Thu, 23 Aug 2012 10:21:12 +0200
7178363: G1: Remove the serial code for PrintGCDetails and make it a special case of the parallel code
Summary: Also reviewed by vitalyd@gmail.com. Introduced the WorkerDataArray class. Fixed some minor logging bugs.
Reviewed-by: johnc, mgerdin
1 /*
2 * Copyright (c) 2001, 2012, 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 */
25 #include "precompiled.hpp"
26 #include "code/icBuffer.hpp"
27 #include "gc_implementation/g1/bufferingOopClosure.hpp"
28 #include "gc_implementation/g1/concurrentG1Refine.hpp"
29 #include "gc_implementation/g1/concurrentG1RefineThread.hpp"
30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
31 #include "gc_implementation/g1/g1AllocRegion.inline.hpp"
32 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
33 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
34 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
35 #include "gc_implementation/g1/g1EvacFailure.hpp"
36 #include "gc_implementation/g1/g1GCPhaseTimes.hpp"
37 #include "gc_implementation/g1/g1Log.hpp"
38 #include "gc_implementation/g1/g1MarkSweep.hpp"
39 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
40 #include "gc_implementation/g1/g1RemSet.inline.hpp"
41 #include "gc_implementation/g1/heapRegion.inline.hpp"
42 #include "gc_implementation/g1/heapRegionRemSet.hpp"
43 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
44 #include "gc_implementation/g1/vm_operations_g1.hpp"
45 #include "gc_implementation/shared/isGCActiveMark.hpp"
46 #include "memory/gcLocker.inline.hpp"
47 #include "memory/genOopClosures.inline.hpp"
48 #include "memory/generationSpec.hpp"
49 #include "memory/referenceProcessor.hpp"
50 #include "oops/oop.inline.hpp"
51 #include "oops/oop.pcgc.inline.hpp"
52 #include "runtime/aprofiler.hpp"
53 #include "runtime/vmThread.hpp"
55 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0;
57 // turn it on so that the contents of the young list (scan-only /
58 // to-be-collected) are printed at "strategic" points before / during
59 // / after the collection --- this is useful for debugging
60 #define YOUNG_LIST_VERBOSE 0
61 // CURRENT STATUS
62 // This file is under construction. Search for "FIXME".
64 // INVARIANTS/NOTES
65 //
66 // All allocation activity covered by the G1CollectedHeap interface is
67 // serialized by acquiring the HeapLock. This happens in mem_allocate
68 // and allocate_new_tlab, which are the "entry" points to the
69 // allocation code from the rest of the JVM. (Note that this does not
70 // apply to TLAB allocation, which is not part of this interface: it
71 // is done by clients of this interface.)
73 // Notes on implementation of parallelism in different tasks.
74 //
75 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism.
76 // The number of GC workers is passed to heap_region_par_iterate_chunked().
77 // It does use run_task() which sets _n_workers in the task.
78 // G1ParTask executes g1_process_strong_roots() ->
79 // SharedHeap::process_strong_roots() which calls eventuall to
80 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses
81 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also
82 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap).
83 //
85 // Local to this file.
87 class RefineCardTableEntryClosure: public CardTableEntryClosure {
88 SuspendibleThreadSet* _sts;
89 G1RemSet* _g1rs;
90 ConcurrentG1Refine* _cg1r;
91 bool _concurrent;
92 public:
93 RefineCardTableEntryClosure(SuspendibleThreadSet* sts,
94 G1RemSet* g1rs,
95 ConcurrentG1Refine* cg1r) :
96 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true)
97 {}
98 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
99 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false);
100 // This path is executed by the concurrent refine or mutator threads,
101 // concurrently, and so we do not care if card_ptr contains references
102 // that point into the collection set.
103 assert(!oops_into_cset, "should be");
105 if (_concurrent && _sts->should_yield()) {
106 // Caller will actually yield.
107 return false;
108 }
109 // Otherwise, we finished successfully; return true.
110 return true;
111 }
112 void set_concurrent(bool b) { _concurrent = b; }
113 };
116 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure {
117 int _calls;
118 G1CollectedHeap* _g1h;
119 CardTableModRefBS* _ctbs;
120 int _histo[256];
121 public:
122 ClearLoggedCardTableEntryClosure() :
123 _calls(0)
124 {
125 _g1h = G1CollectedHeap::heap();
126 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
127 for (int i = 0; i < 256; i++) _histo[i] = 0;
128 }
129 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
130 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
131 _calls++;
132 unsigned char* ujb = (unsigned char*)card_ptr;
133 int ind = (int)(*ujb);
134 _histo[ind]++;
135 *card_ptr = -1;
136 }
137 return true;
138 }
139 int calls() { return _calls; }
140 void print_histo() {
141 gclog_or_tty->print_cr("Card table value histogram:");
142 for (int i = 0; i < 256; i++) {
143 if (_histo[i] != 0) {
144 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]);
145 }
146 }
147 }
148 };
150 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure {
151 int _calls;
152 G1CollectedHeap* _g1h;
153 CardTableModRefBS* _ctbs;
154 public:
155 RedirtyLoggedCardTableEntryClosure() :
156 _calls(0)
157 {
158 _g1h = G1CollectedHeap::heap();
159 _ctbs = (CardTableModRefBS*)_g1h->barrier_set();
160 }
161 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
162 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) {
163 _calls++;
164 *card_ptr = 0;
165 }
166 return true;
167 }
168 int calls() { return _calls; }
169 };
171 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure {
172 public:
173 bool do_card_ptr(jbyte* card_ptr, int worker_i) {
174 *card_ptr = CardTableModRefBS::dirty_card_val();
175 return true;
176 }
177 };
179 YoungList::YoungList(G1CollectedHeap* g1h) :
180 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0),
181 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) {
182 guarantee(check_list_empty(false), "just making sure...");
183 }
185 void YoungList::push_region(HeapRegion *hr) {
186 assert(!hr->is_young(), "should not already be young");
187 assert(hr->get_next_young_region() == NULL, "cause it should!");
189 hr->set_next_young_region(_head);
190 _head = hr;
192 _g1h->g1_policy()->set_region_eden(hr, (int) _length);
193 ++_length;
194 }
196 void YoungList::add_survivor_region(HeapRegion* hr) {
197 assert(hr->is_survivor(), "should be flagged as survivor region");
198 assert(hr->get_next_young_region() == NULL, "cause it should!");
200 hr->set_next_young_region(_survivor_head);
201 if (_survivor_head == NULL) {
202 _survivor_tail = hr;
203 }
204 _survivor_head = hr;
205 ++_survivor_length;
206 }
208 void YoungList::empty_list(HeapRegion* list) {
209 while (list != NULL) {
210 HeapRegion* next = list->get_next_young_region();
211 list->set_next_young_region(NULL);
212 list->uninstall_surv_rate_group();
213 list->set_not_young();
214 list = next;
215 }
216 }
218 void YoungList::empty_list() {
219 assert(check_list_well_formed(), "young list should be well formed");
221 empty_list(_head);
222 _head = NULL;
223 _length = 0;
225 empty_list(_survivor_head);
226 _survivor_head = NULL;
227 _survivor_tail = NULL;
228 _survivor_length = 0;
230 _last_sampled_rs_lengths = 0;
232 assert(check_list_empty(false), "just making sure...");
233 }
235 bool YoungList::check_list_well_formed() {
236 bool ret = true;
238 uint length = 0;
239 HeapRegion* curr = _head;
240 HeapRegion* last = NULL;
241 while (curr != NULL) {
242 if (!curr->is_young()) {
243 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" "
244 "incorrectly tagged (y: %d, surv: %d)",
245 curr->bottom(), curr->end(),
246 curr->is_young(), curr->is_survivor());
247 ret = false;
248 }
249 ++length;
250 last = curr;
251 curr = curr->get_next_young_region();
252 }
253 ret = ret && (length == _length);
255 if (!ret) {
256 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!");
257 gclog_or_tty->print_cr("### list has %u entries, _length is %u",
258 length, _length);
259 }
261 return ret;
262 }
264 bool YoungList::check_list_empty(bool check_sample) {
265 bool ret = true;
267 if (_length != 0) {
268 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u",
269 _length);
270 ret = false;
271 }
272 if (check_sample && _last_sampled_rs_lengths != 0) {
273 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths");
274 ret = false;
275 }
276 if (_head != NULL) {
277 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head");
278 ret = false;
279 }
280 if (!ret) {
281 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty");
282 }
284 return ret;
285 }
287 void
288 YoungList::rs_length_sampling_init() {
289 _sampled_rs_lengths = 0;
290 _curr = _head;
291 }
293 bool
294 YoungList::rs_length_sampling_more() {
295 return _curr != NULL;
296 }
298 void
299 YoungList::rs_length_sampling_next() {
300 assert( _curr != NULL, "invariant" );
301 size_t rs_length = _curr->rem_set()->occupied();
303 _sampled_rs_lengths += rs_length;
305 // The current region may not yet have been added to the
306 // incremental collection set (it gets added when it is
307 // retired as the current allocation region).
308 if (_curr->in_collection_set()) {
309 // Update the collection set policy information for this region
310 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length);
311 }
313 _curr = _curr->get_next_young_region();
314 if (_curr == NULL) {
315 _last_sampled_rs_lengths = _sampled_rs_lengths;
316 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths);
317 }
318 }
320 void
321 YoungList::reset_auxilary_lists() {
322 guarantee( is_empty(), "young list should be empty" );
323 assert(check_list_well_formed(), "young list should be well formed");
325 // Add survivor regions to SurvRateGroup.
326 _g1h->g1_policy()->note_start_adding_survivor_regions();
327 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */);
329 int young_index_in_cset = 0;
330 for (HeapRegion* curr = _survivor_head;
331 curr != NULL;
332 curr = curr->get_next_young_region()) {
333 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset);
335 // The region is a non-empty survivor so let's add it to
336 // the incremental collection set for the next evacuation
337 // pause.
338 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr);
339 young_index_in_cset += 1;
340 }
341 assert((uint) young_index_in_cset == _survivor_length, "post-condition");
342 _g1h->g1_policy()->note_stop_adding_survivor_regions();
344 _head = _survivor_head;
345 _length = _survivor_length;
346 if (_survivor_head != NULL) {
347 assert(_survivor_tail != NULL, "cause it shouldn't be");
348 assert(_survivor_length > 0, "invariant");
349 _survivor_tail->set_next_young_region(NULL);
350 }
352 // Don't clear the survivor list handles until the start of
353 // the next evacuation pause - we need it in order to re-tag
354 // the survivor regions from this evacuation pause as 'young'
355 // at the start of the next.
357 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */);
359 assert(check_list_well_formed(), "young list should be well formed");
360 }
362 void YoungList::print() {
363 HeapRegion* lists[] = {_head, _survivor_head};
364 const char* names[] = {"YOUNG", "SURVIVOR"};
366 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) {
367 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]);
368 HeapRegion *curr = lists[list];
369 if (curr == NULL)
370 gclog_or_tty->print_cr(" empty");
371 while (curr != NULL) {
372 gclog_or_tty->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d",
373 HR_FORMAT_PARAMS(curr),
374 curr->prev_top_at_mark_start(),
375 curr->next_top_at_mark_start(),
376 curr->age_in_surv_rate_group_cond());
377 curr = curr->get_next_young_region();
378 }
379 }
381 gclog_or_tty->print_cr("");
382 }
384 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr)
385 {
386 // Claim the right to put the region on the dirty cards region list
387 // by installing a self pointer.
388 HeapRegion* next = hr->get_next_dirty_cards_region();
389 if (next == NULL) {
390 HeapRegion* res = (HeapRegion*)
391 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(),
392 NULL);
393 if (res == NULL) {
394 HeapRegion* head;
395 do {
396 // Put the region to the dirty cards region list.
397 head = _dirty_cards_region_list;
398 next = (HeapRegion*)
399 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head);
400 if (next == head) {
401 assert(hr->get_next_dirty_cards_region() == hr,
402 "hr->get_next_dirty_cards_region() != hr");
403 if (next == NULL) {
404 // The last region in the list points to itself.
405 hr->set_next_dirty_cards_region(hr);
406 } else {
407 hr->set_next_dirty_cards_region(next);
408 }
409 }
410 } while (next != head);
411 }
412 }
413 }
415 HeapRegion* G1CollectedHeap::pop_dirty_cards_region()
416 {
417 HeapRegion* head;
418 HeapRegion* hr;
419 do {
420 head = _dirty_cards_region_list;
421 if (head == NULL) {
422 return NULL;
423 }
424 HeapRegion* new_head = head->get_next_dirty_cards_region();
425 if (head == new_head) {
426 // The last region.
427 new_head = NULL;
428 }
429 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list,
430 head);
431 } while (hr != head);
432 assert(hr != NULL, "invariant");
433 hr->set_next_dirty_cards_region(NULL);
434 return hr;
435 }
437 void G1CollectedHeap::stop_conc_gc_threads() {
438 _cg1r->stop();
439 _cmThread->stop();
440 }
442 #ifdef ASSERT
443 // A region is added to the collection set as it is retired
444 // so an address p can point to a region which will be in the
445 // collection set but has not yet been retired. This method
446 // therefore is only accurate during a GC pause after all
447 // regions have been retired. It is used for debugging
448 // to check if an nmethod has references to objects that can
449 // be move during a partial collection. Though it can be
450 // inaccurate, it is sufficient for G1 because the conservative
451 // implementation of is_scavengable() for G1 will indicate that
452 // all nmethods must be scanned during a partial collection.
453 bool G1CollectedHeap::is_in_partial_collection(const void* p) {
454 HeapRegion* hr = heap_region_containing(p);
455 return hr != NULL && hr->in_collection_set();
456 }
457 #endif
459 // Returns true if the reference points to an object that
460 // can move in an incremental collecction.
461 bool G1CollectedHeap::is_scavengable(const void* p) {
462 G1CollectedHeap* g1h = G1CollectedHeap::heap();
463 G1CollectorPolicy* g1p = g1h->g1_policy();
464 HeapRegion* hr = heap_region_containing(p);
465 if (hr == NULL) {
466 // perm gen (or null)
467 return false;
468 } else {
469 return !hr->isHumongous();
470 }
471 }
473 void G1CollectedHeap::check_ct_logs_at_safepoint() {
474 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
475 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
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();
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();
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.");
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.");
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 }
515 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
516 }
518 // Private class members.
520 G1CollectedHeap* G1CollectedHeap::_g1h;
522 // Private methods.
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();
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 }
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 }
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 }
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");
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");
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 }
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");
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();
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 }
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");
669 // Index of last region in the series + 1.
670 uint last = first + num_regions;
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.
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");
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;
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);
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);
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");
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();
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 }
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");
787 assert(first_hr->used() == word_size * HeapWordSize, "invariant");
788 _summary_bytes_used += first_hr->used();
789 _humongous_set.add(first_hr);
791 return new_obj;
792 }
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 */);
800 verify_region_sets_optional();
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");
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 }
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");
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 }
846 verify_region_sets_optional();
848 return result;
849 }
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");
855 unsigned int dummy_gc_count_before;
856 return attempt_allocation(word_size, &dummy_gc_count_before);
857 }
859 HeapWord*
860 G1CollectedHeap::mem_allocate(size_t word_size,
861 bool* gc_overhead_limit_was_exceeded) {
862 assert_heap_not_locked_and_not_at_safepoint();
864 // Loop until the allocation is satisified, or unsatisfied after GC.
865 for (int try_count = 1; /* we'll return */; try_count += 1) {
866 unsigned int gc_count_before;
868 HeapWord* result = NULL;
869 if (!isHumongous(word_size)) {
870 result = attempt_allocation(word_size, &gc_count_before);
871 } else {
872 result = attempt_allocation_humongous(word_size, &gc_count_before);
873 }
874 if (result != NULL) {
875 return result;
876 }
878 // Create the garbage collection operation...
879 VM_G1CollectForAllocation op(gc_count_before, word_size);
880 // ...and get the VM thread to execute it.
881 VMThread::execute(&op);
883 if (op.prologue_succeeded() && op.pause_succeeded()) {
884 // If the operation was successful we'll return the result even
885 // if it is NULL. If the allocation attempt failed immediately
886 // after a Full GC, it's unlikely we'll be able to allocate now.
887 HeapWord* result = op.result();
888 if (result != NULL && !isHumongous(word_size)) {
889 // Allocations that take place on VM operations do not do any
890 // card dirtying and we have to do it here. We only have to do
891 // this for non-humongous allocations, though.
892 dirty_young_block(result, word_size);
893 }
894 return result;
895 } else {
896 assert(op.result() == NULL,
897 "the result should be NULL if the VM op did not succeed");
898 }
900 // Give a warning if we seem to be looping forever.
901 if ((QueuedAllocationWarningCount > 0) &&
902 (try_count % QueuedAllocationWarningCount == 0)) {
903 warning("G1CollectedHeap::mem_allocate retries %d times", try_count);
904 }
905 }
907 ShouldNotReachHere();
908 return NULL;
909 }
911 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size,
912 unsigned int *gc_count_before_ret) {
913 // Make sure you read the note in attempt_allocation_humongous().
915 assert_heap_not_locked_and_not_at_safepoint();
916 assert(!isHumongous(word_size), "attempt_allocation_slow() should not "
917 "be called for humongous allocation requests");
919 // We should only get here after the first-level allocation attempt
920 // (attempt_allocation()) failed to allocate.
922 // We will loop until a) we manage to successfully perform the
923 // allocation or b) we successfully schedule a collection which
924 // fails to perform the allocation. b) is the only case when we'll
925 // return NULL.
926 HeapWord* result = NULL;
927 for (int try_count = 1; /* we'll return */; try_count += 1) {
928 bool should_try_gc;
929 unsigned int gc_count_before;
931 {
932 MutexLockerEx x(Heap_lock);
934 result = _mutator_alloc_region.attempt_allocation_locked(word_size,
935 false /* bot_updates */);
936 if (result != NULL) {
937 return result;
938 }
940 // If we reach here, attempt_allocation_locked() above failed to
941 // allocate a new region. So the mutator alloc region should be NULL.
942 assert(_mutator_alloc_region.get() == NULL, "only way to get here");
944 if (GC_locker::is_active_and_needs_gc()) {
945 if (g1_policy()->can_expand_young_list()) {
946 // No need for an ergo verbose message here,
947 // can_expand_young_list() does this when it returns true.
948 result = _mutator_alloc_region.attempt_allocation_force(word_size,
949 false /* bot_updates */);
950 if (result != NULL) {
951 return result;
952 }
953 }
954 should_try_gc = false;
955 } else {
956 // The GCLocker may not be active but the GCLocker initiated
957 // GC may not yet have been performed (GCLocker::needs_gc()
958 // returns true). In this case we do not try this GC and
959 // wait until the GCLocker initiated GC is performed, and
960 // then retry the allocation.
961 if (GC_locker::needs_gc()) {
962 should_try_gc = false;
963 } else {
964 // Read the GC count while still holding the Heap_lock.
965 gc_count_before = total_collections();
966 should_try_gc = true;
967 }
968 }
969 }
971 if (should_try_gc) {
972 bool succeeded;
973 result = do_collection_pause(word_size, gc_count_before, &succeeded);
974 if (result != NULL) {
975 assert(succeeded, "only way to get back a non-NULL result");
976 return result;
977 }
979 if (succeeded) {
980 // If we get here we successfully scheduled a collection which
981 // failed to allocate. No point in trying to allocate
982 // further. We'll just return NULL.
983 MutexLockerEx x(Heap_lock);
984 *gc_count_before_ret = total_collections();
985 return NULL;
986 }
987 } else {
988 // The GCLocker is either active or the GCLocker initiated
989 // GC has not yet been performed. Stall until it is and
990 // then retry the allocation.
991 GC_locker::stall_until_clear();
992 }
994 // We can reach here if we were unsuccessul in scheduling a
995 // collection (because another thread beat us to it) or if we were
996 // stalled due to the GC locker. In either can we should retry the
997 // allocation attempt in case another thread successfully
998 // performed a collection and reclaimed enough space. We do the
999 // first attempt (without holding the Heap_lock) here and the
1000 // follow-on attempt will be at the start of the next loop
1001 // iteration (after taking the Heap_lock).
1002 result = _mutator_alloc_region.attempt_allocation(word_size,
1003 false /* bot_updates */);
1004 if (result != NULL) {
1005 return result;
1006 }
1008 // Give a warning if we seem to be looping forever.
1009 if ((QueuedAllocationWarningCount > 0) &&
1010 (try_count % QueuedAllocationWarningCount == 0)) {
1011 warning("G1CollectedHeap::attempt_allocation_slow() "
1012 "retries %d times", try_count);
1013 }
1014 }
1016 ShouldNotReachHere();
1017 return NULL;
1018 }
1020 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size,
1021 unsigned int * gc_count_before_ret) {
1022 // The structure of this method has a lot of similarities to
1023 // attempt_allocation_slow(). The reason these two were not merged
1024 // into a single one is that such a method would require several "if
1025 // allocation is not humongous do this, otherwise do that"
1026 // conditional paths which would obscure its flow. In fact, an early
1027 // version of this code did use a unified method which was harder to
1028 // follow and, as a result, it had subtle bugs that were hard to
1029 // track down. So keeping these two methods separate allows each to
1030 // be more readable. It will be good to keep these two in sync as
1031 // much as possible.
1033 assert_heap_not_locked_and_not_at_safepoint();
1034 assert(isHumongous(word_size), "attempt_allocation_humongous() "
1035 "should only be called for humongous allocations");
1037 // Humongous objects can exhaust the heap quickly, so we should check if we
1038 // need to start a marking cycle at each humongous object allocation. We do
1039 // the check before we do the actual allocation. The reason for doing it
1040 // before the allocation is that we avoid having to keep track of the newly
1041 // allocated memory while we do a GC.
1042 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation",
1043 word_size)) {
1044 collect(GCCause::_g1_humongous_allocation);
1045 }
1047 // We will loop until a) we manage to successfully perform the
1048 // allocation or b) we successfully schedule a collection which
1049 // fails to perform the allocation. b) is the only case when we'll
1050 // return NULL.
1051 HeapWord* result = NULL;
1052 for (int try_count = 1; /* we'll return */; try_count += 1) {
1053 bool should_try_gc;
1054 unsigned int gc_count_before;
1056 {
1057 MutexLockerEx x(Heap_lock);
1059 // Given that humongous objects are not allocated in young
1060 // regions, we'll first try to do the allocation without doing a
1061 // collection hoping that there's enough space in the heap.
1062 result = humongous_obj_allocate(word_size);
1063 if (result != NULL) {
1064 return result;
1065 }
1067 if (GC_locker::is_active_and_needs_gc()) {
1068 should_try_gc = false;
1069 } else {
1070 // The GCLocker may not be active but the GCLocker initiated
1071 // GC may not yet have been performed (GCLocker::needs_gc()
1072 // returns true). In this case we do not try this GC and
1073 // wait until the GCLocker initiated GC is performed, and
1074 // then retry the allocation.
1075 if (GC_locker::needs_gc()) {
1076 should_try_gc = false;
1077 } else {
1078 // Read the GC count while still holding the Heap_lock.
1079 gc_count_before = total_collections();
1080 should_try_gc = true;
1081 }
1082 }
1083 }
1085 if (should_try_gc) {
1086 // If we failed to allocate the humongous object, we should try to
1087 // do a collection pause (if we're allowed) in case it reclaims
1088 // enough space for the allocation to succeed after the pause.
1090 bool succeeded;
1091 result = do_collection_pause(word_size, gc_count_before, &succeeded);
1092 if (result != NULL) {
1093 assert(succeeded, "only way to get back a non-NULL result");
1094 return result;
1095 }
1097 if (succeeded) {
1098 // If we get here we successfully scheduled a collection which
1099 // failed to allocate. No point in trying to allocate
1100 // further. We'll just return NULL.
1101 MutexLockerEx x(Heap_lock);
1102 *gc_count_before_ret = total_collections();
1103 return NULL;
1104 }
1105 } else {
1106 // The GCLocker is either active or the GCLocker initiated
1107 // GC has not yet been performed. Stall until it is and
1108 // then retry the allocation.
1109 GC_locker::stall_until_clear();
1110 }
1112 // We can reach here if we were unsuccessul in scheduling a
1113 // collection (because another thread beat us to it) or if we were
1114 // stalled due to the GC locker. In either can we should retry the
1115 // allocation attempt in case another thread successfully
1116 // performed a collection and reclaimed enough space. Give a
1117 // warning if we seem to be looping forever.
1119 if ((QueuedAllocationWarningCount > 0) &&
1120 (try_count % QueuedAllocationWarningCount == 0)) {
1121 warning("G1CollectedHeap::attempt_allocation_humongous() "
1122 "retries %d times", try_count);
1123 }
1124 }
1126 ShouldNotReachHere();
1127 return NULL;
1128 }
1130 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size,
1131 bool expect_null_mutator_alloc_region) {
1132 assert_at_safepoint(true /* should_be_vm_thread */);
1133 assert(_mutator_alloc_region.get() == NULL ||
1134 !expect_null_mutator_alloc_region,
1135 "the current alloc region was unexpectedly found to be non-NULL");
1137 if (!isHumongous(word_size)) {
1138 return _mutator_alloc_region.attempt_allocation_locked(word_size,
1139 false /* bot_updates */);
1140 } else {
1141 HeapWord* result = humongous_obj_allocate(word_size);
1142 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) {
1143 g1_policy()->set_initiate_conc_mark_if_possible();
1144 }
1145 return result;
1146 }
1148 ShouldNotReachHere();
1149 }
1151 class PostMCRemSetClearClosure: public HeapRegionClosure {
1152 G1CollectedHeap* _g1h;
1153 ModRefBarrierSet* _mr_bs;
1154 public:
1155 PostMCRemSetClearClosure(G1CollectedHeap* g1h, ModRefBarrierSet* mr_bs) :
1156 _g1h(g1h), _mr_bs(mr_bs) { }
1157 bool doHeapRegion(HeapRegion* r) {
1158 if (r->continuesHumongous()) {
1159 return false;
1160 }
1161 _g1h->reset_gc_time_stamps(r);
1162 HeapRegionRemSet* hrrs = r->rem_set();
1163 if (hrrs != NULL) hrrs->clear();
1164 // You might think here that we could clear just the cards
1165 // corresponding to the used region. But no: if we leave a dirty card
1166 // in a region we might allocate into, then it would prevent that card
1167 // from being enqueued, and cause it to be missed.
1168 // Re: the performance cost: we shouldn't be doing full GC anyway!
1169 _mr_bs->clear(MemRegion(r->bottom(), r->end()));
1170 return false;
1171 }
1172 };
1174 void G1CollectedHeap::clear_rsets_post_compaction() {
1175 PostMCRemSetClearClosure rs_clear(this, mr_bs());
1176 heap_region_iterate(&rs_clear);
1177 }
1179 class RebuildRSOutOfRegionClosure: public HeapRegionClosure {
1180 G1CollectedHeap* _g1h;
1181 UpdateRSOopClosure _cl;
1182 int _worker_i;
1183 public:
1184 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) :
1185 _cl(g1->g1_rem_set(), worker_i),
1186 _worker_i(worker_i),
1187 _g1h(g1)
1188 { }
1190 bool doHeapRegion(HeapRegion* r) {
1191 if (!r->continuesHumongous()) {
1192 _cl.set_from(r);
1193 r->oop_iterate(&_cl);
1194 }
1195 return false;
1196 }
1197 };
1199 class ParRebuildRSTask: public AbstractGangTask {
1200 G1CollectedHeap* _g1;
1201 public:
1202 ParRebuildRSTask(G1CollectedHeap* g1)
1203 : AbstractGangTask("ParRebuildRSTask"),
1204 _g1(g1)
1205 { }
1207 void work(uint worker_id) {
1208 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id);
1209 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id,
1210 _g1->workers()->active_workers(),
1211 HeapRegion::RebuildRSClaimValue);
1212 }
1213 };
1215 class PostCompactionPrinterClosure: public HeapRegionClosure {
1216 private:
1217 G1HRPrinter* _hr_printer;
1218 public:
1219 bool doHeapRegion(HeapRegion* hr) {
1220 assert(!hr->is_young(), "not expecting to find young regions");
1221 // We only generate output for non-empty regions.
1222 if (!hr->is_empty()) {
1223 if (!hr->isHumongous()) {
1224 _hr_printer->post_compaction(hr, G1HRPrinter::Old);
1225 } else if (hr->startsHumongous()) {
1226 if (hr->region_num() == 1) {
1227 // single humongous region
1228 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous);
1229 } else {
1230 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous);
1231 }
1232 } else {
1233 assert(hr->continuesHumongous(), "only way to get here");
1234 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous);
1235 }
1236 }
1237 return false;
1238 }
1240 PostCompactionPrinterClosure(G1HRPrinter* hr_printer)
1241 : _hr_printer(hr_printer) { }
1242 };
1244 void G1CollectedHeap::print_hrs_post_compaction() {
1245 PostCompactionPrinterClosure cl(hr_printer());
1246 heap_region_iterate(&cl);
1247 }
1249 double G1CollectedHeap::verify(bool guard, const char* msg) {
1250 double verify_time_ms = 0.0;
1252 if (guard && total_collections() >= VerifyGCStartAt) {
1253 double verify_start = os::elapsedTime();
1254 HandleMark hm; // Discard invalid handles created during verification
1255 gclog_or_tty->print(msg);
1256 prepare_for_verify();
1257 Universe::verify(false /* silent */, VerifyOption_G1UsePrevMarking);
1258 verify_time_ms = (os::elapsedTime() - verify_start) * 1000;
1259 }
1261 return verify_time_ms;
1262 }
1264 void G1CollectedHeap::verify_before_gc() {
1265 double verify_time_ms = verify(VerifyBeforeGC, " VerifyBeforeGC:");
1266 g1_policy()->phase_times()->record_verify_before_time_ms(verify_time_ms);
1267 }
1269 void G1CollectedHeap::verify_after_gc() {
1270 double verify_time_ms = verify(VerifyAfterGC, " VerifyAfterGC:");
1271 g1_policy()->phase_times()->record_verify_after_time_ms(verify_time_ms);
1272 }
1274 bool G1CollectedHeap::do_collection(bool explicit_gc,
1275 bool clear_all_soft_refs,
1276 size_t word_size) {
1277 assert_at_safepoint(true /* should_be_vm_thread */);
1279 if (GC_locker::check_active_before_gc()) {
1280 return false;
1281 }
1283 SvcGCMarker sgcm(SvcGCMarker::FULL);
1284 ResourceMark rm;
1286 print_heap_before_gc();
1288 HRSPhaseSetter x(HRSPhaseFullGC);
1289 verify_region_sets_optional();
1291 const bool do_clear_all_soft_refs = clear_all_soft_refs ||
1292 collector_policy()->should_clear_all_soft_refs();
1294 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy());
1296 {
1297 IsGCActiveMark x;
1299 // Timing
1300 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant");
1301 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps);
1302 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
1304 TraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, gclog_or_tty);
1305 TraceCollectorStats tcs(g1mm()->full_collection_counters());
1306 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause());
1308 double start = os::elapsedTime();
1309 g1_policy()->record_full_collection_start();
1311 // Note: When we have a more flexible GC logging framework that
1312 // allows us to add optional attributes to a GC log record we
1313 // could consider timing and reporting how long we wait in the
1314 // following two methods.
1315 wait_while_free_regions_coming();
1316 // If we start the compaction before the CM threads finish
1317 // scanning the root regions we might trip them over as we'll
1318 // be moving objects / updating references. So let's wait until
1319 // they are done. By telling them to abort, they should complete
1320 // early.
1321 _cm->root_regions()->abort();
1322 _cm->root_regions()->wait_until_scan_finished();
1323 append_secondary_free_list_if_not_empty_with_lock();
1325 gc_prologue(true);
1326 increment_total_collections(true /* full gc */);
1327 increment_old_marking_cycles_started();
1329 size_t g1h_prev_used = used();
1330 assert(used() == recalculate_used(), "Should be equal");
1332 verify_before_gc();
1334 pre_full_gc_dump();
1336 COMPILER2_PRESENT(DerivedPointerTable::clear());
1338 // Disable discovery and empty the discovered lists
1339 // for the CM ref processor.
1340 ref_processor_cm()->disable_discovery();
1341 ref_processor_cm()->abandon_partial_discovery();
1342 ref_processor_cm()->verify_no_references_recorded();
1344 // Abandon current iterations of concurrent marking and concurrent
1345 // refinement, if any are in progress. We have to do this before
1346 // wait_until_scan_finished() below.
1347 concurrent_mark()->abort();
1349 // Make sure we'll choose a new allocation region afterwards.
1350 release_mutator_alloc_region();
1351 abandon_gc_alloc_regions();
1352 g1_rem_set()->cleanupHRRS();
1354 // We should call this after we retire any currently active alloc
1355 // regions so that all the ALLOC / RETIRE events are generated
1356 // before the start GC event.
1357 _hr_printer.start_gc(true /* full */, (size_t) total_collections());
1359 // We may have added regions to the current incremental collection
1360 // set between the last GC or pause and now. We need to clear the
1361 // incremental collection set and then start rebuilding it afresh
1362 // after this full GC.
1363 abandon_collection_set(g1_policy()->inc_cset_head());
1364 g1_policy()->clear_incremental_cset();
1365 g1_policy()->stop_incremental_cset_building();
1367 tear_down_region_sets(false /* free_list_only */);
1368 g1_policy()->set_gcs_are_young(true);
1370 // See the comments in g1CollectedHeap.hpp and
1371 // G1CollectedHeap::ref_processing_init() about
1372 // how reference processing currently works in G1.
1374 // Temporarily make discovery by the STW ref processor single threaded (non-MT).
1375 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false);
1377 // Temporarily clear the STW ref processor's _is_alive_non_header field.
1378 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL);
1380 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
1381 ref_processor_stw()->setup_policy(do_clear_all_soft_refs);
1383 // Do collection work
1384 {
1385 HandleMark hm; // Discard invalid handles created during gc
1386 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs);
1387 }
1389 assert(free_regions() == 0, "we should not have added any free regions");
1390 rebuild_region_sets(false /* free_list_only */);
1392 // Enqueue any discovered reference objects that have
1393 // not been removed from the discovered lists.
1394 ref_processor_stw()->enqueue_discovered_references();
1396 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
1398 MemoryService::track_memory_usage();
1400 verify_after_gc();
1402 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
1403 ref_processor_stw()->verify_no_references_recorded();
1405 // Note: since we've just done a full GC, concurrent
1406 // marking is no longer active. Therefore we need not
1407 // re-enable reference discovery for the CM ref processor.
1408 // That will be done at the start of the next marking cycle.
1409 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition");
1410 ref_processor_cm()->verify_no_references_recorded();
1412 reset_gc_time_stamp();
1413 // Since everything potentially moved, we will clear all remembered
1414 // sets, and clear all cards. Later we will rebuild remebered
1415 // sets. We will also reset the GC time stamps of the regions.
1416 clear_rsets_post_compaction();
1417 check_gc_time_stamps();
1419 // Resize the heap if necessary.
1420 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size);
1422 if (_hr_printer.is_active()) {
1423 // We should do this after we potentially resize the heap so
1424 // that all the COMMIT / UNCOMMIT events are generated before
1425 // the end GC event.
1427 print_hrs_post_compaction();
1428 _hr_printer.end_gc(true /* full */, (size_t) total_collections());
1429 }
1431 if (_cg1r->use_cache()) {
1432 _cg1r->clear_and_record_card_counts();
1433 _cg1r->clear_hot_cache();
1434 }
1436 // Rebuild remembered sets of all regions.
1437 if (G1CollectedHeap::use_parallel_gc_threads()) {
1438 uint n_workers =
1439 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
1440 workers()->active_workers(),
1441 Threads::number_of_non_daemon_threads());
1442 assert(UseDynamicNumberOfGCThreads ||
1443 n_workers == workers()->total_workers(),
1444 "If not dynamic should be using all the workers");
1445 workers()->set_active_workers(n_workers);
1446 // Set parallel threads in the heap (_n_par_threads) only
1447 // before a parallel phase and always reset it to 0 after
1448 // the phase so that the number of parallel threads does
1449 // no get carried forward to a serial phase where there
1450 // may be code that is "possibly_parallel".
1451 set_par_threads(n_workers);
1453 ParRebuildRSTask rebuild_rs_task(this);
1454 assert(check_heap_region_claim_values(
1455 HeapRegion::InitialClaimValue), "sanity check");
1456 assert(UseDynamicNumberOfGCThreads ||
1457 workers()->active_workers() == workers()->total_workers(),
1458 "Unless dynamic should use total workers");
1459 // Use the most recent number of active workers
1460 assert(workers()->active_workers() > 0,
1461 "Active workers not properly set");
1462 set_par_threads(workers()->active_workers());
1463 workers()->run_task(&rebuild_rs_task);
1464 set_par_threads(0);
1465 assert(check_heap_region_claim_values(
1466 HeapRegion::RebuildRSClaimValue), "sanity check");
1467 reset_heap_region_claim_values();
1468 } else {
1469 RebuildRSOutOfRegionClosure rebuild_rs(this);
1470 heap_region_iterate(&rebuild_rs);
1471 }
1473 if (G1Log::fine()) {
1474 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity());
1475 }
1477 if (true) { // FIXME
1478 // Ask the permanent generation to adjust size for full collections
1479 perm()->compute_new_size();
1480 }
1482 // Start a new incremental collection set for the next pause
1483 assert(g1_policy()->collection_set() == NULL, "must be");
1484 g1_policy()->start_incremental_cset_building();
1486 // Clear the _cset_fast_test bitmap in anticipation of adding
1487 // regions to the incremental collection set for the next
1488 // evacuation pause.
1489 clear_cset_fast_test();
1491 init_mutator_alloc_region();
1493 double end = os::elapsedTime();
1494 g1_policy()->record_full_collection_end();
1496 #ifdef TRACESPINNING
1497 ParallelTaskTerminator::print_termination_counts();
1498 #endif
1500 gc_epilogue(true);
1502 // Discard all rset updates
1503 JavaThread::dirty_card_queue_set().abandon_logs();
1504 assert(!G1DeferredRSUpdate
1505 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
1507 _young_list->reset_sampled_info();
1508 // At this point there should be no regions in the
1509 // entire heap tagged as young.
1510 assert( check_young_list_empty(true /* check_heap */),
1511 "young list should be empty at this point");
1513 // Update the number of full collections that have been completed.
1514 increment_old_marking_cycles_completed(false /* concurrent */);
1516 _hrs.verify_optional();
1517 verify_region_sets_optional();
1519 print_heap_after_gc();
1521 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
1522 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
1523 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
1524 // before any GC notifications are raised.
1525 g1mm()->update_sizes();
1526 }
1528 post_full_gc_dump();
1530 return true;
1531 }
1533 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) {
1534 // do_collection() will return whether it succeeded in performing
1535 // the GC. Currently, there is no facility on the
1536 // do_full_collection() API to notify the caller than the collection
1537 // did not succeed (e.g., because it was locked out by the GC
1538 // locker). So, right now, we'll ignore the return value.
1539 bool dummy = do_collection(true, /* explicit_gc */
1540 clear_all_soft_refs,
1541 0 /* word_size */);
1542 }
1544 // This code is mostly copied from TenuredGeneration.
1545 void
1546 G1CollectedHeap::
1547 resize_if_necessary_after_full_collection(size_t word_size) {
1548 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check");
1550 // Include the current allocation, if any, and bytes that will be
1551 // pre-allocated to support collections, as "used".
1552 const size_t used_after_gc = used();
1553 const size_t capacity_after_gc = capacity();
1554 const size_t free_after_gc = capacity_after_gc - used_after_gc;
1556 // This is enforced in arguments.cpp.
1557 assert(MinHeapFreeRatio <= MaxHeapFreeRatio,
1558 "otherwise the code below doesn't make sense");
1560 // We don't have floating point command-line arguments
1561 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0;
1562 const double maximum_used_percentage = 1.0 - minimum_free_percentage;
1563 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0;
1564 const double minimum_used_percentage = 1.0 - maximum_free_percentage;
1566 const size_t min_heap_size = collector_policy()->min_heap_byte_size();
1567 const size_t max_heap_size = collector_policy()->max_heap_byte_size();
1569 // We have to be careful here as these two calculations can overflow
1570 // 32-bit size_t's.
1571 double used_after_gc_d = (double) used_after_gc;
1572 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage;
1573 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage;
1575 // Let's make sure that they are both under the max heap size, which
1576 // by default will make them fit into a size_t.
1577 double desired_capacity_upper_bound = (double) max_heap_size;
1578 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d,
1579 desired_capacity_upper_bound);
1580 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d,
1581 desired_capacity_upper_bound);
1583 // We can now safely turn them into size_t's.
1584 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d;
1585 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d;
1587 // This assert only makes sense here, before we adjust them
1588 // with respect to the min and max heap size.
1589 assert(minimum_desired_capacity <= maximum_desired_capacity,
1590 err_msg("minimum_desired_capacity = "SIZE_FORMAT", "
1591 "maximum_desired_capacity = "SIZE_FORMAT,
1592 minimum_desired_capacity, maximum_desired_capacity));
1594 // Should not be greater than the heap max size. No need to adjust
1595 // it with respect to the heap min size as it's a lower bound (i.e.,
1596 // we'll try to make the capacity larger than it, not smaller).
1597 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size);
1598 // Should not be less than the heap min size. No need to adjust it
1599 // with respect to the heap max size as it's an upper bound (i.e.,
1600 // we'll try to make the capacity smaller than it, not greater).
1601 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size);
1603 if (capacity_after_gc < minimum_desired_capacity) {
1604 // Don't expand unless it's significant
1605 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc;
1606 ergo_verbose4(ErgoHeapSizing,
1607 "attempt heap expansion",
1608 ergo_format_reason("capacity lower than "
1609 "min desired capacity after Full GC")
1610 ergo_format_byte("capacity")
1611 ergo_format_byte("occupancy")
1612 ergo_format_byte_perc("min desired capacity"),
1613 capacity_after_gc, used_after_gc,
1614 minimum_desired_capacity, (double) MinHeapFreeRatio);
1615 expand(expand_bytes);
1617 // No expansion, now see if we want to shrink
1618 } else if (capacity_after_gc > maximum_desired_capacity) {
1619 // Capacity too large, compute shrinking size
1620 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity;
1621 ergo_verbose4(ErgoHeapSizing,
1622 "attempt heap shrinking",
1623 ergo_format_reason("capacity higher than "
1624 "max desired capacity after Full GC")
1625 ergo_format_byte("capacity")
1626 ergo_format_byte("occupancy")
1627 ergo_format_byte_perc("max desired capacity"),
1628 capacity_after_gc, used_after_gc,
1629 maximum_desired_capacity, (double) MaxHeapFreeRatio);
1630 shrink(shrink_bytes);
1631 }
1632 }
1635 HeapWord*
1636 G1CollectedHeap::satisfy_failed_allocation(size_t word_size,
1637 bool* succeeded) {
1638 assert_at_safepoint(true /* should_be_vm_thread */);
1640 *succeeded = true;
1641 // Let's attempt the allocation first.
1642 HeapWord* result =
1643 attempt_allocation_at_safepoint(word_size,
1644 false /* expect_null_mutator_alloc_region */);
1645 if (result != NULL) {
1646 assert(*succeeded, "sanity");
1647 return result;
1648 }
1650 // In a G1 heap, we're supposed to keep allocation from failing by
1651 // incremental pauses. Therefore, at least for now, we'll favor
1652 // expansion over collection. (This might change in the future if we can
1653 // do something smarter than full collection to satisfy a failed alloc.)
1654 result = expand_and_allocate(word_size);
1655 if (result != NULL) {
1656 assert(*succeeded, "sanity");
1657 return result;
1658 }
1660 // Expansion didn't work, we'll try to do a Full GC.
1661 bool gc_succeeded = do_collection(false, /* explicit_gc */
1662 false, /* clear_all_soft_refs */
1663 word_size);
1664 if (!gc_succeeded) {
1665 *succeeded = false;
1666 return NULL;
1667 }
1669 // Retry the allocation
1670 result = attempt_allocation_at_safepoint(word_size,
1671 true /* expect_null_mutator_alloc_region */);
1672 if (result != NULL) {
1673 assert(*succeeded, "sanity");
1674 return result;
1675 }
1677 // Then, try a Full GC that will collect all soft references.
1678 gc_succeeded = do_collection(false, /* explicit_gc */
1679 true, /* clear_all_soft_refs */
1680 word_size);
1681 if (!gc_succeeded) {
1682 *succeeded = false;
1683 return NULL;
1684 }
1686 // Retry the allocation once more
1687 result = attempt_allocation_at_safepoint(word_size,
1688 true /* expect_null_mutator_alloc_region */);
1689 if (result != NULL) {
1690 assert(*succeeded, "sanity");
1691 return result;
1692 }
1694 assert(!collector_policy()->should_clear_all_soft_refs(),
1695 "Flag should have been handled and cleared prior to this point");
1697 // What else? We might try synchronous finalization later. If the total
1698 // space available is large enough for the allocation, then a more
1699 // complete compaction phase than we've tried so far might be
1700 // appropriate.
1701 assert(*succeeded, "sanity");
1702 return NULL;
1703 }
1705 // Attempting to expand the heap sufficiently
1706 // to support an allocation of the given "word_size". If
1707 // successful, perform the allocation and return the address of the
1708 // allocated block, or else "NULL".
1710 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) {
1711 assert_at_safepoint(true /* should_be_vm_thread */);
1713 verify_region_sets_optional();
1715 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes);
1716 ergo_verbose1(ErgoHeapSizing,
1717 "attempt heap expansion",
1718 ergo_format_reason("allocation request failed")
1719 ergo_format_byte("allocation request"),
1720 word_size * HeapWordSize);
1721 if (expand(expand_bytes)) {
1722 _hrs.verify_optional();
1723 verify_region_sets_optional();
1724 return attempt_allocation_at_safepoint(word_size,
1725 false /* expect_null_mutator_alloc_region */);
1726 }
1727 return NULL;
1728 }
1730 void G1CollectedHeap::update_committed_space(HeapWord* old_end,
1731 HeapWord* new_end) {
1732 assert(old_end != new_end, "don't call this otherwise");
1733 assert((HeapWord*) _g1_storage.high() == new_end, "invariant");
1735 // Update the committed mem region.
1736 _g1_committed.set_end(new_end);
1737 // Tell the card table about the update.
1738 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed);
1739 // Tell the BOT about the update.
1740 _bot_shared->resize(_g1_committed.word_size());
1741 }
1743 bool G1CollectedHeap::expand(size_t expand_bytes) {
1744 size_t old_mem_size = _g1_storage.committed_size();
1745 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes);
1746 aligned_expand_bytes = align_size_up(aligned_expand_bytes,
1747 HeapRegion::GrainBytes);
1748 ergo_verbose2(ErgoHeapSizing,
1749 "expand the heap",
1750 ergo_format_byte("requested expansion amount")
1751 ergo_format_byte("attempted expansion amount"),
1752 expand_bytes, aligned_expand_bytes);
1754 // First commit the memory.
1755 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1756 bool successful = _g1_storage.expand_by(aligned_expand_bytes);
1757 if (successful) {
1758 // Then propagate this update to the necessary data structures.
1759 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1760 update_committed_space(old_end, new_end);
1762 FreeRegionList expansion_list("Local Expansion List");
1763 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list);
1764 assert(mr.start() == old_end, "post-condition");
1765 // mr might be a smaller region than what was requested if
1766 // expand_by() was unable to allocate the HeapRegion instances
1767 assert(mr.end() <= new_end, "post-condition");
1769 size_t actual_expand_bytes = mr.byte_size();
1770 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition");
1771 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(),
1772 "post-condition");
1773 if (actual_expand_bytes < aligned_expand_bytes) {
1774 // We could not expand _hrs to the desired size. In this case we
1775 // need to shrink the committed space accordingly.
1776 assert(mr.end() < new_end, "invariant");
1778 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes;
1779 // First uncommit the memory.
1780 _g1_storage.shrink_by(diff_bytes);
1781 // Then propagate this update to the necessary data structures.
1782 update_committed_space(new_end, mr.end());
1783 }
1784 _free_list.add_as_tail(&expansion_list);
1786 if (_hr_printer.is_active()) {
1787 HeapWord* curr = mr.start();
1788 while (curr < mr.end()) {
1789 HeapWord* curr_end = curr + HeapRegion::GrainWords;
1790 _hr_printer.commit(curr, curr_end);
1791 curr = curr_end;
1792 }
1793 assert(curr == mr.end(), "post-condition");
1794 }
1795 g1_policy()->record_new_heap_size(n_regions());
1796 } else {
1797 ergo_verbose0(ErgoHeapSizing,
1798 "did not expand the heap",
1799 ergo_format_reason("heap expansion operation failed"));
1800 // The expansion of the virtual storage space was unsuccessful.
1801 // Let's see if it was because we ran out of swap.
1802 if (G1ExitOnExpansionFailure &&
1803 _g1_storage.uncommitted_size() >= aligned_expand_bytes) {
1804 // We had head room...
1805 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion");
1806 }
1807 }
1808 return successful;
1809 }
1811 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) {
1812 size_t old_mem_size = _g1_storage.committed_size();
1813 size_t aligned_shrink_bytes =
1814 ReservedSpace::page_align_size_down(shrink_bytes);
1815 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes,
1816 HeapRegion::GrainBytes);
1817 uint num_regions_deleted = 0;
1818 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted);
1819 HeapWord* old_end = (HeapWord*) _g1_storage.high();
1820 assert(mr.end() == old_end, "post-condition");
1822 ergo_verbose3(ErgoHeapSizing,
1823 "shrink the heap",
1824 ergo_format_byte("requested shrinking amount")
1825 ergo_format_byte("aligned shrinking amount")
1826 ergo_format_byte("attempted shrinking amount"),
1827 shrink_bytes, aligned_shrink_bytes, mr.byte_size());
1828 if (mr.byte_size() > 0) {
1829 if (_hr_printer.is_active()) {
1830 HeapWord* curr = mr.end();
1831 while (curr > mr.start()) {
1832 HeapWord* curr_end = curr;
1833 curr -= HeapRegion::GrainWords;
1834 _hr_printer.uncommit(curr, curr_end);
1835 }
1836 assert(curr == mr.start(), "post-condition");
1837 }
1839 _g1_storage.shrink_by(mr.byte_size());
1840 HeapWord* new_end = (HeapWord*) _g1_storage.high();
1841 assert(mr.start() == new_end, "post-condition");
1843 _expansion_regions += num_regions_deleted;
1844 update_committed_space(old_end, new_end);
1845 HeapRegionRemSet::shrink_heap(n_regions());
1846 g1_policy()->record_new_heap_size(n_regions());
1847 } else {
1848 ergo_verbose0(ErgoHeapSizing,
1849 "did not shrink the heap",
1850 ergo_format_reason("heap shrinking operation failed"));
1851 }
1852 }
1854 void G1CollectedHeap::shrink(size_t shrink_bytes) {
1855 verify_region_sets_optional();
1857 // We should only reach here at the end of a Full GC which means we
1858 // should not not be holding to any GC alloc regions. The method
1859 // below will make sure of that and do any remaining clean up.
1860 abandon_gc_alloc_regions();
1862 // Instead of tearing down / rebuilding the free lists here, we
1863 // could instead use the remove_all_pending() method on free_list to
1864 // remove only the ones that we need to remove.
1865 tear_down_region_sets(true /* free_list_only */);
1866 shrink_helper(shrink_bytes);
1867 rebuild_region_sets(true /* free_list_only */);
1869 _hrs.verify_optional();
1870 verify_region_sets_optional();
1871 }
1873 // Public methods.
1875 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
1876 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
1877 #endif // _MSC_VER
1880 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
1881 SharedHeap(policy_),
1882 _g1_policy(policy_),
1883 _dirty_card_queue_set(false),
1884 _into_cset_dirty_card_queue_set(false),
1885 _is_alive_closure_cm(this),
1886 _is_alive_closure_stw(this),
1887 _ref_processor_cm(NULL),
1888 _ref_processor_stw(NULL),
1889 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)),
1890 _bot_shared(NULL),
1891 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL),
1892 _evac_failure_scan_stack(NULL) ,
1893 _mark_in_progress(false),
1894 _cg1r(NULL), _summary_bytes_used(0),
1895 _g1mm(NULL),
1896 _refine_cte_cl(NULL),
1897 _full_collection(false),
1898 _free_list("Master Free List"),
1899 _secondary_free_list("Secondary Free List"),
1900 _old_set("Old Set"),
1901 _humongous_set("Master Humongous Set"),
1902 _free_regions_coming(false),
1903 _young_list(new YoungList(this)),
1904 _gc_time_stamp(0),
1905 _retained_old_gc_alloc_region(NULL),
1906 _survivor_plab_stats(YoungPLABSize, PLABWeight),
1907 _old_plab_stats(OldPLABSize, PLABWeight),
1908 _expand_heap_after_alloc_failure(true),
1909 _surviving_young_words(NULL),
1910 _old_marking_cycles_started(0),
1911 _old_marking_cycles_completed(0),
1912 _in_cset_fast_test(NULL),
1913 _in_cset_fast_test_base(NULL),
1914 _dirty_cards_region_list(NULL),
1915 _worker_cset_start_region(NULL),
1916 _worker_cset_start_region_time_stamp(NULL) {
1917 _g1h = this; // To catch bugs.
1918 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
1919 vm_exit_during_initialization("Failed necessary allocation.");
1920 }
1922 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2;
1924 int n_queues = MAX2((int)ParallelGCThreads, 1);
1925 _task_queues = new RefToScanQueueSet(n_queues);
1927 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets();
1928 assert(n_rem_sets > 0, "Invariant.");
1930 HeapRegionRemSetIterator** iter_arr =
1931 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues, mtGC);
1932 for (int i = 0; i < n_queues; i++) {
1933 iter_arr[i] = new HeapRegionRemSetIterator();
1934 }
1935 _rem_set_iterator = iter_arr;
1937 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC);
1938 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC);
1940 for (int i = 0; i < n_queues; i++) {
1941 RefToScanQueue* q = new RefToScanQueue();
1942 q->initialize();
1943 _task_queues->register_queue(i, q);
1944 }
1946 clear_cset_start_regions();
1948 guarantee(_task_queues != NULL, "task_queues allocation failure.");
1949 #ifdef SPARC
1950 // Issue a stern warning, but allow use for experimentation and debugging.
1951 if (VM_Version::is_sun4v() && UseMemSetInBOT) {
1952 assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error");
1953 warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability"
1954 " on sun4v; please understand that you are using at your own risk!");
1955 }
1956 #endif
1957 }
1959 jint G1CollectedHeap::initialize() {
1960 CollectedHeap::pre_initialize();
1961 os::enable_vtime();
1963 G1Log::init();
1965 // Necessary to satisfy locking discipline assertions.
1967 MutexLocker x(Heap_lock);
1969 // We have to initialize the printer before committing the heap, as
1970 // it will be used then.
1971 _hr_printer.set_active(G1PrintHeapRegions);
1973 // While there are no constraints in the GC code that HeapWordSize
1974 // be any particular value, there are multiple other areas in the
1975 // system which believe this to be true (e.g. oop->object_size in some
1976 // cases incorrectly returns the size in wordSize units rather than
1977 // HeapWordSize).
1978 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize");
1980 size_t init_byte_size = collector_policy()->initial_heap_byte_size();
1981 size_t max_byte_size = collector_policy()->max_heap_byte_size();
1983 // Ensure that the sizes are properly aligned.
1984 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap");
1985 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap");
1987 _cg1r = new ConcurrentG1Refine();
1989 // Reserve the maximum.
1990 PermanentGenerationSpec* pgs = collector_policy()->permanent_generation();
1991 // Includes the perm-gen.
1993 // When compressed oops are enabled, the preferred heap base
1994 // is calculated by subtracting the requested size from the
1995 // 32Gb boundary and using the result as the base address for
1996 // heap reservation. If the requested size is not aligned to
1997 // HeapRegion::GrainBytes (i.e. the alignment that is passed
1998 // into the ReservedHeapSpace constructor) then the actual
1999 // base of the reserved heap may end up differing from the
2000 // address that was requested (i.e. the preferred heap base).
2001 // If this happens then we could end up using a non-optimal
2002 // compressed oops mode.
2004 // Since max_byte_size is aligned to the size of a heap region (checked
2005 // above), we also need to align the perm gen size as it might not be.
2006 const size_t total_reserved = max_byte_size +
2007 align_size_up(pgs->max_size(), HeapRegion::GrainBytes);
2008 Universe::check_alignment(total_reserved, HeapRegion::GrainBytes, "g1 heap and perm");
2010 char* addr = Universe::preferred_heap_base(total_reserved, Universe::UnscaledNarrowOop);
2012 ReservedHeapSpace heap_rs(total_reserved, HeapRegion::GrainBytes,
2013 UseLargePages, addr);
2015 if (UseCompressedOops) {
2016 if (addr != NULL && !heap_rs.is_reserved()) {
2017 // Failed to reserve at specified address - the requested memory
2018 // region is taken already, for example, by 'java' launcher.
2019 // Try again to reserver heap higher.
2020 addr = Universe::preferred_heap_base(total_reserved, Universe::ZeroBasedNarrowOop);
2022 ReservedHeapSpace heap_rs0(total_reserved, HeapRegion::GrainBytes,
2023 UseLargePages, addr);
2025 if (addr != NULL && !heap_rs0.is_reserved()) {
2026 // Failed to reserve at specified address again - give up.
2027 addr = Universe::preferred_heap_base(total_reserved, Universe::HeapBasedNarrowOop);
2028 assert(addr == NULL, "");
2030 ReservedHeapSpace heap_rs1(total_reserved, HeapRegion::GrainBytes,
2031 UseLargePages, addr);
2032 heap_rs = heap_rs1;
2033 } else {
2034 heap_rs = heap_rs0;
2035 }
2036 }
2037 }
2039 if (!heap_rs.is_reserved()) {
2040 vm_exit_during_initialization("Could not reserve enough space for object heap");
2041 return JNI_ENOMEM;
2042 }
2044 // It is important to do this in a way such that concurrent readers can't
2045 // temporarily think somethings in the heap. (I've actually seen this
2046 // happen in asserts: DLD.)
2047 _reserved.set_word_size(0);
2048 _reserved.set_start((HeapWord*)heap_rs.base());
2049 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size()));
2051 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes);
2053 // Create the gen rem set (and barrier set) for the entire reserved region.
2054 _rem_set = collector_policy()->create_rem_set(_reserved, 2);
2055 set_barrier_set(rem_set()->bs());
2056 if (barrier_set()->is_a(BarrierSet::ModRef)) {
2057 _mr_bs = (ModRefBarrierSet*)_barrier_set;
2058 } else {
2059 vm_exit_during_initialization("G1 requires a mod ref bs.");
2060 return JNI_ENOMEM;
2061 }
2063 // Also create a G1 rem set.
2064 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) {
2065 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs());
2066 } else {
2067 vm_exit_during_initialization("G1 requires a cardtable mod ref bs.");
2068 return JNI_ENOMEM;
2069 }
2071 // Carve out the G1 part of the heap.
2073 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size);
2074 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(),
2075 g1_rs.size()/HeapWordSize);
2076 ReservedSpace perm_gen_rs = heap_rs.last_part(max_byte_size);
2078 _perm_gen = pgs->init(perm_gen_rs, pgs->init_size(), rem_set());
2080 _g1_storage.initialize(g1_rs, 0);
2081 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0);
2082 _hrs.initialize((HeapWord*) _g1_reserved.start(),
2083 (HeapWord*) _g1_reserved.end(),
2084 _expansion_regions);
2086 // 6843694 - ensure that the maximum region index can fit
2087 // in the remembered set structures.
2088 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1;
2089 guarantee((max_regions() - 1) <= max_region_idx, "too many regions");
2091 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1;
2092 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized");
2093 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region,
2094 "too many cards per region");
2096 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1);
2098 _bot_shared = new G1BlockOffsetSharedArray(_reserved,
2099 heap_word_size(init_byte_size));
2101 _g1h = this;
2103 _in_cset_fast_test_length = max_regions();
2104 _in_cset_fast_test_base =
2105 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC);
2107 // We're biasing _in_cset_fast_test to avoid subtracting the
2108 // beginning of the heap every time we want to index; basically
2109 // it's the same with what we do with the card table.
2110 _in_cset_fast_test = _in_cset_fast_test_base -
2111 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes);
2113 // Clear the _cset_fast_test bitmap in anticipation of adding
2114 // regions to the incremental collection set for the first
2115 // evacuation pause.
2116 clear_cset_fast_test();
2118 // Create the ConcurrentMark data structure and thread.
2119 // (Must do this late, so that "max_regions" is defined.)
2120 _cm = new ConcurrentMark(heap_rs, max_regions());
2121 _cmThread = _cm->cmThread();
2123 // Initialize the from_card cache structure of HeapRegionRemSet.
2124 HeapRegionRemSet::init_heap(max_regions());
2126 // Now expand into the initial heap size.
2127 if (!expand(init_byte_size)) {
2128 vm_exit_during_initialization("Failed to allocate initial heap.");
2129 return JNI_ENOMEM;
2130 }
2132 // Perform any initialization actions delegated to the policy.
2133 g1_policy()->init();
2135 _refine_cte_cl =
2136 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(),
2137 g1_rem_set(),
2138 concurrent_g1_refine());
2139 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl);
2141 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon,
2142 SATB_Q_FL_lock,
2143 G1SATBProcessCompletedThreshold,
2144 Shared_SATB_Q_lock);
2146 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2147 DirtyCardQ_FL_lock,
2148 concurrent_g1_refine()->yellow_zone(),
2149 concurrent_g1_refine()->red_zone(),
2150 Shared_DirtyCardQ_lock);
2152 if (G1DeferredRSUpdate) {
2153 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
2154 DirtyCardQ_FL_lock,
2155 -1, // never trigger processing
2156 -1, // no limit on length
2157 Shared_DirtyCardQ_lock,
2158 &JavaThread::dirty_card_queue_set());
2159 }
2161 // Initialize the card queue set used to hold cards containing
2162 // references into the collection set.
2163 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon,
2164 DirtyCardQ_FL_lock,
2165 -1, // never trigger processing
2166 -1, // no limit on length
2167 Shared_DirtyCardQ_lock,
2168 &JavaThread::dirty_card_queue_set());
2170 // In case we're keeping closure specialization stats, initialize those
2171 // counts and that mechanism.
2172 SpecializationStats::clear();
2174 // Do later initialization work for concurrent refinement.
2175 _cg1r->init();
2177 // Here we allocate the dummy full region that is required by the
2178 // G1AllocRegion class. If we don't pass an address in the reserved
2179 // space here, lots of asserts fire.
2181 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */,
2182 _g1_reserved.start());
2183 // We'll re-use the same region whether the alloc region will
2184 // require BOT updates or not and, if it doesn't, then a non-young
2185 // region will complain that it cannot support allocations without
2186 // BOT updates. So we'll tag the dummy region as young to avoid that.
2187 dummy_region->set_young();
2188 // Make sure it's full.
2189 dummy_region->set_top(dummy_region->end());
2190 G1AllocRegion::setup(this, dummy_region);
2192 init_mutator_alloc_region();
2194 // Do create of the monitoring and management support so that
2195 // values in the heap have been properly initialized.
2196 _g1mm = new G1MonitoringSupport(this);
2198 return JNI_OK;
2199 }
2201 void G1CollectedHeap::ref_processing_init() {
2202 // Reference processing in G1 currently works as follows:
2203 //
2204 // * There are two reference processor instances. One is
2205 // used to record and process discovered references
2206 // during concurrent marking; the other is used to
2207 // record and process references during STW pauses
2208 // (both full and incremental).
2209 // * Both ref processors need to 'span' the entire heap as
2210 // the regions in the collection set may be dotted around.
2211 //
2212 // * For the concurrent marking ref processor:
2213 // * Reference discovery is enabled at initial marking.
2214 // * Reference discovery is disabled and the discovered
2215 // references processed etc during remarking.
2216 // * Reference discovery is MT (see below).
2217 // * Reference discovery requires a barrier (see below).
2218 // * Reference processing may or may not be MT
2219 // (depending on the value of ParallelRefProcEnabled
2220 // and ParallelGCThreads).
2221 // * A full GC disables reference discovery by the CM
2222 // ref processor and abandons any entries on it's
2223 // discovered lists.
2224 //
2225 // * For the STW processor:
2226 // * Non MT discovery is enabled at the start of a full GC.
2227 // * Processing and enqueueing during a full GC is non-MT.
2228 // * During a full GC, references are processed after marking.
2229 //
2230 // * Discovery (may or may not be MT) is enabled at the start
2231 // of an incremental evacuation pause.
2232 // * References are processed near the end of a STW evacuation pause.
2233 // * For both types of GC:
2234 // * Discovery is atomic - i.e. not concurrent.
2235 // * Reference discovery will not need a barrier.
2237 SharedHeap::ref_processing_init();
2238 MemRegion mr = reserved_region();
2240 // Concurrent Mark ref processor
2241 _ref_processor_cm =
2242 new ReferenceProcessor(mr, // span
2243 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2244 // mt processing
2245 (int) ParallelGCThreads,
2246 // degree of mt processing
2247 (ParallelGCThreads > 1) || (ConcGCThreads > 1),
2248 // mt discovery
2249 (int) MAX2(ParallelGCThreads, ConcGCThreads),
2250 // degree of mt discovery
2251 false,
2252 // Reference discovery is not atomic
2253 &_is_alive_closure_cm,
2254 // is alive closure
2255 // (for efficiency/performance)
2256 true);
2257 // Setting next fields of discovered
2258 // lists requires a barrier.
2260 // STW ref processor
2261 _ref_processor_stw =
2262 new ReferenceProcessor(mr, // span
2263 ParallelRefProcEnabled && (ParallelGCThreads > 1),
2264 // mt processing
2265 MAX2((int)ParallelGCThreads, 1),
2266 // degree of mt processing
2267 (ParallelGCThreads > 1),
2268 // mt discovery
2269 MAX2((int)ParallelGCThreads, 1),
2270 // degree of mt discovery
2271 true,
2272 // Reference discovery is atomic
2273 &_is_alive_closure_stw,
2274 // is alive closure
2275 // (for efficiency/performance)
2276 false);
2277 // Setting next fields of discovered
2278 // lists requires a barrier.
2279 }
2281 size_t G1CollectedHeap::capacity() const {
2282 return _g1_committed.byte_size();
2283 }
2285 void G1CollectedHeap::reset_gc_time_stamps(HeapRegion* hr) {
2286 assert(!hr->continuesHumongous(), "pre-condition");
2287 hr->reset_gc_time_stamp();
2288 if (hr->startsHumongous()) {
2289 uint first_index = hr->hrs_index() + 1;
2290 uint last_index = hr->last_hc_index();
2291 for (uint i = first_index; i < last_index; i += 1) {
2292 HeapRegion* chr = region_at(i);
2293 assert(chr->continuesHumongous(), "sanity");
2294 chr->reset_gc_time_stamp();
2295 }
2296 }
2297 }
2299 #ifndef PRODUCT
2300 class CheckGCTimeStampsHRClosure : public HeapRegionClosure {
2301 private:
2302 unsigned _gc_time_stamp;
2303 bool _failures;
2305 public:
2306 CheckGCTimeStampsHRClosure(unsigned gc_time_stamp) :
2307 _gc_time_stamp(gc_time_stamp), _failures(false) { }
2309 virtual bool doHeapRegion(HeapRegion* hr) {
2310 unsigned region_gc_time_stamp = hr->get_gc_time_stamp();
2311 if (_gc_time_stamp != region_gc_time_stamp) {
2312 gclog_or_tty->print_cr("Region "HR_FORMAT" has GC time stamp = %d, "
2313 "expected %d", HR_FORMAT_PARAMS(hr),
2314 region_gc_time_stamp, _gc_time_stamp);
2315 _failures = true;
2316 }
2317 return false;
2318 }
2320 bool failures() { return _failures; }
2321 };
2323 void G1CollectedHeap::check_gc_time_stamps() {
2324 CheckGCTimeStampsHRClosure cl(_gc_time_stamp);
2325 heap_region_iterate(&cl);
2326 guarantee(!cl.failures(), "all GC time stamps should have been reset");
2327 }
2328 #endif // PRODUCT
2330 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl,
2331 DirtyCardQueue* into_cset_dcq,
2332 bool concurrent,
2333 int worker_i) {
2334 // Clean cards in the hot card cache
2335 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq);
2337 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
2338 int n_completed_buffers = 0;
2339 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) {
2340 n_completed_buffers++;
2341 }
2342 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers);
2343 dcqs.clear_n_completed_buffers();
2344 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!");
2345 }
2348 // Computes the sum of the storage used by the various regions.
2350 size_t G1CollectedHeap::used() const {
2351 assert(Heap_lock->owner() != NULL,
2352 "Should be owned on this thread's behalf.");
2353 size_t result = _summary_bytes_used;
2354 // Read only once in case it is set to NULL concurrently
2355 HeapRegion* hr = _mutator_alloc_region.get();
2356 if (hr != NULL)
2357 result += hr->used();
2358 return result;
2359 }
2361 size_t G1CollectedHeap::used_unlocked() const {
2362 size_t result = _summary_bytes_used;
2363 return result;
2364 }
2366 class SumUsedClosure: public HeapRegionClosure {
2367 size_t _used;
2368 public:
2369 SumUsedClosure() : _used(0) {}
2370 bool doHeapRegion(HeapRegion* r) {
2371 if (!r->continuesHumongous()) {
2372 _used += r->used();
2373 }
2374 return false;
2375 }
2376 size_t result() { return _used; }
2377 };
2379 size_t G1CollectedHeap::recalculate_used() const {
2380 SumUsedClosure blk;
2381 heap_region_iterate(&blk);
2382 return blk.result();
2383 }
2385 size_t G1CollectedHeap::unsafe_max_alloc() {
2386 if (free_regions() > 0) return HeapRegion::GrainBytes;
2387 // otherwise, is there space in the current allocation region?
2389 // We need to store the current allocation region in a local variable
2390 // here. The problem is that this method doesn't take any locks and
2391 // there may be other threads which overwrite the current allocation
2392 // region field. attempt_allocation(), for example, sets it to NULL
2393 // and this can happen *after* the NULL check here but before the call
2394 // to free(), resulting in a SIGSEGV. Note that this doesn't appear
2395 // to be a problem in the optimized build, since the two loads of the
2396 // current allocation region field are optimized away.
2397 HeapRegion* hr = _mutator_alloc_region.get();
2398 if (hr == NULL) {
2399 return 0;
2400 }
2401 return hr->free();
2402 }
2404 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) {
2405 switch (cause) {
2406 case GCCause::_gc_locker: return GCLockerInvokesConcurrent;
2407 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent;
2408 case GCCause::_g1_humongous_allocation: return true;
2409 default: return false;
2410 }
2411 }
2413 #ifndef PRODUCT
2414 void G1CollectedHeap::allocate_dummy_regions() {
2415 // Let's fill up most of the region
2416 size_t word_size = HeapRegion::GrainWords - 1024;
2417 // And as a result the region we'll allocate will be humongous.
2418 guarantee(isHumongous(word_size), "sanity");
2420 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) {
2421 // Let's use the existing mechanism for the allocation
2422 HeapWord* dummy_obj = humongous_obj_allocate(word_size);
2423 if (dummy_obj != NULL) {
2424 MemRegion mr(dummy_obj, word_size);
2425 CollectedHeap::fill_with_object(mr);
2426 } else {
2427 // If we can't allocate once, we probably cannot allocate
2428 // again. Let's get out of the loop.
2429 break;
2430 }
2431 }
2432 }
2433 #endif // !PRODUCT
2435 void G1CollectedHeap::increment_old_marking_cycles_started() {
2436 assert(_old_marking_cycles_started == _old_marking_cycles_completed ||
2437 _old_marking_cycles_started == _old_marking_cycles_completed + 1,
2438 err_msg("Wrong marking cycle count (started: %d, completed: %d)",
2439 _old_marking_cycles_started, _old_marking_cycles_completed));
2441 _old_marking_cycles_started++;
2442 }
2444 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) {
2445 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag);
2447 // We assume that if concurrent == true, then the caller is a
2448 // concurrent thread that was joined the Suspendible Thread
2449 // Set. If there's ever a cheap way to check this, we should add an
2450 // assert here.
2452 // Given that this method is called at the end of a Full GC or of a
2453 // concurrent cycle, and those can be nested (i.e., a Full GC can
2454 // interrupt a concurrent cycle), the number of full collections
2455 // completed should be either one (in the case where there was no
2456 // nesting) or two (when a Full GC interrupted a concurrent cycle)
2457 // behind the number of full collections started.
2459 // This is the case for the inner caller, i.e. a Full GC.
2460 assert(concurrent ||
2461 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) ||
2462 (_old_marking_cycles_started == _old_marking_cycles_completed + 2),
2463 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u "
2464 "is inconsistent with _old_marking_cycles_completed = %u",
2465 _old_marking_cycles_started, _old_marking_cycles_completed));
2467 // This is the case for the outer caller, i.e. the concurrent cycle.
2468 assert(!concurrent ||
2469 (_old_marking_cycles_started == _old_marking_cycles_completed + 1),
2470 err_msg("for outer caller (concurrent cycle): "
2471 "_old_marking_cycles_started = %u "
2472 "is inconsistent with _old_marking_cycles_completed = %u",
2473 _old_marking_cycles_started, _old_marking_cycles_completed));
2475 _old_marking_cycles_completed += 1;
2477 // We need to clear the "in_progress" flag in the CM thread before
2478 // we wake up any waiters (especially when ExplicitInvokesConcurrent
2479 // is set) so that if a waiter requests another System.gc() it doesn't
2480 // incorrectly see that a marking cyle is still in progress.
2481 if (concurrent) {
2482 _cmThread->clear_in_progress();
2483 }
2485 // This notify_all() will ensure that a thread that called
2486 // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
2487 // and it's waiting for a full GC to finish will be woken up. It is
2488 // waiting in VM_G1IncCollectionPause::doit_epilogue().
2489 FullGCCount_lock->notify_all();
2490 }
2492 void G1CollectedHeap::collect_as_vm_thread(GCCause::Cause cause) {
2493 assert_at_safepoint(true /* should_be_vm_thread */);
2494 GCCauseSetter gcs(this, cause);
2495 switch (cause) {
2496 case GCCause::_heap_inspection:
2497 case GCCause::_heap_dump: {
2498 HandleMark hm;
2499 do_full_collection(false); // don't clear all soft refs
2500 break;
2501 }
2502 default: // XXX FIX ME
2503 ShouldNotReachHere(); // Unexpected use of this function
2504 }
2505 }
2507 void G1CollectedHeap::collect(GCCause::Cause cause) {
2508 assert_heap_not_locked();
2510 unsigned int gc_count_before;
2511 unsigned int old_marking_count_before;
2512 bool retry_gc;
2514 do {
2515 retry_gc = false;
2517 {
2518 MutexLocker ml(Heap_lock);
2520 // Read the GC count while holding the Heap_lock
2521 gc_count_before = total_collections();
2522 old_marking_count_before = _old_marking_cycles_started;
2523 }
2525 if (should_do_concurrent_full_gc(cause)) {
2526 // Schedule an initial-mark evacuation pause that will start a
2527 // concurrent cycle. We're setting word_size to 0 which means that
2528 // we are not requesting a post-GC allocation.
2529 VM_G1IncCollectionPause op(gc_count_before,
2530 0, /* word_size */
2531 true, /* should_initiate_conc_mark */
2532 g1_policy()->max_pause_time_ms(),
2533 cause);
2535 VMThread::execute(&op);
2536 if (!op.pause_succeeded()) {
2537 if (old_marking_count_before == _old_marking_cycles_started) {
2538 retry_gc = op.should_retry_gc();
2539 } else {
2540 // A Full GC happened while we were trying to schedule the
2541 // initial-mark GC. No point in starting a new cycle given
2542 // that the whole heap was collected anyway.
2543 }
2545 if (retry_gc) {
2546 if (GC_locker::is_active_and_needs_gc()) {
2547 GC_locker::stall_until_clear();
2548 }
2549 }
2550 }
2551 } else {
2552 if (cause == GCCause::_gc_locker
2553 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) {
2555 // Schedule a standard evacuation pause. We're setting word_size
2556 // to 0 which means that we are not requesting a post-GC allocation.
2557 VM_G1IncCollectionPause op(gc_count_before,
2558 0, /* word_size */
2559 false, /* should_initiate_conc_mark */
2560 g1_policy()->max_pause_time_ms(),
2561 cause);
2562 VMThread::execute(&op);
2563 } else {
2564 // Schedule a Full GC.
2565 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause);
2566 VMThread::execute(&op);
2567 }
2568 }
2569 } while (retry_gc);
2570 }
2572 bool G1CollectedHeap::is_in(const void* p) const {
2573 if (_g1_committed.contains(p)) {
2574 // Given that we know that p is in the committed space,
2575 // heap_region_containing_raw() should successfully
2576 // return the containing region.
2577 HeapRegion* hr = heap_region_containing_raw(p);
2578 return hr->is_in(p);
2579 } else {
2580 return _perm_gen->as_gen()->is_in(p);
2581 }
2582 }
2584 // Iteration functions.
2586 // Iterates an OopClosure over all ref-containing fields of objects
2587 // within a HeapRegion.
2589 class IterateOopClosureRegionClosure: public HeapRegionClosure {
2590 MemRegion _mr;
2591 OopClosure* _cl;
2592 public:
2593 IterateOopClosureRegionClosure(MemRegion mr, OopClosure* cl)
2594 : _mr(mr), _cl(cl) {}
2595 bool doHeapRegion(HeapRegion* r) {
2596 if (!r->continuesHumongous()) {
2597 r->oop_iterate(_cl);
2598 }
2599 return false;
2600 }
2601 };
2603 void G1CollectedHeap::oop_iterate(OopClosure* cl, bool do_perm) {
2604 IterateOopClosureRegionClosure blk(_g1_committed, cl);
2605 heap_region_iterate(&blk);
2606 if (do_perm) {
2607 perm_gen()->oop_iterate(cl);
2608 }
2609 }
2611 void G1CollectedHeap::oop_iterate(MemRegion mr, OopClosure* cl, bool do_perm) {
2612 IterateOopClosureRegionClosure blk(mr, cl);
2613 heap_region_iterate(&blk);
2614 if (do_perm) {
2615 perm_gen()->oop_iterate(cl);
2616 }
2617 }
2619 // Iterates an ObjectClosure over all objects within a HeapRegion.
2621 class IterateObjectClosureRegionClosure: public HeapRegionClosure {
2622 ObjectClosure* _cl;
2623 public:
2624 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {}
2625 bool doHeapRegion(HeapRegion* r) {
2626 if (! r->continuesHumongous()) {
2627 r->object_iterate(_cl);
2628 }
2629 return false;
2630 }
2631 };
2633 void G1CollectedHeap::object_iterate(ObjectClosure* cl, bool do_perm) {
2634 IterateObjectClosureRegionClosure blk(cl);
2635 heap_region_iterate(&blk);
2636 if (do_perm) {
2637 perm_gen()->object_iterate(cl);
2638 }
2639 }
2641 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) {
2642 // FIXME: is this right?
2643 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap");
2644 }
2646 // Calls a SpaceClosure on a HeapRegion.
2648 class SpaceClosureRegionClosure: public HeapRegionClosure {
2649 SpaceClosure* _cl;
2650 public:
2651 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {}
2652 bool doHeapRegion(HeapRegion* r) {
2653 _cl->do_space(r);
2654 return false;
2655 }
2656 };
2658 void G1CollectedHeap::space_iterate(SpaceClosure* cl) {
2659 SpaceClosureRegionClosure blk(cl);
2660 heap_region_iterate(&blk);
2661 }
2663 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const {
2664 _hrs.iterate(cl);
2665 }
2667 void
2668 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl,
2669 uint worker_id,
2670 uint no_of_par_workers,
2671 jint claim_value) {
2672 const uint regions = n_regions();
2673 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
2674 no_of_par_workers :
2675 1);
2676 assert(UseDynamicNumberOfGCThreads ||
2677 no_of_par_workers == workers()->total_workers(),
2678 "Non dynamic should use fixed number of workers");
2679 // try to spread out the starting points of the workers
2680 const HeapRegion* start_hr =
2681 start_region_for_worker(worker_id, no_of_par_workers);
2682 const uint start_index = start_hr->hrs_index();
2684 // each worker will actually look at all regions
2685 for (uint count = 0; count < regions; ++count) {
2686 const uint index = (start_index + count) % regions;
2687 assert(0 <= index && index < regions, "sanity");
2688 HeapRegion* r = region_at(index);
2689 // we'll ignore "continues humongous" regions (we'll process them
2690 // when we come across their corresponding "start humongous"
2691 // region) and regions already claimed
2692 if (r->claim_value() == claim_value || r->continuesHumongous()) {
2693 continue;
2694 }
2695 // OK, try to claim it
2696 if (r->claimHeapRegion(claim_value)) {
2697 // success!
2698 assert(!r->continuesHumongous(), "sanity");
2699 if (r->startsHumongous()) {
2700 // If the region is "starts humongous" we'll iterate over its
2701 // "continues humongous" first; in fact we'll do them
2702 // first. The order is important. In on case, calling the
2703 // closure on the "starts humongous" region might de-allocate
2704 // and clear all its "continues humongous" regions and, as a
2705 // result, we might end up processing them twice. So, we'll do
2706 // them first (notice: most closures will ignore them anyway) and
2707 // then we'll do the "starts humongous" region.
2708 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) {
2709 HeapRegion* chr = region_at(ch_index);
2711 // if the region has already been claimed or it's not
2712 // "continues humongous" we're done
2713 if (chr->claim_value() == claim_value ||
2714 !chr->continuesHumongous()) {
2715 break;
2716 }
2718 // Noone should have claimed it directly. We can given
2719 // that we claimed its "starts humongous" region.
2720 assert(chr->claim_value() != claim_value, "sanity");
2721 assert(chr->humongous_start_region() == r, "sanity");
2723 if (chr->claimHeapRegion(claim_value)) {
2724 // we should always be able to claim it; noone else should
2725 // be trying to claim this region
2727 bool res2 = cl->doHeapRegion(chr);
2728 assert(!res2, "Should not abort");
2730 // Right now, this holds (i.e., no closure that actually
2731 // does something with "continues humongous" regions
2732 // clears them). We might have to weaken it in the future,
2733 // but let's leave these two asserts here for extra safety.
2734 assert(chr->continuesHumongous(), "should still be the case");
2735 assert(chr->humongous_start_region() == r, "sanity");
2736 } else {
2737 guarantee(false, "we should not reach here");
2738 }
2739 }
2740 }
2742 assert(!r->continuesHumongous(), "sanity");
2743 bool res = cl->doHeapRegion(r);
2744 assert(!res, "Should not abort");
2745 }
2746 }
2747 }
2749 class ResetClaimValuesClosure: public HeapRegionClosure {
2750 public:
2751 bool doHeapRegion(HeapRegion* r) {
2752 r->set_claim_value(HeapRegion::InitialClaimValue);
2753 return false;
2754 }
2755 };
2757 void G1CollectedHeap::reset_heap_region_claim_values() {
2758 ResetClaimValuesClosure blk;
2759 heap_region_iterate(&blk);
2760 }
2762 void G1CollectedHeap::reset_cset_heap_region_claim_values() {
2763 ResetClaimValuesClosure blk;
2764 collection_set_iterate(&blk);
2765 }
2767 #ifdef ASSERT
2768 // This checks whether all regions in the heap have the correct claim
2769 // value. I also piggy-backed on this a check to ensure that the
2770 // humongous_start_region() information on "continues humongous"
2771 // regions is correct.
2773 class CheckClaimValuesClosure : public HeapRegionClosure {
2774 private:
2775 jint _claim_value;
2776 uint _failures;
2777 HeapRegion* _sh_region;
2779 public:
2780 CheckClaimValuesClosure(jint claim_value) :
2781 _claim_value(claim_value), _failures(0), _sh_region(NULL) { }
2782 bool doHeapRegion(HeapRegion* r) {
2783 if (r->claim_value() != _claim_value) {
2784 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2785 "claim value = %d, should be %d",
2786 HR_FORMAT_PARAMS(r),
2787 r->claim_value(), _claim_value);
2788 ++_failures;
2789 }
2790 if (!r->isHumongous()) {
2791 _sh_region = NULL;
2792 } else if (r->startsHumongous()) {
2793 _sh_region = r;
2794 } else if (r->continuesHumongous()) {
2795 if (r->humongous_start_region() != _sh_region) {
2796 gclog_or_tty->print_cr("Region " HR_FORMAT ", "
2797 "HS = "PTR_FORMAT", should be "PTR_FORMAT,
2798 HR_FORMAT_PARAMS(r),
2799 r->humongous_start_region(),
2800 _sh_region);
2801 ++_failures;
2802 }
2803 }
2804 return false;
2805 }
2806 uint failures() { return _failures; }
2807 };
2809 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) {
2810 CheckClaimValuesClosure cl(claim_value);
2811 heap_region_iterate(&cl);
2812 return cl.failures() == 0;
2813 }
2815 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure {
2816 private:
2817 jint _claim_value;
2818 uint _failures;
2820 public:
2821 CheckClaimValuesInCSetHRClosure(jint claim_value) :
2822 _claim_value(claim_value), _failures(0) { }
2824 uint failures() { return _failures; }
2826 bool doHeapRegion(HeapRegion* hr) {
2827 assert(hr->in_collection_set(), "how?");
2828 assert(!hr->isHumongous(), "H-region in CSet");
2829 if (hr->claim_value() != _claim_value) {
2830 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", "
2831 "claim value = %d, should be %d",
2832 HR_FORMAT_PARAMS(hr),
2833 hr->claim_value(), _claim_value);
2834 _failures += 1;
2835 }
2836 return false;
2837 }
2838 };
2840 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) {
2841 CheckClaimValuesInCSetHRClosure cl(claim_value);
2842 collection_set_iterate(&cl);
2843 return cl.failures() == 0;
2844 }
2845 #endif // ASSERT
2847 // Clear the cached CSet starting regions and (more importantly)
2848 // the time stamps. Called when we reset the GC time stamp.
2849 void G1CollectedHeap::clear_cset_start_regions() {
2850 assert(_worker_cset_start_region != NULL, "sanity");
2851 assert(_worker_cset_start_region_time_stamp != NULL, "sanity");
2853 int n_queues = MAX2((int)ParallelGCThreads, 1);
2854 for (int i = 0; i < n_queues; i++) {
2855 _worker_cset_start_region[i] = NULL;
2856 _worker_cset_start_region_time_stamp[i] = 0;
2857 }
2858 }
2860 // Given the id of a worker, obtain or calculate a suitable
2861 // starting region for iterating over the current collection set.
2862 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) {
2863 assert(get_gc_time_stamp() > 0, "should have been updated by now");
2865 HeapRegion* result = NULL;
2866 unsigned gc_time_stamp = get_gc_time_stamp();
2868 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) {
2869 // Cached starting region for current worker was set
2870 // during the current pause - so it's valid.
2871 // Note: the cached starting heap region may be NULL
2872 // (when the collection set is empty).
2873 result = _worker_cset_start_region[worker_i];
2874 assert(result == NULL || result->in_collection_set(), "sanity");
2875 return result;
2876 }
2878 // The cached entry was not valid so let's calculate
2879 // a suitable starting heap region for this worker.
2881 // We want the parallel threads to start their collection
2882 // set iteration at different collection set regions to
2883 // avoid contention.
2884 // If we have:
2885 // n collection set regions
2886 // p threads
2887 // Then thread t will start at region floor ((t * n) / p)
2889 result = g1_policy()->collection_set();
2890 if (G1CollectedHeap::use_parallel_gc_threads()) {
2891 uint cs_size = g1_policy()->cset_region_length();
2892 uint active_workers = workers()->active_workers();
2893 assert(UseDynamicNumberOfGCThreads ||
2894 active_workers == workers()->total_workers(),
2895 "Unless dynamic should use total workers");
2897 uint end_ind = (cs_size * worker_i) / active_workers;
2898 uint start_ind = 0;
2900 if (worker_i > 0 &&
2901 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) {
2902 // Previous workers starting region is valid
2903 // so let's iterate from there
2904 start_ind = (cs_size * (worker_i - 1)) / active_workers;
2905 result = _worker_cset_start_region[worker_i - 1];
2906 }
2908 for (uint i = start_ind; i < end_ind; i++) {
2909 result = result->next_in_collection_set();
2910 }
2911 }
2913 // Note: the calculated starting heap region may be NULL
2914 // (when the collection set is empty).
2915 assert(result == NULL || result->in_collection_set(), "sanity");
2916 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp,
2917 "should be updated only once per pause");
2918 _worker_cset_start_region[worker_i] = result;
2919 OrderAccess::storestore();
2920 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp;
2921 return result;
2922 }
2924 HeapRegion* G1CollectedHeap::start_region_for_worker(uint worker_i,
2925 uint no_of_par_workers) {
2926 uint worker_num =
2927 G1CollectedHeap::use_parallel_gc_threads() ? no_of_par_workers : 1U;
2928 assert(UseDynamicNumberOfGCThreads ||
2929 no_of_par_workers == workers()->total_workers(),
2930 "Non dynamic should use fixed number of workers");
2931 const uint start_index = n_regions() * worker_i / worker_num;
2932 return region_at(start_index);
2933 }
2935 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
2936 HeapRegion* r = g1_policy()->collection_set();
2937 while (r != NULL) {
2938 HeapRegion* next = r->next_in_collection_set();
2939 if (cl->doHeapRegion(r)) {
2940 cl->incomplete();
2941 return;
2942 }
2943 r = next;
2944 }
2945 }
2947 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r,
2948 HeapRegionClosure *cl) {
2949 if (r == NULL) {
2950 // The CSet is empty so there's nothing to do.
2951 return;
2952 }
2954 assert(r->in_collection_set(),
2955 "Start region must be a member of the collection set.");
2956 HeapRegion* cur = r;
2957 while (cur != NULL) {
2958 HeapRegion* next = cur->next_in_collection_set();
2959 if (cl->doHeapRegion(cur) && false) {
2960 cl->incomplete();
2961 return;
2962 }
2963 cur = next;
2964 }
2965 cur = g1_policy()->collection_set();
2966 while (cur != r) {
2967 HeapRegion* next = cur->next_in_collection_set();
2968 if (cl->doHeapRegion(cur) && false) {
2969 cl->incomplete();
2970 return;
2971 }
2972 cur = next;
2973 }
2974 }
2976 CompactibleSpace* G1CollectedHeap::first_compactible_space() {
2977 return n_regions() > 0 ? region_at(0) : NULL;
2978 }
2981 Space* G1CollectedHeap::space_containing(const void* addr) const {
2982 Space* res = heap_region_containing(addr);
2983 if (res == NULL)
2984 res = perm_gen()->space_containing(addr);
2985 return res;
2986 }
2988 HeapWord* G1CollectedHeap::block_start(const void* addr) const {
2989 Space* sp = space_containing(addr);
2990 if (sp != NULL) {
2991 return sp->block_start(addr);
2992 }
2993 return NULL;
2994 }
2996 size_t G1CollectedHeap::block_size(const HeapWord* addr) const {
2997 Space* sp = space_containing(addr);
2998 assert(sp != NULL, "block_size of address outside of heap");
2999 return sp->block_size(addr);
3000 }
3002 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const {
3003 Space* sp = space_containing(addr);
3004 return sp->block_is_obj(addr);
3005 }
3007 bool G1CollectedHeap::supports_tlab_allocation() const {
3008 return true;
3009 }
3011 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const {
3012 return HeapRegion::GrainBytes;
3013 }
3015 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const {
3016 // Return the remaining space in the cur alloc region, but not less than
3017 // the min TLAB size.
3019 // Also, this value can be at most the humongous object threshold,
3020 // since we can't allow tlabs to grow big enough to accomodate
3021 // humongous objects.
3023 HeapRegion* hr = _mutator_alloc_region.get();
3024 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize;
3025 if (hr == NULL) {
3026 return max_tlab_size;
3027 } else {
3028 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size);
3029 }
3030 }
3032 size_t G1CollectedHeap::max_capacity() const {
3033 return _g1_reserved.byte_size();
3034 }
3036 jlong G1CollectedHeap::millis_since_last_gc() {
3037 // assert(false, "NYI");
3038 return 0;
3039 }
3041 void G1CollectedHeap::prepare_for_verify() {
3042 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
3043 ensure_parsability(false);
3044 }
3045 g1_rem_set()->prepare_for_verify();
3046 }
3048 bool G1CollectedHeap::allocated_since_marking(oop obj, HeapRegion* hr,
3049 VerifyOption vo) {
3050 switch (vo) {
3051 case VerifyOption_G1UsePrevMarking:
3052 return hr->obj_allocated_since_prev_marking(obj);
3053 case VerifyOption_G1UseNextMarking:
3054 return hr->obj_allocated_since_next_marking(obj);
3055 case VerifyOption_G1UseMarkWord:
3056 return false;
3057 default:
3058 ShouldNotReachHere();
3059 }
3060 return false; // keep some compilers happy
3061 }
3063 HeapWord* G1CollectedHeap::top_at_mark_start(HeapRegion* hr, VerifyOption vo) {
3064 switch (vo) {
3065 case VerifyOption_G1UsePrevMarking: return hr->prev_top_at_mark_start();
3066 case VerifyOption_G1UseNextMarking: return hr->next_top_at_mark_start();
3067 case VerifyOption_G1UseMarkWord: return NULL;
3068 default: ShouldNotReachHere();
3069 }
3070 return NULL; // keep some compilers happy
3071 }
3073 bool G1CollectedHeap::is_marked(oop obj, VerifyOption vo) {
3074 switch (vo) {
3075 case VerifyOption_G1UsePrevMarking: return isMarkedPrev(obj);
3076 case VerifyOption_G1UseNextMarking: return isMarkedNext(obj);
3077 case VerifyOption_G1UseMarkWord: return obj->is_gc_marked();
3078 default: ShouldNotReachHere();
3079 }
3080 return false; // keep some compilers happy
3081 }
3083 const char* G1CollectedHeap::top_at_mark_start_str(VerifyOption vo) {
3084 switch (vo) {
3085 case VerifyOption_G1UsePrevMarking: return "PTAMS";
3086 case VerifyOption_G1UseNextMarking: return "NTAMS";
3087 case VerifyOption_G1UseMarkWord: return "NONE";
3088 default: ShouldNotReachHere();
3089 }
3090 return NULL; // keep some compilers happy
3091 }
3093 class VerifyLivenessOopClosure: public OopClosure {
3094 G1CollectedHeap* _g1h;
3095 VerifyOption _vo;
3096 public:
3097 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo):
3098 _g1h(g1h), _vo(vo)
3099 { }
3100 void do_oop(narrowOop *p) { do_oop_work(p); }
3101 void do_oop( oop *p) { do_oop_work(p); }
3103 template <class T> void do_oop_work(T *p) {
3104 oop obj = oopDesc::load_decode_heap_oop(p);
3105 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo),
3106 "Dead object referenced by a not dead object");
3107 }
3108 };
3110 class VerifyObjsInRegionClosure: public ObjectClosure {
3111 private:
3112 G1CollectedHeap* _g1h;
3113 size_t _live_bytes;
3114 HeapRegion *_hr;
3115 VerifyOption _vo;
3116 public:
3117 // _vo == UsePrevMarking -> use "prev" marking information,
3118 // _vo == UseNextMarking -> use "next" marking information,
3119 // _vo == UseMarkWord -> use mark word from object header.
3120 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo)
3121 : _live_bytes(0), _hr(hr), _vo(vo) {
3122 _g1h = G1CollectedHeap::heap();
3123 }
3124 void do_object(oop o) {
3125 VerifyLivenessOopClosure isLive(_g1h, _vo);
3126 assert(o != NULL, "Huh?");
3127 if (!_g1h->is_obj_dead_cond(o, _vo)) {
3128 // If the object is alive according to the mark word,
3129 // then verify that the marking information agrees.
3130 // Note we can't verify the contra-positive of the
3131 // above: if the object is dead (according to the mark
3132 // word), it may not be marked, or may have been marked
3133 // but has since became dead, or may have been allocated
3134 // since the last marking.
3135 if (_vo == VerifyOption_G1UseMarkWord) {
3136 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch");
3137 }
3139 o->oop_iterate(&isLive);
3140 if (!_hr->obj_allocated_since_prev_marking(o)) {
3141 size_t obj_size = o->size(); // Make sure we don't overflow
3142 _live_bytes += (obj_size * HeapWordSize);
3143 }
3144 }
3145 }
3146 size_t live_bytes() { return _live_bytes; }
3147 };
3149 class PrintObjsInRegionClosure : public ObjectClosure {
3150 HeapRegion *_hr;
3151 G1CollectedHeap *_g1;
3152 public:
3153 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) {
3154 _g1 = G1CollectedHeap::heap();
3155 };
3157 void do_object(oop o) {
3158 if (o != NULL) {
3159 HeapWord *start = (HeapWord *) o;
3160 size_t word_sz = o->size();
3161 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT
3162 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n",
3163 (void*) o, word_sz,
3164 _g1->isMarkedPrev(o),
3165 _g1->isMarkedNext(o),
3166 _hr->obj_allocated_since_prev_marking(o));
3167 HeapWord *end = start + word_sz;
3168 HeapWord *cur;
3169 int *val;
3170 for (cur = start; cur < end; cur++) {
3171 val = (int *) cur;
3172 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val);
3173 }
3174 }
3175 }
3176 };
3178 class VerifyRegionClosure: public HeapRegionClosure {
3179 private:
3180 bool _par;
3181 VerifyOption _vo;
3182 bool _failures;
3183 public:
3184 // _vo == UsePrevMarking -> use "prev" marking information,
3185 // _vo == UseNextMarking -> use "next" marking information,
3186 // _vo == UseMarkWord -> use mark word from object header.
3187 VerifyRegionClosure(bool par, VerifyOption vo)
3188 : _par(par),
3189 _vo(vo),
3190 _failures(false) {}
3192 bool failures() {
3193 return _failures;
3194 }
3196 bool doHeapRegion(HeapRegion* r) {
3197 if (!r->continuesHumongous()) {
3198 bool failures = false;
3199 r->verify(_vo, &failures);
3200 if (failures) {
3201 _failures = true;
3202 } else {
3203 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo);
3204 r->object_iterate(¬_dead_yet_cl);
3205 if (_vo != VerifyOption_G1UseNextMarking) {
3206 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) {
3207 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] "
3208 "max_live_bytes "SIZE_FORMAT" "
3209 "< calculated "SIZE_FORMAT,
3210 r->bottom(), r->end(),
3211 r->max_live_bytes(),
3212 not_dead_yet_cl.live_bytes());
3213 _failures = true;
3214 }
3215 } else {
3216 // When vo == UseNextMarking we cannot currently do a sanity
3217 // check on the live bytes as the calculation has not been
3218 // finalized yet.
3219 }
3220 }
3221 }
3222 return false; // stop the region iteration if we hit a failure
3223 }
3224 };
3226 class VerifyRootsClosure: public OopsInGenClosure {
3227 private:
3228 G1CollectedHeap* _g1h;
3229 VerifyOption _vo;
3230 bool _failures;
3231 public:
3232 // _vo == UsePrevMarking -> use "prev" marking information,
3233 // _vo == UseNextMarking -> use "next" marking information,
3234 // _vo == UseMarkWord -> use mark word from object header.
3235 VerifyRootsClosure(VerifyOption vo) :
3236 _g1h(G1CollectedHeap::heap()),
3237 _vo(vo),
3238 _failures(false) { }
3240 bool failures() { return _failures; }
3242 template <class T> void do_oop_nv(T* p) {
3243 T heap_oop = oopDesc::load_heap_oop(p);
3244 if (!oopDesc::is_null(heap_oop)) {
3245 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
3246 if (_g1h->is_obj_dead_cond(obj, _vo)) {
3247 gclog_or_tty->print_cr("Root location "PTR_FORMAT" "
3248 "points to dead obj "PTR_FORMAT, p, (void*) obj);
3249 if (_vo == VerifyOption_G1UseMarkWord) {
3250 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark()));
3251 }
3252 obj->print_on(gclog_or_tty);
3253 _failures = true;
3254 }
3255 }
3256 }
3258 void do_oop(oop* p) { do_oop_nv(p); }
3259 void do_oop(narrowOop* p) { do_oop_nv(p); }
3260 };
3262 // This is the task used for parallel heap verification.
3264 class G1ParVerifyTask: public AbstractGangTask {
3265 private:
3266 G1CollectedHeap* _g1h;
3267 VerifyOption _vo;
3268 bool _failures;
3270 public:
3271 // _vo == UsePrevMarking -> use "prev" marking information,
3272 // _vo == UseNextMarking -> use "next" marking information,
3273 // _vo == UseMarkWord -> use mark word from object header.
3274 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) :
3275 AbstractGangTask("Parallel verify task"),
3276 _g1h(g1h),
3277 _vo(vo),
3278 _failures(false) { }
3280 bool failures() {
3281 return _failures;
3282 }
3284 void work(uint worker_id) {
3285 HandleMark hm;
3286 VerifyRegionClosure blk(true, _vo);
3287 _g1h->heap_region_par_iterate_chunked(&blk, worker_id,
3288 _g1h->workers()->active_workers(),
3289 HeapRegion::ParVerifyClaimValue);
3290 if (blk.failures()) {
3291 _failures = true;
3292 }
3293 }
3294 };
3296 void G1CollectedHeap::verify(bool silent) {
3297 verify(silent, VerifyOption_G1UsePrevMarking);
3298 }
3300 void G1CollectedHeap::verify(bool silent,
3301 VerifyOption vo) {
3302 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) {
3303 if (!silent) { gclog_or_tty->print("Roots (excluding permgen) "); }
3304 VerifyRootsClosure rootsCl(vo);
3306 assert(Thread::current()->is_VM_thread(),
3307 "Expected to be executed serially by the VM thread at this point");
3309 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false);
3311 // We apply the relevant closures to all the oops in the
3312 // system dictionary, the string table and the code cache.
3313 const int so = SO_AllClasses | SO_Strings | SO_CodeCache;
3315 process_strong_roots(true, // activate StrongRootsScope
3316 true, // we set "collecting perm gen" to true,
3317 // so we don't reset the dirty cards in the perm gen.
3318 ScanningOption(so), // roots scanning options
3319 &rootsCl,
3320 &blobsCl,
3321 &rootsCl);
3323 // If we're verifying after the marking phase of a Full GC then we can't
3324 // treat the perm gen as roots into the G1 heap. Some of the objects in
3325 // the perm gen may be dead and hence not marked. If one of these dead
3326 // objects is considered to be a root then we may end up with a false
3327 // "Root location <x> points to dead ob <y>" failure.
3328 if (vo != VerifyOption_G1UseMarkWord) {
3329 // Since we used "collecting_perm_gen" == true above, we will not have
3330 // checked the refs from perm into the G1-collected heap. We check those
3331 // references explicitly below. Whether the relevant cards are dirty
3332 // is checked further below in the rem set verification.
3333 if (!silent) { gclog_or_tty->print("Permgen roots "); }
3334 perm_gen()->oop_iterate(&rootsCl);
3335 }
3336 bool failures = rootsCl.failures();
3338 if (vo != VerifyOption_G1UseMarkWord) {
3339 // If we're verifying during a full GC then the region sets
3340 // will have been torn down at the start of the GC. Therefore
3341 // verifying the region sets will fail. So we only verify
3342 // the region sets when not in a full GC.
3343 if (!silent) { gclog_or_tty->print("HeapRegionSets "); }
3344 verify_region_sets();
3345 }
3347 if (!silent) { gclog_or_tty->print("HeapRegions "); }
3348 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) {
3349 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3350 "sanity check");
3352 G1ParVerifyTask task(this, vo);
3353 assert(UseDynamicNumberOfGCThreads ||
3354 workers()->active_workers() == workers()->total_workers(),
3355 "If not dynamic should be using all the workers");
3356 int n_workers = workers()->active_workers();
3357 set_par_threads(n_workers);
3358 workers()->run_task(&task);
3359 set_par_threads(0);
3360 if (task.failures()) {
3361 failures = true;
3362 }
3364 // Checks that the expected amount of parallel work was done.
3365 // The implication is that n_workers is > 0.
3366 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue),
3367 "sanity check");
3369 reset_heap_region_claim_values();
3371 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3372 "sanity check");
3373 } else {
3374 VerifyRegionClosure blk(false, vo);
3375 heap_region_iterate(&blk);
3376 if (blk.failures()) {
3377 failures = true;
3378 }
3379 }
3380 if (!silent) gclog_or_tty->print("RemSet ");
3381 rem_set()->verify();
3383 if (failures) {
3384 gclog_or_tty->print_cr("Heap:");
3385 // It helps to have the per-region information in the output to
3386 // help us track down what went wrong. This is why we call
3387 // print_extended_on() instead of print_on().
3388 print_extended_on(gclog_or_tty);
3389 gclog_or_tty->print_cr("");
3390 #ifndef PRODUCT
3391 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) {
3392 concurrent_mark()->print_reachable("at-verification-failure",
3393 vo, false /* all */);
3394 }
3395 #endif
3396 gclog_or_tty->flush();
3397 }
3398 guarantee(!failures, "there should not have been any failures");
3399 } else {
3400 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) ");
3401 }
3402 }
3404 class PrintRegionClosure: public HeapRegionClosure {
3405 outputStream* _st;
3406 public:
3407 PrintRegionClosure(outputStream* st) : _st(st) {}
3408 bool doHeapRegion(HeapRegion* r) {
3409 r->print_on(_st);
3410 return false;
3411 }
3412 };
3414 void G1CollectedHeap::print_on(outputStream* st) const {
3415 st->print(" %-20s", "garbage-first heap");
3416 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K",
3417 capacity()/K, used_unlocked()/K);
3418 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")",
3419 _g1_storage.low_boundary(),
3420 _g1_storage.high(),
3421 _g1_storage.high_boundary());
3422 st->cr();
3423 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K);
3424 uint young_regions = _young_list->length();
3425 st->print("%u young (" SIZE_FORMAT "K), ", young_regions,
3426 (size_t) young_regions * HeapRegion::GrainBytes / K);
3427 uint survivor_regions = g1_policy()->recorded_survivor_regions();
3428 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions,
3429 (size_t) survivor_regions * HeapRegion::GrainBytes / K);
3430 st->cr();
3431 perm()->as_gen()->print_on(st);
3432 }
3434 void G1CollectedHeap::print_extended_on(outputStream* st) const {
3435 print_on(st);
3437 // Print the per-region information.
3438 st->cr();
3439 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), "
3440 "HS=humongous(starts), HC=humongous(continues), "
3441 "CS=collection set, F=free, TS=gc time stamp, "
3442 "PTAMS=previous top-at-mark-start, "
3443 "NTAMS=next top-at-mark-start)");
3444 PrintRegionClosure blk(st);
3445 heap_region_iterate(&blk);
3446 }
3448 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
3449 if (G1CollectedHeap::use_parallel_gc_threads()) {
3450 workers()->print_worker_threads_on(st);
3451 }
3452 _cmThread->print_on(st);
3453 st->cr();
3454 _cm->print_worker_threads_on(st);
3455 _cg1r->print_worker_threads_on(st);
3456 st->cr();
3457 }
3459 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
3460 if (G1CollectedHeap::use_parallel_gc_threads()) {
3461 workers()->threads_do(tc);
3462 }
3463 tc->do_thread(_cmThread);
3464 _cg1r->threads_do(tc);
3465 }
3467 void G1CollectedHeap::print_tracing_info() const {
3468 // We'll overload this to mean "trace GC pause statistics."
3469 if (TraceGen0Time || TraceGen1Time) {
3470 // The "G1CollectorPolicy" is keeping track of these stats, so delegate
3471 // to that.
3472 g1_policy()->print_tracing_info();
3473 }
3474 if (G1SummarizeRSetStats) {
3475 g1_rem_set()->print_summary_info();
3476 }
3477 if (G1SummarizeConcMark) {
3478 concurrent_mark()->print_summary_info();
3479 }
3480 g1_policy()->print_yg_surv_rate_info();
3481 SpecializationStats::print();
3482 }
3484 #ifndef PRODUCT
3485 // Helpful for debugging RSet issues.
3487 class PrintRSetsClosure : public HeapRegionClosure {
3488 private:
3489 const char* _msg;
3490 size_t _occupied_sum;
3492 public:
3493 bool doHeapRegion(HeapRegion* r) {
3494 HeapRegionRemSet* hrrs = r->rem_set();
3495 size_t occupied = hrrs->occupied();
3496 _occupied_sum += occupied;
3498 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT,
3499 HR_FORMAT_PARAMS(r));
3500 if (occupied == 0) {
3501 gclog_or_tty->print_cr(" RSet is empty");
3502 } else {
3503 hrrs->print();
3504 }
3505 gclog_or_tty->print_cr("----------");
3506 return false;
3507 }
3509 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) {
3510 gclog_or_tty->cr();
3511 gclog_or_tty->print_cr("========================================");
3512 gclog_or_tty->print_cr(msg);
3513 gclog_or_tty->cr();
3514 }
3516 ~PrintRSetsClosure() {
3517 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum);
3518 gclog_or_tty->print_cr("========================================");
3519 gclog_or_tty->cr();
3520 }
3521 };
3523 void G1CollectedHeap::print_cset_rsets() {
3524 PrintRSetsClosure cl("Printing CSet RSets");
3525 collection_set_iterate(&cl);
3526 }
3528 void G1CollectedHeap::print_all_rsets() {
3529 PrintRSetsClosure cl("Printing All RSets");;
3530 heap_region_iterate(&cl);
3531 }
3532 #endif // PRODUCT
3534 G1CollectedHeap* G1CollectedHeap::heap() {
3535 assert(_sh->kind() == CollectedHeap::G1CollectedHeap,
3536 "not a garbage-first heap");
3537 return _g1h;
3538 }
3540 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) {
3541 // always_do_update_barrier = false;
3542 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer");
3543 // Call allocation profiler
3544 AllocationProfiler::iterate_since_last_gc();
3545 // Fill TLAB's and such
3546 ensure_parsability(true);
3547 }
3549 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) {
3550 // FIXME: what is this about?
3551 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled"
3552 // is set.
3553 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(),
3554 "derived pointer present"));
3555 // always_do_update_barrier = true;
3557 // We have just completed a GC. Update the soft reference
3558 // policy with the new heap occupancy
3559 Universe::update_heap_info_at_gc();
3560 }
3562 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size,
3563 unsigned int gc_count_before,
3564 bool* succeeded) {
3565 assert_heap_not_locked_and_not_at_safepoint();
3566 g1_policy()->record_stop_world_start();
3567 VM_G1IncCollectionPause op(gc_count_before,
3568 word_size,
3569 false, /* should_initiate_conc_mark */
3570 g1_policy()->max_pause_time_ms(),
3571 GCCause::_g1_inc_collection_pause);
3572 VMThread::execute(&op);
3574 HeapWord* result = op.result();
3575 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded();
3576 assert(result == NULL || ret_succeeded,
3577 "the result should be NULL if the VM did not succeed");
3578 *succeeded = ret_succeeded;
3580 assert_heap_not_locked();
3581 return result;
3582 }
3584 void
3585 G1CollectedHeap::doConcurrentMark() {
3586 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
3587 if (!_cmThread->in_progress()) {
3588 _cmThread->set_started();
3589 CGC_lock->notify();
3590 }
3591 }
3593 size_t G1CollectedHeap::pending_card_num() {
3594 size_t extra_cards = 0;
3595 JavaThread *curr = Threads::first();
3596 while (curr != NULL) {
3597 DirtyCardQueue& dcq = curr->dirty_card_queue();
3598 extra_cards += dcq.size();
3599 curr = curr->next();
3600 }
3601 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
3602 size_t buffer_size = dcqs.buffer_size();
3603 size_t buffer_num = dcqs.completed_buffers_num();
3605 // PtrQueueSet::buffer_size() and PtrQueue:size() return sizes
3606 // in bytes - not the number of 'entries'. We need to convert
3607 // into a number of cards.
3608 return (buffer_size * buffer_num + extra_cards) / oopSize;
3609 }
3611 size_t G1CollectedHeap::cards_scanned() {
3612 return g1_rem_set()->cardsScanned();
3613 }
3615 void
3616 G1CollectedHeap::setup_surviving_young_words() {
3617 assert(_surviving_young_words == NULL, "pre-condition");
3618 uint array_length = g1_policy()->young_cset_region_length();
3619 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC);
3620 if (_surviving_young_words == NULL) {
3621 vm_exit_out_of_memory(sizeof(size_t) * array_length,
3622 "Not enough space for young surv words summary.");
3623 }
3624 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t));
3625 #ifdef ASSERT
3626 for (uint i = 0; i < array_length; ++i) {
3627 assert( _surviving_young_words[i] == 0, "memset above" );
3628 }
3629 #endif // !ASSERT
3630 }
3632 void
3633 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) {
3634 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
3635 uint array_length = g1_policy()->young_cset_region_length();
3636 for (uint i = 0; i < array_length; ++i) {
3637 _surviving_young_words[i] += surv_young_words[i];
3638 }
3639 }
3641 void
3642 G1CollectedHeap::cleanup_surviving_young_words() {
3643 guarantee( _surviving_young_words != NULL, "pre-condition" );
3644 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC);
3645 _surviving_young_words = NULL;
3646 }
3648 #ifdef ASSERT
3649 class VerifyCSetClosure: public HeapRegionClosure {
3650 public:
3651 bool doHeapRegion(HeapRegion* hr) {
3652 // Here we check that the CSet region's RSet is ready for parallel
3653 // iteration. The fields that we'll verify are only manipulated
3654 // when the region is part of a CSet and is collected. Afterwards,
3655 // we reset these fields when we clear the region's RSet (when the
3656 // region is freed) so they are ready when the region is
3657 // re-allocated. The only exception to this is if there's an
3658 // evacuation failure and instead of freeing the region we leave
3659 // it in the heap. In that case, we reset these fields during
3660 // evacuation failure handling.
3661 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification");
3663 // Here's a good place to add any other checks we'd like to
3664 // perform on CSet regions.
3665 return false;
3666 }
3667 };
3668 #endif // ASSERT
3670 #if TASKQUEUE_STATS
3671 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) {
3672 st->print_raw_cr("GC Task Stats");
3673 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
3674 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
3675 }
3677 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const {
3678 print_taskqueue_stats_hdr(st);
3680 TaskQueueStats totals;
3681 const int n = workers() != NULL ? workers()->total_workers() : 1;
3682 for (int i = 0; i < n; ++i) {
3683 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr();
3684 totals += task_queue(i)->stats;
3685 }
3686 st->print_raw("tot "); totals.print(st); st->cr();
3688 DEBUG_ONLY(totals.verify());
3689 }
3691 void G1CollectedHeap::reset_taskqueue_stats() {
3692 const int n = workers() != NULL ? workers()->total_workers() : 1;
3693 for (int i = 0; i < n; ++i) {
3694 task_queue(i)->stats.reset();
3695 }
3696 }
3697 #endif // TASKQUEUE_STATS
3699 bool
3700 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) {
3701 assert_at_safepoint(true /* should_be_vm_thread */);
3702 guarantee(!is_gc_active(), "collection is not reentrant");
3704 if (GC_locker::check_active_before_gc()) {
3705 return false;
3706 }
3708 SvcGCMarker sgcm(SvcGCMarker::MINOR);
3709 ResourceMark rm;
3711 print_heap_before_gc();
3713 HRSPhaseSetter x(HRSPhaseEvacuation);
3714 verify_region_sets_optional();
3715 verify_dirty_young_regions();
3717 // This call will decide whether this pause is an initial-mark
3718 // pause. If it is, during_initial_mark_pause() will return true
3719 // for the duration of this pause.
3720 g1_policy()->decide_on_conc_mark_initiation();
3722 // We do not allow initial-mark to be piggy-backed on a mixed GC.
3723 assert(!g1_policy()->during_initial_mark_pause() ||
3724 g1_policy()->gcs_are_young(), "sanity");
3726 // We also do not allow mixed GCs during marking.
3727 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity");
3729 // Record whether this pause is an initial mark. When the current
3730 // thread has completed its logging output and it's safe to signal
3731 // the CM thread, the flag's value in the policy has been reset.
3732 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause();
3734 // Inner scope for scope based logging, timers, and stats collection
3735 {
3736 if (g1_policy()->during_initial_mark_pause()) {
3737 // We are about to start a marking cycle, so we increment the
3738 // full collection counter.
3739 increment_old_marking_cycles_started();
3740 }
3741 // if the log level is "finer" is on, we'll print long statistics information
3742 // in the collector policy code, so let's not print this as the output
3743 // is messy if we do.
3744 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps);
3745 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty);
3747 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3748 workers()->active_workers() : 1);
3749 double pause_start_sec = os::elapsedTime();
3750 g1_policy()->phase_times()->note_gc_start(active_workers);
3751 bool initial_mark_gc = g1_policy()->during_initial_mark_pause();
3753 TraceCollectorStats tcs(g1mm()->incremental_collection_counters());
3754 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause());
3756 // If the secondary_free_list is not empty, append it to the
3757 // free_list. No need to wait for the cleanup operation to finish;
3758 // the region allocation code will check the secondary_free_list
3759 // and wait if necessary. If the G1StressConcRegionFreeing flag is
3760 // set, skip this step so that the region allocation code has to
3761 // get entries from the secondary_free_list.
3762 if (!G1StressConcRegionFreeing) {
3763 append_secondary_free_list_if_not_empty_with_lock();
3764 }
3766 assert(check_young_list_well_formed(),
3767 "young list should be well formed");
3769 // Don't dynamically change the number of GC threads this early. A value of
3770 // 0 is used to indicate serial work. When parallel work is done,
3771 // it will be set.
3773 { // Call to jvmpi::post_class_unload_events must occur outside of active GC
3774 IsGCActiveMark x;
3776 gc_prologue(false);
3777 increment_total_collections(false /* full gc */);
3778 increment_gc_time_stamp();
3780 verify_before_gc();
3782 COMPILER2_PRESENT(DerivedPointerTable::clear());
3784 // Please see comment in g1CollectedHeap.hpp and
3785 // G1CollectedHeap::ref_processing_init() to see how
3786 // reference processing currently works in G1.
3788 // Enable discovery in the STW reference processor
3789 ref_processor_stw()->enable_discovery(true /*verify_disabled*/,
3790 true /*verify_no_refs*/);
3792 {
3793 // We want to temporarily turn off discovery by the
3794 // CM ref processor, if necessary, and turn it back on
3795 // on again later if we do. Using a scoped
3796 // NoRefDiscovery object will do this.
3797 NoRefDiscovery no_cm_discovery(ref_processor_cm());
3799 // Forget the current alloc region (we might even choose it to be part
3800 // of the collection set!).
3801 release_mutator_alloc_region();
3803 // We should call this after we retire the mutator alloc
3804 // region(s) so that all the ALLOC / RETIRE events are generated
3805 // before the start GC event.
3806 _hr_printer.start_gc(false /* full */, (size_t) total_collections());
3808 // This timing is only used by the ergonomics to handle our pause target.
3809 // It is unclear why this should not include the full pause. We will
3810 // investigate this in CR 7178365.
3811 //
3812 // Preserving the old comment here if that helps the investigation:
3813 //
3814 // The elapsed time induced by the start time below deliberately elides
3815 // the possible verification above.
3816 double sample_start_time_sec = os::elapsedTime();
3817 size_t start_used_bytes = used();
3819 #if YOUNG_LIST_VERBOSE
3820 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:");
3821 _young_list->print();
3822 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3823 #endif // YOUNG_LIST_VERBOSE
3825 g1_policy()->record_collection_pause_start(sample_start_time_sec,
3826 start_used_bytes);
3828 double scan_wait_start = os::elapsedTime();
3829 // We have to wait until the CM threads finish scanning the
3830 // root regions as it's the only way to ensure that all the
3831 // objects on them have been correctly scanned before we start
3832 // moving them during the GC.
3833 bool waited = _cm->root_regions()->wait_until_scan_finished();
3834 double wait_time_ms = 0.0;
3835 if (waited) {
3836 double scan_wait_end = os::elapsedTime();
3837 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0;
3838 }
3839 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms);
3841 #if YOUNG_LIST_VERBOSE
3842 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:");
3843 _young_list->print();
3844 #endif // YOUNG_LIST_VERBOSE
3846 if (g1_policy()->during_initial_mark_pause()) {
3847 concurrent_mark()->checkpointRootsInitialPre();
3848 }
3849 perm_gen()->save_marks();
3851 #if YOUNG_LIST_VERBOSE
3852 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:");
3853 _young_list->print();
3854 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3855 #endif // YOUNG_LIST_VERBOSE
3857 g1_policy()->finalize_cset(target_pause_time_ms);
3859 _cm->note_start_of_gc();
3860 // We should not verify the per-thread SATB buffers given that
3861 // we have not filtered them yet (we'll do so during the
3862 // GC). We also call this after finalize_cset() to
3863 // ensure that the CSet has been finalized.
3864 _cm->verify_no_cset_oops(true /* verify_stacks */,
3865 true /* verify_enqueued_buffers */,
3866 false /* verify_thread_buffers */,
3867 true /* verify_fingers */);
3869 if (_hr_printer.is_active()) {
3870 HeapRegion* hr = g1_policy()->collection_set();
3871 while (hr != NULL) {
3872 G1HRPrinter::RegionType type;
3873 if (!hr->is_young()) {
3874 type = G1HRPrinter::Old;
3875 } else if (hr->is_survivor()) {
3876 type = G1HRPrinter::Survivor;
3877 } else {
3878 type = G1HRPrinter::Eden;
3879 }
3880 _hr_printer.cset(hr);
3881 hr = hr->next_in_collection_set();
3882 }
3883 }
3885 #ifdef ASSERT
3886 VerifyCSetClosure cl;
3887 collection_set_iterate(&cl);
3888 #endif // ASSERT
3890 setup_surviving_young_words();
3892 // Initialize the GC alloc regions.
3893 init_gc_alloc_regions();
3895 // Actually do the work...
3896 evacuate_collection_set();
3898 // We do this to mainly verify the per-thread SATB buffers
3899 // (which have been filtered by now) since we didn't verify
3900 // them earlier. No point in re-checking the stacks / enqueued
3901 // buffers given that the CSet has not changed since last time
3902 // we checked.
3903 _cm->verify_no_cset_oops(false /* verify_stacks */,
3904 false /* verify_enqueued_buffers */,
3905 true /* verify_thread_buffers */,
3906 true /* verify_fingers */);
3908 free_collection_set(g1_policy()->collection_set());
3909 g1_policy()->clear_collection_set();
3911 cleanup_surviving_young_words();
3913 // Start a new incremental collection set for the next pause.
3914 g1_policy()->start_incremental_cset_building();
3916 // Clear the _cset_fast_test bitmap in anticipation of adding
3917 // regions to the incremental collection set for the next
3918 // evacuation pause.
3919 clear_cset_fast_test();
3921 _young_list->reset_sampled_info();
3923 // Don't check the whole heap at this point as the
3924 // GC alloc regions from this pause have been tagged
3925 // as survivors and moved on to the survivor list.
3926 // Survivor regions will fail the !is_young() check.
3927 assert(check_young_list_empty(false /* check_heap */),
3928 "young list should be empty");
3930 #if YOUNG_LIST_VERBOSE
3931 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:");
3932 _young_list->print();
3933 #endif // YOUNG_LIST_VERBOSE
3935 g1_policy()->record_survivor_regions(_young_list->survivor_length(),
3936 _young_list->first_survivor_region(),
3937 _young_list->last_survivor_region());
3939 _young_list->reset_auxilary_lists();
3941 if (evacuation_failed()) {
3942 _summary_bytes_used = recalculate_used();
3943 } else {
3944 // The "used" of the the collection set have already been subtracted
3945 // when they were freed. Add in the bytes evacuated.
3946 _summary_bytes_used += g1_policy()->bytes_copied_during_gc();
3947 }
3949 if (g1_policy()->during_initial_mark_pause()) {
3950 // We have to do this before we notify the CM threads that
3951 // they can start working to make sure that all the
3952 // appropriate initialization is done on the CM object.
3953 concurrent_mark()->checkpointRootsInitialPost();
3954 set_marking_started();
3955 // Note that we don't actually trigger the CM thread at
3956 // this point. We do that later when we're sure that
3957 // the current thread has completed its logging output.
3958 }
3960 allocate_dummy_regions();
3962 #if YOUNG_LIST_VERBOSE
3963 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:");
3964 _young_list->print();
3965 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty);
3966 #endif // YOUNG_LIST_VERBOSE
3968 init_mutator_alloc_region();
3970 {
3971 size_t expand_bytes = g1_policy()->expansion_amount();
3972 if (expand_bytes > 0) {
3973 size_t bytes_before = capacity();
3974 // No need for an ergo verbose message here,
3975 // expansion_amount() does this when it returns a value > 0.
3976 if (!expand(expand_bytes)) {
3977 // We failed to expand the heap so let's verify that
3978 // committed/uncommitted amount match the backing store
3979 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch");
3980 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch");
3981 }
3982 }
3983 }
3985 // We redo the verificaiton but now wrt to the new CSet which
3986 // has just got initialized after the previous CSet was freed.
3987 _cm->verify_no_cset_oops(true /* verify_stacks */,
3988 true /* verify_enqueued_buffers */,
3989 true /* verify_thread_buffers */,
3990 true /* verify_fingers */);
3991 _cm->note_end_of_gc();
3993 // This timing is only used by the ergonomics to handle our pause target.
3994 // It is unclear why this should not include the full pause. We will
3995 // investigate this in CR 7178365.
3996 double sample_end_time_sec = os::elapsedTime();
3997 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS;
3998 g1_policy()->record_collection_pause_end(pause_time_ms);
4000 MemoryService::track_memory_usage();
4002 // In prepare_for_verify() below we'll need to scan the deferred
4003 // update buffers to bring the RSets up-to-date if
4004 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning
4005 // the update buffers we'll probably need to scan cards on the
4006 // regions we just allocated to (i.e., the GC alloc
4007 // regions). However, during the last GC we called
4008 // set_saved_mark() on all the GC alloc regions, so card
4009 // scanning might skip the [saved_mark_word()...top()] area of
4010 // those regions (i.e., the area we allocated objects into
4011 // during the last GC). But it shouldn't. Given that
4012 // saved_mark_word() is conditional on whether the GC time stamp
4013 // on the region is current or not, by incrementing the GC time
4014 // stamp here we invalidate all the GC time stamps on all the
4015 // regions and saved_mark_word() will simply return top() for
4016 // all the regions. This is a nicer way of ensuring this rather
4017 // than iterating over the regions and fixing them. In fact, the
4018 // GC time stamp increment here also ensures that
4019 // saved_mark_word() will return top() between pauses, i.e.,
4020 // during concurrent refinement. So we don't need the
4021 // is_gc_active() check to decided which top to use when
4022 // scanning cards (see CR 7039627).
4023 increment_gc_time_stamp();
4025 verify_after_gc();
4027 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition");
4028 ref_processor_stw()->verify_no_references_recorded();
4030 // CM reference discovery will be re-enabled if necessary.
4031 }
4033 // We should do this after we potentially expand the heap so
4034 // that all the COMMIT events are generated before the end GC
4035 // event, and after we retire the GC alloc regions so that all
4036 // RETIRE events are generated before the end GC event.
4037 _hr_printer.end_gc(false /* full */, (size_t) total_collections());
4039 if (mark_in_progress()) {
4040 concurrent_mark()->update_g1_committed();
4041 }
4043 #ifdef TRACESPINNING
4044 ParallelTaskTerminator::print_termination_counts();
4045 #endif
4047 gc_epilogue(false);
4049 if (G1Log::fine()) {
4050 if (PrintGCTimeStamps) {
4051 gclog_or_tty->stamp();
4052 gclog_or_tty->print(": ");
4053 }
4055 GCCauseString gc_cause_str = GCCauseString("GC pause", gc_cause())
4056 .append(g1_policy()->gcs_are_young() ? " (young)" : " (mixed)")
4057 .append(initial_mark_gc ? " (initial-mark)" : "");
4059 double pause_time_sec = os::elapsedTime() - pause_start_sec;
4061 if (G1Log::finer()) {
4062 if (evacuation_failed()) {
4063 gc_cause_str.append(" (to-space exhausted)");
4064 }
4065 gclog_or_tty->print_cr("[%s, %3.7f secs]", (const char*)gc_cause_str, pause_time_sec);
4066 g1_policy()->phase_times()->note_gc_end();
4067 g1_policy()->phase_times()->print(pause_time_sec);
4068 g1_policy()->print_detailed_heap_transition();
4069 } else {
4070 if (evacuation_failed()) {
4071 gc_cause_str.append("--");
4072 }
4073 gclog_or_tty->print("[%s", (const char*)gc_cause_str);
4074 g1_policy()->print_heap_transition();
4075 gclog_or_tty->print_cr(", %3.7f secs]", pause_time_sec);
4076 }
4077 }
4078 }
4080 // It is not yet to safe to tell the concurrent mark to
4081 // start as we have some optional output below. We don't want the
4082 // output from the concurrent mark thread interfering with this
4083 // logging output either.
4085 _hrs.verify_optional();
4086 verify_region_sets_optional();
4088 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats());
4089 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats());
4091 print_heap_after_gc();
4093 // We must call G1MonitoringSupport::update_sizes() in the same scoping level
4094 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the
4095 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated
4096 // before any GC notifications are raised.
4097 g1mm()->update_sizes();
4098 }
4100 if (G1SummarizeRSetStats &&
4101 (G1SummarizeRSetStatsPeriod > 0) &&
4102 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) {
4103 g1_rem_set()->print_summary_info();
4104 }
4106 // It should now be safe to tell the concurrent mark thread to start
4107 // without its logging output interfering with the logging output
4108 // that came from the pause.
4110 if (should_start_conc_mark) {
4111 // CAUTION: after the doConcurrentMark() call below,
4112 // the concurrent marking thread(s) could be running
4113 // concurrently with us. Make sure that anything after
4114 // this point does not assume that we are the only GC thread
4115 // running. Note: of course, the actual marking work will
4116 // not start until the safepoint itself is released in
4117 // ConcurrentGCThread::safepoint_desynchronize().
4118 doConcurrentMark();
4119 }
4121 return true;
4122 }
4124 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose)
4125 {
4126 size_t gclab_word_size;
4127 switch (purpose) {
4128 case GCAllocForSurvived:
4129 gclab_word_size = _survivor_plab_stats.desired_plab_sz();
4130 break;
4131 case GCAllocForTenured:
4132 gclab_word_size = _old_plab_stats.desired_plab_sz();
4133 break;
4134 default:
4135 assert(false, "unknown GCAllocPurpose");
4136 gclab_word_size = _old_plab_stats.desired_plab_sz();
4137 break;
4138 }
4140 // Prevent humongous PLAB sizes for two reasons:
4141 // * PLABs are allocated using a similar paths as oops, but should
4142 // never be in a humongous region
4143 // * Allowing humongous PLABs needlessly churns the region free lists
4144 return MIN2(_humongous_object_threshold_in_words, gclab_word_size);
4145 }
4147 void G1CollectedHeap::init_mutator_alloc_region() {
4148 assert(_mutator_alloc_region.get() == NULL, "pre-condition");
4149 _mutator_alloc_region.init();
4150 }
4152 void G1CollectedHeap::release_mutator_alloc_region() {
4153 _mutator_alloc_region.release();
4154 assert(_mutator_alloc_region.get() == NULL, "post-condition");
4155 }
4157 void G1CollectedHeap::init_gc_alloc_regions() {
4158 assert_at_safepoint(true /* should_be_vm_thread */);
4160 _survivor_gc_alloc_region.init();
4161 _old_gc_alloc_region.init();
4162 HeapRegion* retained_region = _retained_old_gc_alloc_region;
4163 _retained_old_gc_alloc_region = NULL;
4165 // We will discard the current GC alloc region if:
4166 // a) it's in the collection set (it can happen!),
4167 // b) it's already full (no point in using it),
4168 // c) it's empty (this means that it was emptied during
4169 // a cleanup and it should be on the free list now), or
4170 // d) it's humongous (this means that it was emptied
4171 // during a cleanup and was added to the free list, but
4172 // has been subseqently used to allocate a humongous
4173 // object that may be less than the region size).
4174 if (retained_region != NULL &&
4175 !retained_region->in_collection_set() &&
4176 !(retained_region->top() == retained_region->end()) &&
4177 !retained_region->is_empty() &&
4178 !retained_region->isHumongous()) {
4179 retained_region->set_saved_mark();
4180 // The retained region was added to the old region set when it was
4181 // retired. We have to remove it now, since we don't allow regions
4182 // we allocate to in the region sets. We'll re-add it later, when
4183 // it's retired again.
4184 _old_set.remove(retained_region);
4185 bool during_im = g1_policy()->during_initial_mark_pause();
4186 retained_region->note_start_of_copying(during_im);
4187 _old_gc_alloc_region.set(retained_region);
4188 _hr_printer.reuse(retained_region);
4189 }
4190 }
4192 void G1CollectedHeap::release_gc_alloc_regions() {
4193 _survivor_gc_alloc_region.release();
4194 // If we have an old GC alloc region to release, we'll save it in
4195 // _retained_old_gc_alloc_region. If we don't
4196 // _retained_old_gc_alloc_region will become NULL. This is what we
4197 // want either way so no reason to check explicitly for either
4198 // condition.
4199 _retained_old_gc_alloc_region = _old_gc_alloc_region.release();
4201 if (ResizePLAB) {
4202 _survivor_plab_stats.adjust_desired_plab_sz();
4203 _old_plab_stats.adjust_desired_plab_sz();
4204 }
4205 }
4207 void G1CollectedHeap::abandon_gc_alloc_regions() {
4208 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition");
4209 assert(_old_gc_alloc_region.get() == NULL, "pre-condition");
4210 _retained_old_gc_alloc_region = NULL;
4211 }
4213 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) {
4214 _drain_in_progress = false;
4215 set_evac_failure_closure(cl);
4216 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4217 }
4219 void G1CollectedHeap::finalize_for_evac_failure() {
4220 assert(_evac_failure_scan_stack != NULL &&
4221 _evac_failure_scan_stack->length() == 0,
4222 "Postcondition");
4223 assert(!_drain_in_progress, "Postcondition");
4224 delete _evac_failure_scan_stack;
4225 _evac_failure_scan_stack = NULL;
4226 }
4228 void G1CollectedHeap::remove_self_forwarding_pointers() {
4229 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4231 G1ParRemoveSelfForwardPtrsTask rsfp_task(this);
4233 if (G1CollectedHeap::use_parallel_gc_threads()) {
4234 set_par_threads();
4235 workers()->run_task(&rsfp_task);
4236 set_par_threads(0);
4237 } else {
4238 rsfp_task.work(0);
4239 }
4241 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity");
4243 // Reset the claim values in the regions in the collection set.
4244 reset_cset_heap_region_claim_values();
4246 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity");
4248 // Now restore saved marks, if any.
4249 if (_objs_with_preserved_marks != NULL) {
4250 assert(_preserved_marks_of_objs != NULL, "Both or none.");
4251 guarantee(_objs_with_preserved_marks->length() ==
4252 _preserved_marks_of_objs->length(), "Both or none.");
4253 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) {
4254 oop obj = _objs_with_preserved_marks->at(i);
4255 markOop m = _preserved_marks_of_objs->at(i);
4256 obj->set_mark(m);
4257 }
4259 // Delete the preserved marks growable arrays (allocated on the C heap).
4260 delete _objs_with_preserved_marks;
4261 delete _preserved_marks_of_objs;
4262 _objs_with_preserved_marks = NULL;
4263 _preserved_marks_of_objs = NULL;
4264 }
4265 }
4267 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) {
4268 _evac_failure_scan_stack->push(obj);
4269 }
4271 void G1CollectedHeap::drain_evac_failure_scan_stack() {
4272 assert(_evac_failure_scan_stack != NULL, "precondition");
4274 while (_evac_failure_scan_stack->length() > 0) {
4275 oop obj = _evac_failure_scan_stack->pop();
4276 _evac_failure_closure->set_region(heap_region_containing(obj));
4277 obj->oop_iterate_backwards(_evac_failure_closure);
4278 }
4279 }
4281 oop
4282 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl,
4283 oop old) {
4284 assert(obj_in_cs(old),
4285 err_msg("obj: "PTR_FORMAT" should still be in the CSet",
4286 (HeapWord*) old));
4287 markOop m = old->mark();
4288 oop forward_ptr = old->forward_to_atomic(old);
4289 if (forward_ptr == NULL) {
4290 // Forward-to-self succeeded.
4292 if (_evac_failure_closure != cl) {
4293 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag);
4294 assert(!_drain_in_progress,
4295 "Should only be true while someone holds the lock.");
4296 // Set the global evac-failure closure to the current thread's.
4297 assert(_evac_failure_closure == NULL, "Or locking has failed.");
4298 set_evac_failure_closure(cl);
4299 // Now do the common part.
4300 handle_evacuation_failure_common(old, m);
4301 // Reset to NULL.
4302 set_evac_failure_closure(NULL);
4303 } else {
4304 // The lock is already held, and this is recursive.
4305 assert(_drain_in_progress, "This should only be the recursive case.");
4306 handle_evacuation_failure_common(old, m);
4307 }
4308 return old;
4309 } else {
4310 // Forward-to-self failed. Either someone else managed to allocate
4311 // space for this object (old != forward_ptr) or they beat us in
4312 // self-forwarding it (old == forward_ptr).
4313 assert(old == forward_ptr || !obj_in_cs(forward_ptr),
4314 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" "
4315 "should not be in the CSet",
4316 (HeapWord*) old, (HeapWord*) forward_ptr));
4317 return forward_ptr;
4318 }
4319 }
4321 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) {
4322 set_evacuation_failed(true);
4324 preserve_mark_if_necessary(old, m);
4326 HeapRegion* r = heap_region_containing(old);
4327 if (!r->evacuation_failed()) {
4328 r->set_evacuation_failed(true);
4329 _hr_printer.evac_failure(r);
4330 }
4332 push_on_evac_failure_scan_stack(old);
4334 if (!_drain_in_progress) {
4335 // prevent recursion in copy_to_survivor_space()
4336 _drain_in_progress = true;
4337 drain_evac_failure_scan_stack();
4338 _drain_in_progress = false;
4339 }
4340 }
4342 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) {
4343 assert(evacuation_failed(), "Oversaving!");
4344 // We want to call the "for_promotion_failure" version only in the
4345 // case of a promotion failure.
4346 if (m->must_be_preserved_for_promotion_failure(obj)) {
4347 if (_objs_with_preserved_marks == NULL) {
4348 assert(_preserved_marks_of_objs == NULL, "Both or none.");
4349 _objs_with_preserved_marks =
4350 new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true);
4351 _preserved_marks_of_objs =
4352 new (ResourceObj::C_HEAP, mtGC) GrowableArray<markOop>(40, true);
4353 }
4354 _objs_with_preserved_marks->push(obj);
4355 _preserved_marks_of_objs->push(m);
4356 }
4357 }
4359 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose,
4360 size_t word_size) {
4361 if (purpose == GCAllocForSurvived) {
4362 HeapWord* result = survivor_attempt_allocation(word_size);
4363 if (result != NULL) {
4364 return result;
4365 } else {
4366 // Let's try to allocate in the old gen in case we can fit the
4367 // object there.
4368 return old_attempt_allocation(word_size);
4369 }
4370 } else {
4371 assert(purpose == GCAllocForTenured, "sanity");
4372 HeapWord* result = old_attempt_allocation(word_size);
4373 if (result != NULL) {
4374 return result;
4375 } else {
4376 // Let's try to allocate in the survivors in case we can fit the
4377 // object there.
4378 return survivor_attempt_allocation(word_size);
4379 }
4380 }
4382 ShouldNotReachHere();
4383 // Trying to keep some compilers happy.
4384 return NULL;
4385 }
4387 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) :
4388 ParGCAllocBuffer(gclab_word_size), _retired(false) { }
4390 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num)
4391 : _g1h(g1h),
4392 _refs(g1h->task_queue(queue_num)),
4393 _dcq(&g1h->dirty_card_queue_set()),
4394 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
4395 _g1_rem(g1h->g1_rem_set()),
4396 _hash_seed(17), _queue_num(queue_num),
4397 _term_attempts(0),
4398 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)),
4399 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)),
4400 _age_table(false),
4401 _strong_roots_time(0), _term_time(0),
4402 _alloc_buffer_waste(0), _undo_waste(0) {
4403 // we allocate G1YoungSurvRateNumRegions plus one entries, since
4404 // we "sacrifice" entry 0 to keep track of surviving bytes for
4405 // non-young regions (where the age is -1)
4406 // We also add a few elements at the beginning and at the end in
4407 // an attempt to eliminate cache contention
4408 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length();
4409 uint array_length = PADDING_ELEM_NUM +
4410 real_length +
4411 PADDING_ELEM_NUM;
4412 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC);
4413 if (_surviving_young_words_base == NULL)
4414 vm_exit_out_of_memory(array_length * sizeof(size_t),
4415 "Not enough space for young surv histo.");
4416 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM;
4417 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t));
4419 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer;
4420 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer;
4422 _start = os::elapsedTime();
4423 }
4425 void
4426 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st)
4427 {
4428 st->print_raw_cr("GC Termination Stats");
4429 st->print_raw_cr(" elapsed --strong roots-- -------termination-------"
4430 " ------waste (KiB)------");
4431 st->print_raw_cr("thr ms ms % ms % attempts"
4432 " total alloc undo");
4433 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------"
4434 " ------- ------- -------");
4435 }
4437 void
4438 G1ParScanThreadState::print_termination_stats(int i,
4439 outputStream* const st) const
4440 {
4441 const double elapsed_ms = elapsed_time() * 1000.0;
4442 const double s_roots_ms = strong_roots_time() * 1000.0;
4443 const double term_ms = term_time() * 1000.0;
4444 st->print_cr("%3d %9.2f %9.2f %6.2f "
4445 "%9.2f %6.2f " SIZE_FORMAT_W(8) " "
4446 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7),
4447 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
4448 term_ms, term_ms * 100 / elapsed_ms, term_attempts(),
4449 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K,
4450 alloc_buffer_waste() * HeapWordSize / K,
4451 undo_waste() * HeapWordSize / K);
4452 }
4454 #ifdef ASSERT
4455 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const {
4456 assert(ref != NULL, "invariant");
4457 assert(UseCompressedOops, "sanity");
4458 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref));
4459 oop p = oopDesc::load_decode_heap_oop(ref);
4460 assert(_g1h->is_in_g1_reserved(p),
4461 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4462 return true;
4463 }
4465 bool G1ParScanThreadState::verify_ref(oop* ref) const {
4466 assert(ref != NULL, "invariant");
4467 if (has_partial_array_mask(ref)) {
4468 // Must be in the collection set--it's already been copied.
4469 oop p = clear_partial_array_mask(ref);
4470 assert(_g1h->obj_in_cs(p),
4471 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4472 } else {
4473 oop p = oopDesc::load_decode_heap_oop(ref);
4474 assert(_g1h->is_in_g1_reserved(p),
4475 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p)));
4476 }
4477 return true;
4478 }
4480 bool G1ParScanThreadState::verify_task(StarTask ref) const {
4481 if (ref.is_narrow()) {
4482 return verify_ref((narrowOop*) ref);
4483 } else {
4484 return verify_ref((oop*) ref);
4485 }
4486 }
4487 #endif // ASSERT
4489 void G1ParScanThreadState::trim_queue() {
4490 assert(_evac_cl != NULL, "not set");
4491 assert(_evac_failure_cl != NULL, "not set");
4492 assert(_partial_scan_cl != NULL, "not set");
4494 StarTask ref;
4495 do {
4496 // Drain the overflow stack first, so other threads can steal.
4497 while (refs()->pop_overflow(ref)) {
4498 deal_with_reference(ref);
4499 }
4501 while (refs()->pop_local(ref)) {
4502 deal_with_reference(ref);
4503 }
4504 } while (!refs()->is_empty());
4505 }
4507 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1,
4508 G1ParScanThreadState* par_scan_state) :
4509 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
4510 _par_scan_state(par_scan_state),
4511 _worker_id(par_scan_state->queue_num()),
4512 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()),
4513 _mark_in_progress(_g1->mark_in_progress()) { }
4515 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4516 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) {
4517 #ifdef ASSERT
4518 HeapRegion* hr = _g1->heap_region_containing(obj);
4519 assert(hr != NULL, "sanity");
4520 assert(!hr->in_collection_set(), "should not mark objects in the CSet");
4521 #endif // ASSERT
4523 // We know that the object is not moving so it's safe to read its size.
4524 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id);
4525 }
4527 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4528 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4529 ::mark_forwarded_object(oop from_obj, oop to_obj) {
4530 #ifdef ASSERT
4531 assert(from_obj->is_forwarded(), "from obj should be forwarded");
4532 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee");
4533 assert(from_obj != to_obj, "should not be self-forwarded");
4535 HeapRegion* from_hr = _g1->heap_region_containing(from_obj);
4536 assert(from_hr != NULL, "sanity");
4537 assert(from_hr->in_collection_set(), "from obj should be in the CSet");
4539 HeapRegion* to_hr = _g1->heap_region_containing(to_obj);
4540 assert(to_hr != NULL, "sanity");
4541 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet");
4542 #endif // ASSERT
4544 // The object might be in the process of being copied by another
4545 // worker so we cannot trust that its to-space image is
4546 // well-formed. So we have to read its size from its from-space
4547 // image which we know should not be changing.
4548 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id);
4549 }
4551 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4552 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4553 ::copy_to_survivor_space(oop old) {
4554 size_t word_sz = old->size();
4555 HeapRegion* from_region = _g1->heap_region_containing_raw(old);
4556 // +1 to make the -1 indexes valid...
4557 int young_index = from_region->young_index_in_cset()+1;
4558 assert( (from_region->is_young() && young_index > 0) ||
4559 (!from_region->is_young() && young_index == 0), "invariant" );
4560 G1CollectorPolicy* g1p = _g1->g1_policy();
4561 markOop m = old->mark();
4562 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age()
4563 : m->age();
4564 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age,
4565 word_sz);
4566 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz);
4567 oop obj = oop(obj_ptr);
4569 if (obj_ptr == NULL) {
4570 // This will either forward-to-self, or detect that someone else has
4571 // installed a forwarding pointer.
4572 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure();
4573 return _g1->handle_evacuation_failure_par(cl, old);
4574 }
4576 // We're going to allocate linearly, so might as well prefetch ahead.
4577 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes);
4579 oop forward_ptr = old->forward_to_atomic(obj);
4580 if (forward_ptr == NULL) {
4581 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz);
4582 if (g1p->track_object_age(alloc_purpose)) {
4583 // We could simply do obj->incr_age(). However, this causes a
4584 // performance issue. obj->incr_age() will first check whether
4585 // the object has a displaced mark by checking its mark word;
4586 // getting the mark word from the new location of the object
4587 // stalls. So, given that we already have the mark word and we
4588 // are about to install it anyway, it's better to increase the
4589 // age on the mark word, when the object does not have a
4590 // displaced mark word. We're not expecting many objects to have
4591 // a displaced marked word, so that case is not optimized
4592 // further (it could be...) and we simply call obj->incr_age().
4594 if (m->has_displaced_mark_helper()) {
4595 // in this case, we have to install the mark word first,
4596 // otherwise obj looks to be forwarded (the old mark word,
4597 // which contains the forward pointer, was copied)
4598 obj->set_mark(m);
4599 obj->incr_age();
4600 } else {
4601 m = m->incr_age();
4602 obj->set_mark(m);
4603 }
4604 _par_scan_state->age_table()->add(obj, word_sz);
4605 } else {
4606 obj->set_mark(m);
4607 }
4609 size_t* surv_young_words = _par_scan_state->surviving_young_words();
4610 surv_young_words[young_index] += word_sz;
4612 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) {
4613 // We keep track of the next start index in the length field of
4614 // the to-space object. The actual length can be found in the
4615 // length field of the from-space object.
4616 arrayOop(obj)->set_length(0);
4617 oop* old_p = set_partial_array_mask(old);
4618 _par_scan_state->push_on_queue(old_p);
4619 } else {
4620 // No point in using the slower heap_region_containing() method,
4621 // given that we know obj is in the heap.
4622 _scanner.set_region(_g1->heap_region_containing_raw(obj));
4623 obj->oop_iterate_backwards(&_scanner);
4624 }
4625 } else {
4626 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz);
4627 obj = forward_ptr;
4628 }
4629 return obj;
4630 }
4632 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object>
4633 template <class T>
4634 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>
4635 ::do_oop_work(T* p) {
4636 oop obj = oopDesc::load_decode_heap_oop(p);
4637 assert(barrier != G1BarrierRS || obj != NULL,
4638 "Precondition: G1BarrierRS implies obj is non-NULL");
4640 assert(_worker_id == _par_scan_state->queue_num(), "sanity");
4642 // here the null check is implicit in the cset_fast_test() test
4643 if (_g1->in_cset_fast_test(obj)) {
4644 oop forwardee;
4645 if (obj->is_forwarded()) {
4646 forwardee = obj->forwardee();
4647 } else {
4648 forwardee = copy_to_survivor_space(obj);
4649 }
4650 assert(forwardee != NULL, "forwardee should not be NULL");
4651 oopDesc::encode_store_heap_oop(p, forwardee);
4652 if (do_mark_object && forwardee != obj) {
4653 // If the object is self-forwarded we don't need to explicitly
4654 // mark it, the evacuation failure protocol will do so.
4655 mark_forwarded_object(obj, forwardee);
4656 }
4658 // When scanning the RS, we only care about objs in CS.
4659 if (barrier == G1BarrierRS) {
4660 _par_scan_state->update_rs(_from, p, _worker_id);
4661 }
4662 } else {
4663 // The object is not in collection set. If we're a root scanning
4664 // closure during an initial mark pause (i.e. do_mark_object will
4665 // be true) then attempt to mark the object.
4666 if (do_mark_object && _g1->is_in_g1_reserved(obj)) {
4667 mark_object(obj);
4668 }
4669 }
4671 if (barrier == G1BarrierEvac && obj != NULL) {
4672 _par_scan_state->update_rs(_from, p, _worker_id);
4673 }
4675 if (do_gen_barrier && obj != NULL) {
4676 par_do_barrier(p);
4677 }
4678 }
4680 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p);
4681 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p);
4683 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) {
4684 assert(has_partial_array_mask(p), "invariant");
4685 oop from_obj = clear_partial_array_mask(p);
4687 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap.");
4688 assert(from_obj->is_objArray(), "must be obj array");
4689 objArrayOop from_obj_array = objArrayOop(from_obj);
4690 // The from-space object contains the real length.
4691 int length = from_obj_array->length();
4693 assert(from_obj->is_forwarded(), "must be forwarded");
4694 oop to_obj = from_obj->forwardee();
4695 assert(from_obj != to_obj, "should not be chunking self-forwarded objects");
4696 objArrayOop to_obj_array = objArrayOop(to_obj);
4697 // We keep track of the next start index in the length field of the
4698 // to-space object.
4699 int next_index = to_obj_array->length();
4700 assert(0 <= next_index && next_index < length,
4701 err_msg("invariant, next index: %d, length: %d", next_index, length));
4703 int start = next_index;
4704 int end = length;
4705 int remainder = end - start;
4706 // We'll try not to push a range that's smaller than ParGCArrayScanChunk.
4707 if (remainder > 2 * ParGCArrayScanChunk) {
4708 end = start + ParGCArrayScanChunk;
4709 to_obj_array->set_length(end);
4710 // Push the remainder before we process the range in case another
4711 // worker has run out of things to do and can steal it.
4712 oop* from_obj_p = set_partial_array_mask(from_obj);
4713 _par_scan_state->push_on_queue(from_obj_p);
4714 } else {
4715 assert(length == end, "sanity");
4716 // We'll process the final range for this object. Restore the length
4717 // so that the heap remains parsable in case of evacuation failure.
4718 to_obj_array->set_length(end);
4719 }
4720 _scanner.set_region(_g1->heap_region_containing_raw(to_obj));
4721 // Process indexes [start,end). It will also process the header
4722 // along with the first chunk (i.e., the chunk with start == 0).
4723 // Note that at this point the length field of to_obj_array is not
4724 // correct given that we are using it to keep track of the next
4725 // start index. oop_iterate_range() (thankfully!) ignores the length
4726 // field and only relies on the start / end parameters. It does
4727 // however return the size of the object which will be incorrect. So
4728 // we have to ignore it even if we wanted to use it.
4729 to_obj_array->oop_iterate_range(&_scanner, start, end);
4730 }
4732 class G1ParEvacuateFollowersClosure : public VoidClosure {
4733 protected:
4734 G1CollectedHeap* _g1h;
4735 G1ParScanThreadState* _par_scan_state;
4736 RefToScanQueueSet* _queues;
4737 ParallelTaskTerminator* _terminator;
4739 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
4740 RefToScanQueueSet* queues() { return _queues; }
4741 ParallelTaskTerminator* terminator() { return _terminator; }
4743 public:
4744 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h,
4745 G1ParScanThreadState* par_scan_state,
4746 RefToScanQueueSet* queues,
4747 ParallelTaskTerminator* terminator)
4748 : _g1h(g1h), _par_scan_state(par_scan_state),
4749 _queues(queues), _terminator(terminator) {}
4751 void do_void();
4753 private:
4754 inline bool offer_termination();
4755 };
4757 bool G1ParEvacuateFollowersClosure::offer_termination() {
4758 G1ParScanThreadState* const pss = par_scan_state();
4759 pss->start_term_time();
4760 const bool res = terminator()->offer_termination();
4761 pss->end_term_time();
4762 return res;
4763 }
4765 void G1ParEvacuateFollowersClosure::do_void() {
4766 StarTask stolen_task;
4767 G1ParScanThreadState* const pss = par_scan_state();
4768 pss->trim_queue();
4770 do {
4771 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) {
4772 assert(pss->verify_task(stolen_task), "sanity");
4773 if (stolen_task.is_narrow()) {
4774 pss->deal_with_reference((narrowOop*) stolen_task);
4775 } else {
4776 pss->deal_with_reference((oop*) stolen_task);
4777 }
4779 // We've just processed a reference and we might have made
4780 // available new entries on the queues. So we have to make sure
4781 // we drain the queues as necessary.
4782 pss->trim_queue();
4783 }
4784 } while (!offer_termination());
4786 pss->retire_alloc_buffers();
4787 }
4789 class G1ParTask : public AbstractGangTask {
4790 protected:
4791 G1CollectedHeap* _g1h;
4792 RefToScanQueueSet *_queues;
4793 ParallelTaskTerminator _terminator;
4794 uint _n_workers;
4796 Mutex _stats_lock;
4797 Mutex* stats_lock() { return &_stats_lock; }
4799 size_t getNCards() {
4800 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1)
4801 / G1BlockOffsetSharedArray::N_bytes;
4802 }
4804 public:
4805 G1ParTask(G1CollectedHeap* g1h,
4806 RefToScanQueueSet *task_queues)
4807 : AbstractGangTask("G1 collection"),
4808 _g1h(g1h),
4809 _queues(task_queues),
4810 _terminator(0, _queues),
4811 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true)
4812 {}
4814 RefToScanQueueSet* queues() { return _queues; }
4816 RefToScanQueue *work_queue(int i) {
4817 return queues()->queue(i);
4818 }
4820 ParallelTaskTerminator* terminator() { return &_terminator; }
4822 virtual void set_for_termination(int active_workers) {
4823 // This task calls set_n_termination() in par_non_clean_card_iterate_work()
4824 // in the young space (_par_seq_tasks) in the G1 heap
4825 // for SequentialSubTasksDone.
4826 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap
4827 // both of which need setting by set_n_termination().
4828 _g1h->SharedHeap::set_n_termination(active_workers);
4829 _g1h->set_n_termination(active_workers);
4830 terminator()->reset_for_reuse(active_workers);
4831 _n_workers = active_workers;
4832 }
4834 void work(uint worker_id) {
4835 if (worker_id >= _n_workers) return; // no work needed this round
4837 double start_time_ms = os::elapsedTime() * 1000.0;
4838 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms);
4840 {
4841 ResourceMark rm;
4842 HandleMark hm;
4844 ReferenceProcessor* rp = _g1h->ref_processor_stw();
4846 G1ParScanThreadState pss(_g1h, worker_id);
4847 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp);
4848 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp);
4849 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp);
4851 pss.set_evac_closure(&scan_evac_cl);
4852 pss.set_evac_failure_closure(&evac_failure_cl);
4853 pss.set_partial_scan_closure(&partial_scan_cl);
4855 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp);
4856 G1ParScanPermClosure only_scan_perm_cl(_g1h, &pss, rp);
4858 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp);
4859 G1ParScanAndMarkPermClosure scan_mark_perm_cl(_g1h, &pss, rp);
4861 OopClosure* scan_root_cl = &only_scan_root_cl;
4862 OopsInHeapRegionClosure* scan_perm_cl = &only_scan_perm_cl;
4864 if (_g1h->g1_policy()->during_initial_mark_pause()) {
4865 // We also need to mark copied objects.
4866 scan_root_cl = &scan_mark_root_cl;
4867 scan_perm_cl = &scan_mark_perm_cl;
4868 }
4870 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss);
4872 pss.start_strong_roots();
4873 _g1h->g1_process_strong_roots(/* not collecting perm */ false,
4874 SharedHeap::SO_AllClasses,
4875 scan_root_cl,
4876 &push_heap_rs_cl,
4877 scan_perm_cl,
4878 worker_id);
4879 pss.end_strong_roots();
4881 {
4882 double start = os::elapsedTime();
4883 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator);
4884 evac.do_void();
4885 double elapsed_ms = (os::elapsedTime()-start)*1000.0;
4886 double term_ms = pss.term_time()*1000.0;
4887 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms);
4888 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts());
4889 }
4890 _g1h->g1_policy()->record_thread_age_table(pss.age_table());
4891 _g1h->update_surviving_young_words(pss.surviving_young_words()+1);
4893 if (ParallelGCVerbose) {
4894 MutexLocker x(stats_lock());
4895 pss.print_termination_stats(worker_id);
4896 }
4898 assert(pss.refs()->is_empty(), "should be empty");
4900 // Close the inner scope so that the ResourceMark and HandleMark
4901 // destructors are executed here and are included as part of the
4902 // "GC Worker Time".
4903 }
4905 double end_time_ms = os::elapsedTime() * 1000.0;
4906 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms);
4907 }
4908 };
4910 // *** Common G1 Evacuation Stuff
4912 // Closures that support the filtering of CodeBlobs scanned during
4913 // external root scanning.
4915 // Closure applied to reference fields in code blobs (specifically nmethods)
4916 // to determine whether an nmethod contains references that point into
4917 // the collection set. Used as a predicate when walking code roots so
4918 // that only nmethods that point into the collection set are added to the
4919 // 'marked' list.
4921 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure {
4923 class G1PointsIntoCSOopClosure : public OopClosure {
4924 G1CollectedHeap* _g1;
4925 bool _points_into_cs;
4926 public:
4927 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) :
4928 _g1(g1), _points_into_cs(false) { }
4930 bool points_into_cs() const { return _points_into_cs; }
4932 template <class T>
4933 void do_oop_nv(T* p) {
4934 if (!_points_into_cs) {
4935 T heap_oop = oopDesc::load_heap_oop(p);
4936 if (!oopDesc::is_null(heap_oop) &&
4937 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) {
4938 _points_into_cs = true;
4939 }
4940 }
4941 }
4943 virtual void do_oop(oop* p) { do_oop_nv(p); }
4944 virtual void do_oop(narrowOop* p) { do_oop_nv(p); }
4945 };
4947 G1CollectedHeap* _g1;
4949 public:
4950 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) :
4951 CodeBlobToOopClosure(cl, true), _g1(g1) { }
4953 virtual void do_code_blob(CodeBlob* cb) {
4954 nmethod* nm = cb->as_nmethod_or_null();
4955 if (nm != NULL && !(nm->test_oops_do_mark())) {
4956 G1PointsIntoCSOopClosure predicate_cl(_g1);
4957 nm->oops_do(&predicate_cl);
4959 if (predicate_cl.points_into_cs()) {
4960 // At least one of the reference fields or the oop relocations
4961 // in the nmethod points into the collection set. We have to
4962 // 'mark' this nmethod.
4963 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob()
4964 // or MarkingCodeBlobClosure::do_code_blob() change.
4965 if (!nm->test_set_oops_do_mark()) {
4966 do_newly_marked_nmethod(nm);
4967 }
4968 }
4969 }
4970 }
4971 };
4973 // This method is run in a GC worker.
4975 void
4976 G1CollectedHeap::
4977 g1_process_strong_roots(bool collecting_perm_gen,
4978 ScanningOption so,
4979 OopClosure* scan_non_heap_roots,
4980 OopsInHeapRegionClosure* scan_rs,
4981 OopsInGenClosure* scan_perm,
4982 int worker_i) {
4984 // First scan the strong roots, including the perm gen.
4985 double ext_roots_start = os::elapsedTime();
4986 double closure_app_time_sec = 0.0;
4988 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots);
4989 BufferingOopsInGenClosure buf_scan_perm(scan_perm);
4990 buf_scan_perm.set_generation(perm_gen());
4992 // Walk the code cache w/o buffering, because StarTask cannot handle
4993 // unaligned oop locations.
4994 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots);
4996 process_strong_roots(false, // no scoping; this is parallel code
4997 collecting_perm_gen, so,
4998 &buf_scan_non_heap_roots,
4999 &eager_scan_code_roots,
5000 &buf_scan_perm);
5002 // Now the CM ref_processor roots.
5003 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) {
5004 // We need to treat the discovered reference lists of the
5005 // concurrent mark ref processor as roots and keep entries
5006 // (which are added by the marking threads) on them live
5007 // until they can be processed at the end of marking.
5008 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots);
5009 }
5011 // Finish up any enqueued closure apps (attributed as object copy time).
5012 buf_scan_non_heap_roots.done();
5013 buf_scan_perm.done();
5015 double obj_copy_time_sec = buf_scan_perm.closure_app_seconds() +
5016 buf_scan_non_heap_roots.closure_app_seconds();
5017 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0);
5019 double ext_root_time_ms =
5020 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0;
5022 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms);
5024 // During conc marking we have to filter the per-thread SATB buffers
5025 // to make sure we remove any oops into the CSet (which will show up
5026 // as implicitly live).
5027 double satb_filtering_ms = 0.0;
5028 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) {
5029 if (mark_in_progress()) {
5030 double satb_filter_start = os::elapsedTime();
5032 JavaThread::satb_mark_queue_set().filter_thread_buffers();
5034 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0;
5035 }
5036 }
5037 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms);
5039 // Now scan the complement of the collection set.
5040 if (scan_rs != NULL) {
5041 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i);
5042 }
5044 _process_strong_tasks->all_tasks_completed();
5045 }
5047 void
5048 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure,
5049 OopClosure* non_root_closure) {
5050 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false);
5051 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure);
5052 }
5054 // Weak Reference Processing support
5056 // An always "is_alive" closure that is used to preserve referents.
5057 // If the object is non-null then it's alive. Used in the preservation
5058 // of referent objects that are pointed to by reference objects
5059 // discovered by the CM ref processor.
5060 class G1AlwaysAliveClosure: public BoolObjectClosure {
5061 G1CollectedHeap* _g1;
5062 public:
5063 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5064 void do_object(oop p) { assert(false, "Do not call."); }
5065 bool do_object_b(oop p) {
5066 if (p != NULL) {
5067 return true;
5068 }
5069 return false;
5070 }
5071 };
5073 bool G1STWIsAliveClosure::do_object_b(oop p) {
5074 // An object is reachable if it is outside the collection set,
5075 // or is inside and copied.
5076 return !_g1->obj_in_cs(p) || p->is_forwarded();
5077 }
5079 // Non Copying Keep Alive closure
5080 class G1KeepAliveClosure: public OopClosure {
5081 G1CollectedHeap* _g1;
5082 public:
5083 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {}
5084 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); }
5085 void do_oop( oop* p) {
5086 oop obj = *p;
5088 if (_g1->obj_in_cs(obj)) {
5089 assert( obj->is_forwarded(), "invariant" );
5090 *p = obj->forwardee();
5091 }
5092 }
5093 };
5095 // Copying Keep Alive closure - can be called from both
5096 // serial and parallel code as long as different worker
5097 // threads utilize different G1ParScanThreadState instances
5098 // and different queues.
5100 class G1CopyingKeepAliveClosure: public OopClosure {
5101 G1CollectedHeap* _g1h;
5102 OopClosure* _copy_non_heap_obj_cl;
5103 OopsInHeapRegionClosure* _copy_perm_obj_cl;
5104 G1ParScanThreadState* _par_scan_state;
5106 public:
5107 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h,
5108 OopClosure* non_heap_obj_cl,
5109 OopsInHeapRegionClosure* perm_obj_cl,
5110 G1ParScanThreadState* pss):
5111 _g1h(g1h),
5112 _copy_non_heap_obj_cl(non_heap_obj_cl),
5113 _copy_perm_obj_cl(perm_obj_cl),
5114 _par_scan_state(pss)
5115 {}
5117 virtual void do_oop(narrowOop* p) { do_oop_work(p); }
5118 virtual void do_oop( oop* p) { do_oop_work(p); }
5120 template <class T> void do_oop_work(T* p) {
5121 oop obj = oopDesc::load_decode_heap_oop(p);
5123 if (_g1h->obj_in_cs(obj)) {
5124 // If the referent object has been forwarded (either copied
5125 // to a new location or to itself in the event of an
5126 // evacuation failure) then we need to update the reference
5127 // field and, if both reference and referent are in the G1
5128 // heap, update the RSet for the referent.
5129 //
5130 // If the referent has not been forwarded then we have to keep
5131 // it alive by policy. Therefore we have copy the referent.
5132 //
5133 // If the reference field is in the G1 heap then we can push
5134 // on the PSS queue. When the queue is drained (after each
5135 // phase of reference processing) the object and it's followers
5136 // will be copied, the reference field set to point to the
5137 // new location, and the RSet updated. Otherwise we need to
5138 // use the the non-heap or perm closures directly to copy
5139 // the refernt object and update the pointer, while avoiding
5140 // updating the RSet.
5142 if (_g1h->is_in_g1_reserved(p)) {
5143 _par_scan_state->push_on_queue(p);
5144 } else {
5145 // The reference field is not in the G1 heap.
5146 if (_g1h->perm_gen()->is_in(p)) {
5147 _copy_perm_obj_cl->do_oop(p);
5148 } else {
5149 _copy_non_heap_obj_cl->do_oop(p);
5150 }
5151 }
5152 }
5153 }
5154 };
5156 // Serial drain queue closure. Called as the 'complete_gc'
5157 // closure for each discovered list in some of the
5158 // reference processing phases.
5160 class G1STWDrainQueueClosure: public VoidClosure {
5161 protected:
5162 G1CollectedHeap* _g1h;
5163 G1ParScanThreadState* _par_scan_state;
5165 G1ParScanThreadState* par_scan_state() { return _par_scan_state; }
5167 public:
5168 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) :
5169 _g1h(g1h),
5170 _par_scan_state(pss)
5171 { }
5173 void do_void() {
5174 G1ParScanThreadState* const pss = par_scan_state();
5175 pss->trim_queue();
5176 }
5177 };
5179 // Parallel Reference Processing closures
5181 // Implementation of AbstractRefProcTaskExecutor for parallel reference
5182 // processing during G1 evacuation pauses.
5184 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
5185 private:
5186 G1CollectedHeap* _g1h;
5187 RefToScanQueueSet* _queues;
5188 FlexibleWorkGang* _workers;
5189 int _active_workers;
5191 public:
5192 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h,
5193 FlexibleWorkGang* workers,
5194 RefToScanQueueSet *task_queues,
5195 int n_workers) :
5196 _g1h(g1h),
5197 _queues(task_queues),
5198 _workers(workers),
5199 _active_workers(n_workers)
5200 {
5201 assert(n_workers > 0, "shouldn't call this otherwise");
5202 }
5204 // Executes the given task using concurrent marking worker threads.
5205 virtual void execute(ProcessTask& task);
5206 virtual void execute(EnqueueTask& task);
5207 };
5209 // Gang task for possibly parallel reference processing
5211 class G1STWRefProcTaskProxy: public AbstractGangTask {
5212 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
5213 ProcessTask& _proc_task;
5214 G1CollectedHeap* _g1h;
5215 RefToScanQueueSet *_task_queues;
5216 ParallelTaskTerminator* _terminator;
5218 public:
5219 G1STWRefProcTaskProxy(ProcessTask& proc_task,
5220 G1CollectedHeap* g1h,
5221 RefToScanQueueSet *task_queues,
5222 ParallelTaskTerminator* terminator) :
5223 AbstractGangTask("Process reference objects in parallel"),
5224 _proc_task(proc_task),
5225 _g1h(g1h),
5226 _task_queues(task_queues),
5227 _terminator(terminator)
5228 {}
5230 virtual void work(uint worker_id) {
5231 // The reference processing task executed by a single worker.
5232 ResourceMark rm;
5233 HandleMark hm;
5235 G1STWIsAliveClosure is_alive(_g1h);
5237 G1ParScanThreadState pss(_g1h, worker_id);
5239 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5240 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5241 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5243 pss.set_evac_closure(&scan_evac_cl);
5244 pss.set_evac_failure_closure(&evac_failure_cl);
5245 pss.set_partial_scan_closure(&partial_scan_cl);
5247 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5248 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL);
5250 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5251 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL);
5253 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5254 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5256 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5257 // We also need to mark copied objects.
5258 copy_non_heap_cl = ©_mark_non_heap_cl;
5259 copy_perm_cl = ©_mark_perm_cl;
5260 }
5262 // Keep alive closure.
5263 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss);
5265 // Complete GC closure
5266 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator);
5268 // Call the reference processing task's work routine.
5269 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue);
5271 // Note we cannot assert that the refs array is empty here as not all
5272 // of the processing tasks (specifically phase2 - pp2_work) execute
5273 // the complete_gc closure (which ordinarily would drain the queue) so
5274 // the queue may not be empty.
5275 }
5276 };
5278 // Driver routine for parallel reference processing.
5279 // Creates an instance of the ref processing gang
5280 // task and has the worker threads execute it.
5281 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) {
5282 assert(_workers != NULL, "Need parallel worker threads.");
5284 ParallelTaskTerminator terminator(_active_workers, _queues);
5285 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator);
5287 _g1h->set_par_threads(_active_workers);
5288 _workers->run_task(&proc_task_proxy);
5289 _g1h->set_par_threads(0);
5290 }
5292 // Gang task for parallel reference enqueueing.
5294 class G1STWRefEnqueueTaskProxy: public AbstractGangTask {
5295 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
5296 EnqueueTask& _enq_task;
5298 public:
5299 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) :
5300 AbstractGangTask("Enqueue reference objects in parallel"),
5301 _enq_task(enq_task)
5302 { }
5304 virtual void work(uint worker_id) {
5305 _enq_task.work(worker_id);
5306 }
5307 };
5309 // Driver routine for parallel reference enqueing.
5310 // Creates an instance of the ref enqueueing gang
5311 // task and has the worker threads execute it.
5313 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
5314 assert(_workers != NULL, "Need parallel worker threads.");
5316 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task);
5318 _g1h->set_par_threads(_active_workers);
5319 _workers->run_task(&enq_task_proxy);
5320 _g1h->set_par_threads(0);
5321 }
5323 // End of weak reference support closures
5325 // Abstract task used to preserve (i.e. copy) any referent objects
5326 // that are in the collection set and are pointed to by reference
5327 // objects discovered by the CM ref processor.
5329 class G1ParPreserveCMReferentsTask: public AbstractGangTask {
5330 protected:
5331 G1CollectedHeap* _g1h;
5332 RefToScanQueueSet *_queues;
5333 ParallelTaskTerminator _terminator;
5334 uint _n_workers;
5336 public:
5337 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) :
5338 AbstractGangTask("ParPreserveCMReferents"),
5339 _g1h(g1h),
5340 _queues(task_queues),
5341 _terminator(workers, _queues),
5342 _n_workers(workers)
5343 { }
5345 void work(uint worker_id) {
5346 ResourceMark rm;
5347 HandleMark hm;
5349 G1ParScanThreadState pss(_g1h, worker_id);
5350 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL);
5351 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL);
5352 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL);
5354 pss.set_evac_closure(&scan_evac_cl);
5355 pss.set_evac_failure_closure(&evac_failure_cl);
5356 pss.set_partial_scan_closure(&partial_scan_cl);
5358 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5361 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL);
5362 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL);
5364 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL);
5365 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL);
5367 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5368 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5370 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5371 // We also need to mark copied objects.
5372 copy_non_heap_cl = ©_mark_non_heap_cl;
5373 copy_perm_cl = ©_mark_perm_cl;
5374 }
5376 // Is alive closure
5377 G1AlwaysAliveClosure always_alive(_g1h);
5379 // Copying keep alive closure. Applied to referent objects that need
5380 // to be copied.
5381 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss);
5383 ReferenceProcessor* rp = _g1h->ref_processor_cm();
5385 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q();
5386 uint stride = MIN2(MAX2(_n_workers, 1U), limit);
5388 // limit is set using max_num_q() - which was set using ParallelGCThreads.
5389 // So this must be true - but assert just in case someone decides to
5390 // change the worker ids.
5391 assert(0 <= worker_id && worker_id < limit, "sanity");
5392 assert(!rp->discovery_is_atomic(), "check this code");
5394 // Select discovered lists [i, i+stride, i+2*stride,...,limit)
5395 for (uint idx = worker_id; idx < limit; idx += stride) {
5396 DiscoveredList& ref_list = rp->discovered_refs()[idx];
5398 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive);
5399 while (iter.has_next()) {
5400 // Since discovery is not atomic for the CM ref processor, we
5401 // can see some null referent objects.
5402 iter.load_ptrs(DEBUG_ONLY(true));
5403 oop ref = iter.obj();
5405 // This will filter nulls.
5406 if (iter.is_referent_alive()) {
5407 iter.make_referent_alive();
5408 }
5409 iter.move_to_next();
5410 }
5411 }
5413 // Drain the queue - which may cause stealing
5414 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator);
5415 drain_queue.do_void();
5416 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure
5417 assert(pss.refs()->is_empty(), "should be");
5418 }
5419 };
5421 // Weak Reference processing during an evacuation pause (part 1).
5422 void G1CollectedHeap::process_discovered_references() {
5423 double ref_proc_start = os::elapsedTime();
5425 ReferenceProcessor* rp = _ref_processor_stw;
5426 assert(rp->discovery_enabled(), "should have been enabled");
5428 // Any reference objects, in the collection set, that were 'discovered'
5429 // by the CM ref processor should have already been copied (either by
5430 // applying the external root copy closure to the discovered lists, or
5431 // by following an RSet entry).
5432 //
5433 // But some of the referents, that are in the collection set, that these
5434 // reference objects point to may not have been copied: the STW ref
5435 // processor would have seen that the reference object had already
5436 // been 'discovered' and would have skipped discovering the reference,
5437 // but would not have treated the reference object as a regular oop.
5438 // As a reult the copy closure would not have been applied to the
5439 // referent object.
5440 //
5441 // We need to explicitly copy these referent objects - the references
5442 // will be processed at the end of remarking.
5443 //
5444 // We also need to do this copying before we process the reference
5445 // objects discovered by the STW ref processor in case one of these
5446 // referents points to another object which is also referenced by an
5447 // object discovered by the STW ref processor.
5449 uint active_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
5450 workers()->active_workers() : 1);
5452 assert(!G1CollectedHeap::use_parallel_gc_threads() ||
5453 active_workers == workers()->active_workers(),
5454 "Need to reset active_workers");
5456 set_par_threads(active_workers);
5457 G1ParPreserveCMReferentsTask keep_cm_referents(this, active_workers, _task_queues);
5459 if (G1CollectedHeap::use_parallel_gc_threads()) {
5460 workers()->run_task(&keep_cm_referents);
5461 } else {
5462 keep_cm_referents.work(0);
5463 }
5465 set_par_threads(0);
5467 // Closure to test whether a referent is alive.
5468 G1STWIsAliveClosure is_alive(this);
5470 // Even when parallel reference processing is enabled, the processing
5471 // of JNI refs is serial and performed serially by the current thread
5472 // rather than by a worker. The following PSS will be used for processing
5473 // JNI refs.
5475 // Use only a single queue for this PSS.
5476 G1ParScanThreadState pss(this, 0);
5478 // We do not embed a reference processor in the copying/scanning
5479 // closures while we're actually processing the discovered
5480 // reference objects.
5481 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL);
5482 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL);
5483 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL);
5485 pss.set_evac_closure(&scan_evac_cl);
5486 pss.set_evac_failure_closure(&evac_failure_cl);
5487 pss.set_partial_scan_closure(&partial_scan_cl);
5489 assert(pss.refs()->is_empty(), "pre-condition");
5491 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL);
5492 G1ParScanPermClosure only_copy_perm_cl(this, &pss, NULL);
5494 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL);
5495 G1ParScanAndMarkPermClosure copy_mark_perm_cl(this, &pss, NULL);
5497 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl;
5498 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl;
5500 if (_g1h->g1_policy()->during_initial_mark_pause()) {
5501 // We also need to mark copied objects.
5502 copy_non_heap_cl = ©_mark_non_heap_cl;
5503 copy_perm_cl = ©_mark_perm_cl;
5504 }
5506 // Keep alive closure.
5507 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_perm_cl, &pss);
5509 // Serial Complete GC closure
5510 G1STWDrainQueueClosure drain_queue(this, &pss);
5512 // Setup the soft refs policy...
5513 rp->setup_policy(false);
5515 if (!rp->processing_is_mt()) {
5516 // Serial reference processing...
5517 rp->process_discovered_references(&is_alive,
5518 &keep_alive,
5519 &drain_queue,
5520 NULL);
5521 } else {
5522 // Parallel reference processing
5523 assert(rp->num_q() == active_workers, "sanity");
5524 assert(active_workers <= rp->max_num_q(), "sanity");
5526 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5527 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor);
5528 }
5530 // We have completed copying any necessary live referent objects
5531 // (that were not copied during the actual pause) so we can
5532 // retire any active alloc buffers
5533 pss.retire_alloc_buffers();
5534 assert(pss.refs()->is_empty(), "both queue and overflow should be empty");
5536 double ref_proc_time = os::elapsedTime() - ref_proc_start;
5537 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0);
5538 }
5540 // Weak Reference processing during an evacuation pause (part 2).
5541 void G1CollectedHeap::enqueue_discovered_references() {
5542 double ref_enq_start = os::elapsedTime();
5544 ReferenceProcessor* rp = _ref_processor_stw;
5545 assert(!rp->discovery_enabled(), "should have been disabled as part of processing");
5547 // Now enqueue any remaining on the discovered lists on to
5548 // the pending list.
5549 if (!rp->processing_is_mt()) {
5550 // Serial reference processing...
5551 rp->enqueue_discovered_references();
5552 } else {
5553 // Parallel reference enqueuing
5555 uint active_workers = (ParallelGCThreads > 0 ? workers()->active_workers() : 1);
5556 assert(active_workers == workers()->active_workers(),
5557 "Need to reset active_workers");
5558 assert(rp->num_q() == active_workers, "sanity");
5559 assert(active_workers <= rp->max_num_q(), "sanity");
5561 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers);
5562 rp->enqueue_discovered_references(&par_task_executor);
5563 }
5565 rp->verify_no_references_recorded();
5566 assert(!rp->discovery_enabled(), "should have been disabled");
5568 // FIXME
5569 // CM's reference processing also cleans up the string and symbol tables.
5570 // Should we do that here also? We could, but it is a serial operation
5571 // and could signicantly increase the pause time.
5573 double ref_enq_time = os::elapsedTime() - ref_enq_start;
5574 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0);
5575 }
5577 void G1CollectedHeap::evacuate_collection_set() {
5578 _expand_heap_after_alloc_failure = true;
5579 set_evacuation_failed(false);
5581 g1_rem_set()->prepare_for_oops_into_collection_set_do();
5582 concurrent_g1_refine()->set_use_cache(false);
5583 concurrent_g1_refine()->clear_hot_cache_claimed_index();
5585 uint n_workers;
5586 if (G1CollectedHeap::use_parallel_gc_threads()) {
5587 n_workers =
5588 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(),
5589 workers()->active_workers(),
5590 Threads::number_of_non_daemon_threads());
5591 assert(UseDynamicNumberOfGCThreads ||
5592 n_workers == workers()->total_workers(),
5593 "If not dynamic should be using all the workers");
5594 workers()->set_active_workers(n_workers);
5595 set_par_threads(n_workers);
5596 } else {
5597 assert(n_par_threads() == 0,
5598 "Should be the original non-parallel value");
5599 n_workers = 1;
5600 }
5602 G1ParTask g1_par_task(this, _task_queues);
5604 init_for_evac_failure(NULL);
5606 rem_set()->prepare_for_younger_refs_iterate(true);
5608 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
5609 double start_par_time_sec = os::elapsedTime();
5610 double end_par_time_sec;
5612 {
5613 StrongRootsScope srs(this);
5615 if (G1CollectedHeap::use_parallel_gc_threads()) {
5616 // The individual threads will set their evac-failure closures.
5617 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
5618 // These tasks use ShareHeap::_process_strong_tasks
5619 assert(UseDynamicNumberOfGCThreads ||
5620 workers()->active_workers() == workers()->total_workers(),
5621 "If not dynamic should be using all the workers");
5622 workers()->run_task(&g1_par_task);
5623 } else {
5624 g1_par_task.set_for_termination(n_workers);
5625 g1_par_task.work(0);
5626 }
5627 end_par_time_sec = os::elapsedTime();
5629 // Closing the inner scope will execute the destructor
5630 // for the StrongRootsScope object. We record the current
5631 // elapsed time before closing the scope so that time
5632 // taken for the SRS destructor is NOT included in the
5633 // reported parallel time.
5634 }
5636 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0;
5637 g1_policy()->phase_times()->record_par_time(par_time_ms);
5639 double code_root_fixup_time_ms =
5640 (os::elapsedTime() - end_par_time_sec) * 1000.0;
5641 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms);
5643 set_par_threads(0);
5645 // Process any discovered reference objects - we have
5646 // to do this _before_ we retire the GC alloc regions
5647 // as we may have to copy some 'reachable' referent
5648 // objects (and their reachable sub-graphs) that were
5649 // not copied during the pause.
5650 process_discovered_references();
5652 // Weak root processing.
5653 // Note: when JSR 292 is enabled and code blobs can contain
5654 // non-perm oops then we will need to process the code blobs
5655 // here too.
5656 {
5657 G1STWIsAliveClosure is_alive(this);
5658 G1KeepAliveClosure keep_alive(this);
5659 JNIHandles::weak_oops_do(&is_alive, &keep_alive);
5660 }
5662 release_gc_alloc_regions();
5663 g1_rem_set()->cleanup_after_oops_into_collection_set_do();
5665 concurrent_g1_refine()->clear_hot_cache();
5666 concurrent_g1_refine()->set_use_cache(true);
5668 finalize_for_evac_failure();
5670 if (evacuation_failed()) {
5671 remove_self_forwarding_pointers();
5672 }
5674 // Enqueue any remaining references remaining on the STW
5675 // reference processor's discovered lists. We need to do
5676 // this after the card table is cleaned (and verified) as
5677 // the act of enqueuing entries on to the pending list
5678 // will log these updates (and dirty their associated
5679 // cards). We need these updates logged to update any
5680 // RSets.
5681 enqueue_discovered_references();
5683 if (G1DeferredRSUpdate) {
5684 RedirtyLoggedCardTableEntryFastClosure redirty;
5685 dirty_card_queue_set().set_closure(&redirty);
5686 dirty_card_queue_set().apply_closure_to_all_completed_buffers();
5688 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set();
5689 dcq.merge_bufferlists(&dirty_card_queue_set());
5690 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
5691 }
5692 COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
5693 }
5695 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr,
5696 size_t* pre_used,
5697 FreeRegionList* free_list,
5698 OldRegionSet* old_proxy_set,
5699 HumongousRegionSet* humongous_proxy_set,
5700 HRRSCleanupTask* hrrs_cleanup_task,
5701 bool par) {
5702 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
5703 if (hr->isHumongous()) {
5704 assert(hr->startsHumongous(), "we should only see starts humongous");
5705 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par);
5706 } else {
5707 _old_set.remove_with_proxy(hr, old_proxy_set);
5708 free_region(hr, pre_used, free_list, par);
5709 }
5710 } else {
5711 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task);
5712 }
5713 }
5715 void G1CollectedHeap::free_region(HeapRegion* hr,
5716 size_t* pre_used,
5717 FreeRegionList* free_list,
5718 bool par) {
5719 assert(!hr->isHumongous(), "this is only for non-humongous regions");
5720 assert(!hr->is_empty(), "the region should not be empty");
5721 assert(free_list != NULL, "pre-condition");
5723 *pre_used += hr->used();
5724 hr->hr_clear(par, true /* clear_space */);
5725 free_list->add_as_head(hr);
5726 }
5728 void G1CollectedHeap::free_humongous_region(HeapRegion* hr,
5729 size_t* pre_used,
5730 FreeRegionList* free_list,
5731 HumongousRegionSet* humongous_proxy_set,
5732 bool par) {
5733 assert(hr->startsHumongous(), "this is only for starts humongous regions");
5734 assert(free_list != NULL, "pre-condition");
5735 assert(humongous_proxy_set != NULL, "pre-condition");
5737 size_t hr_used = hr->used();
5738 size_t hr_capacity = hr->capacity();
5739 size_t hr_pre_used = 0;
5740 _humongous_set.remove_with_proxy(hr, humongous_proxy_set);
5741 // We need to read this before we make the region non-humongous,
5742 // otherwise the information will be gone.
5743 uint last_index = hr->last_hc_index();
5744 hr->set_notHumongous();
5745 free_region(hr, &hr_pre_used, free_list, par);
5747 uint i = hr->hrs_index() + 1;
5748 while (i < last_index) {
5749 HeapRegion* curr_hr = region_at(i);
5750 assert(curr_hr->continuesHumongous(), "invariant");
5751 curr_hr->set_notHumongous();
5752 free_region(curr_hr, &hr_pre_used, free_list, par);
5753 i += 1;
5754 }
5755 assert(hr_pre_used == hr_used,
5756 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" "
5757 "should be the same", hr_pre_used, hr_used));
5758 *pre_used += hr_pre_used;
5759 }
5761 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used,
5762 FreeRegionList* free_list,
5763 OldRegionSet* old_proxy_set,
5764 HumongousRegionSet* humongous_proxy_set,
5765 bool par) {
5766 if (pre_used > 0) {
5767 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL;
5768 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag);
5769 assert(_summary_bytes_used >= pre_used,
5770 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" "
5771 "should be >= pre_used: "SIZE_FORMAT,
5772 _summary_bytes_used, pre_used));
5773 _summary_bytes_used -= pre_used;
5774 }
5775 if (free_list != NULL && !free_list->is_empty()) {
5776 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag);
5777 _free_list.add_as_head(free_list);
5778 }
5779 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) {
5780 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5781 _old_set.update_from_proxy(old_proxy_set);
5782 }
5783 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) {
5784 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag);
5785 _humongous_set.update_from_proxy(humongous_proxy_set);
5786 }
5787 }
5789 class G1ParCleanupCTTask : public AbstractGangTask {
5790 CardTableModRefBS* _ct_bs;
5791 G1CollectedHeap* _g1h;
5792 HeapRegion* volatile _su_head;
5793 public:
5794 G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
5795 G1CollectedHeap* g1h) :
5796 AbstractGangTask("G1 Par Cleanup CT Task"),
5797 _ct_bs(ct_bs), _g1h(g1h) { }
5799 void work(uint worker_id) {
5800 HeapRegion* r;
5801 while (r = _g1h->pop_dirty_cards_region()) {
5802 clear_cards(r);
5803 }
5804 }
5806 void clear_cards(HeapRegion* r) {
5807 // Cards of the survivors should have already been dirtied.
5808 if (!r->is_survivor()) {
5809 _ct_bs->clear(MemRegion(r->bottom(), r->end()));
5810 }
5811 }
5812 };
5814 #ifndef PRODUCT
5815 class G1VerifyCardTableCleanup: public HeapRegionClosure {
5816 G1CollectedHeap* _g1h;
5817 CardTableModRefBS* _ct_bs;
5818 public:
5819 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs)
5820 : _g1h(g1h), _ct_bs(ct_bs) { }
5821 virtual bool doHeapRegion(HeapRegion* r) {
5822 if (r->is_survivor()) {
5823 _g1h->verify_dirty_region(r);
5824 } else {
5825 _g1h->verify_not_dirty_region(r);
5826 }
5827 return false;
5828 }
5829 };
5831 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) {
5832 // All of the region should be clean.
5833 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set();
5834 MemRegion mr(hr->bottom(), hr->end());
5835 ct_bs->verify_not_dirty_region(mr);
5836 }
5838 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) {
5839 // We cannot guarantee that [bottom(),end()] is dirty. Threads
5840 // dirty allocated blocks as they allocate them. The thread that
5841 // retires each region and replaces it with a new one will do a
5842 // maximal allocation to fill in [pre_dummy_top(),end()] but will
5843 // not dirty that area (one less thing to have to do while holding
5844 // a lock). So we can only verify that [bottom(),pre_dummy_top()]
5845 // is dirty.
5846 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5847 MemRegion mr(hr->bottom(), hr->pre_dummy_top());
5848 ct_bs->verify_dirty_region(mr);
5849 }
5851 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) {
5852 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set();
5853 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) {
5854 verify_dirty_region(hr);
5855 }
5856 }
5858 void G1CollectedHeap::verify_dirty_young_regions() {
5859 verify_dirty_young_list(_young_list->first_region());
5860 }
5861 #endif
5863 void G1CollectedHeap::cleanUpCardTable() {
5864 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
5865 double start = os::elapsedTime();
5867 {
5868 // Iterate over the dirty cards region list.
5869 G1ParCleanupCTTask cleanup_task(ct_bs, this);
5871 if (G1CollectedHeap::use_parallel_gc_threads()) {
5872 set_par_threads();
5873 workers()->run_task(&cleanup_task);
5874 set_par_threads(0);
5875 } else {
5876 while (_dirty_cards_region_list) {
5877 HeapRegion* r = _dirty_cards_region_list;
5878 cleanup_task.clear_cards(r);
5879 _dirty_cards_region_list = r->get_next_dirty_cards_region();
5880 if (_dirty_cards_region_list == r) {
5881 // The last region.
5882 _dirty_cards_region_list = NULL;
5883 }
5884 r->set_next_dirty_cards_region(NULL);
5885 }
5886 }
5887 #ifndef PRODUCT
5888 if (G1VerifyCTCleanup || VerifyAfterGC) {
5889 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs);
5890 heap_region_iterate(&cleanup_verifier);
5891 }
5892 #endif
5893 }
5895 double elapsed = os::elapsedTime() - start;
5896 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0);
5897 }
5899 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) {
5900 size_t pre_used = 0;
5901 FreeRegionList local_free_list("Local List for CSet Freeing");
5903 double young_time_ms = 0.0;
5904 double non_young_time_ms = 0.0;
5906 // Since the collection set is a superset of the the young list,
5907 // all we need to do to clear the young list is clear its
5908 // head and length, and unlink any young regions in the code below
5909 _young_list->clear();
5911 G1CollectorPolicy* policy = g1_policy();
5913 double start_sec = os::elapsedTime();
5914 bool non_young = true;
5916 HeapRegion* cur = cs_head;
5917 int age_bound = -1;
5918 size_t rs_lengths = 0;
5920 while (cur != NULL) {
5921 assert(!is_on_master_free_list(cur), "sanity");
5922 if (non_young) {
5923 if (cur->is_young()) {
5924 double end_sec = os::elapsedTime();
5925 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5926 non_young_time_ms += elapsed_ms;
5928 start_sec = os::elapsedTime();
5929 non_young = false;
5930 }
5931 } else {
5932 if (!cur->is_young()) {
5933 double end_sec = os::elapsedTime();
5934 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5935 young_time_ms += elapsed_ms;
5937 start_sec = os::elapsedTime();
5938 non_young = true;
5939 }
5940 }
5942 rs_lengths += cur->rem_set()->occupied();
5944 HeapRegion* next = cur->next_in_collection_set();
5945 assert(cur->in_collection_set(), "bad CS");
5946 cur->set_next_in_collection_set(NULL);
5947 cur->set_in_collection_set(false);
5949 if (cur->is_young()) {
5950 int index = cur->young_index_in_cset();
5951 assert(index != -1, "invariant");
5952 assert((uint) index < policy->young_cset_region_length(), "invariant");
5953 size_t words_survived = _surviving_young_words[index];
5954 cur->record_surv_words_in_group(words_survived);
5956 // At this point the we have 'popped' cur from the collection set
5957 // (linked via next_in_collection_set()) but it is still in the
5958 // young list (linked via next_young_region()). Clear the
5959 // _next_young_region field.
5960 cur->set_next_young_region(NULL);
5961 } else {
5962 int index = cur->young_index_in_cset();
5963 assert(index == -1, "invariant");
5964 }
5966 assert( (cur->is_young() && cur->young_index_in_cset() > -1) ||
5967 (!cur->is_young() && cur->young_index_in_cset() == -1),
5968 "invariant" );
5970 if (!cur->evacuation_failed()) {
5971 MemRegion used_mr = cur->used_region();
5973 // And the region is empty.
5974 assert(!used_mr.is_empty(), "Should not have empty regions in a CS.");
5975 free_region(cur, &pre_used, &local_free_list, false /* par */);
5976 } else {
5977 cur->uninstall_surv_rate_group();
5978 if (cur->is_young()) {
5979 cur->set_young_index_in_cset(-1);
5980 }
5981 cur->set_not_young();
5982 cur->set_evacuation_failed(false);
5983 // The region is now considered to be old.
5984 _old_set.add(cur);
5985 }
5986 cur = next;
5987 }
5989 policy->record_max_rs_lengths(rs_lengths);
5990 policy->cset_regions_freed();
5992 double end_sec = os::elapsedTime();
5993 double elapsed_ms = (end_sec - start_sec) * 1000.0;
5995 if (non_young) {
5996 non_young_time_ms += elapsed_ms;
5997 } else {
5998 young_time_ms += elapsed_ms;
5999 }
6001 update_sets_after_freeing_regions(pre_used, &local_free_list,
6002 NULL /* old_proxy_set */,
6003 NULL /* humongous_proxy_set */,
6004 false /* par */);
6005 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms);
6006 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms);
6007 }
6009 // This routine is similar to the above but does not record
6010 // any policy statistics or update free lists; we are abandoning
6011 // the current incremental collection set in preparation of a
6012 // full collection. After the full GC we will start to build up
6013 // the incremental collection set again.
6014 // This is only called when we're doing a full collection
6015 // and is immediately followed by the tearing down of the young list.
6017 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) {
6018 HeapRegion* cur = cs_head;
6020 while (cur != NULL) {
6021 HeapRegion* next = cur->next_in_collection_set();
6022 assert(cur->in_collection_set(), "bad CS");
6023 cur->set_next_in_collection_set(NULL);
6024 cur->set_in_collection_set(false);
6025 cur->set_young_index_in_cset(-1);
6026 cur = next;
6027 }
6028 }
6030 void G1CollectedHeap::set_free_regions_coming() {
6031 if (G1ConcRegionFreeingVerbose) {
6032 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6033 "setting free regions coming");
6034 }
6036 assert(!free_regions_coming(), "pre-condition");
6037 _free_regions_coming = true;
6038 }
6040 void G1CollectedHeap::reset_free_regions_coming() {
6041 assert(free_regions_coming(), "pre-condition");
6043 {
6044 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6045 _free_regions_coming = false;
6046 SecondaryFreeList_lock->notify_all();
6047 }
6049 if (G1ConcRegionFreeingVerbose) {
6050 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : "
6051 "reset free regions coming");
6052 }
6053 }
6055 void G1CollectedHeap::wait_while_free_regions_coming() {
6056 // Most of the time we won't have to wait, so let's do a quick test
6057 // first before we take the lock.
6058 if (!free_regions_coming()) {
6059 return;
6060 }
6062 if (G1ConcRegionFreeingVerbose) {
6063 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6064 "waiting for free regions");
6065 }
6067 {
6068 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6069 while (free_regions_coming()) {
6070 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag);
6071 }
6072 }
6074 if (G1ConcRegionFreeingVerbose) {
6075 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : "
6076 "done waiting for free regions");
6077 }
6078 }
6080 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) {
6081 assert(heap_lock_held_for_gc(),
6082 "the heap lock should already be held by or for this thread");
6083 _young_list->push_region(hr);
6084 }
6086 class NoYoungRegionsClosure: public HeapRegionClosure {
6087 private:
6088 bool _success;
6089 public:
6090 NoYoungRegionsClosure() : _success(true) { }
6091 bool doHeapRegion(HeapRegion* r) {
6092 if (r->is_young()) {
6093 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young",
6094 r->bottom(), r->end());
6095 _success = false;
6096 }
6097 return false;
6098 }
6099 bool success() { return _success; }
6100 };
6102 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) {
6103 bool ret = _young_list->check_list_empty(check_sample);
6105 if (check_heap) {
6106 NoYoungRegionsClosure closure;
6107 heap_region_iterate(&closure);
6108 ret = ret && closure.success();
6109 }
6111 return ret;
6112 }
6114 class TearDownRegionSetsClosure : public HeapRegionClosure {
6115 private:
6116 OldRegionSet *_old_set;
6118 public:
6119 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { }
6121 bool doHeapRegion(HeapRegion* r) {
6122 if (r->is_empty()) {
6123 // We ignore empty regions, we'll empty the free list afterwards
6124 } else if (r->is_young()) {
6125 // We ignore young regions, we'll empty the young list afterwards
6126 } else if (r->isHumongous()) {
6127 // We ignore humongous regions, we're not tearing down the
6128 // humongous region set
6129 } else {
6130 // The rest should be old
6131 _old_set->remove(r);
6132 }
6133 return false;
6134 }
6136 ~TearDownRegionSetsClosure() {
6137 assert(_old_set->is_empty(), "post-condition");
6138 }
6139 };
6141 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) {
6142 assert_at_safepoint(true /* should_be_vm_thread */);
6144 if (!free_list_only) {
6145 TearDownRegionSetsClosure cl(&_old_set);
6146 heap_region_iterate(&cl);
6148 // Need to do this after the heap iteration to be able to
6149 // recognize the young regions and ignore them during the iteration.
6150 _young_list->empty_list();
6151 }
6152 _free_list.remove_all();
6153 }
6155 class RebuildRegionSetsClosure : public HeapRegionClosure {
6156 private:
6157 bool _free_list_only;
6158 OldRegionSet* _old_set;
6159 FreeRegionList* _free_list;
6160 size_t _total_used;
6162 public:
6163 RebuildRegionSetsClosure(bool free_list_only,
6164 OldRegionSet* old_set, FreeRegionList* free_list) :
6165 _free_list_only(free_list_only),
6166 _old_set(old_set), _free_list(free_list), _total_used(0) {
6167 assert(_free_list->is_empty(), "pre-condition");
6168 if (!free_list_only) {
6169 assert(_old_set->is_empty(), "pre-condition");
6170 }
6171 }
6173 bool doHeapRegion(HeapRegion* r) {
6174 if (r->continuesHumongous()) {
6175 return false;
6176 }
6178 if (r->is_empty()) {
6179 // Add free regions to the free list
6180 _free_list->add_as_tail(r);
6181 } else if (!_free_list_only) {
6182 assert(!r->is_young(), "we should not come across young regions");
6184 if (r->isHumongous()) {
6185 // We ignore humongous regions, we left the humongous set unchanged
6186 } else {
6187 // The rest should be old, add them to the old set
6188 _old_set->add(r);
6189 }
6190 _total_used += r->used();
6191 }
6193 return false;
6194 }
6196 size_t total_used() {
6197 return _total_used;
6198 }
6199 };
6201 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) {
6202 assert_at_safepoint(true /* should_be_vm_thread */);
6204 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list);
6205 heap_region_iterate(&cl);
6207 if (!free_list_only) {
6208 _summary_bytes_used = cl.total_used();
6209 }
6210 assert(_summary_bytes_used == recalculate_used(),
6211 err_msg("inconsistent _summary_bytes_used, "
6212 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT,
6213 _summary_bytes_used, recalculate_used()));
6214 }
6216 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) {
6217 _refine_cte_cl->set_concurrent(concurrent);
6218 }
6220 bool G1CollectedHeap::is_in_closed_subset(const void* p) const {
6221 HeapRegion* hr = heap_region_containing(p);
6222 if (hr == NULL) {
6223 return is_in_permanent(p);
6224 } else {
6225 return hr->is_in(p);
6226 }
6227 }
6229 // Methods for the mutator alloc region
6231 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size,
6232 bool force) {
6233 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6234 assert(!force || g1_policy()->can_expand_young_list(),
6235 "if force is true we should be able to expand the young list");
6236 bool young_list_full = g1_policy()->is_young_list_full();
6237 if (force || !young_list_full) {
6238 HeapRegion* new_alloc_region = new_region(word_size,
6239 false /* do_expand */);
6240 if (new_alloc_region != NULL) {
6241 set_region_short_lived_locked(new_alloc_region);
6242 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full);
6243 return new_alloc_region;
6244 }
6245 }
6246 return NULL;
6247 }
6249 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region,
6250 size_t allocated_bytes) {
6251 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6252 assert(alloc_region->is_young(), "all mutator alloc regions should be young");
6254 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region);
6255 _summary_bytes_used += allocated_bytes;
6256 _hr_printer.retire(alloc_region);
6257 // We update the eden sizes here, when the region is retired,
6258 // instead of when it's allocated, since this is the point that its
6259 // used space has been recored in _summary_bytes_used.
6260 g1mm()->update_eden_size();
6261 }
6263 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size,
6264 bool force) {
6265 return _g1h->new_mutator_alloc_region(word_size, force);
6266 }
6268 void G1CollectedHeap::set_par_threads() {
6269 // Don't change the number of workers. Use the value previously set
6270 // in the workgroup.
6271 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise");
6272 uint n_workers = workers()->active_workers();
6273 assert(UseDynamicNumberOfGCThreads ||
6274 n_workers == workers()->total_workers(),
6275 "Otherwise should be using the total number of workers");
6276 if (n_workers == 0) {
6277 assert(false, "Should have been set in prior evacuation pause.");
6278 n_workers = ParallelGCThreads;
6279 workers()->set_active_workers(n_workers);
6280 }
6281 set_par_threads(n_workers);
6282 }
6284 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region,
6285 size_t allocated_bytes) {
6286 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes);
6287 }
6289 // Methods for the GC alloc regions
6291 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size,
6292 uint count,
6293 GCAllocPurpose ap) {
6294 assert(FreeList_lock->owned_by_self(), "pre-condition");
6296 if (count < g1_policy()->max_regions(ap)) {
6297 HeapRegion* new_alloc_region = new_region(word_size,
6298 true /* do_expand */);
6299 if (new_alloc_region != NULL) {
6300 // We really only need to do this for old regions given that we
6301 // should never scan survivors. But it doesn't hurt to do it
6302 // for survivors too.
6303 new_alloc_region->set_saved_mark();
6304 if (ap == GCAllocForSurvived) {
6305 new_alloc_region->set_survivor();
6306 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor);
6307 } else {
6308 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old);
6309 }
6310 bool during_im = g1_policy()->during_initial_mark_pause();
6311 new_alloc_region->note_start_of_copying(during_im);
6312 return new_alloc_region;
6313 } else {
6314 g1_policy()->note_alloc_region_limit_reached(ap);
6315 }
6316 }
6317 return NULL;
6318 }
6320 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region,
6321 size_t allocated_bytes,
6322 GCAllocPurpose ap) {
6323 bool during_im = g1_policy()->during_initial_mark_pause();
6324 alloc_region->note_end_of_copying(during_im);
6325 g1_policy()->record_bytes_copied_during_gc(allocated_bytes);
6326 if (ap == GCAllocForSurvived) {
6327 young_list()->add_survivor_region(alloc_region);
6328 } else {
6329 _old_set.add(alloc_region);
6330 }
6331 _hr_printer.retire(alloc_region);
6332 }
6334 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size,
6335 bool force) {
6336 assert(!force, "not supported for GC alloc regions");
6337 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived);
6338 }
6340 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region,
6341 size_t allocated_bytes) {
6342 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6343 GCAllocForSurvived);
6344 }
6346 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size,
6347 bool force) {
6348 assert(!force, "not supported for GC alloc regions");
6349 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured);
6350 }
6352 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region,
6353 size_t allocated_bytes) {
6354 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes,
6355 GCAllocForTenured);
6356 }
6357 // Heap region set verification
6359 class VerifyRegionListsClosure : public HeapRegionClosure {
6360 private:
6361 FreeRegionList* _free_list;
6362 OldRegionSet* _old_set;
6363 HumongousRegionSet* _humongous_set;
6364 uint _region_count;
6366 public:
6367 VerifyRegionListsClosure(OldRegionSet* old_set,
6368 HumongousRegionSet* humongous_set,
6369 FreeRegionList* free_list) :
6370 _old_set(old_set), _humongous_set(humongous_set),
6371 _free_list(free_list), _region_count(0) { }
6373 uint region_count() { return _region_count; }
6375 bool doHeapRegion(HeapRegion* hr) {
6376 _region_count += 1;
6378 if (hr->continuesHumongous()) {
6379 return false;
6380 }
6382 if (hr->is_young()) {
6383 // TODO
6384 } else if (hr->startsHumongous()) {
6385 _humongous_set->verify_next_region(hr);
6386 } else if (hr->is_empty()) {
6387 _free_list->verify_next_region(hr);
6388 } else {
6389 _old_set->verify_next_region(hr);
6390 }
6391 return false;
6392 }
6393 };
6395 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index,
6396 HeapWord* bottom) {
6397 HeapWord* end = bottom + HeapRegion::GrainWords;
6398 MemRegion mr(bottom, end);
6399 assert(_g1_reserved.contains(mr), "invariant");
6400 // This might return NULL if the allocation fails
6401 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */);
6402 }
6404 void G1CollectedHeap::verify_region_sets() {
6405 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */);
6407 // First, check the explicit lists.
6408 _free_list.verify();
6409 {
6410 // Given that a concurrent operation might be adding regions to
6411 // the secondary free list we have to take the lock before
6412 // verifying it.
6413 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
6414 _secondary_free_list.verify();
6415 }
6416 _old_set.verify();
6417 _humongous_set.verify();
6419 // If a concurrent region freeing operation is in progress it will
6420 // be difficult to correctly attributed any free regions we come
6421 // across to the correct free list given that they might belong to
6422 // one of several (free_list, secondary_free_list, any local lists,
6423 // etc.). So, if that's the case we will skip the rest of the
6424 // verification operation. Alternatively, waiting for the concurrent
6425 // operation to complete will have a non-trivial effect on the GC's
6426 // operation (no concurrent operation will last longer than the
6427 // interval between two calls to verification) and it might hide
6428 // any issues that we would like to catch during testing.
6429 if (free_regions_coming()) {
6430 return;
6431 }
6433 // Make sure we append the secondary_free_list on the free_list so
6434 // that all free regions we will come across can be safely
6435 // attributed to the free_list.
6436 append_secondary_free_list_if_not_empty_with_lock();
6438 // Finally, make sure that the region accounting in the lists is
6439 // consistent with what we see in the heap.
6440 _old_set.verify_start();
6441 _humongous_set.verify_start();
6442 _free_list.verify_start();
6444 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list);
6445 heap_region_iterate(&cl);
6447 _old_set.verify_end();
6448 _humongous_set.verify_end();
6449 _free_list.verify_end();
6450 }