1 /*
2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "code/icBuffer.hpp"
27 #include "gc_implementation/g1/bufferingOopClosure.hpp"
28 #include "gc_implementation/g1/concurrentG1Refine.hpp"
29 #include "gc_implementation/g1/concurrentG1RefineThread.hpp"
30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
31 #include "gc_implementation/g1/g1AllocRegion.inline.hpp"
32 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
33 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
34 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
35 #include "gc_implementation/g1/g1EvacFailure.hpp"
36 #include "gc_implementation/g1/g1GCPhaseTimes.hpp"
37 #include "gc_implementation/g1/g1Log.hpp"
38 #include "gc_implementation/g1/g1MarkSweep.hpp"
39 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
40 #include "gc_implementation/g1/g1RemSet.inline.hpp"
41 #include "gc_implementation/g1/heapRegion.inline.hpp"
42 #include "gc_implementation/g1/heapRegionRemSet.hpp"
43 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
44 #include "gc_implementation/g1/vm_operations_g1.hpp"
45 #include "gc_implementation/shared/isGCActiveMark.hpp"
46 #include "memory/gcLocker.inline.hpp"
47 #include "memory/genOopClosures.inline.hpp"
48 #include "memory/generationSpec.hpp"
49 #include "memory/referenceProcessor.hpp"
50 #include "oops/oop.inline.hpp"
51 #include "oops/oop.pcgc.inline.hpp"
52 #include "runtime/aprofiler.hpp"
53 #include "runtime/vmThread.hpp"
54
55 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0;
56
57 // turn it on so that the contents of the young list (scan-only /
58 // to-be-collected) are printed at "strategic" points before / during
59 // / after the collection --- this is useful for debugging
60 #define YOUNG_LIST_VERBOSE 0
61 // CURRENT STATUS
62 // This file is under construction. Search for "FIXME".
63
64 // INVARIANTS/NOTES
65 //
66 // All allocation activity covered by the G1CollectedHeap interface is
67 // serialized by acquiring the HeapLock. This happens in mem_allocate
68 // and allocate_new_tlab, which are the "entry" points to the
69 // allocation code from the rest of the JVM. (Note that this does not
70 // apply to TLAB allocation, which is not part of this interface: it
71 // is done by clients of this interface.)
72
73 // Notes on implementation of parallelism in different tasks.
74 //
75 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism.
76 // The number of GC workers is passed to heap_region_par_iterate_chunked().
77 // It does use run_task() which sets _n_workers in the task.
78 // G1ParTask executes g1_process_strong_roots() ->
79 // SharedHeap::process_strong_roots() which calls eventuall to
80 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses
81 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also
82 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap).
83 //
84
85 // Local to this file.
86
87 class RefineCardTableEntryClosure: public CardTableEntryClosure {
88 SuspendibleThreadSet* _sts;
89 G1RemSet* _g1rs;
90 ConcurrentG1Refine* _cg1r;
91 bool _concurrent;
92 public:
93 RefineCardTableEntryClosure(SuspendibleThreadSet* sts,
94 G1RemSet* g1rs,
95 ConcurrentG1Refine* cg1r) :
96 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true)
97 {}
98 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
99 bool oops_into_cset = _g1rs->refine_card(card_ptr, worker_i, false);
100 // This path is executed by the concurrent refine or mutator threads,
101 // concurrently, and so we do not care if card_ptr contains references
102 // that point into the collection set.
103 assert(!oops_into_cset, "should be");
104
105 if (_concurrent && _sts->should_yield()) {
106 // Caller will actually yield.
107 return false;
108 }
109 // Otherwise, we finished successfully; return true.
110 return true;
111 }
112 void set_concurrent(bool b) { _concurrent = b; }
113 };
114
115
116 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure {
117 int _calls;
118 G1CollectedHeap* _g1h;
119 CardTableModRefBS* _ctbs;
120 int _histo[256];
121 public:
122 ClearLoggedCardTableEntryClosure() :
123 _calls(0)
124 {
125 _g1h = G1CollectedHeap::heap();
126 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
127 for (int i = 0; i < 256; i++) _histo[i] = 0;
128 }
129 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
130 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
131 _calls++;
132 unsigned char* ujb = (unsigned char*)card_ptr;
133 int ind = (int)(*ujb);
134 _histo[ind]++;
135 *card_ptr = -1;
136 }
137 return true;
138 }
139 int calls() { return _calls; }
140 void print_histo() {
141 gclog_or_tty->print_cr("Card table value histogram:");
142 for (int i = 0; i < 256; i++) {
143 if (_histo[i] != 0) {
144 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]);
145 }
146 }
147 }
148 };
149
150 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure {
151 int _calls;
152 G1CollectedHeap* _g1h;
153 CardTableModRefBS* _ctbs;
154 public:
155 RedirtyLoggedCardTableEntryClosure() :
156 _calls(0)
157 {
158 _g1h = G1CollectedHeap::heap();
159 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
160 }
161 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
162 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
163 _calls++;
164 *card_ptr = 0;
165 }
166 return true;
167 }
168 int calls() { return _calls; }
169 };
170
171 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure {
172 public:
173 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
174 *card_ptr = CardTableModRefBS::dirty_card_val();
175 return true;
176 }
177 };
178
179 YoungList::YoungList(G1CollectedHeap* g1h) :
180 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0),
181 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) {
182 guarantee(check_list_empty(false), "just making sure...");
183 }
184
185 void YoungList::push_region(HeapRegion *hr) {
186 assert(!hr->is_young(), "should not already be young");
187 assert(hr->get_next_young_region() == NULL, "cause it should!");
188
189 hr->set_next_young_region(_head);
190 _head = hr;
191
192 _g1h->g1_policy()->set_region_eden(hr, (int) _length);
193 ++_length;
194 }
195
196 void YoungList::add_survivor_region(HeapRegion* hr) {
197 assert(hr->is_survivor(), "should be flagged as survivor region");
198 assert(hr->get_next_young_region() == NULL, "cause it should!");
199
200 hr->set_next_young_region(_survivor_head);
201 if (_survivor_head == NULL) {
202 _survivor_tail = hr;
203 }
204 _survivor_head = hr;
205 ++_survivor_length;
206 }
207
208 void YoungList::empty_list(HeapRegion* list) {
209 while (list != NULL) {
210 HeapRegion* next = list->get_next_young_region();
211 list->set_next_young_region(NULL);
212 list->uninstall_surv_rate_group();
213 list->set_not_young();
214 list = next;
215 }
216 }
217
218 void YoungList::empty_list() {
219 assert(check_list_well_formed(), "young list should be well formed");
220
221 empty_list(_head);
222 _head = NULL;
223 _length = 0;
224
225 empty_list(_survivor_head);
226 _survivor_head = NULL;
227 _survivor_tail = NULL;
228 _survivor_length = 0;
229
230 _last_sampled_rs_lengths = 0;
231
232 assert(check_list_empty(false), "just making sure...");
233 }
234
235 bool YoungList::check_list_well_formed() {
236 bool ret = true;
237
238 uint length = 0;
239 HeapRegion* curr = _head;
240 HeapRegion* last = NULL;
241 while (curr != NULL) {
242 if (!curr->is_young()) {
243 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" "
244 "incorrectly tagged (y: %d, surv: %d)",
245 curr->bottom(), curr->end(),
246 curr->is_young(), curr->is_survivor());
247 ret = false;
248 }
249 ++length;
250 last = curr;
251 curr = curr->get_next_young_region();
252 }
253 ret = ret && (length == _length);
254
255 if (!ret) {
256 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!");
257 gclog_or_tty->print_cr("### list has %u entries, _length is %u",
258 length, _length);
259 }
260
261 return ret;
262 }
263
264 bool YoungList::check_list_empty(bool check_sample) {
265 bool ret = true;
266
267 if (_length != 0) {
268 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u",
269 _length);
270 ret = false;
271 }
272 if (check_sample && _last_sampled_rs_lengths != 0) {
273 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths");
274 ret = false;
275 }
276 if (_head != NULL) {
277 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head");
278 ret = false;
279 }
280 if (!ret) {
281 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty");
282 }
283
284 return ret;
285 }
286
287 void
288 YoungList::rs_length_sampling_init() {
289 _sampled_rs_lengths = 0;
290 _curr = _head;
291 }
292
293 bool
294 YoungList::rs_length_sampling_more() {
295 return _curr != NULL;
296 }
297
298 void
299 YoungList::rs_length_sampling_next() {
300 assert( _curr != NULL, "invariant" );
301 size_t rs_length = _curr->rem_set()->occupied();
302
303 _sampled_rs_lengths += rs_length;
304
305 // The current region may not yet have been added to the
306 // incremental collection set (it gets added when it is
307 // retired as the current allocation region).
308 if (_curr->in_collection_set()) {
309 // Update the collection set policy information for this region
310 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length);
311 }
312
313 _curr = _curr->get_next_young_region();
314 if (_curr == NULL) {
315 _last_sampled_rs_lengths = _sampled_rs_lengths;
316 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths);
317 }
318 }
319
320 void
321 YoungList::reset_auxilary_lists() {
322 guarantee( is_empty(), "young list should be empty" );
323 assert(check_list_well_formed(), "young list should be well formed");
324
325 // Add survivor regions to SurvRateGroup.
326 _g1h->g1_policy()->note_start_adding_survivor_regions();
327 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */);
328
329 int young_index_in_cset = 0;
330 for (HeapRegion* curr = _survivor_head;
331 curr != NULL;
332 curr = curr->get_next_young_region()) {
333 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset);
334
335 // The region is a non-empty survivor so let's add it to
336 // the incremental collection set for the next evacuation
337 // pause.
338 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr);
339 young_index_in_cset += 1;
340 }
341 assert((uint) young_index_in_cset == _survivor_length, "post-condition");
342 _g1h->g1_policy()->note_stop_adding_survivor_regions();
343
344 _head = _survivor_head;
345 _length = _survivor_length;
346 if (_survivor_head != NULL) {
347 assert(_survivor_tail != NULL, "cause it shouldn't be");
348 assert(_survivor_length > 0, "invariant");
349 _survivor_tail->set_next_young_region(NULL);
350 }
351
352 // Don't clear the survivor list handles until the start of
353 // the next evacuation pause - we need it in order to re-tag
354 // the survivor regions from this evacuation pause as 'young'
355 // at the start of the next.
356
357 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */);
358
359 assert(check_list_well_formed(), "young list should be well formed");
360 }
361
362 void YoungList::print() {
363 HeapRegion* lists[] = {_head, _survivor_head};
364 const char* names[] = {"YOUNG", "SURVIVOR"};
365
366 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) {
367 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]);
368 HeapRegion *curr = lists[list];
369 if (curr == NULL)
370 gclog_or_tty->print_cr(" empty");
371 while (curr != NULL) {
372 gclog_or_tty->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d",
373 HR_FORMAT_PARAMS(curr),
374 curr->prev_top_at_mark_start(),
375 curr->next_top_at_mark_start(),
376 curr->age_in_surv_rate_group_cond());
377 curr = curr->get_next_young_region();
378 }
379 }
380
381 gclog_or_tty->print_cr("");
382 }
383
384 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr)
385 {
386 // Claim the right to put the region on the dirty cards region list
387 // by installing a self pointer.
388 HeapRegion* next = hr->get_next_dirty_cards_region();
389 if (next == NULL) {
390 HeapRegion* res = (HeapRegion*)
391 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(),
392 NULL);
393 if (res == NULL) {
394 HeapRegion* head;
395 do {
396 // Put the region to the dirty cards region list.
397 head = _dirty_cards_region_list;
398 next = (HeapRegion*)
399 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head);
400 if (next == head) {
401 assert(hr->get_next_dirty_cards_region() == hr,
402 "hr->get_next_dirty_cards_region() != hr");
403 if (next == NULL) {
404 // The last region in the list points to itself.
405 hr->set_next_dirty_cards_region(hr);
406 } else {
407 hr->set_next_dirty_cards_region(next);
408 }
409 }
410 } while (next != head);
411 }
412 }
413 }
414
415 HeapRegion* G1CollectedHeap::pop_dirty_cards_region()
416 {
417 HeapRegion* head;
418 HeapRegion* hr;
419 do {
420 head = _dirty_cards_region_list;
421 if (head == NULL) {
422 return NULL;
423 }
424 HeapRegion* new_head = head->get_next_dirty_cards_region();
425 if (head == new_head) {
426 // The last region.
427 new_head = NULL;
428 }
429 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list,
430 head);
431 } while (hr != head);
432 assert(hr != NULL, "invariant");
433 hr->set_next_dirty_cards_region(NULL);
434 return hr;
435 }
436
437 void G1CollectedHeap::stop_conc_gc_threads() {
438 _cg1r->stop();
439 _cmThread->stop();
440 }
441
442 #ifdef ASSERT
443 // A region is added to the collection set as it is retired
444 // so an address p can point to a region which will be in the
445 // collection set but has not yet been retired. This method
446 // therefore is only accurate during a GC pause after all
447 // regions have been retired. It is used for debugging
448 // to check if an nmethod has references to objects that can
449 // be move during a partial collection. Though it can be
450 // inaccurate, it is sufficient for G1 because the conservative
451 // implementation of is_scavengable() for G1 will indicate that
452 // all nmethods must be scanned during a partial collection.
453 bool G1CollectedHeap::is_in_partial_collection(const void* p) {
454 HeapRegion* hr = heap_region_containing(p);
455 return hr != NULL && hr->in_collection_set();
456 }
457 #endif
458
459 // Returns true if the reference points to an object that
460 // can move in an incremental collecction.
461 bool G1CollectedHeap::is_scavengable(const void* p) {
462 G1CollectedHeap* g1h = G1CollectedHeap::heap();
463 G1CollectorPolicy* g1p = g1h->g1_policy();
464 HeapRegion* hr = heap_region_containing(p);
465 if (hr == NULL) {
466 // null
467 assert(p == NULL, err_msg("Not NULL " PTR_FORMAT ,p));
468 return false;
469 } else {
470 return !hr->isHumongous();
471 }
472 }
473
474 void G1CollectedHeap::check_ct_logs_at_safepoint() {
475 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
476 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
477
478 // Count the dirty cards at the start.
479 CountNonCleanMemRegionClosure count1(this);
480 ct_bs->mod_card_iterate(&count1);
481 int orig_count = count1.n();
482
483 // First clear the logged cards.
484 ClearLoggedCardTableEntryClosure clear;
485 dcqs.set_closure(&clear);
486 dcqs.apply_closure_to_all_completed_buffers();
487 dcqs.iterate_closure_all_threads(false);
488 clear.print_histo();
489
490 // Now ensure that there's no dirty cards.
491 CountNonCleanMemRegionClosure count2(this);
492 ct_bs->mod_card_iterate(&count2);
493 if (count2.n() != 0) {
494 gclog_or_tty->print_cr("Card table has %d entries; %d originally",
495 count2.n(), orig_count);
496 }
497 guarantee(count2.n() == 0, "Card table should be clean.");
498
499 RedirtyLoggedCardTableEntryClosure redirty;
500 JavaThread::dirty_card_queue_set().set_closure(&redirty);
501 dcqs.apply_closure_to_all_completed_buffers();
502 dcqs.iterate_closure_all_threads(false);
503 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.",
504 clear.calls(), orig_count);
505 guarantee(redirty.calls() == clear.calls(),
506 "Or else mechanism is broken.");
507
508 CountNonCleanMemRegionClosure count3(this);
509 ct_bs->mod_card_iterate(&count3);
510 if (count3.n() != orig_count) {
511 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.",
512 orig_count, count3.n());
513 guarantee(count3.n() >= orig_count, "Should have restored them all.");
514 }
515
516 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
517 }
518
519 // Private class members.
520
521 G1CollectedHeap* G1CollectedHeap::_g1h;
522
523 // Private methods.
524
525 HeapRegion*
526 G1CollectedHeap::new_region_try_secondary_free_list() {
527 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
528 while (!_secondary_free_list.is_empty() || free_regions_coming()) {
529 if (!_secondary_free_list.is_empty()) {
530 if (G1ConcRegionFreeingVerbose) {
531 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
532 "secondary_free_list has %u entries",
533 _secondary_free_list.length());
534 }
535 // It looks as if there are free regions available on the
536 // secondary_free_list. Let's move them to the free_list and try
537 // again to allocate from it.
538 append_secondary_free_list();
539
540 assert(!_free_list.is_empty(), "if the secondary_free_list was not "
541 "empty we should have moved at least one entry to the free_list");
542 HeapRegion* res = _free_list.remove_head();
543 if (G1ConcRegionFreeingVerbose) {
544 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
545 "allocated "HR_FORMAT" from secondary_free_list",
546 HR_FORMAT_PARAMS(res));
547 }
548 return res;
549 }
550
551 // Wait here until we get notifed either when (a) there are no
552 // more free regions coming or (b) some regions have been moved on
553 // the secondary_free_list.
554 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
555 }
556
557 if (G1ConcRegionFreeingVerbose) {
558 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
559 "could not allocate from secondary_free_list");
560 }
561 return NULL;
562 }
563
564 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) {
565 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords,
566 "the only time we use this to allocate a humongous region is "
567 "when we are allocating a single humongous region");
568
569 HeapRegion* res;
570 if (G1StressConcRegionFreeing) {
571 if (!_secondary_free_list.is_empty()) {
572 if (G1ConcRegionFreeingVerbose) {
573 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
574 "forced to look at the secondary_free_list");
575 }
576 res = new_region_try_secondary_free_list();
577 if (res != NULL) {
578 return res;
579 }
580 }
581 }
582 res = _free_list.remove_head_or_null();
583 if (res == NULL) {
584 if (G1ConcRegionFreeingVerbose) {
585 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : "
586 "res == NULL, trying the secondary_free_list");
587 }
588 res = new_region_try_secondary_free_list();
589 }
590 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) {
591 // Currently, only attempts to allocate GC alloc regions set
592 // do_expand to true. So, we should only reach here during a
593 // safepoint. If this assumption changes we might have to
594 // reconsider the use of _expand_heap_after_alloc_failure.
595 assert(SafepointSynchronize::is_at_safepoint(), "invariant");
596
597 ergo_verbose1(ErgoHeapSizing,
598 "attempt heap expansion",
599 ergo_format_reason("region allocation request failed")
600 ergo_format_byte("allocation request"),
601 word_size * HeapWordSize);
602 if (expand(word_size * HeapWordSize)) {
603 // Given that expand() succeeded in expanding the heap, and we
604 // always expand the heap by an amount aligned to the heap
605 // region size, the free list should in theory not be empty. So
606 // it would probably be OK to use remove_head(). But the extra
607 // check for NULL is unlikely to be a performance issue here (we
608 // just expanded the heap!) so let's just be conservative and
609 // use remove_head_or_null().
610 res = _free_list.remove_head_or_null();
611 } else {
612 _expand_heap_after_alloc_failure = false;
613 }
614 }
615 return res;
616 }
617
618 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions,
619 size_t word_size) {
620 assert(isHumongous(word_size), "word_size should be humongous");
621 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
622
623 uint first = G1_NULL_HRS_INDEX;
624 if (num_regions == 1) {
625 // Only one region to allocate, no need to go through the slower
626 // path. The caller will attempt the expasion if this fails, so
627 // let's not try to expand here too.
628 HeapRegion* hr = new_region(word_size, false /* do_expand */);
629 if (hr != NULL) {
630 first = hr->hrs_index();
631 } else {
632 first = G1_NULL_HRS_INDEX;
633 }
634 } else {
635 // We can't allocate humongous regions while cleanupComplete() is
636 // running, since some of the regions we find to be empty might not
637 // yet be added to the free list and it is not straightforward to
638 // know which list they are on so that we can remove them. Note
639 // that we only need to do this if we need to allocate more than
640 // one region to satisfy the current humongous allocation
641 // request. If we are only allocating one region we use the common
642 // region allocation code (see above).
643 wait_while_free_regions_coming();
644 append_secondary_free_list_if_not_empty_with_lock();
645
646 if (free_regions() >= num_regions) {
647 first = _hrs.find_contiguous(num_regions);
648 if (first != G1_NULL_HRS_INDEX) {
649 for (uint i = first; i < first + num_regions; ++i) {
650 HeapRegion* hr = region_at(i);
651 assert(hr->is_empty(), "sanity");
652 assert(is_on_master_free_list(hr), "sanity");
653 hr->set_pending_removal(true);
654 }
655 _free_list.remove_all_pending(num_regions);
656 }
657 }
658 }
659 return first;
660 }
661
662 HeapWord*
663 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first,
664 uint num_regions,
665 size_t word_size) {
666 assert(first != G1_NULL_HRS_INDEX, "pre-condition");
667 assert(isHumongous(word_size), "word_size should be humongous");
668 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition");
669
670 // Index of last region in the series + 1.
671 uint last = first + num_regions;
672
673 // We need to initialize the region(s) we just discovered. This is
674 // a bit tricky given that it can happen concurrently with
675 // refinement threads refining cards on these regions and
676 // potentially wanting to refine the BOT as they are scanning
677 // those cards (this can happen shortly after a cleanup; see CR
678 // 6991377). So we have to set up the region(s) carefully and in
679 // a specific order.
680
681 // The word size sum of all the regions we will allocate.
682 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords;
683 assert(word_size <= word_size_sum, "sanity");
684
685 // This will be the "starts humongous" region.
686 HeapRegion* first_hr = region_at(first);
687 // The header of the new object will be placed at the bottom of
688 // the first region.
689 HeapWord* new_obj = first_hr->bottom();
690 // This will be the new end of the first region in the series that
691 // should also match the end of the last region in the seriers.
692 HeapWord* new_end = new_obj + word_size_sum;
693 // This will be the new top of the first region that will reflect
694 // this allocation.
695 HeapWord* new_top = new_obj + word_size;
696
697 // First, we need to zero the header of the space that we will be
698 // allocating. When we update top further down, some refinement
699 // threads might try to scan the region. By zeroing the header we
700 // ensure that any thread that will try to scan the region will
701 // come across the zero klass word and bail out.
702 //
703 // NOTE: It would not have been correct to have used
704 // CollectedHeap::fill_with_object() and make the space look like
705 // an int array. The thread that is doing the allocation will
706 // later update the object header to a potentially different array
707 // type and, for a very short period of time, the klass and length
708 // fields will be inconsistent. This could cause a refinement
709 // thread to calculate the object size incorrectly.
710 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0);
711
712 // We will set up the first region as "starts humongous". This
713 // will also update the BOT covering all the regions to reflect
714 // that there is a single object that starts at the bottom of the
715 // first region.
716 first_hr->set_startsHumongous(new_top, new_end);
717
718 // Then, if there are any, we will set up the "continues
719 // humongous" regions.
720 HeapRegion* hr = NULL;
721 for (uint i = first + 1; i < last; ++i) {
722 hr = region_at(i);
723 hr->set_continuesHumongous(first_hr);
724 }
725 // If we have "continues humongous" regions (hr != NULL), then the
726 // end of the last one should match new_end.
727 assert(hr == NULL || hr->end() == new_end, "sanity");
728
729 // Up to this point no concurrent thread would have been able to
730 // do any scanning on any region in this series. All the top
731 // fields still point to bottom, so the intersection between
732 // [bottom,top] and [card_start,card_end] will be empty. Before we
733 // update the top fields, we'll do a storestore to make sure that
734 // no thread sees the update to top before the zeroing of the
735 // object header and the BOT initialization.
736 OrderAccess::storestore();
737
738 // Now that the BOT and the object header have been initialized,
739 // we can update top of the "starts humongous" region.
740 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(),
741 "new_top should be in this region");
742 first_hr->set_top(new_top);
743 if (_hr_printer.is_active()) {
744 HeapWord* bottom = first_hr->bottom();
745 HeapWord* end = first_hr->orig_end();
746 if ((first + 1) == last) {
747 // the series has a single humongous region
748 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top);
749 } else {
750 // the series has more than one humongous regions
751 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end);
752 }
753 }
754
755 // Now, we will update the top fields of the "continues humongous"
756 // regions. The reason we need to do this is that, otherwise,
757 // these regions would look empty and this will confuse parts of
758 // G1. For example, the code that looks for a consecutive number
759 // of empty regions will consider them empty and try to
760 // re-allocate them. We can extend is_empty() to also include
761 // !continuesHumongous(), but it is easier to just update the top
762 // fields here. The way we set top for all regions (i.e., top ==
763 // end for all regions but the last one, top == new_top for the
764 // last one) is actually used when we will free up the humongous
765 // region in free_humongous_region().
766 hr = NULL;
767 for (uint i = first + 1; i < last; ++i) {
768 hr = region_at(i);
769 if ((i + 1) == last) {
770 // last continues humongous region
771 assert(hr->bottom() < new_top && new_top <= hr->end(),
772 "new_top should fall on this region");
773 hr->set_top(new_top);
774 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top);
775 } else {
776 // not last one
777 assert(new_top > hr->end(), "new_top should be above this region");
778 hr->set_top(hr->end());
779 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end());
780 }
781 }
782 // If we have continues humongous regions (hr != NULL), then the
783 // end of the last one should match new_end and its top should
784 // match new_top.
785 assert(hr == NULL ||
786 (hr->end() == new_end && hr->top() == new_top), "sanity");
787
788 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
789 _summary_bytes_used += first_hr->used();
790 _humongous_set.add(first_hr);
791
792 return new_obj;
793 }
794
795 // If could fit into free regions w/o expansion, try.
796 // Otherwise, if can expand, do so.
797 // Otherwise, if using ex regions might help, try with ex given back.
798 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) {
799 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
800
801 verify_region_sets_optional();
802
803 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords);
804 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords);
805 uint x_num = expansion_regions();
806 uint fs = _hrs.free_suffix();
807 uint first = humongous_obj_allocate_find_first(num_regions, word_size);
808 if (first == G1_NULL_HRS_INDEX) {
809 // The only thing we can do now is attempt expansion.
810 if (fs + x_num >= num_regions) {
811 // If the number of regions we're trying to allocate for this
812 // object is at most the number of regions in the free suffix,
813 // then the call to humongous_obj_allocate_find_first() above
814 // should have succeeded and we wouldn't be here.
815 //
816 // We should only be trying to expand when the free suffix is
817 // not sufficient for the object _and_ we have some expansion
818 // room available.
819 assert(num_regions > fs, "earlier allocation should have succeeded");
820
821 ergo_verbose1(ErgoHeapSizing,
822 "attempt heap expansion",
823 ergo_format_reason("humongous allocation request failed")
824 ergo_format_byte("allocation request"),
825 word_size * HeapWordSize);
826 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) {
827 // Even though the heap was expanded, it might not have
828 // reached the desired size. So, we cannot assume that the
829 // allocation will succeed.
830 first = humongous_obj_allocate_find_first(num_regions, word_size);
831 }
832 }
833 }
834
835 HeapWord* result = NULL;
836 if (first != G1_NULL_HRS_INDEX) {
837 result =
838 humongous_obj_allocate_initialize_regions(first, num_regions, word_size);
839 assert(result != NULL, "it should always return a valid result");
840
841 // A successful humongous object allocation changes the used space
842 // information of the old generation so we need to recalculate the
843 // sizes and update the jstat counters here.
844 g1mm()->update_sizes();
845 }
846
847 verify_region_sets_optional();
848
849 return result;
850 }
851
852 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) {
853 assert_heap_not_locked_and_not_at_safepoint();
854 assert(!isHumongous(word_size), "we do not allow humongous TLABs");
855
856 unsigned int dummy_gc_count_before;
857 int dummy_gclocker_retry_count = 0;
858 return attempt_allocation(word_size, &dummy_gc_count_before, &dummy_gclocker_retry_count);
859 }
860
861 HeapWord*
862 G1CollectedHeap::mem_allocate(size_t word_size,
863 bool* gc_overhead_limit_was_exceeded) {
864 assert_heap_not_locked_and_not_at_safepoint();
865
866 // Loop until the allocation is satisified, or unsatisfied after GC.
867 for (int try_count = 1, gclocker_retry_count = 0; /* we'll return */; try_count += 1) {
868 unsigned int gc_count_before;
869
870 HeapWord* result = NULL;
871 if (!isHumongous(word_size)) {
872 result = attempt_allocation(word_size, &gc_count_before, &gclocker_retry_count);
873 } else {
874 result = attempt_allocation_humongous(word_size, &gc_count_before, &gclocker_retry_count);
875 }
876 if (result != NULL) {
877 return result;
878 }
879
880 // Create the garbage collection operation...
881 VM_G1CollectForAllocation op(gc_count_before, word_size);
882 // ...and get the VM thread to execute it.
883 VMThread::execute(&op);
884
885 if (op.prologue_succeeded() && op.pause_succeeded()) {
886 // If the operation was successful we'll return the result even
887 // if it is NULL. If the allocation attempt failed immediately
888 // after a Full GC, it's unlikely we'll be able to allocate now.
889 HeapWord* result = op.result();
890 if (result != NULL && !isHumongous(word_size)) {
891 // Allocations that take place on VM operations do not do any
892 // card dirtying and we have to do it here. We only have to do
893 // this for non-humongous allocations, though.
894 dirty_young_block(result, word_size);
895 }
896 return result;
897 } else {
898 if (gclocker_retry_count > GCLockerRetryAllocationCount) {
899 return NULL;
900 }
901 assert(op.result() == NULL,
902 "the result should be NULL if the VM op did not succeed");
903 }
904
905 // Give a warning if we seem to be looping forever.
906 if ((QueuedAllocationWarningCount > 0) &&
907 (try_count % QueuedAllocationWarningCount == 0)) {
908 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
909 }
910 }
911
912 ShouldNotReachHere();
913 return NULL;
914 }
915
916 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
917 unsigned int *gc_count_before_ret,
918 int* gclocker_retry_count_ret) {
919 // Make sure you read the note in attempt_allocation_humongous().
920
921 assert_heap_not_locked_and_not_at_safepoint();
922 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
923 "be called for humongous allocation requests");
924
925 // We should only get here after the first-level allocation attempt
926 // (attempt_allocation()) failed to allocate.
927
928 // We will loop until a) we manage to successfully perform the
929 // allocation or b) we successfully schedule a collection which
930 // fails to perform the allocation. b) is the only case when we'll
931 // return NULL.
932 HeapWord* result = NULL;
933 for (int try_count = 1; /* we'll return */; try_count += 1) {
934 bool should_try_gc;
935 unsigned int gc_count_before;
936
937 {
938 MutexLockerEx x(Heap_lock);
939
940 result = _mutator_alloc_region.attempt_allocation_locked(word_size,
941 false /* bot_updates */);
942 if (result != NULL) {
943 return result;
944 }
945
946 // If we reach here, attempt_allocation_locked() above failed to
947 // allocate a new region. So the mutator alloc region should be NULL.
948 assert(_mutator_alloc_region.get() == NULL, "only way to get here");
949
950 if (GC_locker::is_active_and_needs_gc()) {
951 if (g1_policy()->can_expand_young_list()) {
952 // No need for an ergo verbose message here,
953 // can_expand_young_list() does this when it returns true.
954 result = _mutator_alloc_region.attempt_allocation_force(word_size,
955 false /* bot_updates */);
956 if (result != NULL) {
957 return result;
958 }
959 }
960 should_try_gc = false;
961 } else {
962 // The GCLocker may not be active but the GCLocker initiated
963 // GC may not yet have been performed (GCLocker::needs_gc()
964 // returns true). In this case we do not try this GC and
965 // wait until the GCLocker initiated GC is performed, and
966 // then retry the allocation.
967 if (GC_locker::needs_gc()) {
968 should_try_gc = false;
969 } else {
970 // Read the GC count while still holding the Heap_lock.
971 gc_count_before = total_collections();
972 should_try_gc = true;
973 }
974 }
975 }
976
977 if (should_try_gc) {
978 bool succeeded;
979 result = do_collection_pause(word_size, gc_count_before, &succeeded);
980 if (result != NULL) {
981 assert(succeeded, "only way to get back a non-NULL result");
982 return result;
983 }
984
985 if (succeeded) {
986 // If we get here we successfully scheduled a collection which
987 // failed to allocate. No point in trying to allocate
988 // further. We'll just return NULL.
989 MutexLockerEx x(Heap_lock);
990 *gc_count_before_ret = total_collections();
991 return NULL;
992 }
993 } else {
994 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) {
995 MutexLockerEx x(Heap_lock);
996 *gc_count_before_ret = total_collections();
997 return NULL;
998 }
999 // The GCLocker is either active or the GCLocker initiated
1000 // GC has not yet been performed. Stall until it is and
1001 // then retry the allocation.
1002 GC_locker::stall_until_clear();
1003 (*gclocker_retry_count_ret) += 1;
1004 }
1005
1006 // We can reach here if we were unsuccessul in scheduling a
1007 // collection (because another thread beat us to it) or if we were
1008 // stalled due to the GC locker. In either can we should retry the
1009 // allocation attempt in case another thread successfully
1010 // performed a collection and reclaimed enough space. We do the
1011 // first attempt (without holding the Heap_lock) here and the
1012 // follow-on attempt will be at the start of the next loop
1013 // iteration (after taking the Heap_lock).
1014 result = _mutator_alloc_region.attempt_allocation(word_size,
1015 false /* bot_updates */);
1016 if (result != NULL) {
1017 return result;
1018 }
1019
1020 // Give a warning if we seem to be looping forever.
1021 if ((QueuedAllocationWarningCount > 0) &&
1022 (try_count % QueuedAllocationWarningCount == 0)) {
1023 warning("G1CollectedHeap::attempt_allocation_slow() "
1024 "retries %d times", try_count);
1025 }
1026 }
1027
1028 ShouldNotReachHere();
1029 return NULL;
1030 }
1031
1032 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
1033 unsigned int * gc_count_before_ret,
1034 int* gclocker_retry_count_ret) {
1035 // The structure of this method has a lot of similarities to
1036 // attempt_allocation_slow(). The reason these two were not merged
1037 // into a single one is that such a method would require several "if
1038 // allocation is not humongous do this, otherwise do that"
1039 // conditional paths which would obscure its flow. In fact, an early
1040 // version of this code did use a unified method which was harder to
1041 // follow and, as a result, it had subtle bugs that were hard to
1042 // track down. So keeping these two methods separate allows each to
1043 // be more readable. It will be good to keep these two in sync as
1044 // much as possible.
1045
1046 assert_heap_not_locked_and_not_at_safepoint();
1047 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1048 "should only be called for humongous allocations");
1049
1050 // Humongous objects can exhaust the heap quickly, so we should check if we
1051 // need to start a marking cycle at each humongous object allocation. We do
1052 // the check before we do the actual allocation. The reason for doing it
1053 // before the allocation is that we avoid having to keep track of the newly
1054 // allocated memory while we do a GC.
1055 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation",
1056 word_size)) {
1057 collect(GCCause::_g1_humongous_allocation);
1058 }
1059
1060 // We will loop until a) we manage to successfully perform the
1061 // allocation or b) we successfully schedule a collection which
1062 // fails to perform the allocation. b) is the only case when we'll
1063 // return NULL.
1064 HeapWord* result = NULL;
1065 for (int try_count = 1; /* we'll return */; try_count += 1) {
1066 bool should_try_gc;
1067 unsigned int gc_count_before;
1068
1069 {
1070 MutexLockerEx x(Heap_lock);
1071
1072 // Given that humongous objects are not allocated in young
1073 // regions, we'll first try to do the allocation without doing a
1074 // collection hoping that there's enough space in the heap.
1075 result = humongous_obj_allocate(word_size);
1076 if (result != NULL) {
1077 return result;
1078 }
1079
1080 if (GC_locker::is_active_and_needs_gc()) {
1081 should_try_gc = false;
1082 } else {
1083 // The GCLocker may not be active but the GCLocker initiated
1084 // GC may not yet have been performed (GCLocker::needs_gc()
1085 // returns true). In this case we do not try this GC and
1086 // wait until the GCLocker initiated GC is performed, and
1087 // then retry the allocation.
1088 if (GC_locker::needs_gc()) {
1089 should_try_gc = false;
1090 } else {
1091 // Read the GC count while still holding the Heap_lock.
1092 gc_count_before = total_collections();
1093 should_try_gc = true;
1094 }
1095 }
1096 }
1097
1098 if (should_try_gc) {
1099 // If we failed to allocate the humongous object, we should try to
1100 // do a collection pause (if we're allowed) in case it reclaims
1101 // enough space for the allocation to succeed after the pause.
1102
1103 bool succeeded;
1104 result = do_collection_pause(word_size, gc_count_before, &succeeded);
1105 if (result != NULL) {
1106 assert(succeeded, "only way to get back a non-NULL result");
1107 return result;
1108 }
1109
1110 if (succeeded) {
1111 // If we get here we successfully scheduled a collection which
1112 // failed to allocate. No point in trying to allocate
1113 // further. We'll just return NULL.
1114 MutexLockerEx x(Heap_lock);
1115 *gc_count_before_ret = total_collections();
1116 return NULL;
1117 }
1118 } else {
1119 if (*gclocker_retry_count_ret > GCLockerRetryAllocationCount) {
1120 MutexLockerEx x(Heap_lock);
1121 *gc_count_before_ret = total_collections();
1122 return NULL;
1123 }
1124 // The GCLocker is either active or the GCLocker initiated
1125 // GC has not yet been performed. Stall until it is and
1126 // then retry the allocation.
1127 GC_locker::stall_until_clear();
1128 (*gclocker_retry_count_ret) += 1;
1129 }
1130
1131 // We can reach here if we were unsuccessul in scheduling a
1132 // collection (because another thread beat us to it) or if we were
1133 // stalled due to the GC locker. In either can we should retry the
1134 // allocation attempt in case another thread successfully
1135 // performed a collection and reclaimed enough space. Give a
1136 // warning if we seem to be looping forever.
1137
1138 if ((QueuedAllocationWarningCount > 0) &&
1139 (try_count % QueuedAllocationWarningCount == 0)) {
1140 warning("G1CollectedHeap::attempt_allocation_humongous() "
1141 "retries %d times", try_count);
1142 }
1143 }
1144
1145 ShouldNotReachHere();
1146 return NULL;
1147 }
1148
1149 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1150 bool expect_null_mutator_alloc_region) {
1151 assert_at_safepoint(true /* should_be_vm_thread */);
1152 assert(_mutator_alloc_region.get() == NULL ||
1153 !expect_null_mutator_alloc_region,
1154 "the current alloc region was unexpectedly found to be non-NULL");
1155
1156 if (!isHumongous(word_size)) {
1157 return _mutator_alloc_region.attempt_allocation_locked(word_size,
1158 false /* bot_updates */);
1159 } else {
1160 HeapWord* result = humongous_obj_allocate(word_size);
1161 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) {
1162 g1_policy()->set_initiate_conc_mark_if_possible();
1163 }
1164 return result;
1165 }
1166
1167 ShouldNotReachHere();
1168 }
1169
1170 class PostMCRemSetClearClosure: public HeapRegionClosure {
1171 G1CollectedHeap* _g1h;
1172 ModRefBarrierSet* _mr_bs;
1173 public:
1174 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) :
1175 _g1h(g1h), _mr_bs(mr_bs) {}
1176
1177 bool doHeapRegion(HeapRegion* r) {
1178 if (r->continuesHumongous()) {
1179 return false;
1180 }
1181 _g1h->reset_gc_time_stamps(r);
1182 HeapRegionRemSet* hrrs = r->rem_set();
1183 if (hrrs != NULL) hrrs->clear();
1184 // You might think here that we could clear just the cards
1185 // corresponding to the used region. But no: if we leave a dirty card
1186 // in a region we might allocate into, then it would prevent that card
1187 // from being enqueued, and cause it to be missed.
1188 // Re: the performance cost: we shouldn't be doing full GC anyway!
1189 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1190 return false;
1191 }
1192 };
1193
1194 void G1CollectedHeap::clear_rsets_post_compaction() {
1195 PostMCRemSetClearClosure rs_clear(this, mr_bs());
1196 heap_region_iterate(&rs_clear);
1197 }
1198
1199 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1200 G1CollectedHeap* _g1h;
1201 UpdateRSOopClosure _cl;
1202 int _worker_i;
1203 public:
1204 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1205 _cl(g1->g1_rem_set(), worker_i),
1206 _worker_i(worker_i),
1207 _g1h(g1)
1208 { }
1209
1210 bool doHeapRegion(HeapRegion* r) {
1211 if (!r->continuesHumongous()) {
1212 _cl.set_from(r);
1213 r->oop_iterate(&_cl);
1214 }
1215 return false;
1216 }
1217 };
1218
1219 class ParRebuildRSTask: public AbstractGangTask {
1220 G1CollectedHeap* _g1;
1221 public:
1222 ParRebuildRSTask(G1CollectedHeap* g1)
1223 : AbstractGangTask("ParRebuildRSTask"),
1224 _g1(g1)
1225 { }
1226
1227 void work(uint worker_id) {
1228 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id);
1229 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id,
1230 _g1->workers()->active_workers(),
1231 HeapRegion::RebuildRSClaimValue);
1232 }
1233 };
1234
1235 class PostCompactionPrinterClosure: public HeapRegionClosure {
1236 private:
1237 G1HRPrinter* _hr_printer;
1238 public:
1239 bool doHeapRegion(HeapRegion* hr) {
1240 assert(!hr->is_young(), "not expecting to find young regions");
1241 // We only generate output for non-empty regions.
1242 if (!hr->is_empty()) {
1243 if (!hr->isHumongous()) {
1244 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1245 } else if (hr->startsHumongous()) {
1246 if (hr->region_num() == 1) {
1247 // single humongous region
1248 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1249 } else {
1250 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1251 }
1252 } else {
1253 assert(hr->continuesHumongous(), "only way to get here");
1254 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1255 }
1256 }
1257 return false;
1258 }
1259
1260 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1261 : _hr_printer(hr_printer) { }
1262 };
1263
1264 void G1CollectedHeap::print_hrs_post_compaction() {
1265 PostCompactionPrinterClosure cl(hr_printer());
1266 heap_region_iterate(&cl);
1267 }
1268
1269 bool G1CollectedHeap::do_collection(bool explicit_gc,
1270 bool clear_all_soft_refs,
1271 size_t word_size) {
1272 assert_at_safepoint(true /* should_be_vm_thread */);
1273
1274 if (GC_locker::check_active_before_gc()) {
1275 return false;
1276 }
1277
1278 SvcGCMarker sgcm(SvcGCMarker::FULL);
1279 ResourceMark rm;
1280
1281 print_heap_before_gc();
1282
1283 size_t metadata_prev_used = MetaspaceAux::allocated_used_bytes();
1284
1285 HRSPhaseSetter x(HRSPhaseFullGC);
1286 verify_region_sets_optional();
1287
1288 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1289 collector_policy()->should_clear_all_soft_refs();
1290
1291 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1292
1293 {
1294 IsGCActiveMark x;
1295
1296 // Timing
1297 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant");
1298 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps);
1299 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
1300
1301 {
1302 TraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, gclog_or_tty);
1303 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1304 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1305
1306 double start = os::elapsedTime();
1307 g1_policy()->record_full_collection_start();
1308
1309 // Note: When we have a more flexible GC logging framework that
1310 // allows us to add optional attributes to a GC log record we
1311 // could consider timing and reporting how long we wait in the
1312 // following two methods.
1313 wait_while_free_regions_coming();
1314 // If we start the compaction before the CM threads finish
1315 // scanning the root regions we might trip them over as we'll
1316 // be moving objects / updating references. So let's wait until
1317 // they are done. By telling them to abort, they should complete
1318 // early.
1319 _cm->root_regions()->abort();
1320 _cm->root_regions()->wait_until_scan_finished();
1321 append_secondary_free_list_if_not_empty_with_lock();
1322
1323 gc_prologue(true);
1324 increment_total_collections(true /* full gc */);
1325 increment_old_marking_cycles_started();
1326
1327 assert(used() == recalculate_used(), "Should be equal");
1328
1329 verify_before_gc();
1330
1331 pre_full_gc_dump();
1332
1333 COMPILER2_PRESENT(DerivedPointerTable::clear());
1334
1335 // Disable discovery and empty the discovered lists
1336 // for the CM ref processor.
1337 ref_processor_cm()->disable_discovery();
1338 ref_processor_cm()->abandon_partial_discovery();
1339 ref_processor_cm()->verify_no_references_recorded();
1340
1341 // Abandon current iterations of concurrent marking and concurrent
1342 // refinement, if any are in progress. We have to do this before
1343 // wait_until_scan_finished() below.
1344 concurrent_mark()->abort();
1345
1346 // Make sure we'll choose a new allocation region afterwards.
1347 release_mutator_alloc_region();
1348 abandon_gc_alloc_regions();
1349 g1_rem_set()->cleanupHRRS();
1350
1351 // We should call this after we retire any currently active alloc
1352 // regions so that all the ALLOC / RETIRE events are generated
1353 // before the start GC event.
1354 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1355
1356 // We may have added regions to the current incremental collection
1357 // set between the last GC or pause and now. We need to clear the
1358 // incremental collection set and then start rebuilding it afresh
1359 // after this full GC.
1360 abandon_collection_set(g1_policy()->inc_cset_head());
1361 g1_policy()->clear_incremental_cset();
1362 g1_policy()->stop_incremental_cset_building();
1363
1364 tear_down_region_sets(false /* free_list_only */);
1365 g1_policy()->set_gcs_are_young(true);
1366
1367 // See the comments in g1CollectedHeap.hpp and
1368 // G1CollectedHeap::ref_processing_init() about
1369 // how reference processing currently works in G1.
1370
1371 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1372 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1373
1374 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1375 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1376
1377 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1378 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1379
1380 // Do collection work
1381 {
1382 HandleMark hm; // Discard invalid handles created during gc
1383 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1384 }
1385
1386 assert(free_regions() == 0, "we should not have added any free regions");
1387 rebuild_region_sets(false /* free_list_only */);
1388
1389 // Enqueue any discovered reference objects that have
1390 // not been removed from the discovered lists.
1391 ref_processor_stw()->enqueue_discovered_references();
1392
1393 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1394
1395 MemoryService::track_memory_usage();
1396
1397 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1398 ref_processor_stw()->verify_no_references_recorded();
1399
1400 // Delete metaspaces for unloaded class loaders and clean up loader_data graph
1401 ClassLoaderDataGraph::purge();
1402 MetaspaceAux::verify_metrics();
1403
1404 // Note: since we've just done a full GC, concurrent
1405 // marking is no longer active. Therefore we need not
1406 // re-enable reference discovery for the CM ref processor.
1407 // That will be done at the start of the next marking cycle.
1408 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1409 ref_processor_cm()->verify_no_references_recorded();
1410
1411 reset_gc_time_stamp();
1412 // Since everything potentially moved, we will clear all remembered
1413 // sets, and clear all cards. Later we will rebuild remebered
1414 // sets. We will also reset the GC time stamps of the regions.
1415 clear_rsets_post_compaction();
1416 check_gc_time_stamps();
1417
1418 // Resize the heap if necessary.
1419 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1420
1421 if (_hr_printer.is_active()) {
1422 // We should do this after we potentially resize the heap so
1423 // that all the COMMIT / UNCOMMIT events are generated before
1424 // the end GC event.
1425
1426 print_hrs_post_compaction();
1427 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1428 }
1429
1430 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache();
1431 if (hot_card_cache->use_cache()) {
1432 hot_card_cache->reset_card_counts();
1433 hot_card_cache->reset_hot_cache();
1434 }
1435
1436 // Rebuild remembered sets of all regions.
1437 if (G1CollectedHeap::use_parallel_gc_threads()) {
1438 uint n_workers =
1439 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
1440 workers()->active_workers(),
1441 Threads::number_of_non_daemon_threads());
1442 assert(UseDynamicNumberOfGCThreads ||
1443 n_workers == workers()->total_workers(),
1444 "If not dynamic should be using all the workers");
1445 workers()->set_active_workers(n_workers);
1446 // Set parallel threads in the heap (_n_par_threads) only
1447 // before a parallel phase and always reset it to 0 after
1448 // the phase so that the number of parallel threads does
1449 // no get carried forward to a serial phase where there
1450 // may be code that is "possibly_parallel".
1451 set_par_threads(n_workers);
1452
1453 ParRebuildRSTask rebuild_rs_task(this);
1454 assert(check_heap_region_claim_values(
1455 HeapRegion::InitialClaimValue), "sanity check");
1456 assert(UseDynamicNumberOfGCThreads ||
1457 workers()->active_workers() == workers()->total_workers(),
1458 "Unless dynamic should use total workers");
1459 // Use the most recent number of active workers
1460 assert(workers()->active_workers() > 0,
1461 "Active workers not properly set");
1462 set_par_threads(workers()->active_workers());
1463 workers()->run_task(&rebuild_rs_task);
1464 set_par_threads(0);
1465 assert(check_heap_region_claim_values(
1466 HeapRegion::RebuildRSClaimValue), "sanity check");
1467 reset_heap_region_claim_values();
1468 } else {
1469 RebuildRSOutOfRegionClosure rebuild_rs(this);
1470 heap_region_iterate(&rebuild_rs);
1471 }
1472
1473 if (true) { // FIXME
1474 MetaspaceGC::compute_new_size();
1475 }
1476
1477 #ifdef TRACESPINNING
1478 ParallelTaskTerminator::print_termination_counts();
1479 #endif
1480
1481 // Discard all rset updates
1482 JavaThread::dirty_card_queue_set().abandon_logs();
1483 assert(!G1DeferredRSUpdate
1484 || (G1DeferredRSUpdate &&
1485 (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
1486
1487 _young_list->reset_sampled_info();
1488 // At this point there should be no regions in the
1489 // entire heap tagged as young.
1490 assert(check_young_list_empty(true /* check_heap */),
1491 "young list should be empty at this point");
1492
1493 // Update the number of full collections that have been completed.
1494 increment_old_marking_cycles_completed(false /* concurrent */);
1495
1496 _hrs.verify_optional();
1497 verify_region_sets_optional();
1498
1499 verify_after_gc();
1500
1501 // Start a new incremental collection set for the next pause
1502 assert(g1_policy()->collection_set() == NULL, "must be");
1503 g1_policy()->start_incremental_cset_building();
1504
1505 // Clear the _cset_fast_test bitmap in anticipation of adding
1506 // regions to the incremental collection set for the next
1507 // evacuation pause.
1508 clear_cset_fast_test();
1509
1510 init_mutator_alloc_region();
1511
1512 double end = os::elapsedTime();
1513 g1_policy()->record_full_collection_end();
1514
1515 if (G1Log::fine()) {
1516 g1_policy()->print_heap_transition();
1517 }
1518
1519 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
1520 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
1521 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
1522 // before any GC notifications are raised.
1523 g1mm()->update_sizes();
1524
1525 gc_epilogue(true);
1526 }
1527
1528 if (G1Log::finer()) {
1529 g1_policy()->print_detailed_heap_transition(true /* full */);
1530 }
1531
1532 print_heap_after_gc();
1533
1534 post_full_gc_dump();
1535 }
1536
1537 return true;
1538 }
1539
1540 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1541 // do_collection() will return whether it succeeded in performing
1542 // the GC. Currently, there is no facility on the
1543 // do_full_collection() API to notify the caller than the collection
1544 // did not succeed (e.g., because it was locked out by the GC
1545 // locker). So, right now, we'll ignore the return value.
1546 bool dummy = do_collection(true, /* explicit_gc */
1547 clear_all_soft_refs,
1548 0 /* word_size */);
1549 }
1550
1551 // This code is mostly copied from TenuredGeneration.
1552 void
1553 G1CollectedHeap::
1554 resize_if_necessary_after_full_collection(size_t word_size) {
1555 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check");
1556
1557 // Include the current allocation, if any, and bytes that will be
1558 // pre-allocated to support collections, as "used".
1559 const size_t used_after_gc = used();
1560 const size_t capacity_after_gc = capacity();
1561 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1562
1563 // This is enforced in arguments.cpp.
1564 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1565 "otherwise the code below doesn't make sense");
1566
1567 // We don't have floating point command-line arguments
1568 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1569 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1570 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1571 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1572
1573 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1574 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1575
1576 // We have to be careful here as these two calculations can overflow
1577 // 32-bit size_t's.
1578 double used_after_gc_d = (double) used_after_gc;
1579 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1580 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1581
1582 // Let's make sure that they are both under the max heap size, which
1583 // by default will make them fit into a size_t.
1584 double desired_capacity_upper_bound = (double) max_heap_size;
1585 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1586 desired_capacity_upper_bound);
1587 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1588 desired_capacity_upper_bound);
1589
1590 // We can now safely turn them into size_t's.
1591 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1592 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1593
1594 // This assert only makes sense here, before we adjust them
1595 // with respect to the min and max heap size.
1596 assert(minimum_desired_capacity <= maximum_desired_capacity,
1597 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1598 "maximum_desired_capacity = "SIZE_FORMAT,
1599 minimum_desired_capacity, maximum_desired_capacity));
1600
1601 // Should not be greater than the heap max size. No need to adjust
1602 // it with respect to the heap min size as it's a lower bound (i.e.,
1603 // we'll try to make the capacity larger than it, not smaller).
1604 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1605 // Should not be less than the heap min size. No need to adjust it
1606 // with respect to the heap max size as it's an upper bound (i.e.,
1607 // we'll try to make the capacity smaller than it, not greater).
1608 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1609
1610 if (capacity_after_gc < minimum_desired_capacity) {
1611 // Don't expand unless it's significant
1612 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1613 ergo_verbose4(ErgoHeapSizing,
1614 "attempt heap expansion",
1615 ergo_format_reason("capacity lower than "
1616 "min desired capacity after Full GC")
1617 ergo_format_byte("capacity")
1618 ergo_format_byte("occupancy")
1619 ergo_format_byte_perc("min desired capacity"),
1620 capacity_after_gc, used_after_gc,
1621 minimum_desired_capacity, (double) MinHeapFreeRatio);
1622 expand(expand_bytes);
1623
1624 // No expansion, now see if we want to shrink
1625 } else if (capacity_after_gc > maximum_desired_capacity) {
1626 // Capacity too large, compute shrinking size
1627 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1628 ergo_verbose4(ErgoHeapSizing,
1629 "attempt heap shrinking",
1630 ergo_format_reason("capacity higher than "
1631 "max desired capacity after Full GC")
1632 ergo_format_byte("capacity")
1633 ergo_format_byte("occupancy")
1634 ergo_format_byte_perc("max desired capacity"),
1635 capacity_after_gc, used_after_gc,
1636 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1637 shrink(shrink_bytes);
1638 }
1639 }
1640
1641
1642 HeapWord*
1643 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1644 bool* succeeded) {
1645 assert_at_safepoint(true /* should_be_vm_thread */);
1646
1647 *succeeded = true;
1648 // Let's attempt the allocation first.
1649 HeapWord* result =
1650 attempt_allocation_at_safepoint(word_size,
1651 false /* expect_null_mutator_alloc_region */);
1652 if (result != NULL) {
1653 assert(*succeeded, "sanity");
1654 return result;
1655 }
1656
1657 // In a G1 heap, we're supposed to keep allocation from failing by
1658 // incremental pauses. Therefore, at least for now, we'll favor
1659 // expansion over collection. (This might change in the future if we can
1660 // do something smarter than full collection to satisfy a failed alloc.)
1661 result = expand_and_allocate(word_size);
1662 if (result != NULL) {
1663 assert(*succeeded, "sanity");
1664 return result;
1665 }
1666
1667 // Expansion didn't work, we'll try to do a Full GC.
1668 bool gc_succeeded = do_collection(false, /* explicit_gc */
1669 false, /* clear_all_soft_refs */
1670 word_size);
1671 if (!gc_succeeded) {
1672 *succeeded = false;
1673 return NULL;
1674 }
1675
1676 // Retry the allocation
1677 result = attempt_allocation_at_safepoint(word_size,
1678 true /* expect_null_mutator_alloc_region */);
1679 if (result != NULL) {
1680 assert(*succeeded, "sanity");
1681 return result;
1682 }
1683
1684 // Then, try a Full GC that will collect all soft references.
1685 gc_succeeded = do_collection(false, /* explicit_gc */
1686 true, /* clear_all_soft_refs */
1687 word_size);
1688 if (!gc_succeeded) {
1689 *succeeded = false;
1690 return NULL;
1691 }
1692
1693 // Retry the allocation once more
1694 result = attempt_allocation_at_safepoint(word_size,
1695 true /* expect_null_mutator_alloc_region */);
1696 if (result != NULL) {
1697 assert(*succeeded, "sanity");
1698 return result;
1699 }
1700
1701 assert(!collector_policy()->should_clear_all_soft_refs(),
1702 "Flag should have been handled and cleared prior to this point");
1703
1704 // What else? We might try synchronous finalization later. If the total
1705 // space available is large enough for the allocation, then a more
1706 // complete compaction phase than we've tried so far might be
1707 // appropriate.
1708 assert(*succeeded, "sanity");
1709 return NULL;
1710 }
1711
1712 // Attempting to expand the heap sufficiently
1713 // to support an allocation of the given "word_size". If
1714 // successful, perform the allocation and return the address of the
1715 // allocated block, or else "NULL".
1716
1717 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) {
1718 assert_at_safepoint(true /* should_be_vm_thread */);
1719
1720 verify_region_sets_optional();
1721
1722 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1723 ergo_verbose1(ErgoHeapSizing,
1724 "attempt heap expansion",
1725 ergo_format_reason("allocation request failed")
1726 ergo_format_byte("allocation request"),
1727 word_size * HeapWordSize);
1728 if (expand(expand_bytes)) {
1729 _hrs.verify_optional();
1730 verify_region_sets_optional();
1731 return attempt_allocation_at_safepoint(word_size,
1732 false /* expect_null_mutator_alloc_region */);
1733 }
1734 return NULL;
1735 }
1736
1737 void G1CollectedHeap::update_committed_space(HeapWord* old_end,
1738 HeapWord* new_end) {
1739 assert(old_end != new_end, "don't call this otherwise");
1740 assert((HeapWord*) _g1_storage.high() == new_end, "invariant");
1741
1742 // Update the committed mem region.
1743 _g1_committed.set_end(new_end);
1744 // Tell the card table about the update.
1745 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed);
1746 // Tell the BOT about the update.
1747 _bot_shared->resize(_g1_committed.word_size());
1748 // Tell the hot card cache about the update
1749 _cg1r->hot_card_cache()->resize_card_counts(capacity());
1750 }
1751
1752 bool G1CollectedHeap::expand(size_t expand_bytes) {
1753 size_t old_mem_size = _g1_storage.committed_size();
1754 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1755 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1756 HeapRegion::GrainBytes);
1757 ergo_verbose2(ErgoHeapSizing,
1758 "expand the heap",
1759 ergo_format_byte("requested expansion amount")
1760 ergo_format_byte("attempted expansion amount"),
1761 expand_bytes, aligned_expand_bytes);
1762
1763 // First commit the memory.
1764 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1765 bool successful = _g1_storage.expand_by(aligned_expand_bytes);
1766 if (successful) {
1767 // Then propagate this update to the necessary data structures.
1768 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1769 update_committed_space(old_end, new_end);
1770
1771 FreeRegionList expansion_list("Local Expansion List");
1772 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list);
1773 assert(mr.start() == old_end, "post-condition");
1774 // mr might be a smaller region than what was requested if
1775 // expand_by() was unable to allocate the HeapRegion instances
1776 assert(mr.end() <= new_end, "post-condition");
1777
1778 size_t actual_expand_bytes = mr.byte_size();
1779 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1780 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(),
1781 "post-condition");
1782 if (actual_expand_bytes < aligned_expand_bytes) {
1783 // We could not expand _hrs to the desired size. In this case we
1784 // need to shrink the committed space accordingly.
1785 assert(mr.end() < new_end, "invariant");
1786
1787 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes;
1788 // First uncommit the memory.
1789 _g1_storage.shrink_by(diff_bytes);
1790 // Then propagate this update to the necessary data structures.
1791 update_committed_space(new_end, mr.end());
1792 }
1793 _free_list.add_as_tail(&expansion_list);
1794
1795 if (_hr_printer.is_active()) {
1796 HeapWord* curr = mr.start();
1797 while (curr < mr.end()) {
1798 HeapWord* curr_end = curr + HeapRegion::GrainWords;
1799 _hr_printer.commit(curr, curr_end);
1800 curr = curr_end;
1801 }
1802 assert(curr == mr.end(), "post-condition");
1803 }
1804 g1_policy()->record_new_heap_size(n_regions());
1805 } else {
1806 ergo_verbose0(ErgoHeapSizing,
1807 "did not expand the heap",
1808 ergo_format_reason("heap expansion operation failed"));
1809 // The expansion of the virtual storage space was unsuccessful.
1810 // Let's see if it was because we ran out of swap.
1811 if (G1ExitOnExpansionFailure &&
1812 _g1_storage.uncommitted_size() >= aligned_expand_bytes) {
1813 // We had head room...
1814 vm_exit_out_of_memory(aligned_expand_bytes, OOM_MMAP_ERROR, "G1 heap expansion");
1815 }
1816 }
1817 return successful;
1818 }
1819
1820 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1821 size_t old_mem_size = _g1_storage.committed_size();
1822 size_t aligned_shrink_bytes =
1823 ReservedSpace::page_align_size_down(shrink_bytes);
1824 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1825 HeapRegion::GrainBytes);
1826 uint num_regions_to_remove = (uint)(shrink_bytes / HeapRegion::GrainBytes);
1827
1828 uint num_regions_removed = _hrs.shrink_by(num_regions_to_remove);
1829 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1830 size_t shrunk_bytes = num_regions_removed * HeapRegion::GrainBytes;
1831
1832 ergo_verbose3(ErgoHeapSizing,
1833 "shrink the heap",
1834 ergo_format_byte("requested shrinking amount")
1835 ergo_format_byte("aligned shrinking amount")
1836 ergo_format_byte("attempted shrinking amount"),
1837 shrink_bytes, aligned_shrink_bytes, shrunk_bytes);
1838 if (num_regions_removed > 0) {
1839 _g1_storage.shrink_by(shrunk_bytes);
1840 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1841
1842 if (_hr_printer.is_active()) {
1843 HeapWord* curr = old_end;
1844 while (curr > new_end) {
1845 HeapWord* curr_end = curr;
1846 curr -= HeapRegion::GrainWords;
1847 _hr_printer.uncommit(curr, curr_end);
1848 }
1849 }
1850
1851 _expansion_regions += num_regions_removed;
1852 update_committed_space(old_end, new_end);
1853 HeapRegionRemSet::shrink_heap(n_regions());
1854 g1_policy()->record_new_heap_size(n_regions());
1855 } else {
1856 ergo_verbose0(ErgoHeapSizing,
1857 "did not shrink the heap",
1858 ergo_format_reason("heap shrinking operation failed"));
1859 }
1860 }
1861
1862 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1863 verify_region_sets_optional();
1864
1865 // We should only reach here at the end of a Full GC which means we
1866 // should not not be holding to any GC alloc regions. The method
1867 // below will make sure of that and do any remaining clean up.
1868 abandon_gc_alloc_regions();
1869
1870 // Instead of tearing down / rebuilding the free lists here, we
1871 // could instead use the remove_all_pending() method on free_list to
1872 // remove only the ones that we need to remove.
1873 tear_down_region_sets(true /* free_list_only */);
1874 shrink_helper(shrink_bytes);
1875 rebuild_region_sets(true /* free_list_only */);
1876
1877 _hrs.verify_optional();
1878 verify_region_sets_optional();
1879 }
1880
1881 // Public methods.
1882
1883 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1884 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1885 #endif // _MSC_VER
1886
1887
1888 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1889 SharedHeap(policy_),
1890 _g1_policy(policy_),
1891 _dirty_card_queue_set(false),
1892 _into_cset_dirty_card_queue_set(false),
1893 _is_alive_closure_cm(this),
1894 _is_alive_closure_stw(this),
1895 _ref_processor_cm(NULL),
1896 _ref_processor_stw(NULL),
1897 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1898 _bot_shared(NULL),
1899 _evac_failure_scan_stack(NULL) ,
1900 _mark_in_progress(false),
1901 _cg1r(NULL), _summary_bytes_used(0),
1902 _g1mm(NULL),
1903 _refine_cte_cl(NULL),
1904 _full_collection(false),
1905 _free_list("Master Free List"),
1906 _secondary_free_list("Secondary Free List"),
1907 _old_set("Old Set"),
1908 _humongous_set("Master Humongous Set"),
1909 _free_regions_coming(false),
1910 _young_list(new YoungList(this)),
1911 _gc_time_stamp(0),
1912 _retained_old_gc_alloc_region(NULL),
1913 _survivor_plab_stats(YoungPLABSize, PLABWeight),
1914 _old_plab_stats(OldPLABSize, PLABWeight),
1915 _expand_heap_after_alloc_failure(true),
1916 _surviving_young_words(NULL),
1917 _old_marking_cycles_started(0),
1918 _old_marking_cycles_completed(0),
1919 _in_cset_fast_test(NULL),
1920 _in_cset_fast_test_base(NULL),
1921 _dirty_cards_region_list(NULL),
1922 _worker_cset_start_region(NULL),
1923 _worker_cset_start_region_time_stamp(NULL) {
1924 _g1h = this; // To catch bugs.
1925 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1926 vm_exit_during_initialization("Failed necessary allocation.");
1927 }
1928
1929 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1930
1931 int n_queues = MAX2((int)ParallelGCThreads, 1);
1932 _task_queues = new RefToScanQueueSet(n_queues);
1933
1934 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1935 assert(n_rem_sets > 0, "Invariant.");
1936
1937 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC);
1938 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC);
1939
1940 for (int i = 0; i < n_queues; i++) {
1941 RefToScanQueue* q = new RefToScanQueue();
1942 q->initialize();
1943 _task_queues->register_queue(i, q);
1944 }
1945
1946 clear_cset_start_regions();
1947
1948 // Initialize the G1EvacuationFailureALot counters and flags.
1949 NOT_PRODUCT(reset_evacuation_should_fail();)
1950
1951 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1952 }
1953
1954 jint G1CollectedHeap::initialize() {
1955 CollectedHeap::pre_initialize();
1956 os::enable_vtime();
1957
1958 G1Log::init();
1959
1960 // Necessary to satisfy locking discipline assertions.
1961
1962 MutexLocker x(Heap_lock);
1963
1964 // We have to initialize the printer before committing the heap, as
1965 // it will be used then.
1966 _hr_printer.set_active(G1PrintHeapRegions);
1967
1968 // While there are no constraints in the GC code that HeapWordSize
1969 // be any particular value, there are multiple other areas in the
1970 // system which believe this to be true (e.g. oop->object_size in some
1971 // cases incorrectly returns the size in wordSize units rather than
1972 // HeapWordSize).
1973 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1974
1975 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1976 size_t max_byte_size = collector_policy()->max_heap_byte_size();
1977
1978 // Ensure that the sizes are properly aligned.
1979 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
1980 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
1981
1982 _cg1r = new ConcurrentG1Refine(this);
1983
1984 // Reserve the maximum.
1985
1986 // When compressed oops are enabled, the preferred heap base
1987 // is calculated by subtracting the requested size from the
1988 // 32Gb boundary and using the result as the base address for
1989 // heap reservation. If the requested size is not aligned to
1990 // HeapRegion::GrainBytes (i.e. the alignment that is passed
1991 // into the ReservedHeapSpace constructor) then the actual
1992 // base of the reserved heap may end up differing from the
1993 // address that was requested (i.e. the preferred heap base).
1994 // If this happens then we could end up using a non-optimal
1995 // compressed oops mode.
1996
1997 // Since max_byte_size is aligned to the size of a heap region (checked
1998 // above).
1999 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
2000
2001 ReservedSpace heap_rs = Universe::reserve_heap(max_byte_size,
2002 HeapRegion::GrainBytes);
2003
2004 // It is important to do this in a way such that concurrent readers can't
2005 // temporarily think somethings in the heap. (I've actually seen this
2006 // happen in asserts: DLD.)
2007 _reserved.set_word_size(0);
2008 _reserved.set_start((HeapWord*)heap_rs.base());
2009 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
2010
2011 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes);
2012
2013 // Create the gen rem set (and barrier set) for the entire reserved region.
2014 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
2015 set_barrier_set(rem_set()->bs());
2016 if (barrier_set()->is_a(BarrierSet::ModRef)) {
2017 _mr_bs = (ModRefBarrierSet*)_barrier_set;
2018 } else {
2019 vm_exit_during_initialization("G1 requires a mod ref bs.");
2020 return JNI_ENOMEM;
2021 }
2022
2023 // Also create a G1 rem set.
2024 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
2025 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
2026 } else {
2027 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
2028 return JNI_ENOMEM;
2029 }
2030
2031 // Carve out the G1 part of the heap.
2032
2033 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
2034 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
2035 g1_rs.size()/HeapWordSize);
2036
2037 _g1_storage.initialize(g1_rs, 0);
2038 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
2039 _hrs.initialize((HeapWord*) _g1_reserved.start(),
2040 (HeapWord*) _g1_reserved.end(),
2041 _expansion_regions);
2042
2043 // Do later initialization work for concurrent refinement.
2044 _cg1r->init();
2045
2046 // 6843694 - ensure that the maximum region index can fit
2047 // in the remembered set structures.
2048 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
2049 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
2050
2051 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
2052 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
2053 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
2054 "too many cards per region");
2055
2056 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
2057
2058 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
2059 heap_word_size(init_byte_size));
2060
2061 _g1h = this;
2062
2063 _in_cset_fast_test_length = max_regions();
2064 _in_cset_fast_test_base =
2065 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC);
2066
2067 // We're biasing _in_cset_fast_test to avoid subtracting the
2068 // beginning of the heap every time we want to index; basically
2069 // it's the same with what we do with the card table.
2070 _in_cset_fast_test = _in_cset_fast_test_base -
2071 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
2072
2073 // Clear the _cset_fast_test bitmap in anticipation of adding
2074 // regions to the incremental collection set for the first
2075 // evacuation pause.
2076 clear_cset_fast_test();
2077
2078 // Create the ConcurrentMark data structure and thread.
2079 // (Must do this late, so that "max_regions" is defined.)
2080 _cm = new ConcurrentMark(this, heap_rs);
2081 if (_cm == NULL || !_cm->completed_initialization()) {
2082 vm_shutdown_during_initialization("Could not create/initialize ConcurrentMark");
2083 return JNI_ENOMEM;
2084 }
2085 _cmThread = _cm->cmThread();
2086
2087 // Initialize the from_card cache structure of HeapRegionRemSet.
2088 HeapRegionRemSet::init_heap(max_regions());
2089
2090 // Now expand into the initial heap size.
2091 if (!expand(init_byte_size)) {
2092 vm_shutdown_during_initialization("Failed to allocate initial heap.");
2093 return JNI_ENOMEM;
2094 }
2095
2096 // Perform any initialization actions delegated to the policy.
2097 g1_policy()->init();
2098
2099 _refine_cte_cl =
2100 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2101 g1_rem_set(),
2102 concurrent_g1_refine());
2103 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2104
2105 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2106 SATB_Q_FL_lock,
2107 G1SATBProcessCompletedThreshold,
2108 Shared_SATB_Q_lock);
2109
2110 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2111 DirtyCardQ_FL_lock,
2112 concurrent_g1_refine()->yellow_zone(),
2113 concurrent_g1_refine()->red_zone(),
2114 Shared_DirtyCardQ_lock);
2115
2116 if (G1DeferredRSUpdate) {
2117 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2118 DirtyCardQ_FL_lock,
2119 -1, // never trigger processing
2120 -1, // no limit on length
2121 Shared_DirtyCardQ_lock,
2122 &JavaThread::dirty_card_queue_set());
2123 }
2124
2125 // Initialize the card queue set used to hold cards containing
2126 // references into the collection set.
2127 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2128 DirtyCardQ_FL_lock,
2129 -1, // never trigger processing
2130 -1, // no limit on length
2131 Shared_DirtyCardQ_lock,
2132 &JavaThread::dirty_card_queue_set());
2133
2134 // In case we're keeping closure specialization stats, initialize those
2135 // counts and that mechanism.
2136 SpecializationStats::clear();
2137
2138 // Here we allocate the dummy full region that is required by the
2139 // G1AllocRegion class. If we don't pass an address in the reserved
2140 // space here, lots of asserts fire.
2141
2142 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2143 _g1_reserved.start());
2144 // We'll re-use the same region whether the alloc region will
2145 // require BOT updates or not and, if it doesn't, then a non-young
2146 // region will complain that it cannot support allocations without
2147 // BOT updates. So we'll tag the dummy region as young to avoid that.
2148 dummy_region->set_young();
2149 // Make sure it's full.
2150 dummy_region->set_top(dummy_region->end());
2151 G1AllocRegion::setup(this, dummy_region);
2152
2153 init_mutator_alloc_region();
2154
2155 // Do create of the monitoring and management support so that
2156 // values in the heap have been properly initialized.
2157 _g1mm = new G1MonitoringSupport(this);
2158
2159 return JNI_OK;
2160 }
2161
2162 void G1CollectedHeap::ref_processing_init() {
2163 // Reference processing in G1 currently works as follows:
2164 //
2165 // * There are two reference processor instances. One is
2166 // used to record and process discovered references
2167 // during concurrent marking; the other is used to
2168 // record and process references during STW pauses
2169 // (both full and incremental).
2170 // * Both ref processors need to 'span' the entire heap as
2171 // the regions in the collection set may be dotted around.
2172 //
2173 // * For the concurrent marking ref processor:
2174 // * Reference discovery is enabled at initial marking.
2175 // * Reference discovery is disabled and the discovered
2176 // references processed etc during remarking.
2177 // * Reference discovery is MT (see below).
2178 // * Reference discovery requires a barrier (see below).
2179 // * Reference processing may or may not be MT
2180 // (depending on the value of ParallelRefProcEnabled
2181 // and ParallelGCThreads).
2182 // * A full GC disables reference discovery by the CM
2183 // ref processor and abandons any entries on it's
2184 // discovered lists.
2185 //
2186 // * For the STW processor:
2187 // * Non MT discovery is enabled at the start of a full GC.
2188 // * Processing and enqueueing during a full GC is non-MT.
2189 // * During a full GC, references are processed after marking.
2190 //
2191 // * Discovery (may or may not be MT) is enabled at the start
2192 // of an incremental evacuation pause.
2193 // * References are processed near the end of a STW evacuation pause.
2194 // * For both types of GC:
2195 // * Discovery is atomic - i.e. not concurrent.
2196 // * Reference discovery will not need a barrier.
2197
2198 SharedHeap::ref_processing_init();
2199 MemRegion mr = reserved_region();
2200
2201 // Concurrent Mark ref processor
2202 _ref_processor_cm =
2203 new ReferenceProcessor(mr, // span
2204 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2205 // mt processing
2206 (int) ParallelGCThreads,
2207 // degree of mt processing
2208 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2209 // mt discovery
2210 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2211 // degree of mt discovery
2212 false,
2213 // Reference discovery is not atomic
2214 &_is_alive_closure_cm,
2215 // is alive closure
2216 // (for efficiency/performance)
2217 true);
2218 // Setting next fields of discovered
2219 // lists requires a barrier.
2220
2221 // STW ref processor
2222 _ref_processor_stw =
2223 new ReferenceProcessor(mr, // span
2224 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2225 // mt processing
2226 MAX2((int)ParallelGCThreads, 1),
2227 // degree of mt processing
2228 (ParallelGCThreads > 1),
2229 // mt discovery
2230 MAX2((int)ParallelGCThreads, 1),
2231 // degree of mt discovery
2232 true,
2233 // Reference discovery is atomic
2234 &_is_alive_closure_stw,
2235 // is alive closure
2236 // (for efficiency/performance)
2237 false);
2238 // Setting next fields of discovered
2239 // lists requires a barrier.
2240 }
2241
2242 size_t G1CollectedHeap::capacity() const {
2243 return _g1_committed.byte_size();
2244 }
2245
2246 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) {
2247 assert(!hr->continuesHumongous(), "pre-condition");
2248 hr->reset_gc_time_stamp();
2249 if (hr->startsHumongous()) {
2250 uint first_index = hr->hrs_index() + 1;
2251 uint last_index = hr->last_hc_index();
2252 for (uint i = first_index; i < last_index; i += 1) {
2253 HeapRegion* chr = region_at(i);
2254 assert(chr->continuesHumongous(), "sanity");
2255 chr->reset_gc_time_stamp();
2256 }
2257 }
2258 }
2259
2260 #ifndef PRODUCT
2261 class CheckGCTimeStampsHRClosure : public HeapRegionClosure {
2262 private:
2263 unsigned _gc_time_stamp;
2264 bool _failures;
2265
2266 public:
2267 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) :
2268 _gc_time_stamp(gc_time_stamp), _failures(false) { }
2269
2270 virtual bool doHeapRegion(HeapRegion* hr) {
2271 unsigned region_gc_time_stamp = hr->get_gc_time_stamp();
2272 if (_gc_time_stamp != region_gc_time_stamp) {
2273 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, "
2274 "expected %d", HR_FORMAT_PARAMS(hr),
2275 region_gc_time_stamp, _gc_time_stamp);
2276 _failures = true;
2277 }
2278 return false;
2279 }
2280
2281 bool failures() { return _failures; }
2282 };
2283
2284 void G1CollectedHeap::check_gc_time_stamps() {
2285 CheckGCTimeStampsHRClosure cl(_gc_time_stamp);
2286 heap_region_iterate(&cl);
2287 guarantee(!cl.failures(), "all GC time stamps should have been reset");
2288 }
2289 #endif // PRODUCT
2290
2291 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2292 DirtyCardQueue* into_cset_dcq,
2293 bool concurrent,
2294 int worker_i) {
2295 // Clean cards in the hot card cache
2296 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache();
2297 hot_card_cache->drain(worker_i, g1_rem_set(), into_cset_dcq);
2298
2299 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2300 int n_completed_buffers = 0;
2301 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2302 n_completed_buffers++;
2303 }
2304 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers);
2305 dcqs.clear_n_completed_buffers();
2306 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2307 }
2308
2309
2310 // Computes the sum of the storage used by the various regions.
2311
2312 size_t G1CollectedHeap::used() const {
2313 assert(Heap_lock->owner() != NULL,
2314 "Should be owned on this thread's behalf.");
2315 size_t result = _summary_bytes_used;
2316 // Read only once in case it is set to NULL concurrently
2317 HeapRegion* hr = _mutator_alloc_region.get();
2318 if (hr != NULL)
2319 result += hr->used();
2320 return result;
2321 }
2322
2323 size_t G1CollectedHeap::used_unlocked() const {
2324 size_t result = _summary_bytes_used;
2325 return result;
2326 }
2327
2328 class SumUsedClosure: public HeapRegionClosure {
2329 size_t _used;
2330 public:
2331 SumUsedClosure() : _used(0) {}
2332 bool doHeapRegion(HeapRegion* r) {
2333 if (!r->continuesHumongous()) {
2334 _used += r->used();
2335 }
2336 return false;
2337 }
2338 size_t result() { return _used; }
2339 };
2340
2341 size_t G1CollectedHeap::recalculate_used() const {
2342 SumUsedClosure blk;
2343 heap_region_iterate(&blk);
2344 return blk.result();
2345 }
2346
2347 size_t G1CollectedHeap::unsafe_max_alloc() {
2348 if (free_regions() > 0) return HeapRegion::GrainBytes;
2349 // otherwise, is there space in the current allocation region?
2350
2351 // We need to store the current allocation region in a local variable
2352 // here. The problem is that this method doesn't take any locks and
2353 // there may be other threads which overwrite the current allocation
2354 // region field. attempt_allocation(), for example, sets it to NULL
2355 // and this can happen *after* the NULL check here but before the call
2356 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2357 // to be a problem in the optimized build, since the two loads of the
2358 // current allocation region field are optimized away.
2359 HeapRegion* hr = _mutator_alloc_region.get();
2360 if (hr == NULL) {
2361 return 0;
2362 }
2363 return hr->free();
2364 }
2365
2366 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2367 switch (cause) {
2368 case GCCause::_gc_locker: return GCLockerInvokesConcurrent;
2369 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent;
2370 case GCCause::_g1_humongous_allocation: return true;
2371 default: return false;
2372 }
2373 }
2374
2375 #ifndef PRODUCT
2376 void G1CollectedHeap::allocate_dummy_regions() {
2377 // Let's fill up most of the region
2378 size_t word_size = HeapRegion::GrainWords - 1024;
2379 // And as a result the region we'll allocate will be humongous.
2380 guarantee(isHumongous(word_size), "sanity");
2381
2382 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2383 // Let's use the existing mechanism for the allocation
2384 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2385 if (dummy_obj != NULL) {
2386 MemRegion mr(dummy_obj, word_size);
2387 CollectedHeap::fill_with_object(mr);
2388 } else {
2389 // If we can't allocate once, we probably cannot allocate
2390 // again. Let's get out of the loop.
2391 break;
2392 }
2393 }
2394 }
2395 #endif // !PRODUCT
2396
2397 void G1CollectedHeap::increment_old_marking_cycles_started() {
2398 assert(_old_marking_cycles_started == _old_marking_cycles_completed ||
2399 _old_marking_cycles_started == _old_marking_cycles_completed + 1,
2400 err_msg("Wrong marking cycle count (started: %d, completed: %d)",
2401 _old_marking_cycles_started, _old_marking_cycles_completed));
2402
2403 _old_marking_cycles_started++;
2404 }
2405
2406 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) {
2407 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2408
2409 // We assume that if concurrent == true, then the caller is a
2410 // concurrent thread that was joined the Suspendible Thread
2411 // Set. If there's ever a cheap way to check this, we should add an
2412 // assert here.
2413
2414 // Given that this method is called at the end of a Full GC or of a
2415 // concurrent cycle, and those can be nested (i.e., a Full GC can
2416 // interrupt a concurrent cycle), the number of full collections
2417 // completed should be either one (in the case where there was no
2418 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2419 // behind the number of full collections started.
2420
2421 // This is the case for the inner caller, i.e. a Full GC.
2422 assert(concurrent ||
2423 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) ||
2424 (_old_marking_cycles_started == _old_marking_cycles_completed + 2),
2425 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u "
2426 "is inconsistent with _old_marking_cycles_completed = %u",
2427 _old_marking_cycles_started, _old_marking_cycles_completed));
2428
2429 // This is the case for the outer caller, i.e. the concurrent cycle.
2430 assert(!concurrent ||
2431 (_old_marking_cycles_started == _old_marking_cycles_completed + 1),
2432 err_msg("for outer caller (concurrent cycle): "
2433 "_old_marking_cycles_started = %u "
2434 "is inconsistent with _old_marking_cycles_completed = %u",
2435 _old_marking_cycles_started, _old_marking_cycles_completed));
2436
2437 _old_marking_cycles_completed += 1;
2438
2439 // We need to clear the "in_progress" flag in the CM thread before
2440 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2441 // is set) so that if a waiter requests another System.gc() it doesn't
2442 // incorrectly see that a marking cyle is still in progress.
2443 if (concurrent) {
2444 _cmThread->clear_in_progress();
2445 }
2446
2447 // This notify_all() will ensure that a thread that called
2448 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2449 // and it's waiting for a full GC to finish will be woken up. It is
2450 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2451 FullGCCount_lock->notify_all();
2452 }
2453
2454 void G1CollectedHeap::collect(GCCause::Cause cause) {
2455 assert_heap_not_locked();
2456
2457 unsigned int gc_count_before;
2458 unsigned int old_marking_count_before;
2459 bool retry_gc;
2460
2461 do {
2462 retry_gc = false;
2463
2464 {
2465 MutexLocker ml(Heap_lock);
2466
2467 // Read the GC count while holding the Heap_lock
2468 gc_count_before = total_collections();
2469 old_marking_count_before = _old_marking_cycles_started;
2470 }
2471
2472 if (should_do_concurrent_full_gc(cause)) {
2473 // Schedule an initial-mark evacuation pause that will start a
2474 // concurrent cycle. We're setting word_size to 0 which means that
2475 // we are not requesting a post-GC allocation.
2476 VM_G1IncCollectionPause op(gc_count_before,
2477 0, /* word_size */
2478 true, /* should_initiate_conc_mark */
2479 g1_policy()->max_pause_time_ms(),
2480 cause);
2481
2482 VMThread::execute(&op);
2483 if (!op.pause_succeeded()) {
2484 if (old_marking_count_before == _old_marking_cycles_started) {
2485 retry_gc = op.should_retry_gc();
2486 } else {
2487 // A Full GC happened while we were trying to schedule the
2488 // initial-mark GC. No point in starting a new cycle given
2489 // that the whole heap was collected anyway.
2490 }
2491
2492 if (retry_gc) {
2493 if (GC_locker::is_active_and_needs_gc()) {
2494 GC_locker::stall_until_clear();
2495 }
2496 }
2497 }
2498 } else {
2499 if (cause == GCCause::_gc_locker
2500 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2501
2502 // Schedule a standard evacuation pause. We're setting word_size
2503 // to 0 which means that we are not requesting a post-GC allocation.
2504 VM_G1IncCollectionPause op(gc_count_before,
2505 0, /* word_size */
2506 false, /* should_initiate_conc_mark */
2507 g1_policy()->max_pause_time_ms(),
2508 cause);
2509 VMThread::execute(&op);
2510 } else {
2511 // Schedule a Full GC.
2512 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause);
2513 VMThread::execute(&op);
2514 }
2515 }
2516 } while (retry_gc);
2517 }
2518
2519 bool G1CollectedHeap::is_in(const void* p) const {
2520 if (_g1_committed.contains(p)) {
2521 // Given that we know that p is in the committed space,
2522 // heap_region_containing_raw() should successfully
2523 // return the containing region.
2524 HeapRegion* hr = heap_region_containing_raw(p);
2525 return hr->is_in(p);
2526 } else {
2527 return false;
2528 }
2529 }
2530
2531 // Iteration functions.
2532
2533 // Iterates an OopClosure over all ref-containing fields of objects
2534 // within a HeapRegion.
2535
2536 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2537 MemRegion _mr;
2538 ExtendedOopClosure* _cl;
2539 public:
2540 IterateOopClosureRegionClosure(MemRegion mr, ExtendedOopClosure* cl)
2541 : _mr(mr), _cl(cl) {}
2542 bool doHeapRegion(HeapRegion* r) {
2543 if (!r->continuesHumongous()) {
2544 r->oop_iterate(_cl);
2545 }
2546 return false;
2547 }
2548 };
2549
2550 void G1CollectedHeap::oop_iterate(ExtendedOopClosure* cl) {
2551 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2552 heap_region_iterate(&blk);
2553 }
2554
2555 void G1CollectedHeap::oop_iterate(MemRegion mr, ExtendedOopClosure* cl) {
2556 IterateOopClosureRegionClosure blk(mr, cl);
2557 heap_region_iterate(&blk);
2558 }
2559
2560 // Iterates an ObjectClosure over all objects within a HeapRegion.
2561
2562 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2563 ObjectClosure* _cl;
2564 public:
2565 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2566 bool doHeapRegion(HeapRegion* r) {
2567 if (! r->continuesHumongous()) {
2568 r->object_iterate(_cl);
2569 }
2570 return false;
2571 }
2572 };
2573
2574 void G1CollectedHeap::object_iterate(ObjectClosure* cl) {
2575 IterateObjectClosureRegionClosure blk(cl);
2576 heap_region_iterate(&blk);
2577 }
2578
2579 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2580 // FIXME: is this right?
2581 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2582 }
2583
2584 // Calls a SpaceClosure on a HeapRegion.
2585
2586 class SpaceClosureRegionClosure: public HeapRegionClosure {
2587 SpaceClosure* _cl;
2588 public:
2589 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2590 bool doHeapRegion(HeapRegion* r) {
2591 _cl->do_space(r);
2592 return false;
2593 }
2594 };
2595
2596 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2597 SpaceClosureRegionClosure blk(cl);
2598 heap_region_iterate(&blk);
2599 }
2600
2601 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2602 _hrs.iterate(cl);
2603 }
2604
2605 void
2606 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2607 uint worker_id,
2608 uint no_of_par_workers,
2609 jint claim_value) {
2610 const uint regions = n_regions();
2611 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
2612 no_of_par_workers :
2613 1);
2614 assert(UseDynamicNumberOfGCThreads ||
2615 no_of_par_workers == workers()->total_workers(),
2616 "Non dynamic should use fixed number of workers");
2617 // try to spread out the starting points of the workers
2618 const HeapRegion* start_hr =
2619 start_region_for_worker(worker_id, no_of_par_workers);
2620 const uint start_index = start_hr->hrs_index();
2621
2622 // each worker will actually look at all regions
2623 for (uint count = 0; count < regions; ++count) {
2624 const uint index = (start_index + count) % regions;
2625 assert(0 <= index && index < regions, "sanity");
2626 HeapRegion* r = region_at(index);
2627 // we'll ignore "continues humongous" regions (we'll process them
2628 // when we come across their corresponding "start humongous"
2629 // region) and regions already claimed
2630 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2631 continue;
2632 }
2633 // OK, try to claim it
2634 if (r->claimHeapRegion(claim_value)) {
2635 // success!
2636 assert(!r->continuesHumongous(), "sanity");
2637 if (r->startsHumongous()) {
2638 // If the region is "starts humongous" we'll iterate over its
2639 // "continues humongous" first; in fact we'll do them
2640 // first. The order is important. In on case, calling the
2641 // closure on the "starts humongous" region might de-allocate
2642 // and clear all its "continues humongous" regions and, as a
2643 // result, we might end up processing them twice. So, we'll do
2644 // them first (notice: most closures will ignore them anyway) and
2645 // then we'll do the "starts humongous" region.
2646 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) {
2647 HeapRegion* chr = region_at(ch_index);
2648
2649 // if the region has already been claimed or it's not
2650 // "continues humongous" we're done
2651 if (chr->claim_value() == claim_value ||
2652 !chr->continuesHumongous()) {
2653 break;
2654 }
2655
2656 // Noone should have claimed it directly. We can given
2657 // that we claimed its "starts humongous" region.
2658 assert(chr->claim_value() != claim_value, "sanity");
2659 assert(chr->humongous_start_region() == r, "sanity");
2660
2661 if (chr->claimHeapRegion(claim_value)) {
2662 // we should always be able to claim it; noone else should
2663 // be trying to claim this region
2664
2665 bool res2 = cl->doHeapRegion(chr);
2666 assert(!res2, "Should not abort");
2667
2668 // Right now, this holds (i.e., no closure that actually
2669 // does something with "continues humongous" regions
2670 // clears them). We might have to weaken it in the future,
2671 // but let's leave these two asserts here for extra safety.
2672 assert(chr->continuesHumongous(), "should still be the case");
2673 assert(chr->humongous_start_region() == r, "sanity");
2674 } else {
2675 guarantee(false, "we should not reach here");
2676 }
2677 }
2678 }
2679
2680 assert(!r->continuesHumongous(), "sanity");
2681 bool res = cl->doHeapRegion(r);
2682 assert(!res, "Should not abort");
2683 }
2684 }
2685 }
2686
2687 class ResetClaimValuesClosure: public HeapRegionClosure {
2688 public:
2689 bool doHeapRegion(HeapRegion* r) {
2690 r->set_claim_value(HeapRegion::InitialClaimValue);
2691 return false;
2692 }
2693 };
2694
2695 void G1CollectedHeap::reset_heap_region_claim_values() {
2696 ResetClaimValuesClosure blk;
2697 heap_region_iterate(&blk);
2698 }
2699
2700 void G1CollectedHeap::reset_cset_heap_region_claim_values() {
2701 ResetClaimValuesClosure blk;
2702 collection_set_iterate(&blk);
2703 }
2704
2705 #ifdef ASSERT
2706 // This checks whether all regions in the heap have the correct claim
2707 // value. I also piggy-backed on this a check to ensure that the
2708 // humongous_start_region() information on "continues humongous"
2709 // regions is correct.
2710
2711 class CheckClaimValuesClosure : public HeapRegionClosure {
2712 private:
2713 jint _claim_value;
2714 uint _failures;
2715 HeapRegion* _sh_region;
2716
2717 public:
2718 CheckClaimValuesClosure(jint claim_value) :
2719 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2720 bool doHeapRegion(HeapRegion* r) {
2721 if (r->claim_value() != _claim_value) {
2722 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2723 "claim value = %d, should be %d",
2724 HR_FORMAT_PARAMS(r),
2725 r->claim_value(), _claim_value);
2726 ++_failures;
2727 }
2728 if (!r->isHumongous()) {
2729 _sh_region = NULL;
2730 } else if (r->startsHumongous()) {
2731 _sh_region = r;
2732 } else if (r->continuesHumongous()) {
2733 if (r->humongous_start_region() != _sh_region) {
2734 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2735 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2736 HR_FORMAT_PARAMS(r),
2737 r->humongous_start_region(),
2738 _sh_region);
2739 ++_failures;
2740 }
2741 }
2742 return false;
2743 }
2744 uint failures() { return _failures; }
2745 };
2746
2747 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2748 CheckClaimValuesClosure cl(claim_value);
2749 heap_region_iterate(&cl);
2750 return cl.failures() == 0;
2751 }
2752
2753 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure {
2754 private:
2755 jint _claim_value;
2756 uint _failures;
2757
2758 public:
2759 CheckClaimValuesInCSetHRClosure(jint claim_value) :
2760 _claim_value(claim_value), _failures(0) { }
2761
2762 uint failures() { return _failures; }
2763
2764 bool doHeapRegion(HeapRegion* hr) {
2765 assert(hr->in_collection_set(), "how?");
2766 assert(!hr->isHumongous(), "H-region in CSet");
2767 if (hr->claim_value() != _claim_value) {
2768 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", "
2769 "claim value = %d, should be %d",
2770 HR_FORMAT_PARAMS(hr),
2771 hr->claim_value(), _claim_value);
2772 _failures += 1;
2773 }
2774 return false;
2775 }
2776 };
2777
2778 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) {
2779 CheckClaimValuesInCSetHRClosure cl(claim_value);
2780 collection_set_iterate(&cl);
2781 return cl.failures() == 0;
2782 }
2783 #endif // ASSERT
2784
2785 // Clear the cached CSet starting regions and (more importantly)
2786 // the time stamps. Called when we reset the GC time stamp.
2787 void G1CollectedHeap::clear_cset_start_regions() {
2788 assert(_worker_cset_start_region != NULL, "sanity");
2789 assert(_worker_cset_start_region_time_stamp != NULL, "sanity");
2790
2791 int n_queues = MAX2((int)ParallelGCThreads, 1);
2792 for (int i = 0; i < n_queues; i++) {
2793 _worker_cset_start_region[i] = NULL;
2794 _worker_cset_start_region_time_stamp[i] = 0;
2795 }
2796 }
2797
2798 // Given the id of a worker, obtain or calculate a suitable
2799 // starting region for iterating over the current collection set.
2800 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) {
2801 assert(get_gc_time_stamp() > 0, "should have been updated by now");
2802
2803 HeapRegion* result = NULL;
2804 unsigned gc_time_stamp = get_gc_time_stamp();
2805
2806 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) {
2807 // Cached starting region for current worker was set
2808 // during the current pause - so it's valid.
2809 // Note: the cached starting heap region may be NULL
2810 // (when the collection set is empty).
2811 result = _worker_cset_start_region[worker_i];
2812 assert(result == NULL || result->in_collection_set(), "sanity");
2813 return result;
2814 }
2815
2816 // The cached entry was not valid so let's calculate
2817 // a suitable starting heap region for this worker.
2818
2819 // We want the parallel threads to start their collection
2820 // set iteration at different collection set regions to
2821 // avoid contention.
2822 // If we have:
2823 // n collection set regions
2824 // p threads
2825 // Then thread t will start at region floor ((t * n) / p)
2826
2827 result = g1_policy()->collection_set();
2828 if (G1CollectedHeap::use_parallel_gc_threads()) {
2829 uint cs_size = g1_policy()->cset_region_length();
2830 uint active_workers = workers()->active_workers();
2831 assert(UseDynamicNumberOfGCThreads ||
2832 active_workers == workers()->total_workers(),
2833 "Unless dynamic should use total workers");
2834
2835 uint end_ind = (cs_size * worker_i) / active_workers;
2836 uint start_ind = 0;
2837
2838 if (worker_i > 0 &&
2839 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) {
2840 // Previous workers starting region is valid
2841 // so let's iterate from there
2842 start_ind = (cs_size * (worker_i - 1)) / active_workers;
2843 result = _worker_cset_start_region[worker_i - 1];
2844 }
2845
2846 for (uint i = start_ind; i < end_ind; i++) {
2847 result = result->next_in_collection_set();
2848 }
2849 }
2850
2851 // Note: the calculated starting heap region may be NULL
2852 // (when the collection set is empty).
2853 assert(result == NULL || result->in_collection_set(), "sanity");
2854 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp,
2855 "should be updated only once per pause");
2856 _worker_cset_start_region[worker_i] = result;
2857 OrderAccess::storestore();
2858 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp;
2859 return result;
2860 }
2861
2862 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i,
2863 uint no_of_par_workers) {
2864 uint worker_num =
2865 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U;
2866 assert(UseDynamicNumberOfGCThreads ||
2867 no_of_par_workers == workers()->total_workers(),
2868 "Non dynamic should use fixed number of workers");
2869 const uint start_index = n_regions() * worker_i / worker_num;
2870 return region_at(start_index);
2871 }
2872
2873 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2874 HeapRegion* r = g1_policy()->collection_set();
2875 while (r != NULL) {
2876 HeapRegion* next = r->next_in_collection_set();
2877 if (cl->doHeapRegion(r)) {
2878 cl->incomplete();
2879 return;
2880 }
2881 r = next;
2882 }
2883 }
2884
2885 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2886 HeapRegionClosure *cl) {
2887 if (r == NULL) {
2888 // The CSet is empty so there's nothing to do.
2889 return;
2890 }
2891
2892 assert(r->in_collection_set(),
2893 "Start region must be a member of the collection set.");
2894 HeapRegion* cur = r;
2895 while (cur != NULL) {
2896 HeapRegion* next = cur->next_in_collection_set();
2897 if (cl->doHeapRegion(cur) && false) {
2898 cl->incomplete();
2899 return;
2900 }
2901 cur = next;
2902 }
2903 cur = g1_policy()->collection_set();
2904 while (cur != r) {
2905 HeapRegion* next = cur->next_in_collection_set();
2906 if (cl->doHeapRegion(cur) && false) {
2907 cl->incomplete();
2908 return;
2909 }
2910 cur = next;
2911 }
2912 }
2913
2914 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2915 return n_regions() > 0 ? region_at(0) : NULL;
2916 }
2917
2918
2919 Space* G1CollectedHeap::space_containing(const void* addr) const {
2920 Space* res = heap_region_containing(addr);
2921 return res;
2922 }
2923
2924 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2925 Space* sp = space_containing(addr);
2926 if (sp != NULL) {
2927 return sp->block_start(addr);
2928 }
2929 return NULL;
2930 }
2931
2932 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2933 Space* sp = space_containing(addr);
2934 assert(sp != NULL, "block_size of address outside of heap");
2935 return sp->block_size(addr);
2936 }
2937
2938 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
2939 Space* sp = space_containing(addr);
2940 return sp->block_is_obj(addr);
2941 }
2942
2943 bool G1CollectedHeap::supports_tlab_allocation() const {
2944 return true;
2945 }
2946
2947 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
2948 return HeapRegion::GrainBytes;
2949 }
2950
2951 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
2952 // Return the remaining space in the cur alloc region, but not less than
2953 // the min TLAB size.
2954
2955 // Also, this value can be at most the humongous object threshold,
2956 // since we can't allow tlabs to grow big enough to accomodate
2957 // humongous objects.
2958
2959 HeapRegion* hr = _mutator_alloc_region.get();
2960 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
2961 if (hr == NULL) {
2962 return max_tlab_size;
2963 } else {
2964 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
2965 }
2966 }
2967
2968 size_t G1CollectedHeap::max_capacity() const {
2969 return _g1_reserved.byte_size();
2970 }
2971
2972 jlong G1CollectedHeap::millis_since_last_gc() {
2973 // assert(false, "NYI");
2974 return 0;
2975 }
2976
2977 void G1CollectedHeap::prepare_for_verify() {
2978 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
2979 ensure_parsability(false);
2980 }
2981 g1_rem_set()->prepare_for_verify();
2982 }
2983
2984 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr,
2985 VerifyOption vo) {
2986 switch (vo) {
2987 case VerifyOption_G1UsePrevMarking:
2988 return hr->obj_allocated_since_prev_marking(obj);
2989 case VerifyOption_G1UseNextMarking:
2990 return hr->obj_allocated_since_next_marking(obj);
2991 case VerifyOption_G1UseMarkWord:
2992 return false;
2993 default:
2994 ShouldNotReachHere();
2995 }
2996 return false; // keep some compilers happy
2997 }
2998
2999 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) {
3000 switch (vo) {
3001 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start();
3002 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start();
3003 case VerifyOption_G1UseMarkWord: return NULL;
3004 default: ShouldNotReachHere();
3005 }
3006 return NULL; // keep some compilers happy
3007 }
3008
3009 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) {
3010 switch (vo) {
3011 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj);
3012 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj);
3013 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked();
3014 default: ShouldNotReachHere();
3015 }
3016 return false; // keep some compilers happy
3017 }
3018
3019 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) {
3020 switch (vo) {
3021 case VerifyOption_G1UsePrevMarking: return "PTAMS";
3022 case VerifyOption_G1UseNextMarking: return "NTAMS";
3023 case VerifyOption_G1UseMarkWord: return "NONE";
3024 default: ShouldNotReachHere();
3025 }
3026 return NULL; // keep some compilers happy
3027 }
3028
3029 // TODO: VerifyRootsClosure extends OopsInGenClosure so that we can
3030 // pass it as the perm_blk to SharedHeap::process_strong_roots.
3031 // When process_strong_roots stop calling perm_blk->younger_refs_iterate
3032 // we can change this closure to extend the simpler OopClosure.
3033 class VerifyRootsClosure: public OopsInGenClosure {
3034 private:
3035 G1CollectedHeap* _g1h;
3036 VerifyOption _vo;
3037 bool _failures;
3038 public:
3039 // _vo == UsePrevMarking -> use "prev" marking information,
3040 // _vo == UseNextMarking -> use "next" marking information,
3041 // _vo == UseMarkWord -> use mark word from object header.
3042 VerifyRootsClosure(VerifyOption vo) :
3043 _g1h(G1CollectedHeap::heap()),
3044 _vo(vo),
3045 _failures(false) { }
3046
3047 bool failures() { return _failures; }
3048
3049 template <class T> void do_oop_nv(T* p) {
3050 T heap_oop = oopDesc::load_heap_oop(p);
3051 if (!oopDesc::is_null(heap_oop)) {
3052 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
3053 if (_g1h->is_obj_dead_cond(obj, _vo)) {
3054 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
3055 "points to dead obj "PTR_FORMAT, p, (void*) obj);
3056 if (_vo == VerifyOption_G1UseMarkWord) {
3057 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
3058 }
3059 obj->print_on(gclog_or_tty);
3060 _failures = true;
3061 }
3062 }
3063 }
3064
3065 void do_oop(oop* p) { do_oop_nv(p); }
3066 void do_oop(narrowOop* p) { do_oop_nv(p); }
3067 };
3068
3069 class YoungRefCounterClosure : public OopClosure {
3070 G1CollectedHeap* _g1h;
3071 int _count;
3072 public:
3073 YoungRefCounterClosure(G1CollectedHeap* g1h) : _g1h(g1h), _count(0) {}
3074 void do_oop(oop* p) { if (_g1h->is_in_young(*p)) { _count++; } }
3075 void do_oop(narrowOop* p) { ShouldNotReachHere(); }
3076
3077 int count() { return _count; }
3078 void reset_count() { _count = 0; };
3079 };
3080
3081 class VerifyKlassClosure: public KlassClosure {
3082 YoungRefCounterClosure _young_ref_counter_closure;
3083 OopClosure *_oop_closure;
3084 public:
3085 VerifyKlassClosure(G1CollectedHeap* g1h, OopClosure* cl) : _young_ref_counter_closure(g1h), _oop_closure(cl) {}
3086 void do_klass(Klass* k) {
3087 k->oops_do(_oop_closure);
3088
3089 _young_ref_counter_closure.reset_count();
3090 k->oops_do(&_young_ref_counter_closure);
3091 if (_young_ref_counter_closure.count() > 0) {
3092 guarantee(k->has_modified_oops(), err_msg("Klass %p, has young refs but is not dirty.", k));
3093 }
3094 }
3095 };
3096
3097 class VerifyLivenessOopClosure: public OopClosure {
3098 G1CollectedHeap* _g1h;
3099 VerifyOption _vo;
3100 public:
3101 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
3102 _g1h(g1h), _vo(vo)
3103 { }
3104 void do_oop(narrowOop *p) { do_oop_work(p); }
3105 void do_oop( oop *p) { do_oop_work(p); }
3106
3107 template <class T> void do_oop_work(T *p) {
3108 oop obj = oopDesc::load_decode_heap_oop(p);
3109 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
3110 "Dead object referenced by a not dead object");
3111 }
3112 };
3113
3114 class VerifyObjsInRegionClosure: public ObjectClosure {
3115 private:
3116 G1CollectedHeap* _g1h;
3117 size_t _live_bytes;
3118 HeapRegion *_hr;
3119 VerifyOption _vo;
3120 public:
3121 // _vo == UsePrevMarking -> use "prev" marking information,
3122 // _vo == UseNextMarking -> use "next" marking information,
3123 // _vo == UseMarkWord -> use mark word from object header.
3124 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
3125 : _live_bytes(0), _hr(hr), _vo(vo) {
3126 _g1h = G1CollectedHeap::heap();
3127 }
3128 void do_object(oop o) {
3129 VerifyLivenessOopClosure isLive(_g1h, _vo);
3130 assert(o != NULL, "Huh?");
3131 if (!_g1h->is_obj_dead_cond(o, _vo)) {
3132 // If the object is alive according to the mark word,
3133 // then verify that the marking information agrees.
3134 // Note we can't verify the contra-positive of the
3135 // above: if the object is dead (according to the mark
3136 // word), it may not be marked, or may have been marked
3137 // but has since became dead, or may have been allocated
3138 // since the last marking.
3139 if (_vo == VerifyOption_G1UseMarkWord) {
3140 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
3141 }
3142
3143 o->oop_iterate_no_header(&isLive);
3144 if (!_hr->obj_allocated_since_prev_marking(o)) {
3145 size_t obj_size = o->size(); // Make sure we don't overflow
3146 _live_bytes += (obj_size * HeapWordSize);
3147 }
3148 }
3149 }
3150 size_t live_bytes() { return _live_bytes; }
3151 };
3152
3153 class PrintObjsInRegionClosure : public ObjectClosure {
3154 HeapRegion *_hr;
3155 G1CollectedHeap *_g1;
3156 public:
3157 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
3158 _g1 = G1CollectedHeap::heap();
3159 };
3160
3161 void do_object(oop o) {
3162 if (o != NULL) {
3163 HeapWord *start = (HeapWord *) o;
3164 size_t word_sz = o->size();
3165 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
3166 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
3167 (void*) o, word_sz,
3168 _g1->isMarkedPrev(o),
3169 _g1->isMarkedNext(o),
3170 _hr->obj_allocated_since_prev_marking(o));
3171 HeapWord *end = start + word_sz;
3172 HeapWord *cur;
3173 int *val;
3174 for (cur = start; cur < end; cur++) {
3175 val = (int *) cur;
3176 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
3177 }
3178 }
3179 }
3180 };
3181
3182 class VerifyRegionClosure: public HeapRegionClosure {
3183 private:
3184 bool _par;
3185 VerifyOption _vo;
3186 bool _failures;
3187 public:
3188 // _vo == UsePrevMarking -> use "prev" marking information,
3189 // _vo == UseNextMarking -> use "next" marking information,
3190 // _vo == UseMarkWord -> use mark word from object header.
3191 VerifyRegionClosure(bool par, VerifyOption vo)
3192 : _par(par),
3193 _vo(vo),
3194 _failures(false) {}
3195
3196 bool failures() {
3197 return _failures;
3198 }
3199
3200 bool doHeapRegion(HeapRegion* r) {
3201 if (!r->continuesHumongous()) {
3202 bool failures = false;
3203 r->verify(_vo, &failures);
3204 if (failures) {
3205 _failures = true;
3206 } else {
3207 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
3208 r->object_iterate(¬_dead_yet_cl);
3209 if (_vo != VerifyOption_G1UseNextMarking) {
3210 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
3211 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
3212 "max_live_bytes "SIZE_FORMAT" "
3213 "< calculated "SIZE_FORMAT,
3214 r->bottom(), r->end(),
3215 r->max_live_bytes(),
3216 not_dead_yet_cl.live_bytes());
3217 _failures = true;
3218 }
3219 } else {
3220 // When vo == UseNextMarking we cannot currently do a sanity
3221 // check on the live bytes as the calculation has not been
3222 // finalized yet.
3223 }
3224 }
3225 }
3226 return false; // stop the region iteration if we hit a failure
3227 }
3228 };
3229
3230 // This is the task used for parallel verification of the heap regions
3231
3232 class G1ParVerifyTask: public AbstractGangTask {
3233 private:
3234 G1CollectedHeap* _g1h;
3235 VerifyOption _vo;
3236 bool _failures;
3237
3238 public:
3239 // _vo == UsePrevMarking -> use "prev" marking information,
3240 // _vo == UseNextMarking -> use "next" marking information,
3241 // _vo == UseMarkWord -> use mark word from object header.
3242 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) :
3243 AbstractGangTask("Parallel verify task"),
3244 _g1h(g1h),
3245 _vo(vo),
3246 _failures(false) { }
3247
3248 bool failures() {
3249 return _failures;
3250 }
3251
3252 void work(uint worker_id) {
3253 HandleMark hm;
3254 VerifyRegionClosure blk(true, _vo);
3255 _g1h->heap_region_par_iterate_chunked(&blk, worker_id,
3256 _g1h->workers()->active_workers(),
3257 HeapRegion::ParVerifyClaimValue);
3258 if (blk.failures()) {
3259 _failures = true;
3260 }
3261 }
3262 };
3263
3264 void G1CollectedHeap::verify(bool silent, VerifyOption vo) {
3265 if (SafepointSynchronize::is_at_safepoint()) {
3266 assert(Thread::current()->is_VM_thread(),
3267 "Expected to be executed serially by the VM thread at this point");
3268
3269 if (!silent) { gclog_or_tty->print("Roots "); }
3270 VerifyRootsClosure rootsCl(vo);
3271 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
3272 VerifyKlassClosure klassCl(this, &rootsCl);
3273
3274 // We apply the relevant closures to all the oops in the
3275 // system dictionary, the string table and the code cache.
3276 const int so = SO_AllClasses | SO_Strings | SO_CodeCache;
3277
3278 // Need cleared claim bits for the strong roots processing
3279 ClassLoaderDataGraph::clear_claimed_marks();
3280
3281 process_strong_roots(true, // activate StrongRootsScope
3282 false, // we set "is scavenging" to false,
3283 // so we don't reset the dirty cards.
3284 ScanningOption(so), // roots scanning options
3285 &rootsCl,
3286 &blobsCl,
3287 &klassCl
3288 );
3289
3290 bool failures = rootsCl.failures();
3291
3292 if (vo != VerifyOption_G1UseMarkWord) {
3293 // If we're verifying during a full GC then the region sets
3294 // will have been torn down at the start of the GC. Therefore
3295 // verifying the region sets will fail. So we only verify
3296 // the region sets when not in a full GC.
3297 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
3298 verify_region_sets();
3299 }
3300
3301 if (!silent) { gclog_or_tty->print("HeapRegions "); }
3302 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
3303 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3304 "sanity check");
3305
3306 G1ParVerifyTask task(this, vo);
3307 assert(UseDynamicNumberOfGCThreads ||
3308 workers()->active_workers() == workers()->total_workers(),
3309 "If not dynamic should be using all the workers");
3310 int n_workers = workers()->active_workers();
3311 set_par_threads(n_workers);
3312 workers()->run_task(&task);
3313 set_par_threads(0);
3314 if (task.failures()) {
3315 failures = true;
3316 }
3317
3318 // Checks that the expected amount of parallel work was done.
3319 // The implication is that n_workers is > 0.
3320 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
3321 "sanity check");
3322
3323 reset_heap_region_claim_values();
3324
3325 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3326 "sanity check");
3327 } else {
3328 VerifyRegionClosure blk(false, vo);
3329 heap_region_iterate(&blk);
3330 if (blk.failures()) {
3331 failures = true;
3332 }
3333 }
3334 if (!silent) gclog_or_tty->print("RemSet ");
3335 rem_set()->verify();
3336
3337 if (failures) {
3338 gclog_or_tty->print_cr("Heap:");
3339 // It helps to have the per-region information in the output to
3340 // help us track down what went wrong. This is why we call
3341 // print_extended_on() instead of print_on().
3342 print_extended_on(gclog_or_tty);
3343 gclog_or_tty->print_cr("");
3344 #ifndef PRODUCT
3345 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3346 concurrent_mark()->print_reachable("at-verification-failure",
3347 vo, false /* all */);
3348 }
3349 #endif
3350 gclog_or_tty->flush();
3351 }
3352 guarantee(!failures, "there should not have been any failures");
3353 } else {
3354 if (!silent)
3355 gclog_or_tty->print("(SKIPPING roots, heapRegionSets, heapRegions, remset) ");
3356 }
3357 }
3358
3359 void G1CollectedHeap::verify(bool silent) {
3360 verify(silent, VerifyOption_G1UsePrevMarking);
3361 }
3362
3363 double G1CollectedHeap::verify(bool guard, const char* msg) {
3364 double verify_time_ms = 0.0;
3365
3366 if (guard && total_collections() >= VerifyGCStartAt) {
3367 double verify_start = os::elapsedTime();
3368 HandleMark hm; // Discard invalid handles created during verification
3369 prepare_for_verify();
3370 Universe::verify(VerifyOption_G1UsePrevMarking, msg);
3371 verify_time_ms = (os::elapsedTime() - verify_start) * 1000;
3372 }
3373
3374 return verify_time_ms;
3375 }
3376
3377 void G1CollectedHeap::verify_before_gc() {
3378 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:");
3379 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms);
3380 }
3381
3382 void G1CollectedHeap::verify_after_gc() {
3383 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:");
3384 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms);
3385 }
3386
3387 class PrintRegionClosure: public HeapRegionClosure {
3388 outputStream* _st;
3389 public:
3390 PrintRegionClosure(outputStream* st) : _st(st) {}
3391 bool doHeapRegion(HeapRegion* r) {
3392 r->print_on(_st);
3393 return false;
3394 }
3395 };
3396
3397 void G1CollectedHeap::print_on(outputStream* st) const {
3398 st->print(" %-20s", "garbage-first heap");
3399 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3400 capacity()/K, used_unlocked()/K);
3401 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3402 _g1_storage.low_boundary(),
3403 _g1_storage.high(),
3404 _g1_storage.high_boundary());
3405 st->cr();
3406 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3407 uint young_regions = _young_list->length();
3408 st->print("%u young (" SIZE_FORMAT "K), ", young_regions,
3409 (size_t) young_regions * HeapRegion::GrainBytes / K);
3410 uint survivor_regions = g1_policy()->recorded_survivor_regions();
3411 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions,
3412 (size_t) survivor_regions * HeapRegion::GrainBytes / K);
3413 st->cr();
3414 MetaspaceAux::print_on(st);
3415 }
3416
3417 void G1CollectedHeap::print_extended_on(outputStream* st) const {
3418 print_on(st);
3419
3420 // Print the per-region information.
3421 st->cr();
3422 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), "
3423 "HS=humongous(starts), HC=humongous(continues), "
3424 "CS=collection set, F=free, TS=gc time stamp, "
3425 "PTAMS=previous top-at-mark-start, "
3426 "NTAMS=next top-at-mark-start)");
3427 PrintRegionClosure blk(st);
3428 heap_region_iterate(&blk);
3429 }
3430
3431 void G1CollectedHeap::print_on_error(outputStream* st) const {
3432 this->CollectedHeap::print_on_error(st);
3433
3434 if (_cm != NULL) {
3435 st->cr();
3436 _cm->print_on_error(st);
3437 }
3438 }
3439
3440 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3441 if (G1CollectedHeap::use_parallel_gc_threads()) {
3442 workers()->print_worker_threads_on(st);
3443 }
3444 _cmThread->print_on(st);
3445 st->cr();
3446 _cm->print_worker_threads_on(st);
3447 _cg1r->print_worker_threads_on(st);
3448 }
3449
3450 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3451 if (G1CollectedHeap::use_parallel_gc_threads()) {
3452 workers()->threads_do(tc);
3453 }
3454 tc->do_thread(_cmThread);
3455 _cg1r->threads_do(tc);
3456 }
3457
3458 void G1CollectedHeap::print_tracing_info() const {
3459 // We'll overload this to mean "trace GC pause statistics."
3460 if (TraceGen0Time || TraceGen1Time) {
3461 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3462 // to that.
3463 g1_policy()->print_tracing_info();
3464 }
3465 if (G1SummarizeRSetStats) {
3466 g1_rem_set()->print_summary_info();
3467 }
3468 if (G1SummarizeConcMark) {
3469 concurrent_mark()->print_summary_info();
3470 }
3471 g1_policy()->print_yg_surv_rate_info();
3472 SpecializationStats::print();
3473 }
3474
3475 #ifndef PRODUCT
3476 // Helpful for debugging RSet issues.
3477
3478 class PrintRSetsClosure : public HeapRegionClosure {
3479 private:
3480 const char* _msg;
3481 size_t _occupied_sum;
3482
3483 public:
3484 bool doHeapRegion(HeapRegion* r) {
3485 HeapRegionRemSet* hrrs = r->rem_set();
3486 size_t occupied = hrrs->occupied();
3487 _occupied_sum += occupied;
3488
3489 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3490 HR_FORMAT_PARAMS(r));
3491 if (occupied == 0) {
3492 gclog_or_tty->print_cr(" RSet is empty");
3493 } else {
3494 hrrs->print();
3495 }
3496 gclog_or_tty->print_cr("----------");
3497 return false;
3498 }
3499
3500 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3501 gclog_or_tty->cr();
3502 gclog_or_tty->print_cr("========================================");
3503 gclog_or_tty->print_cr(msg);
3504 gclog_or_tty->cr();
3505 }
3506
3507 ~PrintRSetsClosure() {
3508 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3509 gclog_or_tty->print_cr("========================================");
3510 gclog_or_tty->cr();
3511 }
3512 };
3513
3514 void G1CollectedHeap::print_cset_rsets() {
3515 PrintRSetsClosure cl("Printing CSet RSets");
3516 collection_set_iterate(&cl);
3517 }
3518
3519 void G1CollectedHeap::print_all_rsets() {
3520 PrintRSetsClosure cl("Printing All RSets");;
3521 heap_region_iterate(&cl);
3522 }
3523 #endif // PRODUCT
3524
3525 G1CollectedHeap* G1CollectedHeap::heap() {
3526 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3527 "not a garbage-first heap");
3528 return _g1h;
3529 }
3530
3531 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3532 // always_do_update_barrier = false;
3533 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3534 // Call allocation profiler
3535 AllocationProfiler::iterate_since_last_gc();
3536 // Fill TLAB's and such
3537 ensure_parsability(true);
3538 }
3539
3540 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3541
3542 if (G1SummarizeRSetStats &&
3543 (G1SummarizeRSetStatsPeriod > 0) &&
3544 // we are at the end of the GC. Total collections has already been increased.
3545 ((total_collections() - 1) % G1SummarizeRSetStatsPeriod == 0)) {
3546 g1_rem_set()->print_periodic_summary_info();
3547 }
3548
3549 // FIXME: what is this about?
3550 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3551 // is set.
3552 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3553 "derived pointer present"));
3554 // always_do_update_barrier = true;
3555
3556 // We have just completed a GC. Update the soft reference
3557 // policy with the new heap occupancy
3558 Universe::update_heap_info_at_gc();
3559 }
3560
3561 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3562 unsigned int gc_count_before,
3563 bool* succeeded) {
3564 assert_heap_not_locked_and_not_at_safepoint();
3565 g1_policy()->record_stop_world_start();
3566 VM_G1IncCollectionPause op(gc_count_before,
3567 word_size,
3568 false, /* should_initiate_conc_mark */
3569 g1_policy()->max_pause_time_ms(),
3570 GCCause::_g1_inc_collection_pause);
3571 VMThread::execute(&op);
3572
3573 HeapWord* result = op.result();
3574 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3575 assert(result == NULL || ret_succeeded,
3576 "the result should be NULL if the VM did not succeed");
3577 *succeeded = ret_succeeded;
3578
3579 assert_heap_not_locked();
3580 return result;
3581 }
3582
3583 void
3584 G1CollectedHeap::doConcurrentMark() {
3585 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3586 if (!_cmThread->in_progress()) {
3587 _cmThread->set_started();
3588 CGC_lock->notify();
3589 }
3590 }
3591
3592 size_t G1CollectedHeap::pending_card_num() {
3593 size_t extra_cards = 0;
3594 JavaThread *curr = Threads::first();
3595 while (curr != NULL) {
3596 DirtyCardQueue& dcq = curr->dirty_card_queue();
3597 extra_cards += dcq.size();
3598 curr = curr->next();
3599 }
3600 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3601 size_t buffer_size = dcqs.buffer_size();
3602 size_t buffer_num = dcqs.completed_buffers_num();
3603
3604 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes
3605 // in bytes - not the number of 'entries'. We need to convert
3606 // into a number of cards.
3607 return (buffer_size * buffer_num + extra_cards) / oopSize;
3608 }
3609
3610 size_t G1CollectedHeap::cards_scanned() {
3611 return g1_rem_set()->cardsScanned();
3612 }
3613
3614 void
3615 G1CollectedHeap::setup_surviving_young_words() {
3616 assert(_surviving_young_words == NULL, "pre-condition");
3617 uint array_length = g1_policy()->young_cset_region_length();
3618 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC);
3619 if (_surviving_young_words == NULL) {
3620 vm_exit_out_of_memory(sizeof(size_t) * array_length, OOM_MALLOC_ERROR,
3621 "Not enough space for young surv words summary.");
3622 }
3623 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t));
3624 #ifdef ASSERT
3625 for (uint i = 0; i < array_length; ++i) {
3626 assert( _surviving_young_words[i] == 0, "memset above" );
3627 }
3628 #endif // !ASSERT
3629 }
3630
3631 void
3632 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3633 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3634 uint array_length = g1_policy()->young_cset_region_length();
3635 for (uint i = 0; i < array_length; ++i) {
3636 _surviving_young_words[i] += surv_young_words[i];
3637 }
3638 }
3639
3640 void
3641 G1CollectedHeap::cleanup_surviving_young_words() {
3642 guarantee( _surviving_young_words != NULL, "pre-condition" );
3643 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC);
3644 _surviving_young_words = NULL;
3645 }
3646
3647 #ifdef ASSERT
3648 class VerifyCSetClosure: public HeapRegionClosure {
3649 public:
3650 bool doHeapRegion(HeapRegion* hr) {
3651 // Here we check that the CSet region's RSet is ready for parallel
3652 // iteration. The fields that we'll verify are only manipulated
3653 // when the region is part of a CSet and is collected. Afterwards,
3654 // we reset these fields when we clear the region's RSet (when the
3655 // region is freed) so they are ready when the region is
3656 // re-allocated. The only exception to this is if there's an
3657 // evacuation failure and instead of freeing the region we leave
3658 // it in the heap. In that case, we reset these fields during
3659 // evacuation failure handling.
3660 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3661
3662 // Here's a good place to add any other checks we'd like to
3663 // perform on CSet regions.
3664 return false;
3665 }
3666 };
3667 #endif // ASSERT
3668
3669 #if TASKQUEUE_STATS
3670 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3671 st->print_raw_cr("GC Task Stats");
3672 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3673 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3674 }
3675
3676 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3677 print_taskqueue_stats_hdr(st);
3678
3679 TaskQueueStats totals;
3680 const int n = workers() != NULL ? workers()->total_workers() : 1;
3681 for (int i = 0; i < n; ++i) {
3682 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3683 totals += task_queue(i)->stats;
3684 }
3685 st->print_raw("tot "); totals.print(st); st->cr();
3686
3687 DEBUG_ONLY(totals.verify());
3688 }
3689
3690 void G1CollectedHeap::reset_taskqueue_stats() {
3691 const int n = workers() != NULL ? workers()->total_workers() : 1;
3692 for (int i = 0; i < n; ++i) {
3693 task_queue(i)->stats.reset();
3694 }
3695 }
3696 #endif // TASKQUEUE_STATS
3697
3698 void G1CollectedHeap::log_gc_header() {
3699 if (!G1Log::fine()) {
3700 return;
3701 }
3702
3703 gclog_or_tty->date_stamp(PrintGCDateStamps);
3704 gclog_or_tty->stamp(PrintGCTimeStamps);
3705
3706 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause())
3707 .append(g1_policy()->gcs_are_young() ? "(young)" : "(mixed)")
3708 .append(g1_policy()->during_initial_mark_pause() ? " (initial-mark)" : "");
3709
3710 gclog_or_tty->print("[%s", (const char*)gc_cause_str);
3711 }
3712
3713 void G1CollectedHeap::log_gc_footer(double pause_time_sec) {
3714 if (!G1Log::fine()) {
3715 return;
3716 }
3717
3718 if (G1Log::finer()) {
3719 if (evacuation_failed()) {
3720 gclog_or_tty->print(" (to-space exhausted)");
3721 }
3722 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3723 g1_policy()->phase_times()->note_gc_end();
3724 g1_policy()->phase_times()->print(pause_time_sec);
3725 g1_policy()->print_detailed_heap_transition();
3726 } else {
3727 if (evacuation_failed()) {
3728 gclog_or_tty->print("--");
3729 }
3730 g1_policy()->print_heap_transition();
3731 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
3732 }
3733 gclog_or_tty->flush();
3734 }
3735
3736 bool
3737 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3738 assert_at_safepoint(true /* should_be_vm_thread */);
3739 guarantee(!is_gc_active(), "collection is not reentrant");
3740
3741 if (GC_locker::check_active_before_gc()) {
3742 return false;
3743 }
3744
3745 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3746 ResourceMark rm;
3747
3748 print_heap_before_gc();
3749
3750 HRSPhaseSetter x(HRSPhaseEvacuation);
3751 verify_region_sets_optional();
3752 verify_dirty_young_regions();
3753
3754 // This call will decide whether this pause is an initial-mark
3755 // pause. If it is, during_initial_mark_pause() will return true
3756 // for the duration of this pause.
3757 g1_policy()->decide_on_conc_mark_initiation();
3758
3759 // We do not allow initial-mark to be piggy-backed on a mixed GC.
3760 assert(!g1_policy()->during_initial_mark_pause() ||
3761 g1_policy()->gcs_are_young(), "sanity");
3762
3763 // We also do not allow mixed GCs during marking.
3764 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity");
3765
3766 // Record whether this pause is an initial mark. When the current
3767 // thread has completed its logging output and it's safe to signal
3768 // the CM thread, the flag's value in the policy has been reset.
3769 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause();
3770
3771 // Inner scope for scope based logging, timers, and stats collection
3772 {
3773 if (g1_policy()->during_initial_mark_pause()) {
3774 // We are about to start a marking cycle, so we increment the
3775 // full collection counter.
3776 increment_old_marking_cycles_started();
3777 }
3778 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
3779
3780 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3781 workers()->active_workers() : 1);
3782 double pause_start_sec = os::elapsedTime();
3783 g1_policy()->phase_times()->note_gc_start(active_workers);
3784 log_gc_header();
3785
3786 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3787 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3788
3789 // If the secondary_free_list is not empty, append it to the
3790 // free_list. No need to wait for the cleanup operation to finish;
3791 // the region allocation code will check the secondary_free_list
3792 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3793 // set, skip this step so that the region allocation code has to
3794 // get entries from the secondary_free_list.
3795 if (!G1StressConcRegionFreeing) {
3796 append_secondary_free_list_if_not_empty_with_lock();
3797 }
3798
3799 assert(check_young_list_well_formed(),
3800 "young list should be well formed");
3801
3802 // Don't dynamically change the number of GC threads this early. A value of
3803 // 0 is used to indicate serial work. When parallel work is done,
3804 // it will be set.
3805
3806 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3807 IsGCActiveMark x;
3808
3809 gc_prologue(false);
3810 increment_total_collections(false /* full gc */);
3811 increment_gc_time_stamp();
3812
3813 verify_before_gc();
3814
3815 COMPILER2_PRESENT(DerivedPointerTable::clear());
3816
3817 // Please see comment in g1CollectedHeap.hpp and
3818 // G1CollectedHeap::ref_processing_init() to see how
3819 // reference processing currently works in G1.
3820
3821 // Enable discovery in the STW reference processor
3822 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3823 true /*verify_no_refs*/);
3824
3825 {
3826 // We want to temporarily turn off discovery by the
3827 // CM ref processor, if necessary, and turn it back on
3828 // on again later if we do. Using a scoped
3829 // NoRefDiscovery object will do this.
3830 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3831
3832 // Forget the current alloc region (we might even choose it to be part
3833 // of the collection set!).
3834 release_mutator_alloc_region();
3835
3836 // We should call this after we retire the mutator alloc
3837 // region(s) so that all the ALLOC / RETIRE events are generated
3838 // before the start GC event.
3839 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3840
3841 // This timing is only used by the ergonomics to handle our pause target.
3842 // It is unclear why this should not include the full pause. We will
3843 // investigate this in CR 7178365.
3844 //
3845 // Preserving the old comment here if that helps the investigation:
3846 //
3847 // The elapsed time induced by the start time below deliberately elides
3848 // the possible verification above.
3849 double sample_start_time_sec = os::elapsedTime();
3850
3851 #if YOUNG_LIST_VERBOSE
3852 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3853 _young_list->print();
3854 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3855 #endif // YOUNG_LIST_VERBOSE
3856
3857 g1_policy()->record_collection_pause_start(sample_start_time_sec);
3858
3859 double scan_wait_start = os::elapsedTime();
3860 // We have to wait until the CM threads finish scanning the
3861 // root regions as it's the only way to ensure that all the
3862 // objects on them have been correctly scanned before we start
3863 // moving them during the GC.
3864 bool waited = _cm->root_regions()->wait_until_scan_finished();
3865 double wait_time_ms = 0.0;
3866 if (waited) {
3867 double scan_wait_end = os::elapsedTime();
3868 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0;
3869 }
3870 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms);
3871
3872 #if YOUNG_LIST_VERBOSE
3873 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3874 _young_list->print();
3875 #endif // YOUNG_LIST_VERBOSE
3876
3877 if (g1_policy()->during_initial_mark_pause()) {
3878 concurrent_mark()->checkpointRootsInitialPre();
3879 }
3880
3881 #if YOUNG_LIST_VERBOSE
3882 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3883 _young_list->print();
3884 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3885 #endif // YOUNG_LIST_VERBOSE
3886
3887 g1_policy()->finalize_cset(target_pause_time_ms);
3888
3889 _cm->note_start_of_gc();
3890 // We should not verify the per-thread SATB buffers given that
3891 // we have not filtered them yet (we'll do so during the
3892 // GC). We also call this after finalize_cset() to
3893 // ensure that the CSet has been finalized.
3894 _cm->verify_no_cset_oops(true /* verify_stacks */,
3895 true /* verify_enqueued_buffers */,
3896 false /* verify_thread_buffers */,
3897 true /* verify_fingers */);
3898
3899 if (_hr_printer.is_active()) {
3900 HeapRegion* hr = g1_policy()->collection_set();
3901 while (hr != NULL) {
3902 G1HRPrinter::RegionType type;
3903 if (!hr->is_young()) {
3904 type = G1HRPrinter::Old;
3905 } else if (hr->is_survivor()) {
3906 type = G1HRPrinter::Survivor;
3907 } else {
3908 type = G1HRPrinter::Eden;
3909 }
3910 _hr_printer.cset(hr);
3911 hr = hr->next_in_collection_set();
3912 }
3913 }
3914
3915 #ifdef ASSERT
3916 VerifyCSetClosure cl;
3917 collection_set_iterate(&cl);
3918 #endif // ASSERT
3919
3920 setup_surviving_young_words();
3921
3922 // Initialize the GC alloc regions.
3923 init_gc_alloc_regions();
3924
3925 // Actually do the work...
3926 evacuate_collection_set();
3927
3928 // We do this to mainly verify the per-thread SATB buffers
3929 // (which have been filtered by now) since we didn't verify
3930 // them earlier. No point in re-checking the stacks / enqueued
3931 // buffers given that the CSet has not changed since last time
3932 // we checked.
3933 _cm->verify_no_cset_oops(false /* verify_stacks */,
3934 false /* verify_enqueued_buffers */,
3935 true /* verify_thread_buffers */,
3936 true /* verify_fingers */);
3937
3938 free_collection_set(g1_policy()->collection_set());
3939 g1_policy()->clear_collection_set();
3940
3941 cleanup_surviving_young_words();
3942
3943 // Start a new incremental collection set for the next pause.
3944 g1_policy()->start_incremental_cset_building();
3945
3946 // Clear the _cset_fast_test bitmap in anticipation of adding
3947 // regions to the incremental collection set for the next
3948 // evacuation pause.
3949 clear_cset_fast_test();
3950
3951 _young_list->reset_sampled_info();
3952
3953 // Don't check the whole heap at this point as the
3954 // GC alloc regions from this pause have been tagged
3955 // as survivors and moved on to the survivor list.
3956 // Survivor regions will fail the !is_young() check.
3957 assert(check_young_list_empty(false /* check_heap */),
3958 "young list should be empty");
3959
3960 #if YOUNG_LIST_VERBOSE
3961 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3962 _young_list->print();
3963 #endif // YOUNG_LIST_VERBOSE
3964
3965 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3966 _young_list->first_survivor_region(),
3967 _young_list->last_survivor_region());
3968
3969 _young_list->reset_auxilary_lists();
3970
3971 if (evacuation_failed()) {
3972 _summary_bytes_used = recalculate_used();
3973 } else {
3974 // The "used" of the the collection set have already been subtracted
3975 // when they were freed. Add in the bytes evacuated.
3976 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3977 }
3978
3979 if (g1_policy()->during_initial_mark_pause()) {
3980 // We have to do this before we notify the CM threads that
3981 // they can start working to make sure that all the
3982 // appropriate initialization is done on the CM object.
3983 concurrent_mark()->checkpointRootsInitialPost();
3984 set_marking_started();
3985 // Note that we don't actually trigger the CM thread at
3986 // this point. We do that later when we're sure that
3987 // the current thread has completed its logging output.
3988 }
3989
3990 allocate_dummy_regions();
3991
3992 #if YOUNG_LIST_VERBOSE
3993 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3994 _young_list->print();
3995 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3996 #endif // YOUNG_LIST_VERBOSE
3997
3998 init_mutator_alloc_region();
3999
4000 {
4001 size_t expand_bytes = g1_policy()->expansion_amount();
4002 if (expand_bytes > 0) {
4003 size_t bytes_before = capacity();
4004 // No need for an ergo verbose message here,
4005 // expansion_amount() does this when it returns a value > 0.
4006 if (!expand(expand_bytes)) {
4007 // We failed to expand the heap so let's verify that
4008 // committed/uncommitted amount match the backing store
4009 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
4010 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
4011 }
4012 }
4013 }
4014
4015 // We redo the verificaiton but now wrt to the new CSet which
4016 // has just got initialized after the previous CSet was freed.
4017 _cm->verify_no_cset_oops(true /* verify_stacks */,
4018 true /* verify_enqueued_buffers */,
4019 true /* verify_thread_buffers */,
4020 true /* verify_fingers */);
4021 _cm->note_end_of_gc();
4022
4023 // This timing is only used by the ergonomics to handle our pause target.
4024 // It is unclear why this should not include the full pause. We will
4025 // investigate this in CR 7178365.
4026 double sample_end_time_sec = os::elapsedTime();
4027 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS;
4028 g1_policy()->record_collection_pause_end(pause_time_ms);
4029
4030 MemoryService::track_memory_usage();
4031
4032 // In prepare_for_verify() below we'll need to scan the deferred
4033 // update buffers to bring the RSets up-to-date if
4034 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
4035 // the update buffers we'll probably need to scan cards on the
4036 // regions we just allocated to (i.e., the GC alloc
4037 // regions). However, during the last GC we called
4038 // set_saved_mark() on all the GC alloc regions, so card
4039 // scanning might skip the [saved_mark_word()...top()] area of
4040 // those regions (i.e., the area we allocated objects into
4041 // during the last GC). But it shouldn't. Given that
4042 // saved_mark_word() is conditional on whether the GC time stamp
4043 // on the region is current or not, by incrementing the GC time
4044 // stamp here we invalidate all the GC time stamps on all the
4045 // regions and saved_mark_word() will simply return top() for
4046 // all the regions. This is a nicer way of ensuring this rather
4047 // than iterating over the regions and fixing them. In fact, the
4048 // GC time stamp increment here also ensures that
4049 // saved_mark_word() will return top() between pauses, i.e.,
4050 // during concurrent refinement. So we don't need the
4051 // is_gc_active() check to decided which top to use when
4052 // scanning cards (see CR 7039627).
4053 increment_gc_time_stamp();
4054
4055 verify_after_gc();
4056
4057 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
4058 ref_processor_stw()->verify_no_references_recorded();
4059
4060 // CM reference discovery will be re-enabled if necessary.
4061 }
4062
4063 // We should do this after we potentially expand the heap so
4064 // that all the COMMIT events are generated before the end GC
4065 // event, and after we retire the GC alloc regions so that all
4066 // RETIRE events are generated before the end GC event.
4067 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
4068
4069 if (mark_in_progress()) {
4070 concurrent_mark()->update_g1_committed();
4071 }
4072
4073 #ifdef TRACESPINNING
4074 ParallelTaskTerminator::print_termination_counts();
4075 #endif
4076
4077 gc_epilogue(false);
4078 }
4079
4080 // Print the remainder of the GC log output.
4081 log_gc_footer(os::elapsedTime() - pause_start_sec);
4082
4083 // It is not yet to safe to tell the concurrent mark to
4084 // start as we have some optional output below. We don't want the
4085 // output from the concurrent mark thread interfering with this
4086 // logging output either.
4087
4088 _hrs.verify_optional();
4089 verify_region_sets_optional();
4090
4091 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
4092 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
4093
4094 print_heap_after_gc();
4095
4096 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
4097 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
4098 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
4099 // before any GC notifications are raised.
4100 g1mm()->update_sizes();
4101 }
4102
4103 // It should now be safe to tell the concurrent mark thread to start
4104 // without its logging output interfering with the logging output
4105 // that came from the pause.
4106
4107 if (should_start_conc_mark) {
4108 // CAUTION: after the doConcurrentMark() call below,
4109 // the concurrent marking thread(s) could be running
4110 // concurrently with us. Make sure that anything after
4111 // this point does not assume that we are the only GC thread
4112 // running. Note: of course, the actual marking work will
4113 // not start until the safepoint itself is released in
4114 // ConcurrentGCThread::safepoint_desynchronize().
4115 doConcurrentMark();
4116 }
4117
4118 return true;
4119 }
4120
4121 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
4122 {
4123 size_t gclab_word_size;
4124 switch (purpose) {
4125 case GCAllocForSurvived:
4126 gclab_word_size = _survivor_plab_stats.desired_plab_sz();
4127 break;
4128 case GCAllocForTenured:
4129 gclab_word_size = _old_plab_stats.desired_plab_sz();
4130 break;
4131 default:
4132 assert(false, "unknown GCAllocPurpose");
4133 gclab_word_size = _old_plab_stats.desired_plab_sz();
4134 break;
4135 }
4136
4137 // Prevent humongous PLAB sizes for two reasons:
4138 // * PLABs are allocated using a similar paths as oops, but should
4139 // never be in a humongous region
4140 // * Allowing humongous PLABs needlessly churns the region free lists
4141 return MIN2(_humongous_object_threshold_in_words, gclab_word_size);
4142 }
4143
4144 void G1CollectedHeap::init_mutator_alloc_region() {
4145 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
4146 _mutator_alloc_region.init();
4147 }
4148
4149 void G1CollectedHeap::release_mutator_alloc_region() {
4150 _mutator_alloc_region.release();
4151 assert(_mutator_alloc_region.get() == NULL, "post-condition");
4152 }
4153
4154 void G1CollectedHeap::init_gc_alloc_regions() {
4155 assert_at_safepoint(true /* should_be_vm_thread */);
4156
4157 _survivor_gc_alloc_region.init();
4158 _old_gc_alloc_region.init();
4159 HeapRegion* retained_region = _retained_old_gc_alloc_region;
4160 _retained_old_gc_alloc_region = NULL;
4161
4162 // We will discard the current GC alloc region if:
4163 // a) it's in the collection set (it can happen!),
4164 // b) it's already full (no point in using it),
4165 // c) it's empty (this means that it was emptied during
4166 // a cleanup and it should be on the free list now), or
4167 // d) it's humongous (this means that it was emptied
4168 // during a cleanup and was added to the free list, but
4169 // has been subseqently used to allocate a humongous
4170 // object that may be less than the region size).
4171 if (retained_region != NULL &&
4172 !retained_region->in_collection_set() &&
4173 !(retained_region->top() == retained_region->end()) &&
4174 !retained_region->is_empty() &&
4175 !retained_region->isHumongous()) {
4176 retained_region->set_saved_mark();
4177 // The retained region was added to the old region set when it was
4178 // retired. We have to remove it now, since we don't allow regions
4179 // we allocate to in the region sets. We'll re-add it later, when
4180 // it's retired again.
4181 _old_set.remove(retained_region);
4182 bool during_im = g1_policy()->during_initial_mark_pause();
4183 retained_region->note_start_of_copying(during_im);
4184 _old_gc_alloc_region.set(retained_region);
4185 _hr_printer.reuse(retained_region);
4186 }
4187 }
4188
4189 void G1CollectedHeap::release_gc_alloc_regions(uint no_of_gc_workers) {
4190 _survivor_gc_alloc_region.release();
4191 // If we have an old GC alloc region to release, we'll save it in
4192 // _retained_old_gc_alloc_region. If we don't
4193 // _retained_old_gc_alloc_region will become NULL. This is what we
4194 // want either way so no reason to check explicitly for either
4195 // condition.
4196 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
4197
4198 if (ResizePLAB) {
4199 _survivor_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4200 _old_plab_stats.adjust_desired_plab_sz(no_of_gc_workers);
4201 }
4202 }
4203
4204 void G1CollectedHeap::abandon_gc_alloc_regions() {
4205 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
4206 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
4207 _retained_old_gc_alloc_region = NULL;
4208 }
4209
4210 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
4211 _drain_in_progress = false;
4212 set_evac_failure_closure(cl);
4213 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4214 }
4215
4216 void G1CollectedHeap::finalize_for_evac_failure() {
4217 assert(_evac_failure_scan_stack != NULL &&
4218 _evac_failure_scan_stack->length() == 0,
4219 "Postcondition");
4220 assert(!_drain_in_progress, "Postcondition");
4221 delete _evac_failure_scan_stack;
4222 _evac_failure_scan_stack = NULL;
4223 }
4224
4225 void G1CollectedHeap::remove_self_forwarding_pointers() {
4226 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4227
4228 G1ParRemoveSelfForwardPtrsTask rsfp_task(this);
4229
4230 if (G1CollectedHeap::use_parallel_gc_threads()) {
4231 set_par_threads();
4232 workers()->run_task(&rsfp_task);
4233 set_par_threads(0);
4234 } else {
4235 rsfp_task.work(0);
4236 }
4237
4238 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity");
4239
4240 // Reset the claim values in the regions in the collection set.
4241 reset_cset_heap_region_claim_values();
4242
4243 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4244
4245 // Now restore saved marks, if any.
4246 assert(_objs_with_preserved_marks.size() ==
4247 _preserved_marks_of_objs.size(), "Both or none.");
4248 while (!_objs_with_preserved_marks.is_empty()) {
4249 oop obj = _objs_with_preserved_marks.pop();
4250 markOop m = _preserved_marks_of_objs.pop();
4251 obj->set_mark(m);
4252 }
4253 _objs_with_preserved_marks.clear(true);
4254 _preserved_marks_of_objs.clear(true);
4255 }
4256
4257 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
4258 _evac_failure_scan_stack->push(obj);
4259 }
4260
4261 void G1CollectedHeap::drain_evac_failure_scan_stack() {
4262 assert(_evac_failure_scan_stack != NULL, "precondition");
4263
4264 while (_evac_failure_scan_stack->length() > 0) {
4265 oop obj = _evac_failure_scan_stack->pop();
4266 _evac_failure_closure->set_region(heap_region_containing(obj));
4267 obj->oop_iterate_backwards(_evac_failure_closure);
4268 }
4269 }
4270
4271 oop
4272 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4273 oop old) {
4274 assert(obj_in_cs(old),
4275 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4276 (HeapWord*) old));
4277 markOop m = old->mark();
4278 oop forward_ptr = old->forward_to_atomic(old);
4279 if (forward_ptr == NULL) {
4280 // Forward-to-self succeeded.
4281
4282 if (_evac_failure_closure != cl) {
4283 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4284 assert(!_drain_in_progress,
4285 "Should only be true while someone holds the lock.");
4286 // Set the global evac-failure closure to the current thread's.
4287 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4288 set_evac_failure_closure(cl);
4289 // Now do the common part.
4290 handle_evacuation_failure_common(old, m);
4291 // Reset to NULL.
4292 set_evac_failure_closure(NULL);
4293 } else {
4294 // The lock is already held, and this is recursive.
4295 assert(_drain_in_progress, "This should only be the recursive case.");
4296 handle_evacuation_failure_common(old, m);
4297 }
4298 return old;
4299 } else {
4300 // Forward-to-self failed. Either someone else managed to allocate
4301 // space for this object (old != forward_ptr) or they beat us in
4302 // self-forwarding it (old == forward_ptr).
4303 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4304 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4305 "should not be in the CSet",
4306 (HeapWord*) old, (HeapWord*) forward_ptr));
4307 return forward_ptr;
4308 }
4309 }
4310
4311 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4312 set_evacuation_failed(true);
4313
4314 preserve_mark_if_necessary(old, m);
4315
4316 HeapRegion* r = heap_region_containing(old);
4317 if (!r->evacuation_failed()) {
4318 r->set_evacuation_failed(true);
4319 _hr_printer.evac_failure(r);
4320 }
4321
4322 push_on_evac_failure_scan_stack(old);
4323
4324 if (!_drain_in_progress) {
4325 // prevent recursion in copy_to_survivor_space()
4326 _drain_in_progress = true;
4327 drain_evac_failure_scan_stack();
4328 _drain_in_progress = false;
4329 }
4330 }
4331
4332 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4333 assert(evacuation_failed(), "Oversaving!");
4334 // We want to call the "for_promotion_failure" version only in the
4335 // case of a promotion failure.
4336 if (m->must_be_preserved_for_promotion_failure(obj)) {
4337 _objs_with_preserved_marks.push(obj);
4338 _preserved_marks_of_objs.push(m);
4339 }
4340 }
4341
4342 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4343 size_t word_size) {
4344 if (purpose == GCAllocForSurvived) {
4345 HeapWord* result = survivor_attempt_allocation(word_size);
4346 if (result != NULL) {
4347 return result;
4348 } else {
4349 // Let's try to allocate in the old gen in case we can fit the
4350 // object there.
4351 return old_attempt_allocation(word_size);
4352 }
4353 } else {
4354 assert(purpose == GCAllocForTenured, "sanity");
4355 HeapWord* result = old_attempt_allocation(word_size);
4356 if (result != NULL) {
4357 return result;
4358 } else {
4359 // Let's try to allocate in the survivors in case we can fit the
4360 // object there.
4361 return survivor_attempt_allocation(word_size);
4362 }
4363 }
4364
4365 ShouldNotReachHere();
4366 // Trying to keep some compilers happy.
4367 return NULL;
4368 }
4369
4370 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4371 ParGCAllocBuffer(gclab_word_size), _retired(false) { }
4372
4373 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num)
4374 : _g1h(g1h),
4375 _refs(g1h->task_queue(queue_num)),
4376 _dcq(&g1h->dirty_card_queue_set()),
4377 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4378 _g1_rem(g1h->g1_rem_set()),
4379 _hash_seed(17), _queue_num(queue_num),
4380 _term_attempts(0),
4381 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4382 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4383 _age_table(false),
4384 _strong_roots_time(0), _term_time(0),
4385 _alloc_buffer_waste(0), _undo_waste(0) {
4386 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4387 // we "sacrifice" entry 0 to keep track of surviving bytes for
4388 // non-young regions (where the age is -1)
4389 // We also add a few elements at the beginning and at the end in
4390 // an attempt to eliminate cache contention
4391 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
4392 uint array_length = PADDING_ELEM_NUM +
4393 real_length +
4394 PADDING_ELEM_NUM;
4395 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
4396 if (_surviving_young_words_base == NULL)
4397 vm_exit_out_of_memory(array_length * sizeof(size_t), OOM_MALLOC_ERROR,
4398 "Not enough space for young surv histo.");
4399 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4400 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t));
4401
4402 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4403 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4404
4405 _start = os::elapsedTime();
4406 }
4407
4408 void
4409 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4410 {
4411 st->print_raw_cr("GC Termination Stats");
4412 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4413 " ------waste (KiB)------");
4414 st->print_raw_cr("thr ms ms % ms % attempts"
4415 " total alloc undo");
4416 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4417 " ------- ------- -------");
4418 }
4419
4420 void
4421 G1ParScanThreadState::print_termination_stats(int i,
4422 outputStream* const st) const
4423 {
4424 const double elapsed_ms = elapsed_time() * 1000.0;
4425 const double s_roots_ms = strong_roots_time() * 1000.0;
4426 const double term_ms = term_time() * 1000.0;
4427 st->print_cr("%3d %9.2f %9.2f %6.2f "
4428 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4429 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4430 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4431 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4432 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4433 alloc_buffer_waste() * HeapWordSize / K,
4434 undo_waste() * HeapWordSize / K);
4435 }
4436
4437 #ifdef ASSERT
4438 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4439 assert(ref != NULL, "invariant");
4440 assert(UseCompressedOops, "sanity");
4441 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4442 oop p = oopDesc::load_decode_heap_oop(ref);
4443 assert(_g1h->is_in_g1_reserved(p),
4444 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4445 return true;
4446 }
4447
4448 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4449 assert(ref != NULL, "invariant");
4450 if (has_partial_array_mask(ref)) {
4451 // Must be in the collection set--it's already been copied.
4452 oop p = clear_partial_array_mask(ref);
4453 assert(_g1h->obj_in_cs(p),
4454 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4455 } else {
4456 oop p = oopDesc::load_decode_heap_oop(ref);
4457 assert(_g1h->is_in_g1_reserved(p),
4458 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4459 }
4460 return true;
4461 }
4462
4463 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4464 if (ref.is_narrow()) {
4465 return verify_ref((narrowOop*) ref);
4466 } else {
4467 return verify_ref((oop*) ref);
4468 }
4469 }
4470 #endif // ASSERT
4471
4472 void G1ParScanThreadState::trim_queue() {
4473 assert(_evac_cl != NULL, "not set");
4474 assert(_evac_failure_cl != NULL, "not set");
4475 assert(_partial_scan_cl != NULL, "not set");
4476
4477 StarTask ref;
4478 do {
4479 // Drain the overflow stack first, so other threads can steal.
4480 while (refs()->pop_overflow(ref)) {
4481 deal_with_reference(ref);
4482 }
4483
4484 while (refs()->pop_local(ref)) {
4485 deal_with_reference(ref);
4486 }
4487 } while (!refs()->is_empty());
4488 }
4489
4490 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1,
4491 G1ParScanThreadState* par_scan_state) :
4492 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4493 _par_scan_state(par_scan_state),
4494 _worker_id(par_scan_state->queue_num()),
4495 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4496 _mark_in_progress(_g1->mark_in_progress()) { }
4497
4498 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4499 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) {
4500 #ifdef ASSERT
4501 HeapRegion* hr = _g1->heap_region_containing(obj);
4502 assert(hr != NULL, "sanity");
4503 assert(!hr->in_collection_set(), "should not mark objects in the CSet");
4504 #endif // ASSERT
4505
4506 // We know that the object is not moving so it's safe to read its size.
4507 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id);
4508 }
4509
4510 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4511 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4512 ::mark_forwarded_object(oop from_obj, oop to_obj) {
4513 #ifdef ASSERT
4514 assert(from_obj->is_forwarded(), "from obj should be forwarded");
4515 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee");
4516 assert(from_obj != to_obj, "should not be self-forwarded");
4517
4518 HeapRegion* from_hr = _g1->heap_region_containing(from_obj);
4519 assert(from_hr != NULL, "sanity");
4520 assert(from_hr->in_collection_set(), "from obj should be in the CSet");
4521
4522 HeapRegion* to_hr = _g1->heap_region_containing(to_obj);
4523 assert(to_hr != NULL, "sanity");
4524 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet");
4525 #endif // ASSERT
4526
4527 // The object might be in the process of being copied by another
4528 // worker so we cannot trust that its to-space image is
4529 // well-formed. So we have to read its size from its from-space
4530 // image which we know should not be changing.
4531 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id);
4532 }
4533
4534 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4535 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4536 ::copy_to_survivor_space(oop old) {
4537 size_t word_sz = old->size();
4538 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4539 // +1 to make the -1 indexes valid...
4540 int young_index = from_region->young_index_in_cset()+1;
4541 assert( (from_region->is_young() && young_index > 0) ||
4542 (!from_region->is_young() && young_index == 0), "invariant" );
4543 G1CollectorPolicy* g1p = _g1->g1_policy();
4544 markOop m = old->mark();
4545 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
4546 : m->age();
4547 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4548 word_sz);
4549 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4550 #ifndef PRODUCT
4551 // Should this evacuation fail?
4552 if (_g1->evacuation_should_fail()) {
4553 if (obj_ptr != NULL) {
4554 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4555 obj_ptr = NULL;
4556 }
4557 }
4558 #endif // !PRODUCT
4559
4560 if (obj_ptr == NULL) {
4561 // This will either forward-to-self, or detect that someone else has
4562 // installed a forwarding pointer.
4563 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4564 return _g1->handle_evacuation_failure_par(cl, old);
4565 }
4566
4567 oop obj = oop(obj_ptr);
4568
4569 // We're going to allocate linearly, so might as well prefetch ahead.
4570 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4571
4572 oop forward_ptr = old->forward_to_atomic(obj);
4573 if (forward_ptr == NULL) {
4574 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4575 if (g1p->track_object_age(alloc_purpose)) {
4576 // We could simply do obj->incr_age(). However, this causes a
4577 // performance issue. obj->incr_age() will first check whether
4578 // the object has a displaced mark by checking its mark word;
4579 // getting the mark word from the new location of the object
4580 // stalls. So, given that we already have the mark word and we
4581 // are about to install it anyway, it's better to increase the
4582 // age on the mark word, when the object does not have a
4583 // displaced mark word. We're not expecting many objects to have
4584 // a displaced marked word, so that case is not optimized
4585 // further (it could be...) and we simply call obj->incr_age().
4586
4587 if (m->has_displaced_mark_helper()) {
4588 // in this case, we have to install the mark word first,
4589 // otherwise obj looks to be forwarded (the old mark word,
4590 // which contains the forward pointer, was copied)
4591 obj->set_mark(m);
4592 obj->incr_age();
4593 } else {
4594 m = m->incr_age();
4595 obj->set_mark(m);
4596 }
4597 _par_scan_state->age_table()->add(obj, word_sz);
4598 } else {
4599 obj->set_mark(m);
4600 }
4601
4602 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4603 surv_young_words[young_index] += word_sz;
4604
4605 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4606 // We keep track of the next start index in the length field of
4607 // the to-space object. The actual length can be found in the
4608 // length field of the from-space object.
4609 arrayOop(obj)->set_length(0);
4610 oop* old_p = set_partial_array_mask(old);
4611 _par_scan_state->push_on_queue(old_p);
4612 } else {
4613 // No point in using the slower heap_region_containing() method,
4614 // given that we know obj is in the heap.
4615 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4616 obj->oop_iterate_backwards(&_scanner);
4617 }
4618 } else {
4619 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4620 obj = forward_ptr;
4621 }
4622 return obj;
4623 }
4624
4625 template <class T>
4626 void G1ParCopyHelper::do_klass_barrier(T* p, oop new_obj) {
4627 if (_g1->heap_region_containing_raw(new_obj)->is_young()) {
4628 _scanned_klass->record_modified_oops();
4629 }
4630 }
4631
4632 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4633 template <class T>
4634 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4635 ::do_oop_work(T* p) {
4636 oop obj = oopDesc::load_decode_heap_oop(p);
4637 assert(barrier != G1BarrierRS || obj != NULL,
4638 "Precondition: G1BarrierRS implies obj is non-NULL");
4639
4640 assert(_worker_id == _par_scan_state->queue_num(), "sanity");
4641
4642 // here the null check is implicit in the cset_fast_test() test
4643 if (_g1->in_cset_fast_test(obj)) {
4644 oop forwardee;
4645 if (obj->is_forwarded()) {
4646 forwardee = obj->forwardee();
4647 } else {
4648 forwardee = copy_to_survivor_space(obj);
4649 }
4650 assert(forwardee != NULL, "forwardee should not be NULL");
4651 oopDesc::encode_store_heap_oop(p, forwardee);
4652 if (do_mark_object && forwardee != obj) {
4653 // If the object is self-forwarded we don't need to explicitly
4654 // mark it, the evacuation failure protocol will do so.
4655 mark_forwarded_object(obj, forwardee);
4656 }
4657
4658 // When scanning the RS, we only care about objs in CS.
4659 if (barrier == G1BarrierRS) {
4660 _par_scan_state->update_rs(_from, p, _worker_id);
4661 } else if (barrier == G1BarrierKlass) {
4662 do_klass_barrier(p, forwardee);
4663 }
4664 } else {
4665 // The object is not in collection set. If we're a root scanning
4666 // closure during an initial mark pause (i.e. do_mark_object will
4667 // be true) then attempt to mark the object.
4668 if (do_mark_object && _g1->is_in_g1_reserved(obj)) {
4669 mark_object(obj);
4670 }
4671 }
4672
4673 if (barrier == G1BarrierEvac && obj != NULL) {
4674 _par_scan_state->update_rs(_from, p, _worker_id);
4675 }
4676
4677 if (do_gen_barrier && obj != NULL) {
4678 par_do_barrier(p);
4679 }
4680 }
4681
4682 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4683 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4684
4685 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4686 assert(has_partial_array_mask(p), "invariant");
4687 oop from_obj = clear_partial_array_mask(p);
4688
4689 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap.");
4690 assert(from_obj->is_objArray(), "must be obj array");
4691 objArrayOop from_obj_array = objArrayOop(from_obj);
4692 // The from-space object contains the real length.
4693 int length = from_obj_array->length();
4694
4695 assert(from_obj->is_forwarded(), "must be forwarded");
4696 oop to_obj = from_obj->forwardee();
4697 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
4698 objArrayOop to_obj_array = objArrayOop(to_obj);
4699 // We keep track of the next start index in the length field of the
4700 // to-space object.
4701 int next_index = to_obj_array->length();
4702 assert(0 <= next_index && next_index < length,
4703 err_msg("invariant, next index: %d, length: %d", next_index, length));
4704
4705 int start = next_index;
4706 int end = length;
4707 int remainder = end - start;
4708 // We'll try not to push a range that's smaller than ParGCArrayScanChunk.
4709 if (remainder > 2 * ParGCArrayScanChunk) {
4710 end = start + ParGCArrayScanChunk;
4711 to_obj_array->set_length(end);
4712 // Push the remainder before we process the range in case another
4713 // worker has run out of things to do and can steal it.
4714 oop* from_obj_p = set_partial_array_mask(from_obj);
4715 _par_scan_state->push_on_queue(from_obj_p);
4716 } else {
4717 assert(length == end, "sanity");
4718 // We'll process the final range for this object. Restore the length
4719 // so that the heap remains parsable in case of evacuation failure.
4720 to_obj_array->set_length(end);
4721 }
4722 _scanner.set_region(_g1->heap_region_containing_raw(to_obj));
4723 // Process indexes [start,end). It will also process the header
4724 // along with the first chunk (i.e., the chunk with start == 0).
4725 // Note that at this point the length field of to_obj_array is not
4726 // correct given that we are using it to keep track of the next
4727 // start index. oop_iterate_range() (thankfully!) ignores the length
4728 // field and only relies on the start / end parameters. It does
4729 // however return the size of the object which will be incorrect. So
4730 // we have to ignore it even if we wanted to use it.
4731 to_obj_array->oop_iterate_range(&_scanner, start, end);
4732 }
4733
4734 class G1ParEvacuateFollowersClosure : public VoidClosure {
4735 protected:
4736 G1CollectedHeap* _g1h;
4737 G1ParScanThreadState* _par_scan_state;
4738 RefToScanQueueSet* _queues;
4739 ParallelTaskTerminator* _terminator;
4740
4741 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4742 RefToScanQueueSet* queues() { return _queues; }
4743 ParallelTaskTerminator* terminator() { return _terminator; }
4744
4745 public:
4746 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4747 G1ParScanThreadState* par_scan_state,
4748 RefToScanQueueSet* queues,
4749 ParallelTaskTerminator* terminator)
4750 : _g1h(g1h), _par_scan_state(par_scan_state),
4751 _queues(queues), _terminator(terminator) {}
4752
4753 void do_void();
4754
4755 private:
4756 inline bool offer_termination();
4757 };
4758
4759 bool G1ParEvacuateFollowersClosure::offer_termination() {
4760 G1ParScanThreadState* const pss = par_scan_state();
4761 pss->start_term_time();
4762 const bool res = terminator()->offer_termination();
4763 pss->end_term_time();
4764 return res;
4765 }
4766
4767 void G1ParEvacuateFollowersClosure::do_void() {
4768 StarTask stolen_task;
4769 G1ParScanThreadState* const pss = par_scan_state();
4770 pss->trim_queue();
4771
4772 do {
4773 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4774 assert(pss->verify_task(stolen_task), "sanity");
4775 if (stolen_task.is_narrow()) {
4776 pss->deal_with_reference((narrowOop*) stolen_task);
4777 } else {
4778 pss->deal_with_reference((oop*) stolen_task);
4779 }
4780
4781 // We've just processed a reference and we might have made
4782 // available new entries on the queues. So we have to make sure
4783 // we drain the queues as necessary.
4784 pss->trim_queue();
4785 }
4786 } while (!offer_termination());
4787
4788 pss->retire_alloc_buffers();
4789 }
4790
4791 class G1KlassScanClosure : public KlassClosure {
4792 G1ParCopyHelper* _closure;
4793 bool _process_only_dirty;
4794 int _count;
4795 public:
4796 G1KlassScanClosure(G1ParCopyHelper* closure, bool process_only_dirty)
4797 : _process_only_dirty(process_only_dirty), _closure(closure), _count(0) {}
4798 void do_klass(Klass* klass) {
4799 // If the klass has not been dirtied we know that there's
4800 // no references into the young gen and we can skip it.
4801 if (!_process_only_dirty || klass->has_modified_oops()) {
4802 // Clean the klass since we're going to scavenge all the metadata.
4803 klass->clear_modified_oops();
4804
4805 // Tell the closure that this klass is the Klass to scavenge
4806 // and is the one to dirty if oops are left pointing into the young gen.
4807 _closure->set_scanned_klass(klass);
4808
4809 klass->oops_do(_closure);
4810
4811 _closure->set_scanned_klass(NULL);
4812 }
4813 _count++;
4814 }
4815 };
4816
4817 class G1ParTask : public AbstractGangTask {
4818 protected:
4819 G1CollectedHeap* _g1h;
4820 RefToScanQueueSet *_queues;
4821 ParallelTaskTerminator _terminator;
4822 uint _n_workers;
4823
4824 Mutex _stats_lock;
4825 Mutex* stats_lock() { return &_stats_lock; }
4826
4827 size_t getNCards() {
4828 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4829 / G1BlockOffsetSharedArray::N_bytes;
4830 }
4831
4832 public:
4833 G1ParTask(G1CollectedHeap* g1h,
4834 RefToScanQueueSet *task_queues)
4835 : AbstractGangTask("G1 collection"),
4836 _g1h(g1h),
4837 _queues(task_queues),
4838 _terminator(0, _queues),
4839 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true)
4840 {}
4841
4842 RefToScanQueueSet* queues() { return _queues; }
4843
4844 RefToScanQueue *work_queue(int i) {
4845 return queues()->queue(i);
4846 }
4847
4848 ParallelTaskTerminator* terminator() { return &_terminator; }
4849
4850 virtual void set_for_termination(int active_workers) {
4851 // This task calls set_n_termination() in par_non_clean_card_iterate_work()
4852 // in the young space (_par_seq_tasks) in the G1 heap
4853 // for SequentialSubTasksDone.
4854 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap
4855 // both of which need setting by set_n_termination().
4856 _g1h->SharedHeap::set_n_termination(active_workers);
4857 _g1h->set_n_termination(active_workers);
4858 terminator()->reset_for_reuse(active_workers);
4859 _n_workers = active_workers;
4860 }
4861
4862 void work(uint worker_id) {
4863 if (worker_id >= _n_workers) return; // no work needed this round
4864
4865 double start_time_ms = os::elapsedTime() * 1000.0;
4866 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms);
4867
4868 {
4869 ResourceMark rm;
4870 HandleMark hm;
4871
4872 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4873
4874 G1ParScanThreadState pss(_g1h, worker_id);
4875 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4876 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4877 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4878
4879 pss.set_evac_closure(&scan_evac_cl);
4880 pss.set_evac_failure_closure(&evac_failure_cl);
4881 pss.set_partial_scan_closure(&partial_scan_cl);
4882
4883 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4884 G1ParScanMetadataClosure only_scan_metadata_cl(_g1h, &pss, rp);
4885
4886 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4887 G1ParScanAndMarkMetadataClosure scan_mark_metadata_cl(_g1h, &pss, rp);
4888
4889 bool only_young = _g1h->g1_policy()->gcs_are_young();
4890 G1KlassScanClosure scan_mark_klasses_cl_s(&scan_mark_metadata_cl, false);
4891 G1KlassScanClosure only_scan_klasses_cl_s(&only_scan_metadata_cl, only_young);
4892
4893 OopClosure* scan_root_cl = &only_scan_root_cl;
4894 G1KlassScanClosure* scan_klasses_cl = &only_scan_klasses_cl_s;
4895
4896 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4897 // We also need to mark copied objects.
4898 scan_root_cl = &scan_mark_root_cl;
4899 scan_klasses_cl = &scan_mark_klasses_cl_s;
4900 }
4901
4902 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4903
4904 int so = SharedHeap::SO_AllClasses | SharedHeap::SO_Strings | SharedHeap::SO_CodeCache;
4905
4906 pss.start_strong_roots();
4907 _g1h->g1_process_strong_roots(/* is scavenging */ true,
4908 SharedHeap::ScanningOption(so),
4909 scan_root_cl,
4910 &push_heap_rs_cl,
4911 scan_klasses_cl,
4912 worker_id);
4913 pss.end_strong_roots();
4914
4915 {
4916 double start = os::elapsedTime();
4917 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4918 evac.do_void();
4919 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4920 double term_ms = pss.term_time()*1000.0;
4921 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms);
4922 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts());
4923 }
4924 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4925 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4926
4927 if (ParallelGCVerbose) {
4928 MutexLocker x(stats_lock());
4929 pss.print_termination_stats(worker_id);
4930 }
4931
4932 assert(pss.refs()->is_empty(), "should be empty");
4933
4934 // Close the inner scope so that the ResourceMark and HandleMark
4935 // destructors are executed here and are included as part of the
4936 // "GC Worker Time".
4937 }
4938
4939 double end_time_ms = os::elapsedTime() * 1000.0;
4940 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms);
4941 }
4942 };
4943
4944 // *** Common G1 Evacuation Stuff
4945
4946 // Closures that support the filtering of CodeBlobs scanned during
4947 // external root scanning.
4948
4949 // Closure applied to reference fields in code blobs (specifically nmethods)
4950 // to determine whether an nmethod contains references that point into
4951 // the collection set. Used as a predicate when walking code roots so
4952 // that only nmethods that point into the collection set are added to the
4953 // 'marked' list.
4954
4955 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure {
4956
4957 class G1PointsIntoCSOopClosure : public OopClosure {
4958 G1CollectedHeap* _g1;
4959 bool _points_into_cs;
4960 public:
4961 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) :
4962 _g1(g1), _points_into_cs(false) { }
4963
4964 bool points_into_cs() const { return _points_into_cs; }
4965
4966 template <class T>
4967 void do_oop_nv(T* p) {
4968 if (!_points_into_cs) {
4969 T heap_oop = oopDesc::load_heap_oop(p);
4970 if (!oopDesc::is_null(heap_oop) &&
4971 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) {
4972 _points_into_cs = true;
4973 }
4974 }
4975 }
4976
4977 virtual void do_oop(oop* p) { do_oop_nv(p); }
4978 virtual void do_oop(narrowOop* p) { do_oop_nv(p); }
4979 };
4980
4981 G1CollectedHeap* _g1;
4982
4983 public:
4984 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) :
4985 CodeBlobToOopClosure(cl, true), _g1(g1) { }
4986
4987 virtual void do_code_blob(CodeBlob* cb) {
4988 nmethod* nm = cb->as_nmethod_or_null();
4989 if (nm != NULL && !(nm->test_oops_do_mark())) {
4990 G1PointsIntoCSOopClosure predicate_cl(_g1);
4991 nm->oops_do(&predicate_cl);
4992
4993 if (predicate_cl.points_into_cs()) {
4994 // At least one of the reference fields or the oop relocations
4995 // in the nmethod points into the collection set. We have to
4996 // 'mark' this nmethod.
4997 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob()
4998 // or MarkingCodeBlobClosure::do_code_blob() change.
4999 if (!nm->test_set_oops_do_mark()) {
5000 do_newly_marked_nmethod(nm);
5001 }
5002 }
5003 }
5004 }
5005 };
5006
5007 // This method is run in a GC worker.
5008
5009 void
5010 G1CollectedHeap::
5011 g1_process_strong_roots(bool is_scavenging,
5012 ScanningOption so,
5013 OopClosure* scan_non_heap_roots,
5014 OopsInHeapRegionClosure* scan_rs,
5015 G1KlassScanClosure* scan_klasses,
5016 int worker_i) {
5017
5018 // First scan the strong roots
5019 double ext_roots_start = os::elapsedTime();
5020 double closure_app_time_sec = 0.0;
5021
5022 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
5023
5024 // Walk the code cache w/o buffering, because StarTask cannot handle
5025 // unaligned oop locations.
5026 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots);
5027
5028 process_strong_roots(false, // no scoping; this is parallel code
5029 is_scavenging, so,
5030 &buf_scan_non_heap_roots,
5031 &eager_scan_code_roots,
5032 scan_klasses
5033 );
5034
5035 // Now the CM ref_processor roots.
5036 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
5037 // We need to treat the discovered reference lists of the
5038 // concurrent mark ref processor as roots and keep entries
5039 // (which are added by the marking threads) on them live
5040 // until they can be processed at the end of marking.
5041 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
5042 }
5043
5044 // Finish up any enqueued closure apps (attributed as object copy time).
5045 buf_scan_non_heap_roots.done();
5046
5047 double obj_copy_time_sec = buf_scan_non_heap_roots.closure_app_seconds();
5048
5049 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
5050
5051 double ext_root_time_ms =
5052 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0;
5053
5054 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
5055
5056 // During conc marking we have to filter the per-thread SATB buffers
5057 // to make sure we remove any oops into the CSet (which will show up
5058 // as implicitly live).
5059 double satb_filtering_ms = 0.0;
5060 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) {
5061 if (mark_in_progress()) {
5062 double satb_filter_start = os::elapsedTime();
5063
5064 JavaThread::satb_mark_queue_set().filter_thread_buffers();
5065
5066 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0;
5067 }
5068 }
5069 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms);
5070
5071 // Now scan the complement of the collection set.
5072 if (scan_rs != NULL) {
5073 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
5074 }
5075 _process_strong_tasks->all_tasks_completed();
5076 }
5077
5078 void
5079 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure) {
5080 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
5081 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs);
5082 }
5083
5084 // Weak Reference Processing support
5085
5086 // An always "is_alive" closure that is used to preserve referents.
5087 // If the object is non-null then it's alive. Used in the preservation
5088 // of referent objects that are pointed to by reference objects
5089 // discovered by the CM ref processor.
5090 class G1AlwaysAliveClosure: public BoolObjectClosure {
5091 G1CollectedHeap* _g1;
5092 public:
5093 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5094 bool do_object_b(oop p) {
5095 if (p != NULL) {
5096 return true;
5097 }
5098 return false;
5099 }
5100 };
5101
5102 bool G1STWIsAliveClosure::do_object_b(oop p) {
5103 // An object is reachable if it is outside the collection set,
5104 // or is inside and copied.
5105 return !_g1->obj_in_cs(p) || p->is_forwarded();
5106 }
5107
5108 // Non Copying Keep Alive closure
5109 class G1KeepAliveClosure: public OopClosure {
5110 G1CollectedHeap* _g1;
5111 public:
5112 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5113 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
5114 void do_oop( oop* p) {
5115 oop obj = *p;
5116
5117 if (_g1->obj_in_cs(obj)) {
5118 assert( obj->is_forwarded(), "invariant" );
5119 *p = obj->forwardee();
5120 }
5121 }
5122 };
5123
5124 // Copying Keep Alive closure - can be called from both
5125 // serial and parallel code as long as different worker
5126 // threads utilize different G1ParScanThreadState instances
5127 // and different queues.
5128
5129 class G1CopyingKeepAliveClosure: public OopClosure {
5130 G1CollectedHeap* _g1h;
5131 OopClosure* _copy_non_heap_obj_cl;
5132 OopsInHeapRegionClosure* _copy_metadata_obj_cl;
5133 G1ParScanThreadState* _par_scan_state;
5134
5135 public:
5136 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
5137 OopClosure* non_heap_obj_cl,
5138 OopsInHeapRegionClosure* metadata_obj_cl,
5139 G1ParScanThreadState* pss):
5140 _g1h(g1h),
5141 _copy_non_heap_obj_cl(non_heap_obj_cl),
5142 _copy_metadata_obj_cl(metadata_obj_cl),
5143 _par_scan_state(pss)
5144 {}
5145
5146 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
5147 virtual void do_oop( oop* p) { do_oop_work(p); }
5148
5149 template <class T> void do_oop_work(T* p) {
5150 oop obj = oopDesc::load_decode_heap_oop(p);
5151
5152 if (_g1h->obj_in_cs(obj)) {
5153 // If the referent object has been forwarded (either copied
5154 // to a new location or to itself in the event of an
5155 // evacuation failure) then we need to update the reference
5156 // field and, if both reference and referent are in the G1
5157 // heap, update the RSet for the referent.
5158 //
5159 // If the referent has not been forwarded then we have to keep
5160 // it alive by policy. Therefore we have copy the referent.
5161 //
5162 // If the reference field is in the G1 heap then we can push
5163 // on the PSS queue. When the queue is drained (after each
5164 // phase of reference processing) the object and it's followers
5165 // will be copied, the reference field set to point to the
5166 // new location, and the RSet updated. Otherwise we need to
5167 // use the the non-heap or metadata closures directly to copy
5168 // the refernt object and update the pointer, while avoiding
5169 // updating the RSet.
5170
5171 if (_g1h->is_in_g1_reserved(p)) {
5172 _par_scan_state->push_on_queue(p);
5173 } else {
5174 assert(!ClassLoaderDataGraph::contains((address)p),
5175 err_msg("Otherwise need to call _copy_metadata_obj_cl->do_oop(p) "
5176 PTR_FORMAT, p));
5177 _copy_non_heap_obj_cl->do_oop(p);
5178 }
5179 }
5180 }
5181 };
5182
5183 // Serial drain queue closure. Called as the 'complete_gc'
5184 // closure for each discovered list in some of the
5185 // reference processing phases.
5186
5187 class G1STWDrainQueueClosure: public VoidClosure {
5188 protected:
5189 G1CollectedHeap* _g1h;
5190 G1ParScanThreadState* _par_scan_state;
5191
5192 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
5193
5194 public:
5195 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
5196 _g1h(g1h),
5197 _par_scan_state(pss)
5198 { }
5199
5200 void do_void() {
5201 G1ParScanThreadState* const pss = par_scan_state();
5202 pss->trim_queue();
5203 }
5204 };
5205
5206 // Parallel Reference Processing closures
5207
5208 // Implementation of AbstractRefProcTaskExecutor for parallel reference
5209 // processing during G1 evacuation pauses.
5210
5211 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
5212 private:
5213 G1CollectedHeap* _g1h;
5214 RefToScanQueueSet* _queues;
5215 FlexibleWorkGang* _workers;
5216 int _active_workers;
5217
5218 public:
5219 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
5220 FlexibleWorkGang* workers,
5221 RefToScanQueueSet *task_queues,
5222 int n_workers) :
5223 _g1h(g1h),
5224 _queues(task_queues),
5225 _workers(workers),
5226 _active_workers(n_workers)
5227 {
5228 assert(n_workers > 0, "shouldn't call this otherwise");
5229 }
5230
5231 // Executes the given task using concurrent marking worker threads.
5232 virtual void execute(ProcessTask& task);
5233 virtual void execute(EnqueueTask& task);
5234 };
5235
5236 // Gang task for possibly parallel reference processing
5237
5238 class G1STWRefProcTaskProxy: public AbstractGangTask {
5239 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5240 ProcessTask& _proc_task;
5241 G1CollectedHeap* _g1h;
5242 RefToScanQueueSet *_task_queues;
5243 ParallelTaskTerminator* _terminator;
5244
5245 public:
5246 G1STWRefProcTaskProxy(ProcessTask& proc_task,
5247 G1CollectedHeap* g1h,
5248 RefToScanQueueSet *task_queues,
5249 ParallelTaskTerminator* terminator) :
5250 AbstractGangTask("Process reference objects in parallel"),
5251 _proc_task(proc_task),
5252 _g1h(g1h),
5253 _task_queues(task_queues),
5254 _terminator(terminator)
5255 {}
5256
5257 virtual void work(uint worker_id) {
5258 // The reference processing task executed by a single worker.
5259 ResourceMark rm;
5260 HandleMark hm;
5261
5262 G1STWIsAliveClosure is_alive(_g1h);
5263
5264 G1ParScanThreadState pss(_g1h, worker_id);
5265
5266 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5267 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5268 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5269
5270 pss.set_evac_closure(&scan_evac_cl);
5271 pss.set_evac_failure_closure(&evac_failure_cl);
5272 pss.set_partial_scan_closure(&partial_scan_cl);
5273
5274 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5275 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5276
5277 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5278 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5279
5280 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5281 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5282
5283 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5284 // We also need to mark copied objects.
5285 copy_non_heap_cl = ©_mark_non_heap_cl;
5286 copy_metadata_cl = ©_mark_metadata_cl;
5287 }
5288
5289 // Keep alive closure.
5290 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5291
5292 // Complete GC closure
5293 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
5294
5295 // Call the reference processing task's work routine.
5296 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue);
5297
5298 // Note we cannot assert that the refs array is empty here as not all
5299 // of the processing tasks (specifically phase2 - pp2_work) execute
5300 // the complete_gc closure (which ordinarily would drain the queue) so
5301 // the queue may not be empty.
5302 }
5303 };
5304
5305 // Driver routine for parallel reference processing.
5306 // Creates an instance of the ref processing gang
5307 // task and has the worker threads execute it.
5308 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
5309 assert(_workers != NULL, "Need parallel worker threads.");
5310
5311 ParallelTaskTerminator terminator(_active_workers, _queues);
5312 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
5313
5314 _g1h->set_par_threads(_active_workers);
5315 _workers->run_task(&proc_task_proxy);
5316 _g1h->set_par_threads(0);
5317 }
5318
5319 // Gang task for parallel reference enqueueing.
5320
5321 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
5322 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5323 EnqueueTask& _enq_task;
5324
5325 public:
5326 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
5327 AbstractGangTask("Enqueue reference objects in parallel"),
5328 _enq_task(enq_task)
5329 { }
5330
5331 virtual void work(uint worker_id) {
5332 _enq_task.work(worker_id);
5333 }
5334 };
5335
5336 // Driver routine for parallel reference enqueing.
5337 // Creates an instance of the ref enqueueing gang
5338 // task and has the worker threads execute it.
5339
5340 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5341 assert(_workers != NULL, "Need parallel worker threads.");
5342
5343 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5344
5345 _g1h->set_par_threads(_active_workers);
5346 _workers->run_task(&enq_task_proxy);
5347 _g1h->set_par_threads(0);
5348 }
5349
5350 // End of weak reference support closures
5351
5352 // Abstract task used to preserve (i.e. copy) any referent objects
5353 // that are in the collection set and are pointed to by reference
5354 // objects discovered by the CM ref processor.
5355
5356 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5357 protected:
5358 G1CollectedHeap* _g1h;
5359 RefToScanQueueSet *_queues;
5360 ParallelTaskTerminator _terminator;
5361 uint _n_workers;
5362
5363 public:
5364 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5365 AbstractGangTask("ParPreserveCMReferents"),
5366 _g1h(g1h),
5367 _queues(task_queues),
5368 _terminator(workers, _queues),
5369 _n_workers(workers)
5370 { }
5371
5372 void work(uint worker_id) {
5373 ResourceMark rm;
5374 HandleMark hm;
5375
5376 G1ParScanThreadState pss(_g1h, worker_id);
5377 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5378 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5379 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5380
5381 pss.set_evac_closure(&scan_evac_cl);
5382 pss.set_evac_failure_closure(&evac_failure_cl);
5383 pss.set_partial_scan_closure(&partial_scan_cl);
5384
5385 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5386
5387
5388 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5389 G1ParScanMetadataClosure only_copy_metadata_cl(_g1h, &pss, NULL);
5390
5391 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5392 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(_g1h, &pss, NULL);
5393
5394 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5395 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5396
5397 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5398 // We also need to mark copied objects.
5399 copy_non_heap_cl = ©_mark_non_heap_cl;
5400 copy_metadata_cl = ©_mark_metadata_cl;
5401 }
5402
5403 // Is alive closure
5404 G1AlwaysAliveClosure always_alive(_g1h);
5405
5406 // Copying keep alive closure. Applied to referent objects that need
5407 // to be copied.
5408 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_metadata_cl, &pss);
5409
5410 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5411
5412 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5413 uint stride = MIN2(MAX2(_n_workers, 1U), limit);
5414
5415 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5416 // So this must be true - but assert just in case someone decides to
5417 // change the worker ids.
5418 assert(0 <= worker_id && worker_id < limit, "sanity");
5419 assert(!rp->discovery_is_atomic(), "check this code");
5420
5421 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5422 for (uint idx = worker_id; idx < limit; idx += stride) {
5423 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5424
5425 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5426 while (iter.has_next()) {
5427 // Since discovery is not atomic for the CM ref processor, we
5428 // can see some null referent objects.
5429 iter.load_ptrs(DEBUG_ONLY(true));
5430 oop ref = iter.obj();
5431
5432 // This will filter nulls.
5433 if (iter.is_referent_alive()) {
5434 iter.make_referent_alive();
5435 }
5436 iter.move_to_next();
5437 }
5438 }
5439
5440 // Drain the queue - which may cause stealing
5441 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5442 drain_queue.do_void();
5443 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5444 assert(pss.refs()->is_empty(), "should be");
5445 }
5446 };
5447
5448 // Weak Reference processing during an evacuation pause (part 1).
5449 void G1CollectedHeap::process_discovered_references(uint no_of_gc_workers) {
5450 double ref_proc_start = os::elapsedTime();
5451
5452 ReferenceProcessor* rp = _ref_processor_stw;
5453 assert(rp->discovery_enabled(), "should have been enabled");
5454
5455 // Any reference objects, in the collection set, that were 'discovered'
5456 // by the CM ref processor should have already been copied (either by
5457 // applying the external root copy closure to the discovered lists, or
5458 // by following an RSet entry).
5459 //
5460 // But some of the referents, that are in the collection set, that these
5461 // reference objects point to may not have been copied: the STW ref
5462 // processor would have seen that the reference object had already
5463 // been 'discovered' and would have skipped discovering the reference,
5464 // but would not have treated the reference object as a regular oop.
5465 // As a reult the copy closure would not have been applied to the
5466 // referent object.
5467 //
5468 // We need to explicitly copy these referent objects - the references
5469 // will be processed at the end of remarking.
5470 //
5471 // We also need to do this copying before we process the reference
5472 // objects discovered by the STW ref processor in case one of these
5473 // referents points to another object which is also referenced by an
5474 // object discovered by the STW ref processor.
5475
5476 assert(!G1CollectedHeap::use_parallel_gc_threads() ||
5477 no_of_gc_workers == workers()->active_workers(),
5478 "Need to reset active GC workers");
5479
5480 set_par_threads(no_of_gc_workers);
5481 G1ParPreserveCMReferentsTask keep_cm_referents(this,
5482 no_of_gc_workers,
5483 _task_queues);
5484
5485 if (G1CollectedHeap::use_parallel_gc_threads()) {
5486 workers()->run_task(&keep_cm_referents);
5487 } else {
5488 keep_cm_referents.work(0);
5489 }
5490
5491 set_par_threads(0);
5492
5493 // Closure to test whether a referent is alive.
5494 G1STWIsAliveClosure is_alive(this);
5495
5496 // Even when parallel reference processing is enabled, the processing
5497 // of JNI refs is serial and performed serially by the current thread
5498 // rather than by a worker. The following PSS will be used for processing
5499 // JNI refs.
5500
5501 // Use only a single queue for this PSS.
5502 G1ParScanThreadState pss(this, 0);
5503
5504 // We do not embed a reference processor in the copying/scanning
5505 // closures while we're actually processing the discovered
5506 // reference objects.
5507 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5508 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5509 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5510
5511 pss.set_evac_closure(&scan_evac_cl);
5512 pss.set_evac_failure_closure(&evac_failure_cl);
5513 pss.set_partial_scan_closure(&partial_scan_cl);
5514
5515 assert(pss.refs()->is_empty(), "pre-condition");
5516
5517 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5518 G1ParScanMetadataClosure only_copy_metadata_cl(this, &pss, NULL);
5519
5520 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5521 G1ParScanAndMarkMetadataClosure copy_mark_metadata_cl(this, &pss, NULL);
5522
5523 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5524 OopsInHeapRegionClosure* copy_metadata_cl = &only_copy_metadata_cl;
5525
5526 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5527 // We also need to mark copied objects.
5528 copy_non_heap_cl = ©_mark_non_heap_cl;
5529 copy_metadata_cl = ©_mark_metadata_cl;
5530 }
5531
5532 // Keep alive closure.
5533 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_metadata_cl, &pss);
5534
5535 // Serial Complete GC closure
5536 G1STWDrainQueueClosure drain_queue(this, &pss);
5537
5538 // Setup the soft refs policy...
5539 rp->setup_policy(false);
5540
5541 if (!rp->processing_is_mt()) {
5542 // Serial reference processing...
5543 rp->process_discovered_references(&is_alive,
5544 &keep_alive,
5545 &drain_queue,
5546 NULL);
5547 } else {
5548 // Parallel reference processing
5549 assert(rp->num_q() == no_of_gc_workers, "sanity");
5550 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5551
5552 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5553 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5554 }
5555
5556 // We have completed copying any necessary live referent objects
5557 // (that were not copied during the actual pause) so we can
5558 // retire any active alloc buffers
5559 pss.retire_alloc_buffers();
5560 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5561
5562 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5563 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0);
5564 }
5565
5566 // Weak Reference processing during an evacuation pause (part 2).
5567 void G1CollectedHeap::enqueue_discovered_references(uint no_of_gc_workers) {
5568 double ref_enq_start = os::elapsedTime();
5569
5570 ReferenceProcessor* rp = _ref_processor_stw;
5571 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5572
5573 // Now enqueue any remaining on the discovered lists on to
5574 // the pending list.
5575 if (!rp->processing_is_mt()) {
5576 // Serial reference processing...
5577 rp->enqueue_discovered_references();
5578 } else {
5579 // Parallel reference enqueuing
5580
5581 assert(no_of_gc_workers == workers()->active_workers(),
5582 "Need to reset active workers");
5583 assert(rp->num_q() == no_of_gc_workers, "sanity");
5584 assert(no_of_gc_workers <= rp->max_num_q(), "sanity");
5585
5586 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, no_of_gc_workers);
5587 rp->enqueue_discovered_references(&par_task_executor);
5588 }
5589
5590 rp->verify_no_references_recorded();
5591 assert(!rp->discovery_enabled(), "should have been disabled");
5592
5593 // FIXME
5594 // CM's reference processing also cleans up the string and symbol tables.
5595 // Should we do that here also? We could, but it is a serial operation
5596 // and could signicantly increase the pause time.
5597
5598 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5599 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0);
5600 }
5601
5602 void G1CollectedHeap::evacuate_collection_set() {
5603 _expand_heap_after_alloc_failure = true;
5604 set_evacuation_failed(false);
5605
5606 // Should G1EvacuationFailureALot be in effect for this GC?
5607 NOT_PRODUCT(set_evacuation_failure_alot_for_current_gc();)
5608
5609 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5610
5611 // Disable the hot card cache.
5612 G1HotCardCache* hot_card_cache = _cg1r->hot_card_cache();
5613 hot_card_cache->reset_hot_cache_claimed_index();
5614 hot_card_cache->set_use_cache(false);
5615
5616 uint n_workers;
5617 if (G1CollectedHeap::use_parallel_gc_threads()) {
5618 n_workers =
5619 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
5620 workers()->active_workers(),
5621 Threads::number_of_non_daemon_threads());
5622 assert(UseDynamicNumberOfGCThreads ||
5623 n_workers == workers()->total_workers(),
5624 "If not dynamic should be using all the workers");
5625 workers()->set_active_workers(n_workers);
5626 set_par_threads(n_workers);
5627 } else {
5628 assert(n_par_threads() == 0,
5629 "Should be the original non-parallel value");
5630 n_workers = 1;
5631 }
5632
5633 G1ParTask g1_par_task(this, _task_queues);
5634
5635 init_for_evac_failure(NULL);
5636
5637 rem_set()->prepare_for_younger_refs_iterate(true);
5638
5639 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5640 double start_par_time_sec = os::elapsedTime();
5641 double end_par_time_sec;
5642
5643 {
5644 StrongRootsScope srs(this);
5645
5646 if (G1CollectedHeap::use_parallel_gc_threads()) {
5647 // The individual threads will set their evac-failure closures.
5648 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5649 // These tasks use ShareHeap::_process_strong_tasks
5650 assert(UseDynamicNumberOfGCThreads ||
5651 workers()->active_workers() == workers()->total_workers(),
5652 "If not dynamic should be using all the workers");
5653 workers()->run_task(&g1_par_task);
5654 } else {
5655 g1_par_task.set_for_termination(n_workers);
5656 g1_par_task.work(0);
5657 }
5658 end_par_time_sec = os::elapsedTime();
5659
5660 // Closing the inner scope will execute the destructor
5661 // for the StrongRootsScope object. We record the current
5662 // elapsed time before closing the scope so that time
5663 // taken for the SRS destructor is NOT included in the
5664 // reported parallel time.
5665 }
5666
5667 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0;
5668 g1_policy()->phase_times()->record_par_time(par_time_ms);
5669
5670 double code_root_fixup_time_ms =
5671 (os::elapsedTime() - end_par_time_sec) * 1000.0;
5672 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms);
5673
5674 set_par_threads(0);
5675
5676 // Process any discovered reference objects - we have
5677 // to do this _before_ we retire the GC alloc regions
5678 // as we may have to copy some 'reachable' referent
5679 // objects (and their reachable sub-graphs) that were
5680 // not copied during the pause.
5681 process_discovered_references(n_workers);
5682
5683 // Weak root processing.
5684 {
5685 G1STWIsAliveClosure is_alive(this);
5686 G1KeepAliveClosure keep_alive(this);
5687 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5688 }
5689
5690 release_gc_alloc_regions(n_workers);
5691 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5692
5693 // Reset and re-enable the hot card cache.
5694 // Note the counts for the cards in the regions in the
5695 // collection set are reset when the collection set is freed.
5696 hot_card_cache->reset_hot_cache();
5697 hot_card_cache->set_use_cache(true);
5698
5699 finalize_for_evac_failure();
5700
5701 if (evacuation_failed()) {
5702 remove_self_forwarding_pointers();
5703
5704 // Reset the G1EvacuationFailureALot counters and flags
5705 // Note: the values are reset only when an actual
5706 // evacuation failure occurs.
5707 NOT_PRODUCT(reset_evacuation_should_fail();)
5708 }
5709
5710 // Enqueue any remaining references remaining on the STW
5711 // reference processor's discovered lists. We need to do
5712 // this after the card table is cleaned (and verified) as
5713 // the act of enqueuing entries on to the pending list
5714 // will log these updates (and dirty their associated
5715 // cards). We need these updates logged to update any
5716 // RSets.
5717 enqueue_discovered_references(n_workers);
5718
5719 if (G1DeferredRSUpdate) {
5720 RedirtyLoggedCardTableEntryFastClosure redirty;
5721 dirty_card_queue_set().set_closure(&redirty);
5722 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5723
5724 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5725 dcq.merge_bufferlists(&dirty_card_queue_set());
5726 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5727 }
5728 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5729 }
5730
5731 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5732 size_t* pre_used,
5733 FreeRegionList* free_list,
5734 OldRegionSet* old_proxy_set,
5735 HumongousRegionSet* humongous_proxy_set,
5736 HRRSCleanupTask* hrrs_cleanup_task,
5737 bool par) {
5738 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5739 if (hr->isHumongous()) {
5740 assert(hr->startsHumongous(), "we should only see starts humongous");
5741 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5742 } else {
5743 _old_set.remove_with_proxy(hr, old_proxy_set);
5744 free_region(hr, pre_used, free_list, par);
5745 }
5746 } else {
5747 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5748 }
5749 }
5750
5751 void G1CollectedHeap::free_region(HeapRegion* hr,
5752 size_t* pre_used,
5753 FreeRegionList* free_list,
5754 bool par) {
5755 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5756 assert(!hr->is_empty(), "the region should not be empty");
5757 assert(free_list != NULL, "pre-condition");
5758
5759 // Clear the card counts for this region.
5760 // Note: we only need to do this if the region is not young
5761 // (since we don't refine cards in young regions).
5762 if (!hr->is_young()) {
5763 _cg1r->hot_card_cache()->reset_card_counts(hr);
5764 }
5765 *pre_used += hr->used();
5766 hr->hr_clear(par, true /* clear_space */);
5767 free_list->add_as_head(hr);
5768 }
5769
5770 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5771 size_t* pre_used,
5772 FreeRegionList* free_list,
5773 HumongousRegionSet* humongous_proxy_set,
5774 bool par) {
5775 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5776 assert(free_list != NULL, "pre-condition");
5777 assert(humongous_proxy_set != NULL, "pre-condition");
5778
5779 size_t hr_used = hr->used();
5780 size_t hr_capacity = hr->capacity();
5781 size_t hr_pre_used = 0;
5782 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5783 // We need to read this before we make the region non-humongous,
5784 // otherwise the information will be gone.
5785 uint last_index = hr->last_hc_index();
5786 hr->set_notHumongous();
5787 free_region(hr, &hr_pre_used, free_list, par);
5788
5789 uint i = hr->hrs_index() + 1;
5790 while (i < last_index) {
5791 HeapRegion* curr_hr = region_at(i);
5792 assert(curr_hr->continuesHumongous(), "invariant");
5793 curr_hr->set_notHumongous();
5794 free_region(curr_hr, &hr_pre_used, free_list, par);
5795 i += 1;
5796 }
5797 assert(hr_pre_used == hr_used,
5798 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5799 "should be the same", hr_pre_used, hr_used));
5800 *pre_used += hr_pre_used;
5801 }
5802
5803 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5804 FreeRegionList* free_list,
5805 OldRegionSet* old_proxy_set,
5806 HumongousRegionSet* humongous_proxy_set,
5807 bool par) {
5808 if (pre_used > 0) {
5809 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5810 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5811 assert(_summary_bytes_used >= pre_used,
5812 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5813 "should be >= pre_used: "SIZE_FORMAT,
5814 _summary_bytes_used, pre_used));
5815 _summary_bytes_used -= pre_used;
5816 }
5817 if (free_list != NULL && !free_list->is_empty()) {
5818 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5819 _free_list.add_as_head(free_list);
5820 }
5821 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) {
5822 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5823 _old_set.update_from_proxy(old_proxy_set);
5824 }
5825 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5826 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5827 _humongous_set.update_from_proxy(humongous_proxy_set);
5828 }
5829 }
5830
5831 class G1ParCleanupCTTask : public AbstractGangTask {
5832 CardTableModRefBS* _ct_bs;
5833 G1CollectedHeap* _g1h;
5834 HeapRegion* volatile _su_head;
5835 public:
5836 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5837 G1CollectedHeap* g1h) :
5838 AbstractGangTask("G1 Par Cleanup CT Task"),
5839 _ct_bs(ct_bs), _g1h(g1h) { }
5840
5841 void work(uint worker_id) {
5842 HeapRegion* r;
5843 while (r = _g1h->pop_dirty_cards_region()) {
5844 clear_cards(r);
5845 }
5846 }
5847
5848 void clear_cards(HeapRegion* r) {
5849 // Cards of the survivors should have already been dirtied.
5850 if (!r->is_survivor()) {
5851 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5852 }
5853 }
5854 };
5855
5856 #ifndef PRODUCT
5857 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5858 G1CollectedHeap* _g1h;
5859 CardTableModRefBS* _ct_bs;
5860 public:
5861 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5862 : _g1h(g1h), _ct_bs(ct_bs) { }
5863 virtual bool doHeapRegion(HeapRegion* r) {
5864 if (r->is_survivor()) {
5865 _g1h->verify_dirty_region(r);
5866 } else {
5867 _g1h->verify_not_dirty_region(r);
5868 }
5869 return false;
5870 }
5871 };
5872
5873 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5874 // All of the region should be clean.
5875 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5876 MemRegion mr(hr->bottom(), hr->end());
5877 ct_bs->verify_not_dirty_region(mr);
5878 }
5879
5880 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5881 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5882 // dirty allocated blocks as they allocate them. The thread that
5883 // retires each region and replaces it with a new one will do a
5884 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5885 // not dirty that area (one less thing to have to do while holding
5886 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5887 // is dirty.
5888 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5889 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5890 ct_bs->verify_dirty_region(mr);
5891 }
5892
5893 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5894 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5895 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5896 verify_dirty_region(hr);
5897 }
5898 }
5899
5900 void G1CollectedHeap::verify_dirty_young_regions() {
5901 verify_dirty_young_list(_young_list->first_region());
5902 }
5903 #endif
5904
5905 void G1CollectedHeap::cleanUpCardTable() {
5906 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5907 double start = os::elapsedTime();
5908
5909 {
5910 // Iterate over the dirty cards region list.
5911 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5912
5913 if (G1CollectedHeap::use_parallel_gc_threads()) {
5914 set_par_threads();
5915 workers()->run_task(&cleanup_task);
5916 set_par_threads(0);
5917 } else {
5918 while (_dirty_cards_region_list) {
5919 HeapRegion* r = _dirty_cards_region_list;
5920 cleanup_task.clear_cards(r);
5921 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5922 if (_dirty_cards_region_list == r) {
5923 // The last region.
5924 _dirty_cards_region_list = NULL;
5925 }
5926 r->set_next_dirty_cards_region(NULL);
5927 }
5928 }
5929 #ifndef PRODUCT
5930 if (G1VerifyCTCleanup || VerifyAfterGC) {
5931 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5932 heap_region_iterate(&cleanup_verifier);
5933 }
5934 #endif
5935 }
5936
5937 double elapsed = os::elapsedTime() - start;
5938 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0);
5939 }
5940
5941 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5942 size_t pre_used = 0;
5943 FreeRegionList local_free_list("Local List for CSet Freeing");
5944
5945 double young_time_ms = 0.0;
5946 double non_young_time_ms = 0.0;
5947
5948 // Since the collection set is a superset of the the young list,
5949 // all we need to do to clear the young list is clear its
5950 // head and length, and unlink any young regions in the code below
5951 _young_list->clear();
5952
5953 G1CollectorPolicy* policy = g1_policy();
5954
5955 double start_sec = os::elapsedTime();
5956 bool non_young = true;
5957
5958 HeapRegion* cur = cs_head;
5959 int age_bound = -1;
5960 size_t rs_lengths = 0;
5961
5962 while (cur != NULL) {
5963 assert(!is_on_master_free_list(cur), "sanity");
5964 if (non_young) {
5965 if (cur->is_young()) {
5966 double end_sec = os::elapsedTime();
5967 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5968 non_young_time_ms += elapsed_ms;
5969
5970 start_sec = os::elapsedTime();
5971 non_young = false;
5972 }
5973 } else {
5974 if (!cur->is_young()) {
5975 double end_sec = os::elapsedTime();
5976 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5977 young_time_ms += elapsed_ms;
5978
5979 start_sec = os::elapsedTime();
5980 non_young = true;
5981 }
5982 }
5983
5984 rs_lengths += cur->rem_set()->occupied();
5985
5986 HeapRegion* next = cur->next_in_collection_set();
5987 assert(cur->in_collection_set(), "bad CS");
5988 cur->set_next_in_collection_set(NULL);
5989 cur->set_in_collection_set(false);
5990
5991 if (cur->is_young()) {
5992 int index = cur->young_index_in_cset();
5993 assert(index != -1, "invariant");
5994 assert((uint) index < policy->young_cset_region_length(), "invariant");
5995 size_t words_survived = _surviving_young_words[index];
5996 cur->record_surv_words_in_group(words_survived);
5997
5998 // At this point the we have 'popped' cur from the collection set
5999 // (linked via next_in_collection_set()) but it is still in the
6000 // young list (linked via next_young_region()). Clear the
6001 // _next_young_region field.
6002 cur->set_next_young_region(NULL);
6003 } else {
6004 int index = cur->young_index_in_cset();
6005 assert(index == -1, "invariant");
6006 }
6007
6008 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
6009 (!cur->is_young() && cur->young_index_in_cset() == -1),
6010 "invariant" );
6011
6012 if (!cur->evacuation_failed()) {
6013 MemRegion used_mr = cur->used_region();
6014
6015 // And the region is empty.
6016 assert(!used_mr.is_empty(), "Should not have empty regions in a CS.");
6017 free_region(cur, &pre_used, &local_free_list, false /* par */);
6018 } else {
6019 cur->uninstall_surv_rate_group();
6020 if (cur->is_young()) {
6021 cur->set_young_index_in_cset(-1);
6022 }
6023 cur->set_not_young();
6024 cur->set_evacuation_failed(false);
6025 // The region is now considered to be old.
6026 _old_set.add(cur);
6027 }
6028 cur = next;
6029 }
6030
6031 policy->record_max_rs_lengths(rs_lengths);
6032 policy->cset_regions_freed();
6033
6034 double end_sec = os::elapsedTime();
6035 double elapsed_ms = (end_sec - start_sec) * 1000.0;
6036
6037 if (non_young) {
6038 non_young_time_ms += elapsed_ms;
6039 } else {
6040 young_time_ms += elapsed_ms;
6041 }
6042
6043 update_sets_after_freeing_regions(pre_used, &local_free_list,
6044 NULL /* old_proxy_set */,
6045 NULL /* humongous_proxy_set */,
6046 false /* par */);
6047 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms);
6048 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms);
6049 }
6050
6051 // This routine is similar to the above but does not record
6052 // any policy statistics or update free lists; we are abandoning
6053 // the current incremental collection set in preparation of a
6054 // full collection. After the full GC we will start to build up
6055 // the incremental collection set again.
6056 // This is only called when we're doing a full collection
6057 // and is immediately followed by the tearing down of the young list.
6058
6059 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
6060 HeapRegion* cur = cs_head;
6061
6062 while (cur != NULL) {
6063 HeapRegion* next = cur->next_in_collection_set();
6064 assert(cur->in_collection_set(), "bad CS");
6065 cur->set_next_in_collection_set(NULL);
6066 cur->set_in_collection_set(false);
6067 cur->set_young_index_in_cset(-1);
6068 cur = next;
6069 }
6070 }
6071
6072 void G1CollectedHeap::set_free_regions_coming() {
6073 if (G1ConcRegionFreeingVerbose) {
6074 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6075 "setting free regions coming");
6076 }
6077
6078 assert(!free_regions_coming(), "pre-condition");
6079 _free_regions_coming = true;
6080 }
6081
6082 void G1CollectedHeap::reset_free_regions_coming() {
6083 assert(free_regions_coming(), "pre-condition");
6084
6085 {
6086 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6087 _free_regions_coming = false;
6088 SecondaryFreeList_lock->notify_all();
6089 }
6090
6091 if (G1ConcRegionFreeingVerbose) {
6092 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6093 "reset free regions coming");
6094 }
6095 }
6096
6097 void G1CollectedHeap::wait_while_free_regions_coming() {
6098 // Most of the time we won't have to wait, so let's do a quick test
6099 // first before we take the lock.
6100 if (!free_regions_coming()) {
6101 return;
6102 }
6103
6104 if (G1ConcRegionFreeingVerbose) {
6105 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6106 "waiting for free regions");
6107 }
6108
6109 {
6110 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6111 while (free_regions_coming()) {
6112 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
6113 }
6114 }
6115
6116 if (G1ConcRegionFreeingVerbose) {
6117 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6118 "done waiting for free regions");
6119 }
6120 }
6121
6122 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
6123 assert(heap_lock_held_for_gc(),
6124 "the heap lock should already be held by or for this thread");
6125 _young_list->push_region(hr);
6126 }
6127
6128 class NoYoungRegionsClosure: public HeapRegionClosure {
6129 private:
6130 bool _success;
6131 public:
6132 NoYoungRegionsClosure() : _success(true) { }
6133 bool doHeapRegion(HeapRegion* r) {
6134 if (r->is_young()) {
6135 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
6136 r->bottom(), r->end());
6137 _success = false;
6138 }
6139 return false;
6140 }
6141 bool success() { return _success; }
6142 };
6143
6144 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
6145 bool ret = _young_list->check_list_empty(check_sample);
6146
6147 if (check_heap) {
6148 NoYoungRegionsClosure closure;
6149 heap_region_iterate(&closure);
6150 ret = ret && closure.success();
6151 }
6152
6153 return ret;
6154 }
6155
6156 class TearDownRegionSetsClosure : public HeapRegionClosure {
6157 private:
6158 OldRegionSet *_old_set;
6159
6160 public:
6161 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { }
6162
6163 bool doHeapRegion(HeapRegion* r) {
6164 if (r->is_empty()) {
6165 // We ignore empty regions, we'll empty the free list afterwards
6166 } else if (r->is_young()) {
6167 // We ignore young regions, we'll empty the young list afterwards
6168 } else if (r->isHumongous()) {
6169 // We ignore humongous regions, we're not tearing down the
6170 // humongous region set
6171 } else {
6172 // The rest should be old
6173 _old_set->remove(r);
6174 }
6175 return false;
6176 }
6177
6178 ~TearDownRegionSetsClosure() {
6179 assert(_old_set->is_empty(), "post-condition");
6180 }
6181 };
6182
6183 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) {
6184 assert_at_safepoint(true /* should_be_vm_thread */);
6185
6186 if (!free_list_only) {
6187 TearDownRegionSetsClosure cl(&_old_set);
6188 heap_region_iterate(&cl);
6189
6190 // Need to do this after the heap iteration to be able to
6191 // recognize the young regions and ignore them during the iteration.
6192 _young_list->empty_list();
6193 }
6194 _free_list.remove_all();
6195 }
6196
6197 class RebuildRegionSetsClosure : public HeapRegionClosure {
6198 private:
6199 bool _free_list_only;
6200 OldRegionSet* _old_set;
6201 FreeRegionList* _free_list;
6202 size_t _total_used;
6203
6204 public:
6205 RebuildRegionSetsClosure(bool free_list_only,
6206 OldRegionSet* old_set, FreeRegionList* free_list) :
6207 _free_list_only(free_list_only),
6208 _old_set(old_set), _free_list(free_list), _total_used(0) {
6209 assert(_free_list->is_empty(), "pre-condition");
6210 if (!free_list_only) {
6211 assert(_old_set->is_empty(), "pre-condition");
6212 }
6213 }
6214
6215 bool doHeapRegion(HeapRegion* r) {
6216 if (r->continuesHumongous()) {
6217 return false;
6218 }
6219
6220 if (r->is_empty()) {
6221 // Add free regions to the free list
6222 _free_list->add_as_tail(r);
6223 } else if (!_free_list_only) {
6224 assert(!r->is_young(), "we should not come across young regions");
6225
6226 if (r->isHumongous()) {
6227 // We ignore humongous regions, we left the humongous set unchanged
6228 } else {
6229 // The rest should be old, add them to the old set
6230 _old_set->add(r);
6231 }
6232 _total_used += r->used();
6233 }
6234
6235 return false;
6236 }
6237
6238 size_t total_used() {
6239 return _total_used;
6240 }
6241 };
6242
6243 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) {
6244 assert_at_safepoint(true /* should_be_vm_thread */);
6245
6246 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list);
6247 heap_region_iterate(&cl);
6248
6249 if (!free_list_only) {
6250 _summary_bytes_used = cl.total_used();
6251 }
6252 assert(_summary_bytes_used == recalculate_used(),
6253 err_msg("inconsistent _summary_bytes_used, "
6254 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT,
6255 _summary_bytes_used, recalculate_used()));
6256 }
6257
6258 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
6259 _refine_cte_cl->set_concurrent(concurrent);
6260 }
6261
6262 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
6263 HeapRegion* hr = heap_region_containing(p);
6264 if (hr == NULL) {
6265 return false;
6266 } else {
6267 return hr->is_in(p);
6268 }
6269 }
6270
6271 // Methods for the mutator alloc region
6272
6273 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
6274 bool force) {
6275 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6276 assert(!force || g1_policy()->can_expand_young_list(),
6277 "if force is true we should be able to expand the young list");
6278 bool young_list_full = g1_policy()->is_young_list_full();
6279 if (force || !young_list_full) {
6280 HeapRegion* new_alloc_region = new_region(word_size,
6281 false /* do_expand */);
6282 if (new_alloc_region != NULL) {
6283 set_region_short_lived_locked(new_alloc_region);
6284 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
6285 return new_alloc_region;
6286 }
6287 }
6288 return NULL;
6289 }
6290
6291 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
6292 size_t allocated_bytes) {
6293 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6294 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
6295
6296 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
6297 _summary_bytes_used += allocated_bytes;
6298 _hr_printer.retire(alloc_region);
6299 // We update the eden sizes here, when the region is retired,
6300 // instead of when it's allocated, since this is the point that its
6301 // used space has been recored in _summary_bytes_used.
6302 g1mm()->update_eden_size();
6303 }
6304
6305 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
6306 bool force) {
6307 return _g1h->new_mutator_alloc_region(word_size, force);
6308 }
6309
6310 void G1CollectedHeap::set_par_threads() {
6311 // Don't change the number of workers. Use the value previously set
6312 // in the workgroup.
6313 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise");
6314 uint n_workers = workers()->active_workers();
6315 assert(UseDynamicNumberOfGCThreads ||
6316 n_workers == workers()->total_workers(),
6317 "Otherwise should be using the total number of workers");
6318 if (n_workers == 0) {
6319 assert(false, "Should have been set in prior evacuation pause.");
6320 n_workers = ParallelGCThreads;
6321 workers()->set_active_workers(n_workers);
6322 }
6323 set_par_threads(n_workers);
6324 }
6325
6326 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
6327 size_t allocated_bytes) {
6328 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
6329 }
6330
6331 // Methods for the GC alloc regions
6332
6333 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
6334 uint count,
6335 GCAllocPurpose ap) {
6336 assert(FreeList_lock->owned_by_self(), "pre-condition");
6337
6338 if (count < g1_policy()->max_regions(ap)) {
6339 HeapRegion* new_alloc_region = new_region(word_size,
6340 true /* do_expand */);
6341 if (new_alloc_region != NULL) {
6342 // We really only need to do this for old regions given that we
6343 // should never scan survivors. But it doesn't hurt to do it
6344 // for survivors too.
6345 new_alloc_region->set_saved_mark();
6346 if (ap == GCAllocForSurvived) {
6347 new_alloc_region->set_survivor();
6348 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
6349 } else {
6350 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
6351 }
6352 bool during_im = g1_policy()->during_initial_mark_pause();
6353 new_alloc_region->note_start_of_copying(during_im);
6354 return new_alloc_region;
6355 } else {
6356 g1_policy()->note_alloc_region_limit_reached(ap);
6357 }
6358 }
6359 return NULL;
6360 }
6361
6362 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
6363 size_t allocated_bytes,
6364 GCAllocPurpose ap) {
6365 bool during_im = g1_policy()->during_initial_mark_pause();
6366 alloc_region->note_end_of_copying(during_im);
6367 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
6368 if (ap == GCAllocForSurvived) {
6369 young_list()->add_survivor_region(alloc_region);
6370 } else {
6371 _old_set.add(alloc_region);
6372 }
6373 _hr_printer.retire(alloc_region);
6374 }
6375
6376 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
6377 bool force) {
6378 assert(!force, "not supported for GC alloc regions");
6379 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
6380 }
6381
6382 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
6383 size_t allocated_bytes) {
6384 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6385 GCAllocForSurvived);
6386 }
6387
6388 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
6389 bool force) {
6390 assert(!force, "not supported for GC alloc regions");
6391 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
6392 }
6393
6394 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
6395 size_t allocated_bytes) {
6396 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6397 GCAllocForTenured);
6398 }
6399 // Heap region set verification
6400
6401 class VerifyRegionListsClosure : public HeapRegionClosure {
6402 private:
6403 FreeRegionList* _free_list;
6404 OldRegionSet* _old_set;
6405 HumongousRegionSet* _humongous_set;
6406 uint _region_count;
6407
6408 public:
6409 VerifyRegionListsClosure(OldRegionSet* old_set,
6410 HumongousRegionSet* humongous_set,
6411 FreeRegionList* free_list) :
6412 _old_set(old_set), _humongous_set(humongous_set),
6413 _free_list(free_list), _region_count(0) { }
6414
6415 uint region_count() { return _region_count; }
6416
6417 bool doHeapRegion(HeapRegion* hr) {
6418 _region_count += 1;
6419
6420 if (hr->continuesHumongous()) {
6421 return false;
6422 }
6423
6424 if (hr->is_young()) {
6425 // TODO
6426 } else if (hr->startsHumongous()) {
6427 _humongous_set->verify_next_region(hr);
6428 } else if (hr->is_empty()) {
6429 _free_list->verify_next_region(hr);
6430 } else {
6431 _old_set->verify_next_region(hr);
6432 }
6433 return false;
6434 }
6435 };
6436
6437 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index,
6438 HeapWord* bottom) {
6439 HeapWord* end = bottom + HeapRegion::GrainWords;
6440 MemRegion mr(bottom, end);
6441 assert(_g1_reserved.contains(mr), "invariant");
6442 // This might return NULL if the allocation fails
6443 return new HeapRegion(hrs_index, _bot_shared, mr);
6444 }
6445
6446 void G1CollectedHeap::verify_region_sets() {
6447 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6448
6449 // First, check the explicit lists.
6450 _free_list.verify();
6451 {
6452 // Given that a concurrent operation might be adding regions to
6453 // the secondary free list we have to take the lock before
6454 // verifying it.
6455 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6456 _secondary_free_list.verify();
6457 }
6458 _old_set.verify();
6459 _humongous_set.verify();
6460
6461 // If a concurrent region freeing operation is in progress it will
6462 // be difficult to correctly attributed any free regions we come
6463 // across to the correct free list given that they might belong to
6464 // one of several (free_list, secondary_free_list, any local lists,
6465 // etc.). So, if that's the case we will skip the rest of the
6466 // verification operation. Alternatively, waiting for the concurrent
6467 // operation to complete will have a non-trivial effect on the GC's
6468 // operation (no concurrent operation will last longer than the
6469 // interval between two calls to verification) and it might hide
6470 // any issues that we would like to catch during testing.
6471 if (free_regions_coming()) {
6472 return;
6473 }
6474
6475 // Make sure we append the secondary_free_list on the free_list so
6476 // that all free regions we will come across can be safely
6477 // attributed to the free_list.
6478 append_secondary_free_list_if_not_empty_with_lock();
6479
6480 // Finally, make sure that the region accounting in the lists is
6481 // consistent with what we see in the heap.
6482 _old_set.verify_start();
6483 _humongous_set.verify_start();
6484 _free_list.verify_start();
6485
6486 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list);
6487 heap_region_iterate(&cl);
6488
6489 _old_set.verify_end();
6490 _humongous_set.verify_end();
6491 _free_list.verify_end();
6492 }