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