion(HeapRegion* hr) {
3648 // Here we check that the CSet region's RSet is ready for parallel
3649 // iteration. The fields that we'll verify are only manipulated
3650 // when the region is part of a CSet and is collected. Afterwards,
3651 // we reset these fields when we clear the region's RSet (when the
3652 // region is freed) so they are ready when the region is
3653 // re-allocated. The only exception to this is if there's an
3654 // evacuation failure and instead of freeing the region we leave
3655 // it in the heap. In that case, we reset these fields during
3656 // evacuation failure handling.
3657 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3658
3659 // Here's a good place to add any other checks we'd like to
3660 // perform on CSet regions.
3661 return false;
3662 }
3663 };
3664 #endif // ASSERT
3665
3666 #if TASKQUEUE_STATS
3667 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3668 st->print_raw_cr("GC Task Stats");
3669 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3670 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3671 }
3672
3673 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3674 print_taskqueue_stats_hdr(st);
3675
3676 TaskQueueStats totals;
3677 const int n = workers() != NULL ? workers()->total_workers() : 1;
3678 for (int i = 0; i < n; ++i) {
3679 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3680 totals += task_queue(i)->stats;
3681 }
3682 st->print_raw("tot "); totals.print(st); st->cr();
3683
3684 DEBUG_ONLY(totals.verify());
3685 }
3686
3687 void G1CollectedHeap::reset_taskqueue_stats() {
3688 const int n = workers() != NULL ? workers()->total_workers() : 1;
3689 for (int i = 0; i < n; ++i) {
3690 task_queue(i)->stats.reset();
3691 }
3692 }
3693 #endif // TASKQUEUE_STATS
3694
3695 void G1CollectedHeap::log_gc_header() {
3696 if (!G1Log::fine()) {
3697 return;
3698 }
3699
3700 gclog_or_tty->date_stamp(PrintGCDateStamps);
3701 gclog_or_tty->stamp(PrintGCTimeStamps);
3702
3703 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause())
3704 .append(g1_policy()->gcs_are_young() ? "(young)" : "(mixed)")
3705 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : "");
3706
3707 gclog_or_tty->print("[%s", (const char*)gc_cause_str);
3708 }
3709
3710 void G1CollectedHeap::log_gc_footer(double pause_time_sec) {
3711 if (!G1Log::fine()) {
3712 return;
3713 }
3714
3715 if (G1Log::finer()) {
3716 if (evacuation_failed()) {
3717 gclog_or_tty->print(" (to-space exhausted)");
3718 }
3719 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3720 g1_policy()->phase_times()->note_gc_end();
3721 g1_policy()->phase_times()->print(pause_time_sec);
3722 g1_policy()->print_detailed_heap_transition();
3723 } else {
3724 if (evacuation_failed()) {
3725 gclog_or_tty->print("--");
3726 }
3727 g1_policy()->print_heap_transition();
3728 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3729 }
3730 gclog_or_tty->flush();
3731 }
3732
3733 bool
3734 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3735 assert_at_safepoint(true /* should_be_vm_thread */);
3736 guarantee(!is_gc_active(), "collection is not reentrant");
3737
3738 if (GC_locker::check_active_before_gc()) {
3739 return false;
3740 }
3741
3742 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3743 ResourceMark rm;
3744
3745 print_heap_before_gc();
3746
3747 HRSPhaseSetter x(HRSPhaseEvacuation);
3748 verify_region_sets_optional();
3749 verify_dirty_young_regions();
3750
3751 // This call will decide whether this pause is an initial-mark
3752 // pause. If it is, during_initial_mark_pause() will return true
3753 // for the duration of this pause.
3754 g1_policy()->decide_on_conc_mark_initiation();
3755
3756 // We do not allow initial-mark to be piggy-backed on a mixed GC.
3757 assert(!g1_policy()->during_initial_mark_pause() ||
3758 g1_policy()->gcs_are_young(), "sanity");
3759
3760 // We also do not allow mixed GCs during marking.
3761 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity");
3762
3763 // Record whether this pause is an initial mark. When the current
3764 // thread has completed its logging output and it's safe to signal
3765 // the CM thread, the flag's value in the policy has been reset.
3766 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause();
3767
3768 // Inner scope for scope based logging, timers, and stats collection
3769 {
3770 if (g1_policy()->during_initial_mark_pause()) {
3771 // We are about to start a marking cycle, so we increment the
3772 // full collection counter.
3773 increment_old_marking_cycles_started();
3774 }
3775 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
3776
3777 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3778 workers()->active_workers() : 1);
3779 double pause_start_sec = os::elapsedTime();
3780 g1_policy()->phase_times()->note_gc_start(active_workers);
3781 log_gc_header();
3782
3783 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3784 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3785
3786 // If the secondary_free_list is not empty, append it to the
3787 // free_list. No need to wait for the cleanup operation to finish;
3788 // the region allocation code will check the secondary_free_list
3789 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3790 // set, skip this step so that the region allocation code has to
3791 // get entries from the secondary_free_list.
3792 if (!G1StressConcRegionFreeing) {
3793 append_secondary_free_list_if_not_empty_with_lock();
3794 }
3795
3796 assert(check_young_list_well_formed(),
3797 "young list should be well formed");
3798
3799 // Don't dynamically change the number of GC threads this early. A value of
3800 // 0 is used to indicate serial work. When parallel work is done,
3801 // it will be set.
3802
3803 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3804 IsGCActiveMark x;
3805
3806 gc_prologue(false);
3807 increment_total_collections(false /* full gc */);
3808 increment_gc_time_stamp();
3809
3810 verify_before_gc();
3811
3812 COMPILER2_PRESENT(DerivedPointerTable::clear());
3813
3814 // Please see comment in g1CollectedHeap.hpp and
3815 // G1CollectedHeap::ref_processing_init() to see how
3816 // reference processing currently works in G1.
3817
3818 // Enable discovery in the STW reference processor
3819 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3820 true /*verify_no_refs*/);
3821
3822 {
3823 // We want to temporarily turn off discovery by the
3824 // CM ref processor, if necessary, and turn it back on
3825 // on again later if we do. Using a scoped
3826 // NoRefDiscovery object will do this.
3827 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3828
3829 // Forget the current alloc region (we might even choose it to be part
3830 // of the collection set!).
3831 release_mutator_alloc_region();
3832
3833 // We should call this after we retire the mutator alloc
3834 // region(s) so that all the ALLOC / RETIRE events are generated
3835 // before the start GC event.
3836 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3837
3838 // This timing is only used by the ergonomics to handle our pause target.
3839 // It is unclear why this should not include the full pause. We will
3840 // investigate this in CR 7178365.
3841 //
3842 // Preserving the old comment here if that helps the investigation:
3843 //
3844 // The elapsed time induced by the start time below deliberately elides
3845 // the possible verification above.
3846 double sample_start_time_sec = os::elapsedTime();
3847
3848 #if YOUNG_LIST_VERBOSE
3849 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3850 _young_list->print();
3851 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3852 #endif // YOUNG_LIST_VERBOSE
3853
3854 g1_policy()->record_collection_pause_start(sample_start_time_sec);
3855
3856 double scan_wait_start = os::elapsedTime();
3857 // We have to wait until the CM threads finish scanning the
3858 // root regions as it's the only way to ensure that all the
3859 // objects on them have been correctly scanned before we start
3860 // moving them during the GC.
3861 bool waited = _cm->root_regions()->wait_until_scan_finished();
3862 double wait_time_ms = 0.0;
3863 if (waited) {
3864 double scan_wait_end = os::elapsedTime();
3865 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0;
3866 }
3867 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms);
3868
3869 #if YOUNG_LIST_VERBOSE
3870 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3871 _young_list->print();
3872 #endif // YOUNG_LIST_VERBOSE
3873
3874 if (g1_policy()->during_initial_mark_pause()) {
3875 concurrent_mark()->checkpointRootsInitialPre();
3876 }
3877
3878 #if YOUNG_LIST_VERBOSE
3879 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3880 _young_list->print();
3881 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3882 #endif // YOUNG_LIST_VERBOSE
3883
3884 g1_policy()->finalize_cset(target_pause_time_ms);
3885
3886 _cm->note_start_of_gc();
3887 // We should not verify the per-thread SATB buffers given that
3888 // we have not filtered them yet (we'll do so during the
3889 // GC). We also call this after finalize_cset() to
3890 // ensure that the CSet has been finalized.
3891 _cm->verify_no_cset_oops(true /* verify_stacks */,
3892 true /* verify_enqueued_buffers */,
3893 false /* verify_thread_buffers */,
3894 true /* verify_fingers */);
3895
3896 if (_hr_printer.is_active()) {
3897 HeapRegion* hr = g1_policy()->collection_set();
3898 while (hr != NULL) {
3899 G1HRPrinter::RegionType type;
3900 if (!hr->is_young()) {
3901 type = G1HRPrinter::Old;
3902 } else if (hr->is_survivor()) {
3903 type = G1HRPrinter::Survivor;
3904 } else {
3905 type = G1HRPrinter::Eden;
3906 }
3907 _hr_printer.cset(hr);
3908 hr = hr->next_in_collection_set();
3909 }
3910 }
3911
3912 #ifdef ASSERT
3913 VerifyCSetClosure cl;
3914 collection_set_iterate(&cl);
3915 #endif // ASSERT
3916
3917 setup_surviving_young_words();
3918
3919 // Initialize the GC alloc regions.
3920 init_gc_alloc_regions();
3921
3922 // Actually do the work...
3923 evacuate_collection_set();
3924
3925 // We do this to mainly verify the per-thread SATB buffers
3926 // (which have been filtered by now) since we didn't verify
3927 // them earlier. No point in re-checking the stacks / enqueued
3928 // buffers given that the CSet has not changed since last time
3929 // we checked.
3930 _cm->verify_no_cset_oops(false /* verify_stacks */,
3931 false /* verify_enqueued_buffers */,
3932 true /* verify_thread_buffers */,
3933 true /* verify_fingers */);
3934
3935 free_collection_set(g1_policy()->collection_set());
3936 g1_policy()->clear_collection_set();
3937
3938 cleanup_surviving_young_words();
3939
3940 // Start a new incremental collection set for the next pause.
3941 g1_policy()->start_incremental_cset_building();
3942
3943 // Clear the _cset_fast_test bitmap in anticipation of adding
3944 // regions to the incremental collection set for the next
3945 // evacuation pause.
3946 clear_cset_fast_test();
3947
3948 _young_list->reset_sampled_info();
3949
3950 // Don't check the whole heap at this point as the
3951 // GC alloc regions from this pause have been tagged
3952 // as survivors and moved on to the survivor list.
3953 // Survivor regions will fail the !is_young() check.
3954 assert(check_young_list_empty(false /* check_heap */),
3955 "young list should be empty");
3956
3957 #if YOUNG_LIST_VERBOSE
3958 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3959 _young_list->print();
3960 #endif // YOUNG_LIST_VERBOSE
3961
3962 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3963 _young_list->first_survivor_region(),
3964 _young_list->last_survivor_region());
3965
3966 _young_list->reset_auxilary_lists();
3967
3968 if (evacuation_failed()) {
3969 _summary_bytes_used = recalculate_used();
3970 } else {
3971 // The "used" of the the collection set have already been subtracted
3972 // when they were freed. Add in the bytes evacuated.
3973 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3974 }
3975
3976 if (g1_policy()->during_initial_mark_pause()) {
3977 // We have to do this before we notify the CM threads that
3978 // they can start working to make sure that all the
3979 // appropriate initialization is done on the CM object.
3980 concurrent_mark()->checkpointRootsInitialPost();
3981 set_marking_started();
3982 // Note that we don't actually trigger the CM thread at
3983 // this point. We do that later when we're sure that
3984 // the current thread has completed its logging output.
3985 }
3986
3987 allocate_dummy_regions();
3988
3989 #if YOUNG_LIST_VERBOSE
3990 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3991 _young_list->print();
3992 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3993 #endif // YOUNG_LIST_VERBOSE
3994
3995 init_mutator_alloc_region();
3996
3997 {
3998 size_t expand_bytes = g1_policy()->expansion_amount();
3999 if (expand_bytes > 0) {
4000 size_t bytes_before = capacity();
4001 // No need for an ergo verbose message here,
4002 // expansion_amount() does this when it returns a value > 0.
4003 if (!expand(expand_bytes)) {
4004 // We failed to expand the heap so let's verify that
4005 // committed/uncommitted amount match the backing store
4006 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
4007 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
4008 }
4009 }
4010 }
4011
4012 // We redo the verificaiton but now wrt to the new CSet which
4013 // has just got initialized after the previous CSet was freed.
4014 _cm->verify_no_cset_oops(true /* verify_stacks */,
4015 true /* verify_enqueued_buffers */,
4016 true /* verify_thread_buffers */,
4017 true /* verify_fingers */);
4018 _cm->note_end_of_gc();
4019
4020 // This timing is only used by the ergonomics to handle our pause target.
4021 // It is unclear why this should not include the full pause. We will
4022 // investigate this in CR 7178365.
4023 double sample_end_time_sec = os::elapsedTime();
4024 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS;
4025 g1_policy()->record_collection_pause_end(pause_time_ms);
4026
4027 MemoryService::track_memory_usage();
4028
4029 // In prepare_for_verify() below we'll need to scan the deferred
4030 // update buffers to bring the RSets up-to-date if
4031 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
4032 // the update buffers we'll probably need to scan cards on the
4033 // regions we just allocated to (i.e., the GC alloc
4034 // regions). However, during the last GC we called
4035 // set_saved_mark() on all the GC alloc regions, so card
4036 // scanning might skip the [saved_mark_word()...top()] area of
4037 // those regions (i.e., the area we allocated objects into
4038 // during the last GC). But it shouldn't. Given that
4039 // saved_mark_word() is conditional on whether the GC time stamp
4040 // on the region is current or not, by incrementing the GC time
4041 // stamp here we invalidate all the GC time stamps on all the
4042 // regions and saved_mark_word() will simply return top() for
4043 // all the regions. This is a nicer way of ensuring this rather
4044 // than iterating over the regions and fixing them. In fact, the
4045 // GC time stamp increment here also ensures that
4046 // saved_mark_word() will return top() between pauses, i.e.,
4047 // during concurrent refinement. So we don't need the
4048 // is_gc_active() check to decided which top to use when
4049 // scanning cards (see CR 7039627).
4050 increment_gc_time_stamp();
4051
4052 verify_after_gc();
4053
4054 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
4055 ref_processor_stw()->verify_no_references_recorded();
4056
4057 // CM reference discovery will be re-enabled if necessary.
4058 }
4059
4060 // We should do this after we potentially expand the heap so
4061 // that all the COMMIT events are generated before the end GC
4062 // event, and after we retire the GC alloc regions so that all
4063 // RETIRE events are generated before the end GC event.
4064 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
4065
4066 if (mark_in_progress()) {
4067 concurrent_mark()->update_g1_committed();
4068 }
4069
4070 #ifdef TRACESPINNING
4071 ParallelTaskTerminator::print_termination_counts();
4072 #endif
4073
4074 gc_epilogue(false);
4075 }
4076
4077 // Print the remainder of the GC log output.
4078 log_gc_footer(os::elapsedTime() - pause_start_sec);
4079
4080 // It is not yet to safe to tell the concurrent mark to
4081 // start as we have some optional output below. We don't want the
4082 // output from the concurrent mark thread interfering with this
4083 // logging output either.
4084
4085 _hrs.verify_optional();
4086 verify_region_sets_optional();
4087
4088 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
4089 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
4090
4091 print_heap_after_gc();
4092
4093 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
4094 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
4095 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
4096 // before any GC notifications are raised.
4097 g1mm()->update_sizes();
4098 }
4099
4100 if (G1SummarizeRSetStats &&
4101 (G1SummarizeRSetStatsPeriod > 0) &&
4102 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
4103 g1_rem_set()->print_summary_info();
4104 }
4105
4106 // It should now be safe to tell the concurrent mark thread to start
4107 // without its logging output interfering with the logging output
4108 // that came from the pause.
4109
4110 if (should_start_conc_mark) {
4111 // CAUTION: after the doConcurrentMark() call below,
4112 // the concurrent marking thread(s) could be running
4113 // concurrently with us. Make sure that anything after
4114 // this point does not assume that we are the only GC thread
4115 // running. Note: of course, the actual marking work will
4116 // not start until the safepoint itself is released in
4117 // ConcurrentGCThread::safepoint_desynchronize().
4118 doConcurrentMark();
4119 }
4120
4121 return true;
4122 }
4123
4124 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
4125 {
4126 size_t gclab_word_size;
4127 switch (purpose) {
4128 case GCAllocForSurvived:
4129 gclab_word_size = _survivor_plab_stats.desired_plab_sz();
4130 break;
4131 case GCAllocForTenured:
4132 gclab_word_size = _old_plab_stats.desired_plab_sz();
4133 break;
4134 default:
4135 assert(false, "unknown GCAllocPurpose");
4136 gclab_word_size = _old_plab_stats.desired_plab_sz();
4137 break;
4138 }
4139
4140 // Prevent humongous PLAB sizes for two reasons:
4141 // * PLABs are allocated using a similar paths as oops, but should
4142 // never be in a humongous region
4143 // * Allowing humongous PLABs needlessly churns the region free lists
4144 return MIN2(_humongous_object_threshold_in_words, gclab_word_size);
4145 }
4146
4147 void G1CollectedHeap::init_mutator_alloc_region() {
4148 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
4149 _mutator_alloc_region.init();
4150 }
4151
4152 void G1CollectedHeap::release_mutator_alloc_region() {
4153 _mutator_alloc_region.release();
4154 assert(_mutator_alloc_region.get() == NULL, "post-condition");
4155 }
4156
4157 void G1CollectedHeap::init_gc_alloc_regions() {
4158 assert_at_safepoint(true /* should_be_vm_thread */);
4159
4160 _survivor_gc_alloc_region.init();
4161 _old_gc_alloc_region.init();
4162 HeapRegion* retained_region = _retained_old_gc_alloc_region;
4163 _retained_old_gc_alloc_region = NULL;
4164
4165 // We will discard the current GC alloc region if:
4166 // a) it's in the collection set (it can happen!),
4167 // b) it's already full (no point in using it),
4168 // c) it's empty (this means that it was emptied during
4169 // a cleanup and it should be on the free list now), or
4170 // d) it's humongous (this means that it was emptied
4171 // during a cleanup and was added to the free list, but
4172 // has been subseqently used to allocate a humongous
4173 // object that may be less than the region size).
4174 if (retained_region != NULL &&
4175 !retained_region->in_collection_set() &&
4176 !(retained_region->top() == retained_region->end()) &&
4177 !retained_region->is_empty() &&
4178 !retained_region->isHumongous()) {
4179 retained_region->set_saved_mark();
4180 // The retained region was added to the old region set when it was
4181 // retired. We have to remove it now, since we don't allow regions
4182 // we allocate to in the region sets. We'll re-add it later, when
4183 // it's retired again.
4184 _old_set.remove(retained_region);
4185 bool during_im = g1_policy()->during_initial_mark_pause();
4186 retained_region->note_start_of_copying(during_im);
4187 _old_gc_alloc_region.set(retained_region);
4188 _hr_printer.reuse(retained_region);
4189 }
4190 }
4191
4192 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers) {
4193 _survivor_gc_alloc_region.release();
4194 // If we have an old GC alloc region to release, we'll save it in
4195 // _retained_old_gc_alloc_region. If we don't
4196 // _retained_old_gc_alloc_region will become NULL. This is what we
4197 // want either way so no reason to check explicitly for either
4198 // condition.
4199 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
4200
4201 if (ResizePLAB) {
4202 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4203 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4204 }
4205 }
4206
4207 void G1CollectedHeap::abandon_gc_alloc_regions() {
4208 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
4209 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
4210 _retained_old_gc_alloc_region = NULL;
4211 }
4212
4213 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
4214 _drain_in_progress = false;
4215 set_evac_failure_closure(cl);
4216 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4217 }
4218
4219 void G1CollectedHeap::finalize_for_evac_failure() {
4220 assert(_evac_failure_scan_stack != NULL &&
4221 _evac_failure_scan_stack->length() == 0,
4222 "Postcondition");
4223 assert(!_drain_in_progress, "Postcondition");
4224 delete _evac_failure_scan_stack;
4225 _evac_failure_scan_stack = NULL;
4226 }
4227
4228 void G1CollectedHeap::remove_self_forwarding_pointers() {
4229 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4230
4231 G1ParRemoveSelfForwardPtrsTask rsfp_task(this);
4232
4233 if (G1CollectedHeap::use_parallel_gc_threads()) {
4234 set_par_threads();
4235 workers()->run_task(&rsfp_task);
4236 set_par_threads(0);
4237 } else {
4238 rsfp_task.work(0);
4239 }
4240
4241 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity");
4242
4243 // Reset the claim values in the regions in the collection set.
4244 reset_cset_heap_region_claim_values();
4245
4246 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4247
4248 // Now restore saved marks, if any.
4249 assert(_objs_with_preserved_marks.size() ==
4250 _preserved_marks_of_objs.size(), "Both or none.");
4251 while (!_objs_with_preserved_marks.is_empty()) {
4252 oop obj = _objs_with_preserved_marks.pop();
4253 markOop m = _preserved_marks_of_objs.pop();
4254 obj->set_mark(m);
4255 }
4256 _objs_with_preserved_marks.clear(true);
4257 _preserved_marks_of_objs.clear(true);
4258 }
4259
4260 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
4261 _evac_failure_scan_stack->push(obj);
4262 }
4263
4264 void G1CollectedHeap::drain_evac_failure_scan_stack() {
4265 assert(_evac_failure_scan_stack != NULL, "precondition");
4266
4267 while (_evac_failure_scan_stack->length() > 0) {
4268 oop obj = _evac_failure_scan_stack->pop();
4269 _evac_failure_closure->set_region(heap_region_containing(obj));
4270 obj->oop_iterate_backwards(_evac_failure_closure);
4271 }
4272 }
4273
4274 oop
4275 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4276 oop old) {
4277 assert(obj_in_cs(old),
4278 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4279 (HeapWord*) old));
4280 markOop m = old->mark();
4281 oop forward_ptr = old->forward_to_atomic(old);
4282 if (forward_ptr == NULL) {
4283 // Forward-to-self succeeded.
4284
4285 if (_evac_failure_closure != cl) {
4286 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4287 assert(!_drain_in_progress,
4288 "Should only be true while someone holds the lock.");
4289 // Set the global evac-failure closure to the current thread's.
4290 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4291 set_evac_failure_closure(cl);
4292 // Now do the common part.
4293 handle_evacuation_failure_common(old, m);
4294 // Reset to NULL.
4295 set_evac_failure_closure(NULL);
4296 } else {
4297 // The lock is already held, and this is recursive.
4298 assert(_drain_in_progress, "This should only be the recursive case.");
4299 handle_evacuation_failure_common(old, m);
4300 }
4301 return old;
4302 } else {
4303 // Forward-to-self failed. Either someone else managed to allocate
4304 // space for this object (old != forward_ptr) or they beat us in
4305 // self-forwarding it (old == forward_ptr).
4306 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4307 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4308 "should not be in the CSet",
4309 (HeapWord*) old, (HeapWord*) forward_ptr));
4310 return forward_ptr;
4311 }
4312 }
4313
4314 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4315 set_evacuation_failed(true);
4316
4317 preserve_mark_if_necessary(old, m);
4318
4319 HeapRegion* r = heap_region_containing(old);
4320 if (!r->evacuation_failed()) {
4321 r->set_evacuation_failed(true);
4322 _hr_printer.evac_failure(r);
4323 }
4324
4325 push_on_evac_failure_scan_stack(old);
4326
4327 if (!_drain_in_progress) {
4328 // prevent recursion in copy_to_survivor_space()
4329 _drain_in_progress = true;
4330 drain_evac_failure_scan_stack();
4331 _drain_in_progress = false;
4332 }
4333 }
4334
4335 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4336 assert(evacuation_failed(), "Oversaving!");
4337 // We want to call the "for_promotion_failure" version only in the
4338 // case of a promotion failure.
4339 if (m->must_be_preserved_for_promotion_failure(obj)) {
4340 _objs_with_preserved_marks.push(obj);
4341 _preserved_marks_of_objs.push(m);
4342 }
4343 }
4344
4345 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4346 size_t word_size) {
4347 if (purpose == GCAllocForSurvived) {
4348 HeapWord* result = survivor_attempt_allocation(word_size);
4349 if (result != NULL) {
4350 return result;
4351 } else {
4352 // Let's try to allocate in the old gen in case we can fit the
4353 // object there.
4354 return old_attempt_allocation(word_size);
4355 }
4356 } else {
4357 assert(purpose == GCAllocForTenured, "sanity");
4358 HeapWord* result = old_attempt_allocation(word_size);
4359 if (result != NULL) {
4360 return result;
4361 } else {
4362 // Let's try to allocate in the survivors in case we can fit the
4363 // object there.
4364 return survivor_attempt_allocation(word_size);
4365 }
4366 }
4367
4368 ShouldNotReachHere();
4369 // Trying to keep some compilers happy.
4370 return NULL;
4371 }
4372
4373 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4374 ParGCAllocBuffer(gclab_word_size), _retired(false) { }
4375
4376 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num)
4377 : _g1h(g1h),
4378 _refs(g1h->task_queue(queue_num)),
4379 _dcq(&g1h->dirty_card_queue_set()),
4380 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4381 _g1_rem(g1h->g1_rem_set()),
4382 _hash_seed(17), _queue_num(queue_num),
4383 _term_attempts(0),
4384 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4385 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4386 _age_table(false),
4387 _strong_roots_time(0), _term_time(0),
4388 _alloc_buffer_waste(0), _undo_waste(0) {
4389 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4390 // we "sacrifice" entry 0 to keep track of surviving bytes for
4391 // non-young regions (where the age is -1)
4392 // We also add a few elements at the beginning and at the end in
4393 // an attempt to eliminate cache contention
4394 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
4395 uint array_length = PADDING_ELEM_NUM +
4396 real_length +
4397 PADDING_ELEM_NUM;
4398 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
4399 if (_surviving_young_words_base == NULL)
4400 vm_exit_out_of_memory(array_length * sizeof(size_t),
4401 "Not enough space for young surv histo.");
4402 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4403 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t));
4404
4405 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4406 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4407
4408 _start = os::elapsedTime();
4409 }
4410
4411 void
4412 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4413 {
4414 st->print_raw_cr("GC Termination Stats");
4415 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4416 " ------waste (KiB)------");
4417 st->print_raw_cr("thr ms ms % ms % attempts"
4418 " total alloc undo");
4419 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4420 " ------- ------- -------");
4421 }
4422
4423 void
4424 G1ParScanThreadState::print_termination_stats(int i,
4425 outputStream* const st) const
4426 {
4427 const double elapsed_ms = elapsed_time() * 1000.0;
4428 const double s_roots_ms = strong_roots_time() * 1000.0;
4429 const double term_ms = term_time() * 1000.0;
4430 st->print_cr("%3d %9.2f %9.2f %6.2f "
4431 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4432 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4433 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4434 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4435 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4436 alloc_buffer_waste() * HeapWordSize / K,
4437 undo_waste() * HeapWordSize / K);
4438 }
4439
4440 #ifdef ASSERT
4441 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4442 assert(ref != NULL, "invariant");
4443 assert(UseCompressedOops, "sanity");
4444 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4445 oop p = oopDesc::load_decode_heap_oop(ref);
4446 assert(_g1h->is_in_g1_reserved(p),
4447 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4448 return true;
4449 }
4450
4451 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4452 assert(ref != NULL, "invariant");
4453 if (has_partial_array_mask(ref)) {
4454 // Must be in the collection set--it's already been copied.
4455 oop p = clear_partial_array_mask(ref);
4456 assert(_g1h->obj_in_cs(p),
4457 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4458 } else {
4459 oop p = oopDesc::load_decode_heap_oop(ref);
4460 assert(_g1h->is_in_g1_reserved(p),
4461 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4462 }
4463 return true;
4464 }
4465
4466 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4467 if (ref.is_narrow()) {
4468 return verify_ref((narrowOop*) ref);
4469 } else {
4470 return verify_ref((oop*) ref);
4471 }
4472 }
4473 #endif // ASSERT
4474
4475 void G1ParScanThreadState::trim_queue() {
4476 assert(_evac_cl != NULL, "not set");
4477 assert(_evac_failure_cl != NULL, "not set");
4478 assert(_partial_scan_cl != NULL, "not set");
4479
4480 StarTask ref;
4481 do {
4482 // Drain the overflow stack first, so other threads can steal.
4483 while (refs()->pop_overflow(ref)) {
4484 deal_with_reference(ref);
4485 }
4486
4487 while (refs()->pop_local(ref)) {
4488 deal_with_reference(ref);
4489 }
4490 } while (!refs()->is_empty());
4491 }
4492
4493 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1,
4494 G1ParScanThreadState* par_scan_state) :
4495 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4496 _par_scan_state(par_scan_state),
4497 _worker_id(par_scan_state->queue_num()),
4498 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4499 _mark_in_progress(_g1->mark_in_progress()) { }
4500
4501 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4502 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) {
4503 #ifdef ASSERT
4504 HeapRegion* hr = _g1->heap_region_containing(obj);
4505 assert(hr != NULL, "sanity");
4506 assert(!hr->in_collection_set(), "should not mark objects in the CSet");
4507 #endif // ASSERT
4508
4509 // We know that the object is not moving so it's safe to read its size.
4510 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id);
4511 }
4512
4513 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4514 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4515 ::mark_forwarded_object(oop from_obj, oop to_obj) {
4516 #ifdef ASSERT
4517 assert(from_obj->is_forwarded(), "from obj should be forwarded");
4518 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee");
4519 assert(from_obj != to_obj, "should not be self-forwarded");
4520
4521 HeapRegion* from_hr = _g1->heap_region_containing(from_obj);
4522 assert(from_hr != NULL, "sanity");
4523 assert(from_hr->in_collection_set(), "from obj should be in the CSet");
4524
4525 HeapRegion* to_hr = _g1->heap_region_containing(to_obj);
4526 assert(to_hr != NULL, "sanity");
4527 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet");
4528 #endif // ASSERT
4529
4530 // The object might be in the process of being copied by another
4531 // worker so we cannot trust that its to-space image is
4532 // well-formed. So we have to read its size from its from-space
4533 // image which we know should not be changing.
4534 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id);
4535 }
4536
4537 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4538 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4539 ::copy_to_survivor_space(oop old) {
4540 size_t word_sz = old->size();
4541 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4542 // +1 to make the -1 indexes valid...
4543 int young_index = from_region->young_index_in_cset()+1;
4544 assert( (from_region->is_young() && young_index > 0) ||
4545 (!from_region->is_young() && young_index == 0), "invariant" );
4546 G1CollectorPolicy* g1p = _g1->g1_policy();
4547 markOop m = old->mark();
4548 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
4549 : m->age();
4550 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4551 word_sz);
4552 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4553 #ifndef PRODUCT
4554 // Should this evacuation fail?
4555 if (_g1->evacuation_should_fail()) {
4556 if (obj_ptr != NULL) {
4557 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4558 obj_ptr = NULL;
4559 }
4560 }
4561 #endif // !PRODUCT
4562
4563 if (obj_ptr == NULL) {
4564 // This will either forward-to-self, or detect that someone else has
4565 // installed a forwarding pointer.
4566 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4567 return _g1->handle_evacuation_failure_par(cl, old);
4568 }
4569
4570 oop obj = oop(obj_ptr);
4571
4572 // We're going to allocate linearly, so might as well prefetch ahead.
4573 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4574
4575 oop forward_ptr = old->forward_to_atomic(obj);
4576 if (forward_ptr == NULL) {
4577 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4578 if (g1p->track_object_age(alloc_purpose)) {
4579 // We could simply do obj->incr_age(). However, this causes a
4580 // performance issue. obj->incr_age() will first check whether
4581 // the object has a displaced mark by checking its mark word;
4582 // getting the mark word from the new location of the object
4583 // stalls. So, given that we already have the mark word and we
4584 // are about to install it anyway, it's better to increase the
4585 // age on the mark word, when the object does not have a
4586 // displaced mark word. We're not expecting many objects to have
4587 // a displaced marked word, so that case is not optimized
4588 // further (it could be...) and we simply call obj->incr_age().
4589
4590 if (m->has_displaced_mark_helper()) {
4591 // in this case, we have to install the mark word first,
4592 // otherwise obj looks to be forwarded (the old mark word,
4593 // which contains the forward pointer, was copied)
4594 obj->set_mark(m);
4595 obj->incr_age();
4596 } else {
4597 m = m->incr_age();
4598 obj->set_mark(m);
4599 }
4600 _par_scan_state->age_table()->add(obj, word_sz);
4601 } else {
4602 obj->set_mark(m);
4603 }
4604
4605 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4606 surv_young_words[young_index] += word_sz;
4607
4608 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4609 // We keep track of the next start index in the length field of
4610 // the to-space object. The actual length can be found in the
4611 // length field of the from-space object.
4612 arrayOop(obj)->set_length(0);
4613 oop* old_p = set_partial_array_mask(old);
4614 _par_scan_state->push_on_queue(old_p);
4615 } else {
4616 // No point in using the slower heap_region_containing() method,
4617 // given that we know obj is in the heap.
4618 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4619 obj->oop_iterate_backwards(&_scanner);
4620 }
4621 } else {
4622 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4623 obj = forward_ptr;
4624 }
4625 return obj;
4626 }
4627
4628 template <class T>
4629 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) {
4630 if (_g1->heap_region_containing_raw(new_obj)->is_young()) {
4631 _scanned_klass->record_modified_oops();
4632 }
4633 }
4634
4635 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4636 template <class T>
4637 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4638 ::do_oop_work(T* p) {
4639 oop obj = oopDesc::load_decode_heap_oop(p);
4640 assert(barrier != G1BarrierRS || obj != NULL,
4641 "Precondition: G1BarrierRS implies obj is non-NULL");
4642
4643 assert(_worker_id == _par_scan_state->queue_num(), "sanity");
4644
4645 // here the null check is implicit in the cset_fast_test() test
4646 if (_g1->in_cset_fast_test(obj)) {
4647 oop forwardee;
4648 if (obj->is_forwarded()) {
4649 forwardee = obj->forwardee();
4650 } else {
4651 forwardee = copy_to_survivor_space(obj);
4652 }
4653 assert(forwardee != NULL, "forwardee should not be NULL");
4654 oopDesc::encode_store_heap_oop(p, forwardee);
4655 if (do_mark_object && forwardee != obj) {
4656 // If the object is self-forwarded we don't need to explicitly
4657 // mark it, the evacuation failure protocol will do so.
4658 mark_forwarded_object(obj, forwardee);
4659 }
4660
4661 // When scanning the RS, we only care about objs in CS.
4662 if (barrier == G1BarrierRS) {
4663 _par_scan_state->update_rs(_from, p, _worker_id);
4664 } else if (barrier == G1BarrierKlass) {
4665 do_klass_barrier(p, forwardee);
4666 }
4667 } else {
4668 // The object is not in collection set. If we're a root scanning
4669 // closure during an initial mark pause (i.e. do_mark_object will
4670 // be true) then attempt to mark the object.
4671 if (do_mark_object && _g1->is_in_g1_reserved(obj)) {
4672 mark_object(obj);
4673 }
4674 }
4675
4676 if (barrier == G1BarrierEvac && obj != NULL) {
4677 _par_scan_state->update_rs(_from, p, _worker_id);
4678 }
4679
4680 if (do_gen_barrier && obj != NULL) {
4681 par_do_barrier(p);
4682 }
4683 }
4684
4685 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4686 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4687
4688 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4689 assert(has_partial_array_mask(p), "invariant");
4690 oop from_obj = clear_partial_array_mask(p);
4691
4692 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap.");
4693 assert(from_obj->is_objArray(), "must be obj array");
4694 objArrayOop from_obj_array = objArrayOop(from_obj);
4695 // The from-space object contains the real length.
4696 int length = from_obj_array->length();
4697
4698 assert(from_obj->is_forwarded(), "must be forwarded");
4699 oop to_obj = from_obj->forwardee();
4700 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
4701 objArrayOop to_obj_array = objArrayOop(to_obj);
4702 // We keep track of the next start index in the length field of the
4703 // to-space object.
4704 int next_index = to_obj_array->length();
4705 assert(0 <= next_index && next_index < length,
4706 err_msg("invariant, next index: %d, length: %d", next_index, length));
4707
4708 int start = next_index;
4709 int end = length;
4710 int remainder = end - start;
4711 // We'll try not to push a range that's smaller than ParGCArrayScanChunk.
4712 if (remainder > 2 * ParGCArrayScanChunk) {
4713 end = start + ParGCArrayScanChunk;
4714 to_obj_array->set_length(end);
4715 // Push the remainder before we process the range in case another
4716 // worker has run out of things to do and can steal it.
4717 oop* from_obj_p = set_partial_array_mask(from_obj);
4718 _par_scan_state->push_on_queue(from_obj_p);
4719 } else {
4720 assert(length == end, "sanity");
4721 // We'll process the final range for this object. Restore the length
4722 // so that the heap remains parsable in case of evacuation failure.
4723 to_obj_array->set_length(end);
4724 }
4725 _scanner.set_region(_g1->heap_region_containing_raw(to_obj));
4726 // Process indexes [start,end). It will also process the header
4727 // along with the first chunk (i.e., the chunk with start == 0).
4728 // Note that at this point the length field of to_obj_array is not
4729 // correct given that we are using it to keep track of the next
4730 // start index. oop_iterate_range() (thankfully!) ignores the length
4731 // field and only relies on the start / end parameters. It does
4732 // however return the size of the object which will be incorrect. So
4733 // we have to ignore it even if we wanted to use it.
4734 to_obj_array->oop_iterate_range(&_scanner, start, end);
4735 }
4736
4737 class G1ParEvacuateFollowersClosure : public VoidClosure {
4738 protected:
4739 G1CollectedHeap* _g1h;
4740 G1ParScanThreadState* _par_scan_state;
4741 RefToScanQueueSet* _queues;
4742 ParallelTaskTerminator* _terminator;
4743
4744 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4745 RefToScanQueueSet* queues() { return _queues; }
4746 ParallelTaskTerminator* terminator() { return _terminator; }
4747
4748 public:
4749 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4750 G1ParScanThreadState* par_scan_state,
4751 RefToScanQueueSet* queues,
4752 ParallelTaskTerminator* terminator)
4753 : _g1h(g1h), _par_scan_state(par_scan_state),
4754 _queues(queues), _terminator(terminator) {}
4755
4756 void do_void();
4757
4758 private:
4759 inline bool offer_termination();
4760 };
4761
4762 bool G1ParEvacuateFollowersClosure::offer_termination() {
4763 G1ParScanThreadState* const pss = par_scan_state();
4764 pss->start_term_time();
4765 const bool res = terminator()->offer_termination();
4766 pss->end_term_time();
4767 return res;
4768 }
4769
4770 void G1ParEvacuateFollowersClosure::do_void() {
4771 StarTask stolen_task;
4772 G1ParScanThreadState* const pss = par_scan_state();
4773 pss->trim_queue();
4774
4775 do {
4776 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4777 assert(pss->verify_task(stolen_task), "sanity");
4778 if (stolen_task.is_narrow()) {
4779 pss->deal_with_reference((narrowOop*) stolen_task);
4780 } else {
4781 pss->deal_with_reference((oop*) stolen_task);
4782 }
4783
4784 // We've just processed a reference and we might have made
4785 // available new entries on the queues. So we have to make sure
4786 // we drain the queues as necessary.
4787 pss->trim_queue();
4788 }
4789 } while (!offer_termination());
4790
4791 pss->retire_alloc_buffers();
4792 }
4793
4794 class G1KlassScanClosure : public KlassClosure {
4795 G1ParCopyHelper* _closure;
4796 bool _process_only_dirty;
4797 int _count;
4798 public:
4799 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty)
4800 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {}
4801 void do_klass(Klass* klass) {
4802 // If the klass has not been dirtied we know that there's
4803 // no references into the young gen and we can skip it.
4804 if (!_process_only_dirty || klass->has_modified_oops()) {
4805 // Clean the klass since we're going to scavenge all the metadata.
4806 klass->clear_modified_oops();
4807
4808 // Tell the closure that this klass is the Klass to scavenge
4809 // and is the one to dirty if oops are left pointing into the young gen.
4810 _closure->set_scanned_klass(klass);
4811
4812 klass->oops_do(_closure);
4813
4814 _closure->set_scanned_klass(NULL);
4815 }
4816 _count++;
4817 }
4818 };
4819
4820 class G1ParTask : public AbstractGangTask {
4821 protected:
4822 G1CollectedHeap* _g1h;
4823 RefToScanQueueSet *_queues;
4824 ParallelTaskTerminator _terminator;
4825 uint _n_workers;
4826
4827 Mutex _stats_lock;
4828 Mutex* stats_lock() { return &_stats_lock; }
4829
4830 size_t getNCards() {
4831 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4832 / G1BlockOffsetSharedArray::N_bytes;
4833 }
4834
4835 public:
4836 G1ParTask(G1CollectedHeap* g1h,
4837 RefToScanQueueSet *task_queues)
4838 : AbstractGangTask("G1 collection"),
4839 _g1h(g1h),
4840 _queues(task_queues),
4841 _terminator(0, _queues),
4842 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true)
4843 {}
4844
4845 RefToScanQueueSet* queues() { return _queues; }
4846
4847 RefToScanQueue *work_queue(int i) {
4848 return queues()->queue(i);
4849 }
4850
4851 ParallelTaskTerminator* terminator() { return &_terminator; }
4852
4853 virtual void set_for_termination(int active_workers) {
4854 // This task calls set_n_termination() in par_non_clean_card_iterate_work()
4855 // in the young space (_par_seq_tasks) in the G1 heap
4856 // for SequentialSubTasksDone.
4857 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap
4858 // both of which need setting by set_n_termination().
4859 _g1h->SharedHeap::set_n_termination(active_workers);
4860 _g1h->set_n_termination(active_workers);
4861 terminator()->reset_for_reuse(active_workers);
4862 _n_workers = active_workers;
4863 }
4864
4865 void work(uint worker_id) {
4866 if (worker_id >= _n_workers) return; // no work needed this round
4867
4868 double start_time_ms = os::elapsedTime() * 1000.0;
4869 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms);
4870
4871 {
4872 ResourceMark rm;
4873 HandleMark hm;
4874
4875 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4876
4877 G1ParScanThreadState pss(_g1h, worker_id);
4878 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4879 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4880 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4881
4882 pss.set_evac_closure(&scan_evac_cl);
4883 pss.set_evac_failure_closure(&evac_failure_cl);
4884 pss.set_partial_scan_closure(&partial_scan_cl);
4885
4886 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4887 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp);
4888
4889 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4890 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp);
4891
4892 bool only_young = _g1h->g1_policy()->gcs_are_young();
4893 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false);
4894 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young);
4895
4896 OopClosure* scan_root_cl = &only_scan_root_cl;
4897 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s;
4898
4899 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4900 // We also need to mark copied objects.
4901 scan_root_cl = &scan_mark_root_cl;
4902 scan_klasses_cl = &scan_mark_klasses_cl_s;
4903 }
4904
4905 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4906
4907 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
4908
4909 pss.start_strong_roots();
4910 _g1h->g1_process_strong_roots(/* is scavenging */ true,
4911 SharedHeap::ScanningOption(so),
4912 scan_root_cl,
4913 &push_heap_rs_cl,
4914 scan_klasses_cl,
4915 worker_id);
4916 pss.end_strong_roots();
4917
4918 {
4919 double start = os::elapsedTime();
4920 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4921 evac.do_void();
4922 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4923 double term_ms = pss.term_time()*1000.0;
4924 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms);
4925 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts());
4926 }
4927 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4928 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4929
4930 if (ParallelGCVerbose) {
4931 MutexLocker x(stats_lock());
4932 pss.print_termination_stats(worker_id);
4933 }
4934
4935 assert(pss.refs()->is_empty(), "should be empty");
4936
4937 // Close the inner scope so that the ResourceMark and HandleMark
4938 // destructors are executed here and are included as part of the
4939 // "GC Worker Time".
4940 }
4941
4942 double end_time_ms = os::elapsedTime() * 1000.0;
4943 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms);
4944 }
4945 };
4946
4947 // *** Common G1 Evacuation Stuff
4948
4949 // Closures that support the filtering of CodeBlobs scanned during
4950 // external root scanning.
4951
4952 // Closure applied to reference fields in code blobs (specifically nmethods)
4953 // to determine whether an nmethod contains references that point into
4954 // the collection set. Used as a predicate when walking code roots so
4955 // that only nmethods that point into the collection set are added to the
4956 // 'marked' list.
4957
4958 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure {
4959
4960 class G1PointsIntoCSOopClosure : public OopClosure {
4961 G1CollectedHeap* _g1;
4962 bool _points_into_cs;
4963 public:
4964 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) :
4965 _g1(g1), _points_into_cs(false) { }
4966
4967 bool points_into_cs() const { return _points_into_cs; }
4968
4969 template <class T>
4970 void do_oop_nv(T* p) {
4971 if (!_points_into_cs) {
4972 T heap_oop = oopDesc::load_heap_oop(p);
4973 if (!oopDesc::is_null(heap_oop) &&
4974 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) {
4975 _points_into_cs = true;
4976 }
4977 }
4978 }
4979
4980 virtual void do_oop(oop* p) { do_oop_nv(p); }
4981 virtual void do_oop(narrowOop* p) { do_oop_nv(p); }
4982 };
4983
4984 G1CollectedHeap* _g1;
4985
4986 public:
4987 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) :
4988 CodeBlobToOopClosure(cl, true), _g1(g1) { }
4989
4990 virtual void do_code_blob(CodeBlob* cb) {
4991 nmethod* nm = cb->as_nmethod_or_null();
4992 if (nm != NULL && !(nm->test_oops_do_mark())) {
4993 G1PointsIntoCSOopClosure predicate_cl(_g1);
4994 nm->oops_do(&predicate_cl);
4995
4996 if (predicate_cl.points_into_cs()) {
4997 // At least one of the reference fields or the oop relocations
4998 // in the nmethod points into the collection set. We have to
4999 // 'mark' this nmethod.
5000 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob()
5001 // or MarkingCodeBlobClosure::do_code_blob() change.
5002 if (!nm->test_set_oops_do_mark()) {
5003 do_newly_marked_nmethod(nm);
5004 }
5005 }
5006 }
5007 }
5008 };
5009
5010 // This method is run in a GC worker.
5011
5012 void
5013 G1CollectedHeap::
5014 g1_process_strong_roots(bool is_scavenging,
5015 ScanningOption so,
5016 OopClosure* scan_non_heap_roots,
5017 OopsInHeapRegionClosure* scan_rs,
5018 G1KlassScanClosure* scan_klasses,
5019 int worker_i) {
5020
5021 // First scan the strong roots
5022 double ext_roots_start = os::elapsedTime();
5023 double closure_app_time_sec = 0.0;
5024
5025 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
5026
5027 // Walk the code cache w/o buffering, because StarTask cannot handle
5028 // unaligned oop locations.
5029 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots);
5030
5031 process_strong_roots(false, // no scoping; this is parallel code
5032 is_scavenging, so,
5033 &buf_scan_non_heap_roots,
5034 &eager_scan_code_roots,
5035 scan_klasses
5036 );
5037
5038 // Now the CM ref_processor roots.
5039 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
5040 // We need to treat the discovered reference lists of the
5041 // concurrent mark ref processor as roots and keep entries
5042 // (which are added by the marking threads) on them live
5043 // until they can be processed at the end of marking.
5044 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
5045 }
5046
5047 // Finish up any enqueued closure apps (attributed as object copy time).
5048 buf_scan_non_heap_roots.done();
5049
5050 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds();
5051
5052 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
5053
5054 double ext_root_time_ms =
5055 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0;
5056
5057 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
5058
5059 // During conc marking we have to filter the per-thread SATB buffers
5060 // to make sure we remove any oops into the CSet (which will show up
5061 // as implicitly live).
5062 double satb_filtering_ms = 0.0;
5063 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) {
5064 if (mark_in_progress()) {
5065 double satb_filter_start = os::elapsedTime();
5066
5067 JavaThread::satb_mark_queue_set().filter_thread_buffers();
5068
5069 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0;
5070 }
5071 }
5072 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms);
5073
5074 // Now scan the complement of the collection set.
5075 if (scan_rs != NULL) {
5076 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
5077 }
5078 _process_strong_tasks->all_tasks_completed();
5079 }
5080
5081 void
5082 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure,
5083 OopClosure* non_root_closure) {
5084 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
5085 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure);
5086 }
5087
5088 // Weak Reference Processing support
5089
5090 // An always "is_alive" closure that is used to preserve referents.
5091 // If the object is non-null then it's alive. Used in the preservation
5092 // of referent objects that are pointed to by reference objects
5093 // discovered by the CM ref processor.
5094 class G1AlwaysAliveClosure: public BoolObjectClosure {
5095 G1CollectedHeap* _g1;
5096 public:
5097 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5098 void do_object(oop p) { assert(false, "Do not call."); }
5099 bool do_object_b(oop p) {
5100 if (p != NULL) {
5101 return true;
5102 }
5103 return false;
5104 }
5105 };
5106
5107 bool G1STWIsAliveClosure::do_object_b(oop p) {
5108 // An object is reachable if it is outside the collection set,
5109 // or is inside and copied.
5110 return !_g1->obj_in_cs(p) || p->is_forwarded();
5111 }
5112
5113 // Non Copying Keep Alive closure
5114 class G1KeepAliveClosure: public OopClosure {
5115 G1CollectedHeap* _g1;
5116 public:
5117 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5118 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
5119 void do_oop( oop* p) {
5120 oop obj = *p;
5121
5122 if (_g1->obj_in_cs(obj)) {
5123 assert( obj->is_forwarded(), "invariant" );
5124 *p = obj->forwardee();
5125 }
5126 }
5127 };
5128
5129 // Copying Keep Alive closure - can be called from both
5130 // serial and parallel code as long as different worker
5131 // threads utilize different G1ParScanThreadState instances
5132 // and different queues.
5133
5134 class G1CopyingKeepAliveClosure: public OopClosure {
5135 G1CollectedHeap* _g1h;
5136 OopClosure* _copy_non_heap_obj_cl;
5137 OopsInHeapRegionClosure* _copy_metadata_obj_cl;
5138 G1ParScanThreadState* _par_scan_state;
5139
5140 public:
5141 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
5142 OopClosure* non_heap_obj_cl,
5143 OopsInHeapRegionClosure* metadata_obj_cl,
5144 G1ParScanThreadState* pss):
5145 _g1h(g1h),
5146 _copy_non_heap_obj_cl(non_heap_obj_cl),
5147 _copy_metadata_obj_cl(metadata_obj_cl),
5148 _par_scan_state(pss)
5149 {}
5150
5151 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
5152 virtual void do_oop( oop* p) { do_oop_work(p); }
5153
5154 template <class T> void do_oop_work(T* p) {
5155 oop obj = oopDesc::load_decode_heap_oop(p);
5156
5157 if (_g1h->obj_in_cs(obj)) {
5158 // If the referent object has been forwarded (either copied
5159 // to a new location or to itself in the event of an
5160 // evacuation failure) then we need to update the reference
5161 // field and, if both reference and referent are in the G1
5162 // heap, update the RSet for the referent.
5163 //
5164 // If the referent has not been forwarded then we have to keep
5165 // it alive by policy. Therefore we have copy the referent.
5166 //
5167 // If the reference field is in the G1 heap then we can push
5168 // on the PSS queue. When the queue is drained (after each
5169 // phase of reference processing) the object and it's followers
5170 // will be copied, the reference field set to point to the
5171 // new location, and the RSet updated. Otherwise we need to
5172 // use the the non-heap or metadata closures directly to copy
5173 // the refernt object and update the pointer, while avoiding
5174 // updating the RSet.
5175
5176 if (_g1h->is_in_g1_reserved(p)) {
5177 _par_scan_state->push_on_queue(p);
5178 } else {
5179 assert(!ClassLoaderDataGraph::contains((address)p),
5180 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) "
5181 PTR_FORMAT, p));
5182 _copy_non_heap_obj_cl->do_oop(p);
5183 }
5184 }
5185 }
5186 };
5187
5188 // Serial drain queue closure. Called as the 'complete_gc'
5189 // closure for each discovered list in some of the
5190 // reference processing phases.
5191
5192 class G1STWDrainQueueClosure: public VoidClosure {
5193 protected:
5194 G1CollectedHeap* _g1h;
5195 G1ParScanThreadState* _par_scan_state;
5196
5197 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
5198
5199 public:
5200 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
5201 _g1h(g1h),
5202 _par_scan_state(pss)
5203 { }
5204
5205 void do_void() {
5206 G1ParScanThreadState* const pss = par_scan_state();
5207 pss->trim_queue();
5208 }
5209 };
5210
5211 // Parallel Reference Processing closures
5212
5213 // Implementation of AbstractRefProcTaskExecutor for parallel reference
5214 // processing during G1 evacuation pauses.
5215
5216 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
5217 private:
5218 G1CollectedHeap* _g1h;
5219 RefToScanQueueSet* _queues;
5220 FlexibleWorkGang* _workers;
5221 int _active_workers;
5222
5223 public:
5224 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
5225 FlexibleWorkGang* workers,
5226 RefToScanQueueSet *task_queues,
5227 int n_workers) :
5228 _g1h(g1h),
5229 _queues(task_queues),
5230 _workers(workers),
5231 _active_workers(n_workers)
5232 {
5233 assert(n_workers > 0, "shouldn't call this otherwise");
5234 }
5235
5236 // Executes the given task using concurrent marking worker threads.
5237 virtual void execute(ProcessTask& task);
5238 virtual void execute(EnqueueTask& task);
5239 };
5240
5241 // Gang task for possibly parallel reference processing
5242
5243 class G1STWRefProcTaskProxy: public AbstractGangTask {
5244 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5245 ProcessTask& _proc_task;
5246 G1CollectedHeap* _g1h;
5247 RefToScanQueueSet *_task_queues;
5248 ParallelTaskTerminator* _terminator;
5249
5250 public:
5251 G1STWRefProcTaskProxy(ProcessTask& proc_task,
5252 G1CollectedHeap* g1h,
5253 RefToScanQueueSet *task_queues,
5254 ParallelTaskTerminator* terminator) :
5255 AbstractGangTask("Process reference objects in parallel"),
5256 _proc_task(proc_task),
5257 _g1h(g1h),
5258 _task_queues(task_queues),
5259 _terminator(terminator)
5260 {}
5261
5262 virtual void work(uint worker_id) {
5263 // The reference processing task executed by a single worker.
5264 ResourceMark rm;
5265 HandleMark hm;
5266
5267 G1STWIsAliveClosure is_alive(_g1h);
5268
5269 G1ParScanThreadState pss(_g1h, worker_id);
5270
5271 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5272 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5273 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5274
5275 pss.set_evac_closure(&scan_evac_cl);
5276 pss.set_evac_failure_closure(&evac_failure_cl);
5277 pss.set_partial_scan_closure(&partial_scan_cl);
5278
5279 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5280 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5281
5282 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5283 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5284
5285 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5286 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5287
5288 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5289 // We also need to mark copied objects.
5290 copy_non_heap_cl = ©_mark_non_heap_cl;
5291 copy_metadata_cl = ©_mark_metadata_cl;
5292 }
5293
5294 // Keep alive closure.
5295 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5296
5297 // Complete GC closure
5298 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
5299
5300 // Call the reference processing task's work routine.
5301 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue);
5302
5303 // Note we cannot assert that the refs array is empty here as not all
5304 // of the processing tasks (specifically phase2 - pp2_work) execute
5305 // the complete_gc closure (which ordinarily would drain the queue) so
5306 // the queue may not be empty.
5307 }
5308 };
5309
5310 // Driver routine for parallel reference processing.
5311 // Creates an instance of the ref processing gang
5312 // task and has the worker threads execute it.
5313 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
5314 assert(_workers != NULL, "Need parallel worker threads.");
5315
5316 ParallelTaskTerminator terminator(_active_workers, _queues);
5317 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
5318
5319 _g1h->set_par_threads(_active_workers);
5320 _workers->run_task(&proc_task_proxy);
5321 _g1h->set_par_threads(0);
5322 }
5323
5324 // Gang task for parallel reference enqueueing.
5325
5326 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
5327 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5328 EnqueueTask& _enq_task;
5329
5330 public:
5331 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
5332 AbstractGangTask("Enqueue reference objects in parallel"),
5333 _enq_task(enq_task)
5334 { }
5335
5336 virtual void work(uint worker_id) {
5337 _enq_task.work(worker_id);
5338 }
5339 };
5340
5341 // Driver routine for parallel reference enqueing.
5342 // Creates an instance of the ref enqueueing gang
5343 // task and has the worker threads execute it.
5344
5345 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5346 assert(_workers != NULL, "Need parallel worker threads.");
5347
5348 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5349
5350 _g1h->set_par_threads(_active_workers);
5351 _workers->run_task(&enq_task_proxy);
5352 _g1h->set_par_threads(0);
5353 }
5354
5355 // End of weak reference support closures
5356
5357 // Abstract task used to preserve (i.e. copy) any referent objects
5358 // that are in the collection set and are pointed to by reference
5359 // objects discovered by the CM ref processor.
5360
5361 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5362 protected:
5363 G1CollectedHeap* _g1h;
5364 RefToScanQueueSet *_queues;
5365 ParallelTaskTerminator _terminator;
5366 uint _n_workers;
5367
5368 public:
5369 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5370 AbstractGangTask("ParPreserveCMReferents"),
5371 _g1h(g1h),
5372 _queues(task_queues),
5373 _terminator(workers, _queues),
5374 _n_workers(workers)
5375 { }
5376
5377 void work(uint worker_id) {
5378 ResourceMark rm;
5379 HandleMark hm;
5380
5381 G1ParScanThreadState pss(_g1h, worker_id);
5382 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5383 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5384 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5385
5386 pss.set_evac_closure(&scan_evac_cl);
5387 pss.set_evac_failure_closure(&evac_failure_cl);
5388 pss.set_partial_scan_closure(&partial_scan_cl);
5389
5390 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5391
5392
5393 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5394 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5395
5396 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5397 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5398
5399 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5400 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5401
5402 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5403 // We also need to mark copied objects.
5404 copy_non_heap_cl = ©_mark_non_heap_cl;
5405 copy_metadata_cl = ©_mark_metadata_cl;
5406 }
5407
5408 // Is alive closure
5409 G1AlwaysAliveClosure always_alive(_g1h);
5410
5411 // Copying keep alive closure. Applied to referent objects that need
5412 // to be copied.
5413 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5414
5415 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5416
5417 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5418 uint stride = MIN2(MAX2(_n_workers, 1U), limit);
5419
5420 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5421 // So this must be true - but assert just in case someone decides to
5422 // change the worker ids.
5423 assert(0 <= worker_id && worker_id < limit, "sanity");
5424 assert(!rp->discovery_is_atomic(), "check this code");
5425
5426 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5427 for (uint idx = worker_id; idx < limit; idx += stride) {
5428 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5429
5430 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5431 while (iter.has_next()) {
5432 // Since discovery is not atomic for the CM ref processor, we
5433 // can see some null referent objects.
5434 iter.load_ptrs(DEBUG_ONLY(true));
5435 oop ref = iter.obj();
5436
5437 // This will filter nulls.
5438 if (iter.is_referent_alive()) {
5439 iter.make_referent_alive();
5440 }
5441 iter.move_to_next();
5442 }
5443 }
5444
5445 // Drain the queue - which may cause stealing
5446 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5447 drain_queue.do_void();
5448 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5449 assert(pss.refs()->is_empty(), "should be");
5450 }
5451 };
5452
5453 // Weak Reference processing during an evacuation pause (part 1).
5454 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) {
5455 double ref_proc_start = os::elapsedTime();
5456
5457 ReferenceProcessor* rp = _ref_processor_stw;
5458 assert(rp->discovery_enabled(), "should have been enabled");
5459
5460 // Any reference objects, in the collection set, that were 'discovered'
5461 // by the CM ref processor should have already been copied (either by
5462 // applying the external root copy closure to the discovered lists, or
5463 // by following an RSet entry).
5464 //
5465 // But some of the referents, that are in the collection set, that these
5466 // reference objects point to may not have been copied: the STW ref
5467 // processor would have seen that the reference object had already
5468 // been 'discovered' and would have skipped discovering the reference,
5469 // but would not have treated the reference object as a regular oop.
5470 // As a reult the copy closure would not have been applied to the
5471 // referent object.
5472 //
5473 // We need to explicitly copy these referent objects - the references
5474 // will be processed at the end of remarking.
5475 //
5476 // We also need to do this copying before we process the reference
5477 // objects discovered by the STW ref processor in case one of these
5478 // referents points to another object which is also referenced by an
5479 // object discovered by the STW ref processor.
5480
5481 assert(!G1CollectedHeap::use_parallel_gc_threads() ||
5482 no_of_gc_workers == workers()->active_workers(),
5483 "Need to reset active GC workers");
5484
5485 set_par_threads(no_of_gc_workers);
5486 G1ParPreserveCMReferentsTask keep_cm_referents(this,
5487 no_of_gc_workers,
5488 _task_queues);
5489
5490 if (G1CollectedHeap::use_parallel_gc_threads()) {
5491 workers()->run_task(&keep_cm_referents);
5492 } else {
5493 keep_cm_referents.work(0);
5494 }
5495
5496 set_par_threads(0);
5497
5498 // Closure to test whether a referent is alive.
5499 G1STWIsAliveClosure is_alive(this);
5500
5501 // Even when parallel reference processing is enabled, the processing
5502 // of JNI refs is serial and performed serially by the current thread
5503 // rather than by a worker. The following PSS will be used for processing
5504 // JNI refs.
5505
5506 // Use only a single queue for this PSS.
5507 G1ParScanThreadState pss(this, 0);
5508
5509 // We do not embed a reference processor in the copying/scanning
5510 // closures while we're actually processing the discovered
5511 // reference objects.
5512 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5513 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5514 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5515
5516 pss.set_evac_closure(&scan_evac_cl);
5517 pss.set_evac_failure_closure(&evac_failure_cl);
5518 pss.set_partial_scan_closure(&partial_scan_cl);
5519
5520 assert(pss.refs()->is_empty(), "pre-condition");
5521
5522 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5523 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL);
5524
5525 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5526 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL);
5527
5528 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5529 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5530
5531 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5532 // We also need to mark copied objects.
5533 copy_non_heap_cl = ©_mark_non_heap_cl;
5534 copy_metadata_cl = ©_mark_metadata_cl;
5535 }
5536
5537 // Keep alive closure.
5538 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss);
5539
5540 // Serial Complete GC closure
5541 G1STWDrainQueueClosure drain_queue(this, &pss);
5542
5543 // Setup the soft refs policy...
5544 rp->setup_policy(false);
5545
5546 if (!rp->processing_is_mt()) {
5547 // Serial reference processing...
5548 rp->process_discovered_references(&is_alive,
5549 &keep_alive,
5550 &drain_queue,
5551 NULL);
5552 } else {
5553 // Parallel reference processing
5554 assert(rp->num_q() == no_of_gc_workers, "sanity");
5555 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5556
5557 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5558 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5559 }
5560
5561 // We have completed copying any necessary live referent objects
5562 // (that were not copied during the actual pause) so we can
5563 // retire any active alloc buffers
5564 pss.retire_alloc_buffers();
5565 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5566
5567 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5568 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0);
5569 }
5570
5571 // Weak Reference processing during an evacuation pause (part 2).
5572 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) {
5573 double ref_enq_start = os::elapsedTime();
5574
5575 ReferenceProcessor* rp = _ref_processor_stw;
5576 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5577
5578 // Now enqueue any remaining on the discovered lists on to
5579 // the pending list.
5580 if (!rp->processing_is_mt()) {
5581 // Serial reference processing...
5582 rp->enqueue_discovered_references();
5583 } else {
5584 // Parallel reference enqueuing
5585
5586 assert(no_of_gc_workers == workers()->active_workers(),
5587 "Need to reset active workers");
5588 assert(rp->num_q() == no_of_gc_workers, "sanity");
5589 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5590
5591 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5592 rp->enqueue_discovered_references(&par_task_executor);
5593 }
5594
5595 rp->verify_no_references_recorded();
5596 assert(!rp->discovery_enabled(), "should have been disabled");
5597
5598 // FIXME
5599 // CM's reference processing also cleans up the string and symbol tables.
5600 // Should we do that here also? We could, but it is a serial operation
5601 // and could signicantly increase the pause time.
5602
5603 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5604 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0);
5605 }
5606
5607 void G1CollectedHeap::evacuate_collection_set() {
5608 _expand_heap_after_alloc_failure = true;
5609 set_evacuation_failed(false);
5610
5611 // Should G1EvacuationFailureALot be in effect for this GC?
5612 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();)
5613
5614 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5615 concurrent_g1_refine()->set_use_cache(false);
5616 concurrent_g1_refine()->clear_hot_cache_claimed_index();
5617
5618 uint n_workers;
5619 if (G1CollectedHeap::use_parallel_gc_threads()) {
5620 n_workers =
5621 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
5622 workers()->active_workers(),
5623 Threads::number_of_non_daemon_threads());
5624 assert(UseDynamicNumberOfGCThreads ||
5625 n_workers == workers()->total_workers(),
5626 "If not dynamic should be using all the workers");
5627 workers()->set_active_workers(n_workers);
5628 set_par_threads(n_workers);
5629 } else {
5630 assert(n_par_threads() == 0,
5631 "Should be the original non-parallel value");
5632 n_workers = 1;
5633 }
5634
5635 G1ParTask g1_par_task(this, _task_queues);
5636
5637 init_for_evac_failure(NULL);
5638
5639 rem_set()->prepare_for_younger_refs_iterate(true);
5640
5641 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5642 double start_par_time_sec = os::elapsedTime();
5643 double end_par_time_sec;
5644
5645 {
5646 StrongRootsScope srs(this);
5647
5648 if (G1CollectedHeap::use_parallel_gc_threads()) {
5649 // The individual threads will set their evac-failure closures.
5650 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5651 // These tasks use ShareHeap::_process_strong_tasks
5652 assert(UseDynamicNumberOfGCThreads ||
5653 workers()->active_workers() == workers()->total_workers(),
5654 "If not dynamic should be using all the workers");
5655 workers()->run_task(&g1_par_task);
5656 } else {
5657 g1_par_task.set_for_termination(n_workers);
5658 g1_par_task.work(0);
5659 }
5660 end_par_time_sec = os::elapsedTime();
5661
5662 // Closing the inner scope will execute the destructor
5663 // for the StrongRootsScope object. We record the current
5664 // elapsed time before closing the scope so that time
5665 // taken for the SRS destructor is NOT included in the
5666 // reported parallel time.
5667 }
5668
5669 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0;
5670 g1_policy()->phase_times()->record_par_time(par_time_ms);
5671
5672 double code_root_fixup_time_ms =
5673 (os::elapsedTime() - end_par_time_sec) * 1000.0;
5674 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms);
5675
5676 set_par_threads(0);
5677
5678 // Process any discovered reference objects - we have
5679 // to do this _before_ we retire the GC alloc regions
5680 // as we may have to copy some 'reachable' referent
5681 // objects (and their reachable sub-graphs) that were
5682 // not copied during the pause.
5683 process_discovered_references(n_workers);
5684
5685 // Weak root processing.
5686 // Note: when JSR 292 is enabled and code blobs can contain
5687 // non-perm oops then we will need to process the code blobs
5688 // here too.
5689 {
5690 G1STWIsAliveClosure is_alive(this);
5691 G1KeepAliveClosure keep_alive(this);
5692 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5693 }
5694
5695 release_gc_alloc_regions(n_workers);
5696 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5697
5698 concurrent_g1_refine()->clear_hot_cache();
5699 concurrent_g1_refine()->set_use_cache(true);
5700
5701 finalize_for_evac_failure();
5702
5703 if (evacuation_failed()) {
5704 remove_self_forwarding_pointers();
5705
5706 // Reset the G1EvacuationFailureALot counters and flags
5707 // Note: the values are reset only when an actual
5708 // evacuation failure occurs.
5709 NOT_PRODUCT(reset_evacuation_should_fail();)
5710 }
5711
5712 // Enqueue any remaining references remaining on the STW
5713 // reference processor's discovered lists. We need to do
5714 // this after the card table is cleaned (and verified) as
5715 // the act of enqueuing entries on to the pending list
5716 // will log these updates (and dirty their associated
5717 // cards). We need these updates logged to update any
5718 // RSets.
5719 enqueue_discovered_references(n_workers);
5720
5721 if (G1DeferredRSUpdate) {
5722 RedirtyLoggedCardTableEntryFastClosure redirty;
5723 dirty_card_queue_set().set_closure(&redirty);
5724 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5725
5726 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5727 dcq.merge_bufferlists(&dirty_card_queue_set());
5728 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5729 }
5730 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5731 }
5732
5733 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5734 size_t* pre_used,
5735 FreeRegionList* free_list,
5736 OldRegionSet* old_proxy_set,
5737 HumongousRegionSet* humongous_proxy_set,
5738 HRRSCleanupTask* hrrs_cleanup_task,
5739 bool par) {
5740 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5741 if (hr->isHumongous()) {
5742 assert(hr->startsHumongous(), "we should only see starts humongous");
5743 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5744 } else {
5745 _old_set.remove_with_proxy(hr, old_proxy_set);
5746 free_region(hr, pre_used, free_list, par);
5747 }
5748 } else {
5749 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5750 }
5751 }
5752
5753 void G1CollectedHeap::free_region(HeapRegion* hr,
5754 size_t* pre_used,
5755 FreeRegionList* free_list,
5756 bool par) {
5757 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5758 assert(!hr->is_empty(), "the region should not be empty");
5759 assert(free_list != NULL, "pre-condition");
5760
5761 *pre_used += hr->used();
5762 hr->hr_clear(par, true /* clear_space */);
5763 free_list->add_as_head(hr);
5764 }
5765
5766 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5767 size_t* pre_used,
5768 FreeRegionList* free_list,
5769 HumongousRegionSet* humongous_proxy_set,
5770 bool par) {
5771 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5772 assert(free_list != NULL, "pre-condition");
5773 assert(humongous_proxy_set != NULL, "pre-condition");
5774
5775 size_t hr_used = hr->used();
5776 size_t hr_capacity = hr->capacity();
5777 size_t hr_pre_used = 0;
5778 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5779 // We need to read this before we make the region non-humongous,
5780 // otherwise the information will be gone.
5781 uint last_index = hr->last_hc_index();
5782 hr->set_notHumongous();
5783 free_region(hr, &hr_pre_used, free_list, par);
5784
5785 uint i = hr->hrs_index() + 1;
5786 while (i < last_index) {
5787 HeapRegion* curr_hr = region_at(i);
5788 assert(curr_hr->continuesHumongous(), "invariant");
5789 curr_hr->set_notHumongous();
5790 free_region(curr_hr, &hr_pre_used, free_list, par);
5791 i += 1;
5792 }
5793 assert(hr_pre_used == hr_used,
5794 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5795 "should be the same", hr_pre_used, hr_used));
5796 *pre_used += hr_pre_used;
5797 }
5798
5799 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5800 FreeRegionList* free_list,
5801 OldRegionSet* old_proxy_set,
5802 HumongousRegionSet* humongous_proxy_set,
5803 bool par) {
5804 if (pre_used > 0) {
5805 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5806 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5807 assert(_summary_bytes_used >= pre_used,
5808 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5809 "should be >= pre_used: "SIZE_FORMAT,
5810 _summary_bytes_used, pre_used));
5811 _summary_bytes_used -= pre_used;
5812 }
5813 if (free_list != NULL && !free_list->is_empty()) {
5814 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5815 _free_list.add_as_head(free_list);
5816 }
5817 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) {
5818 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5819 _old_set.update_from_proxy(old_proxy_set);
5820 }
5821 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5822 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5823 _humongous_set.update_from_proxy(humongous_proxy_set);
5824 }
5825 }
5826
5827 class G1ParCleanupCTTask : public AbstractGangTask {
5828 CardTableModRefBS* _ct_bs;
5829 G1CollectedHeap* _g1h;
5830 HeapRegion* volatile _su_head;
5831 public:
5832 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5833 G1CollectedHeap* g1h) :
5834 AbstractGangTask("G1 Par Cleanup CT Task"),
5835 _ct_bs(ct_bs), _g1h(g1h) { }
5836
5837 void work(uint worker_id) {
5838 HeapRegion* r;
5839 while (r = _g1h->pop_dirty_cards_region()) {
5840 clear_cards(r);
5841 }
5842 }
5843
5844 void clear_cards(HeapRegion* r) {
5845 // Cards of the survivors should have already been dirtied.
5846 if (!r->is_survivor()) {
5847 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5848 }
5849 }
5850 };
5851
5852 #ifndef PRODUCT
5853 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5854 G1CollectedHeap* _g1h;
5855 CardTableModRefBS* _ct_bs;
5856 public:
5857 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5858 : _g1h(g1h), _ct_bs(ct_bs) { }
5859 virtual bool doHeapRegion(HeapRegion* r) {
5860 if (r->is_survivor()) {
5861 _g1h->verify_dirty_region(r);
5862 } else {
5863 _g1h->verify_not_dirty_region(r);
5864 }
5865 return false;
5866 }
5867 };
5868
5869 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5870 // All of the region should be clean.
5871 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5872 MemRegion mr(hr->bottom(), hr->end());
5873 ct_bs->verify_not_dirty_region(mr);
5874 }
5875
5876 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5877 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5878 // dirty allocated blocks as they allocate them. The thread that
5879 // retires each region and replaces it with a new one will do a
5880 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5881 // not dirty that area (one less thing to have to do while holding
5882 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5883 // is dirty.
5884 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5885 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5886 ct_bs->verify_dirty_region(mr);
5887 }
5888
5889 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5890 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5891 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5892 verify_dirty_region(hr);
5893 }
5894 }
5895
5896 void G1CollectedHeap::verify_dirty_young_regions() {
5897 verify_dirty_young_list(_young_list->first_region());
5898 }
5899 #endif
5900
5901 void G1CollectedHeap::cleanUpCardTable() {
5902 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5903 double start = os::elapsedTime();
5904
5905 {
5906 // Iterate over the dirty cards region list.
5907 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5908
5909 if (G1CollectedHeap::use_parallel_gc_threads()) {
5910 set_par_threads();
5911 workers()->run_task(&cleanup_task);
5912 set_par_threads(0);
5913 } else {
5914 while (_dirty_cards_region_list) {
5915 HeapRegion* r = _dirty_cards_region_list;
5916 cleanup_task.clear_cards(r);
5917 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5918 if (_dirty_cards_region_list == r) {
5919 // The last region.
5920 _dirty_cards_region_list = NULL;
5921 }
5922 r->set_next_dirty_cards_region(NULL);
5923 }
5924 }
5925 #ifndef PRODUCT
5926 if (G1VerifyCTCleanup || VerifyAfterGC) {
5927 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5928 heap_region_iterate(&cleanup_verifier);
5929 }
5930 #endif
5931 }
5932
5933 double elapsed = os::elapsedTime() - start;
5934 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0);
5935 }
5936
5937 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5938 size_t pre_used = 0;
5939 FreeRegionList local_free_list("Local List for CSet Freeing");
5940
5941 double young_time_ms = 0.0;
5942 double non_young_time_ms = 0.0;
5943
5944 // Since the collection set is a superset of the the young list,
5945 // all we need to do to clear the young list is clear its
5946 // head and length, and unlink any young regions in the code below
5947 _young_list->clear();
5948
5949 G1CollectorPolicy* policy = g1_policy();
5950
5951 double start_sec = os::elapsedTime();
5952 bool non_young = true;
5953
5954 HeapRegion* cur = cs_head;
5955 int age_bound = -1;
5956 size_t rs_lengths = 0;
5957
5958 while (cur != NULL) {
5959 assert(!is_on_master_free_list(cur), "sanity");
5960 if (non_young) {
5961 if (cur->is_young()) {
5962 double end_sec = os::elapsedTime();
5963 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5964 non_young_time_ms += elapsed_ms;
5965
5966 start_sec = os::elapsedTime();
5967 non_young = false;
5968 }
5969 } else {
5970 if (!cur->is_young()) {
5971 double end_sec = os::elapsedTime();
5972 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5973 young_time_ms += elapsed_ms;
5974
5975 start_sec = os::elapsedTime();
5976 non_young = true;
5977 }
5978 }
5979
5980 rs_lengths += cur->rem_set()->occupied();
5981
5982 HeapRegion* next = cur->next_in_collection_set();
5983 assert(cur->in_collection_set(), "bad CS");
5984 cur->set_next_in_collection_set(NULL);
5985 cur->set_in_collection_set(false);
5986
5987 if (cur->is_young()) {
5988 int index = cur->young_index_in_cset();
5989 assert(index != -1, "invariant");
5990 assert((uint) index < policy->young_cset_region_length(), "invariant");
5991 size_t words_survived = _surviving_young_words[index];
5992 cur->record_surv_words_in_group(words_survived);
5993
5994 // At this point the we have 'popped' cur from the collection set
5995 // (linked via next_in_collection_set()) but it is still in the
5996 // young list (linked via next_young_region()). Clear the
5997 // _next_young_region field.
5998 cur->set_next_young_region(NULL);
5999 } else {
6000 int index = cur->young_index_in_cset();
6001 assert(index == -1, "invariant");
6002 }
6003
6004 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
6005 (!cur->is_young() && cur->young_index_in_cset() == -1),
6006 "invariant" );
6007
6008 if (!cur->evacuation_failed()) {
6009 MemRegion used_mr = cur->used_region();
6010
6011 // And the region is empty.
6012 assert(!used_mr.is_empty(), "Should not have empty regions in a CS.");
6013 free_region(cur, &pre_used, &local_free_list, false /* par */);
6014 } else {
6015 cur->uninstall_surv_rate_group();
6016 if (cur->is_young()) {
6017 cur->set_young_index_in_cset(-1);
6018 }
6019 cur->set_not_young();
6020 cur->set_evacuation_failed(false);
6021 // The region is now considered to be old.
6022 _old_set.add(cur);
6023 }
6024 cur = next;
6025 }
6026
6027 policy->record_max_rs_lengths(rs_lengths);
6028 policy->cset_regions_freed();
6029
6030 double end_sec = os::elapsedTime();
6031 double elapsed_ms = (end_sec - start_sec) * 1000.0;
6032
6033 if (non_young) {
6034 non_young_time_ms += elapsed_ms;
6035 } else {
6036 young_time_ms += elapsed_ms;
6037 }
6038
6039 update_sets_after_freeing_regions(pre_used, &local_free_list,
6040 NULL /* old_proxy_set */,
6041 NULL /* humongous_proxy_set */,
6042 false /* par */);
6043 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms);
6044 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms);
6045 }
6046
6047 // This routine is similar to the above but does not record
6048 // any policy statistics or update free lists; we are abandoning
6049 // the current incremental collection set in preparation of a
6050 // full collection. After the full GC we will start to build up
6051 // the incremental collection set again.
6052 // This is only called when we're doing a full collection
6053 // and is immediately followed by the tearing down of the young list.
6054
6055 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
6056 HeapRegion* cur = cs_head;
6057
6058 while (cur != NULL) {
6059 HeapRegion* next = cur->next_in_collection_set();
6060 assert(cur->in_collection_set(), "bad CS");
6061 cur->set_next_in_collection_set(NULL);
6062 cur->set_in_collection_set(false);
6063 cur->set_young_index_in_cset(-1);
6064 cur = next;
6065 }
6066 }
6067
6068 void G1CollectedHeap::set_free_regions_coming() {
6069 if (G1ConcRegionFreeingVerbose) {
6070 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6071 "setting free regions coming");
6072 }
6073
6074 assert(!free_regions_coming(), "pre-condition");
6075 _free_regions_coming = true;
6076 }
6077
6078 void G1CollectedHeap::reset_free_regions_coming() {
6079 assert(free_regions_coming(), "pre-condition");
6080
6081 {
6082 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6083 _free_regions_coming = false;
6084 SecondaryFreeList_lock->notify_all();
6085 }
6086
6087 if (G1ConcRegionFreeingVerbose) {
6088 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6089 "reset free regions coming");
6090 }
6091 }
6092
6093 void G1CollectedHeap::wait_while_free_regions_coming() {
6094 // Most of the time we won't have to wait, so let's do a quick test
6095 // first before we take the lock.
6096 if (!free_regions_coming()) {
6097 return;
6098 }
6099
6100 if (G1ConcRegionFreeingVerbose) {
6101 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6102 "waiting for free regions");
6103 }
6104
6105 {
6106 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6107 while (free_regions_coming()) {
6108 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
6109 }
6110 }
6111
6112 if (G1ConcRegionFreeingVerbose) {
6113 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6114 "done waiting for free regions");
6115 }
6116 }
6117
6118 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
6119 assert(heap_lock_held_for_gc(),
6120 "the heap lock should already be held by or for this thread");
6121 _young_list->push_region(hr);
6122 }
6123
6124 class NoYoungRegionsClosure: public HeapRegionClosure {
6125 private:
6126 bool _success;
6127 public:
6128 NoYoungRegionsClosure() : _success(true) { }
6129 bool doHeapRegion(HeapRegion* r) {
6130 if (r->is_young()) {
6131 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
6132 r->bottom(), r->end());
6133 _success = false;
6134 }
6135 return false;
6136 }
6137 bool success() { return _success; }
6138 };
6139
6140 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
6141 bool ret = _young_list->check_list_empty(check_sample);
6142
6143 if (check_heap) {
6144 NoYoungRegionsClosure closure;
6145 heap_region_iterate(&closure);
6146 ret = ret && closure.success();
6147 }
6148
6149 return ret;
6150 }
6151
6152 class TearDownRegionSetsClosure : public HeapRegionClosure {
6153 private:
6154 OldRegionSet *_old_set;
6155
6156 public:
6157 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { }
6158
6159 bool doHeapRegion(HeapRegion* r) {
6160 if (r->is_empty()) {
6161 // We ignore empty regions, we'll empty the free list afterwards
6162 } else if (r->is_young()) {
6163 // We ignore young regions, we'll empty the young list afterwards
6164 } else if (r->isHumongous()) {
6165 // We ignore humongous regions, we're not tearing down the
6166 // humongous region set
6167 } else {
6168 // The rest should be old
6169 _old_set->remove(r);
6170 }
6171 return false;
6172 }
6173
6174 ~TearDownRegionSetsClosure() {
6175 assert(_old_set->is_empty(), "post-condition");
6176 }
6177 };
6178
6179 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) {
6180 assert_at_safepoint(true /* should_be_vm_thread */);
6181
6182 if (!free_list_only) {
6183 TearDownRegionSetsClosure cl(&_old_set);
6184 heap_region_iterate(&cl);
6185
6186 // Need to do this after the heap iteration to be able to
6187 // recognize the young regions and ignore them during the iteration.
6188 _young_list->empty_list();
6189 }
6190 _free_list.remove_all();
6191 }
6192
6193 class RebuildRegionSetsClosure : public HeapRegionClosure {
6194 private:
6195 bool _free_list_only;
6196 OldRegionSet* _old_set;
6197 FreeRegionList* _free_list;
6198 size_t _total_used;
6199
6200 public:
6201 RebuildRegionSetsClosure(bool free_list_only,
6202 OldRegionSet* old_set, FreeRegionList* free_list) :
6203 _free_list_only(free_list_only),
6204 _old_set(old_set), _free_list(free_list), _total_used(0) {
6205 assert(_free_list->is_empty(), "pre-condition");
6206 if (!free_list_only) {
6207 assert(_old_set->is_empty(), "pre-condition");
6208 }
6209 }
6210
6211 bool doHeapRegion(HeapRegion* r) {
6212 if (r->continuesHumongous()) {
6213 return false;
6214 }
6215
6216 if (r->is_empty()) {
6217 // Add free regions to the free list
6218 _free_list->add_as_tail(r);
6219 } else if (!_free_list_only) {
6220 assert(!r->is_young(), "we should not come across young regions");
6221
6222 if (r->isHumongous()) {
6223 // We ignore humongous regions, we left the humongous set unchanged
6224 } else {
6225 // The rest should be old, add them to the old set
6226 _old_set->add(r);
6227 }
6228 _total_used += r->used();
6229 }
6230
6231 return false;
6232 }
6233
6234 size_t total_used() {
6235 return _total_used;
6236 }
6237 };
6238
6239 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) {
6240 assert_at_safepoint(true /* should_be_vm_thread */);
6241
6242 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list);
6243 heap_region_iterate(&cl);
6244
6245 if (!free_list_only) {
6246 _summary_bytes_used = cl.total_used();
6247 }
6248 assert(_summary_bytes_used == recalculate_used(),
6249 err_msg("inconsistent _summary_bytes_used, "
6250 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT,
6251 _summary_bytes_used, recalculate_used()));
6252 }
6253
6254 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
6255 _refine_cte_cl->set_concurrent(concurrent);
6256 }
6257
6258 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
6259 HeapRegion* hr = heap_region_containing(p);
6260 if (hr == NULL) {
6261 return false;
6262 } else {
6263 return hr->is_in(p);
6264 }
6265 }
6266
6267 // Methods for the mutator alloc region
6268
6269 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
6270 bool force) {
6271 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6272 assert(!force || g1_policy()->can_expand_young_list(),
6273 "if force is true we should be able to expand the young list");
6274 bool young_list_full = g1_policy()->is_young_list_full();
6275 if (force || !young_list_full) {
6276 HeapRegion* new_alloc_region = new_region(word_size,
6277 false /* do_expand */);
6278 if (new_alloc_region != NULL) {
6279 set_region_short_lived_locked(new_alloc_region);
6280 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
6281 return new_alloc_region;
6282 }
6283 }
6284 return NULL;
6285 }
6286
6287 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
6288 size_t allocated_bytes) {
6289 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6290 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
6291
6292 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
6293 _summary_bytes_used += allocated_bytes;
6294 _hr_printer.retire(alloc_region);
6295 // We update the eden sizes here, when the region is retired,
6296 // instead of when it's allocated, since this is the point that its
6297 // used space has been recored in _summary_bytes_used.
6298 g1mm()->update_eden_size();
6299 }
6300
6301 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
6302 bool force) {
6303 return _g1h->new_mutator_alloc_region(word_size, force);
6304 }
6305
6306 void G1CollectedHeap::set_par_threads() {
6307 // Don't change the number of workers. Use the value previously set
6308 // in the workgroup.
6309 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise");
6310 uint n_workers = workers()->active_workers();
6311 assert(UseDynamicNumberOfGCThreads ||
6312 n_workers == workers()->total_workers(),
6313 "Otherwise should be using the total number of workers");
6314 if (n_workers == 0) {
6315 assert(false, "Should have been set in prior evacuation pause.");
6316 n_workers = ParallelGCThreads;
6317 workers()->set_active_workers(n_workers);
6318 }
6319 set_par_threads(n_workers);
6320 }
6321
6322 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
6323 size_t allocated_bytes) {
6324 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
6325 }
6326
6327 // Methods for the GC alloc regions
6328
6329 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
6330 uint count,
6331 GCAllocPurpose ap) {
6332 assert(FreeList_lock->owned_by_self(), "pre-condition");
6333
6334 if (count < g1_policy()->max_regions(ap)) {
6335 HeapRegion* new_alloc_region = new_region(word_size,
6336 true /* do_expand */);
6337 if (new_alloc_region != NULL) {
6338 // We really only need to do this for old regions given that we
6339 // should never scan survivors. But it doesn't hurt to do it
6340 // for survivors too.
6341 new_alloc_region->set_saved_mark();
6342 if (ap == GCAllocForSurvived) {
6343 new_alloc_region->set_survivor();
6344 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
6345 } else {
6346 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
6347 }
6348 bool during_im = g1_policy()->during_initial_mark_pause();
6349 new_alloc_region->note_start_of_copying(during_im);
6350 return new_alloc_region;
6351 } else {
6352 g1_policy()->note_alloc_region_limit_reached(ap);
6353 }
6354 }
6355 return NULL;
6356 }
6357
6358 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
6359 size_t allocated_bytes,
6360 GCAllocPurpose ap) {
6361 bool during_im = g1_policy()->during_initial_mark_pause();
6362 alloc_region->note_end_of_copying(during_im);
6363 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
6364 if (ap == GCAllocForSurvived) {
6365 young_list()->add_survivor_region(alloc_region);
6366 } else {
6367 _old_set.add(alloc_region);
6368 }
6369 _hr_printer.retire(alloc_region);
6370 }
6371
6372 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
6373 bool force) {
6374 assert(!force, "not supported for GC alloc regions");
6375 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
6376 }
6377
6378 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
6379 size_t allocated_bytes) {
6380 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6381 GCAllocForSurvived);
6382 }
6383
6384 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
6385 bool force) {
6386 assert(!force, "not supported for GC alloc regions");
6387 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
6388 }
6389
6390 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
6391 size_t allocated_bytes) {
6392 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6393 GCAllocForTenured);
6394 }
6395 // Heap region set verification
6396
6397 class VerifyRegionListsClosure : public HeapRegionClosure {
6398 private:
6399 FreeRegionList* _free_list;
6400 OldRegionSet* _old_set;
6401 HumongousRegionSet* _humongous_set;
6402 uint _region_count;
6403
6404 public:
6405 VerifyRegionListsClosure(OldRegionSet* old_set,
6406 HumongousRegionSet* humongous_set,
6407 FreeRegionList* free_list) :
6408 _old_set(old_set), _humongous_set(humongous_set),
6409 _free_list(free_list), _region_count(0) { }
6410
6411 uint region_count() { return _region_count; }
6412
6413 bool doHeapRegion(HeapRegion* hr) {
6414 _region_count += 1;
6415
6416 if (hr->continuesHumongous()) {
6417 return false;
6418 }
6419
6420 if (hr->is_young()) {
6421 // TODO
6422 } else if (hr->startsHumongous()) {
6423 _humongous_set->verify_next_region(hr);
6424 } else if (hr->is_empty()) {
6425 _free_list->verify_next_region(hr);
6426 } else {
6427 _old_set->verify_next_region(hr);
6428 }
6429 return false;
6430 }
6431 };
6432
6433 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index,
6434 HeapWord* bottom) {
6435 HeapWord* end = bottom + HeapRegion::GrainWords;
6436 MemRegion mr(bottom, end);
6437 assert(_g1_reserved.contains(mr), "invariant");
6438 // This might return NULL if the allocation fails
6439 return new HeapRegion(hrs_index, _bot_shared, mr);
6440 }
6441
6442 void G1CollectedHeap::verify_region_sets() {
6443 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6444
6445 // First, check the explicit lists.
6446 _free_list.verify();
6447 {
6448 // Given that a concurrent operation might be adding regions to
6449 // the secondary free list we have to take the lock before
6450 // verifying it.
6451 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6452 _secondary_free_list.verify();
6453 }
6454 _old_set.verify();
6455 _humongous_set.verify();
6456
6457 // If a concurrent region freeing operation is in progress it will
6458 // be difficult to correctly attributed any free regions we come
6459 // across to the correct free list given that they might belong to
6460 // one of several (free_list, secondary_free_list, any local lists,
6461 // etc.). So, if that's the case we will skip the rest of the
6462 // verification operation. Alternatively, waiting for the concurrent
6463 // operation to complete will have a non-trivial effect on the GC's
6464 // operation (no concurrent operation will last longer than the
6465 // interval between two calls to verification) and it might hide
6466 // any issues that we would like to catch during testing.
6467 if (free_regions_coming()) {
6468 return;
6469 }
6470
6471 // Make sure we append the secondary_free_list on the free_list so
6472 // that all free regions we will come across can be safely
6473 // attributed to the free_list.
6474 append_secondary_free_list_if_not_empty_with_lock();
6475
6476 // Finally, make sure that the region accounting in the lists is
6477 // consistent with what we see in the heap.
6478 _old_set.verify_start();
6479 _humongous_set.verify_start();
6480 _free_list.verify_start();
6481
6482 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list);
6483 heap_region_iterate(&cl);
6484
6485 _old_set.verify_end();
6486 _humongous_set.verify_end();
6487 _free_list.verify_end();
6488 }